test(coding-agent): fix compaction test broken by model catalog window drift

[code-yeongyu]

Problem

The Release publish-npm job failed on one test in packages/coding-agent:

agent-session-auto-compaction-queue.test.ts > should trigger threshold compaction for error messages using last successful usage
AssertionError: expected "_runAutoCompaction" to be called with arguments: [ 'threshold', false ]
Number of calls: 0

This blocked publishing v2026.7.5 (nothing was published). It is not a regression in this cycle's code — the test source and all compaction logic are unchanged since v2026.7.4 (whose publish job passed this test).

Root cause

The test uses getModel("anthropic", "claude-sonnet-4-5") and sets up 190K tokens of "last successful usage", expecting threshold compaction to fire. shouldCompact triggers when contextTokens > contextWindow - reserveTokens (default reserveTokens = 16384).

The generated model catalog was refreshed from models.dev, growing Sonnet 4.5's contextWindow from 200000 to 1000000:

• Old: 190000 > 200000 - 16384 = 183616 → true → compaction fired → test passed.
• New: 190000 > 1000000 - 16384 = 983616 → false → no compaction → test failed.

So the test hardcoded an assumption about a live-catalog value. This is the same drift class as the sibling example-model-id fix (#116).

Fix

Derive the near-limit usage from the model's actual contextWindow (window - 8000, comfortably above the 16384 reserve) so the assertion holds for any window size. Behavior is unchanged — only the test fixture is made drift-proof.

Testing

• The failing test now passes; full file 6/6 green (RED before, GREEN after).
• Full npm run check (incl. check:neo) green.

🤖 Generated with Claude Code

---

Summary by cubic

Fix publish-blocking compaction test by deriving near-limit token usage from the model’s contextWindow, making the test resilient to model catalog window changes.

• Bug Fixes
  • Replace hardcoded 190K tokens with model.contextWindow - 8_000 (above the 16,384 reserve) so threshold compaction reliably triggers regardless of window size.
  • No behavior changes; only the test fixture is updated.

^{Written for commit a9609d5. Summary will update on new commits.}

[cubic-dev-ai]

No issues found across 1 file

_{You’re at about 95% of the monthly reviewed-line limit. You may want to disable incremental reviews to conserve quota. Reviews will continue until that limit is exceeded. If you need help avoiding interruptions, please contact [email protected].}

_{Re-trigger cubic}