fix(compression): floor safe_pct at small-context threshold before recommending

[JonthanaHanh]

Summary

The compression feasibility warning calculates safe_pct independently and can recommend a compression.threshold below the 75% small-context floor applied by ContextCompressor._effective_threshold_percent(). Users who configure the recommended value see the same warning repeat next session because the floor raises it back to 75%.

Fixes #67422

Root Cause

In conversation_compression.py, safe_pct is calculated as int((aux_context / main_ctx) * 100) without considering the small-context floor. For context windows below 512K, ContextCompressor applies a hard-coded 75% raise-only floor, but the warning doesn't account for this.

Fix

Import _SMALL_CTX_WINDOW_LIMIT and _SMALL_CTX_THRESHOLD_PERCENT from context_compressor and apply max(safe_pct, floor_pct) so the warning always recommends a threshold the system will actually honor.

Changes

• agent/conversation_compression.py: Added small-context floor check after safe_pct calculation

[alt-glitch]

This was generated by AI during triage.

Related to #67431: both address #67422's invalid small-context threshold recommendation, but #67431's live patch also suppresses the suggestion when the 75% floor still cannot fit the auxiliary context. Maintainer choice is needed.

[teknium1]

Thanks for tracing the mismatch between the warning and the small-context floor. The premise is real on current main, but this patch does not fully make the guidance actionable.

Problems

• agent/conversation_compression.py:312 produces 40% for an 80K auxiliary model and 200K main model. Raising that to 75% at this PR's added line 347 still yields a 150K trigger, which exceeds the 80K auxiliary context. agent/context_compressor.py:1221-1223 applies the percent floor, while agent/context_compressor.py:1253-1265 computes the actual trigger. The same warning can therefore recur after following the revised recommendation.
• The diff adds no regression coverage. Existing generic warning coverage is in tests/run_agent/test_compression_feasibility.py:68-99; the 75% compressor floor is covered separately at tests/agent/test_context_compressor.py:2512-2518.

Suggested changes

• Base whether to offer compression.threshold on the built-in compressor's full recomputed trigger, and omit the option when that trigger cannot fit the auxiliary context.
• Add floored-unviable and viable-boundary tests; keep non-built-in context engines out of built-in threshold-policy assumptions.

This is an automated hermes-sweeper review.

[teknium1]

This changes a 40% recommendation to 75%, but it can still be unusable: for an 80K auxiliary model and 200K main model, the 75% trigger is 150K, so the warning repeats. Please base whether to offer this option on the built-in compressor's full recomputed trigger and omit it when that trigger exceeds aux_context.

[teknium1]

Fixed via #69332 (merged) — salvage of #67431, which mirrors the compressor's full trigger recomputation (effective floor + output-token reservation + degenerate-window guard) rather than clamping alone; when no achievable threshold helps, it now says so instead of recommending a dead number. Your clamp correctly identified the floor interaction and you're credited as an independent fix in the merged PR body. Thanks!