- Clamp the summary request's max_tokens to the model's output limit so a
large STRIX_CONTEXT_SUMMARY_TOKENS can't get the request rejected (which
left the overflowing session uncompacted). Applied consistently to the
input-budget reservation and the request itself.
- Replace the tokenizer-unavailable fallback with the UTF-8 byte length, a
guaranteed upper bound on tokens for byte-level BPE, so budget checks can
never under-count dense history.
Drop the brittle substring matching for context-overflow detection. LiteLLM
already normalises every provider's overflow error to
ContextWindowExceededError and maintains the provider-specific matching
upstream, so is_context_overflow just checks that type.
When instructions, the reserved summary output, and any prior checkpoint
already consume the model's context window, the summary-input budget is
zero. Previously a doomed summary request was still submitted (and
rejected). Skip summarization in that case and log, rather than issue a
request guaranteed to overflow.
The 1,000-token floor on summary-input room could exceed the space
actually left after instructions, the reserved summary output, and any
existing checkpoint. On a small window the summary request then overflowed
and returned nothing, leaving the oversized session uncompacted. Clamp to
the real room instead so the request always fits.
Head+tail-truncate the serialized history before the summary request so a
very large head cannot itself overflow the model context and abort the
recovery it is meant to perform.
Restore cumulative history compaction (removed in the SDK migration) so a
long scan no longer replays an ever-growing transcript until it overflows
the model's context window and the run fails.
- strix/llm/context_budget.py: resolve the model's real input/output token
limits from LiteLLM metadata (128k gpt-4o, 272k gpt-5, 1M claude-sonnet-4,
131k deepseek), with a large configurable fallback for unmapped models and
a chars/4 token-count fallback.
- strix/llm/compaction.py: provider-agnostic compaction via litellm. Keeps a
security-focused structured summary (objective, findings, credentials,
payloads, URLs/paths, work state, dead ends, next move), keeps the most
recent turns by token budget, and snaps the summary boundary so no tool
call is separated from its result. Both triggers: proactive before each
run and reactive compact-and-retry on a real context-overflow error.
- Wire both triggers into the agent run loop next to the existing image
recovery; add replace_session_items() with restore-on-failure.
Env-tunable via STRIX_CONTEXT_* (auto-compact on by default).
Ahmed Allam