fix(compaction): stabilize request-local reduction cache

[codeg-dev]

Summary

• Keep builtin request-local context reduction engaged after the 50% gate is crossed.
• Reset the latch only after an accepted, persisted session_compact event.
• Preserve the provider-native compaction bypass even while the latch is engaged.

The context hook rewrites only the outgoing request; it does not replace stored session history. With a stateless gate, reported usage moving from 501,000 to 499,000 tokens against a one-million-token window alternated reduced and unreduced payload shapes for the same stored history. This change keeps the request shape stable until compaction actually changes that history.

The stateful high/low watermark introduced for emergency pruning in #425 is the closest precedent. Here, an accepted compaction is the release condition because a narrow numeric release band cannot prove that stored history changed.

Fixes #900.

Tests

• Stock v2026.8.14 characterization: the same history at 501,000 then 499,000 reported tokens produces different payload hashes.
• Threshold-control regression: 507 request-local evaluations with 319 crossings verifies stable reduction without treating those counts as payload-scale evidence or attributing overhead to them.
• Payload-scale lifecycle regression: 1,510 eligible tool results and more than one megabyte of serialized history verify that a rejected compaction does not reset the latch and an accepted compaction does.
• Provider-native regression: the native path remains bypassed while the latch is engaged.
• npm run check
• Compaction and regression suite: 63 files, 421 tests passed.
• npm run build
• CLI mock loop with a tool: 4/4 checks passed.
• CLI offline smoke: 8/8 checks passed.

---

Summary by cubic

Stabilizes request-local context reduction by latching once the 50% gate is crossed and releasing only after an accepted, persisted compaction or when session-tree navigation changes the active branch. Previously reduction flipped around 50%, producing alternating payload shapes; now requests keep one reduced shape until history actually changes.

• Adds ContextReductionLatch with createContextReductionLatch and resetContextReductionLatch; shouldApplyContextReduction optionally accepts a latch. Stateless callers are unchanged; pass a latch to keep stable reduction across requests.
• Resets the latch on accepted session_compact and on session_tree navigation; rejected compaction does not reset. Preserves provider-native compaction bypass while latched.
• Primary touch points: packages/coding-agent/src/core/extensions/builtin/compaction/context-reduction.ts and packages/coding-agent/src/core/extensions/builtin/compaction/index.ts.
• Adds targeted regressions for threshold-control stability and payload-scale lifecycle.

^{Written for commit 231d467. Summary will update on new commits.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f11d2061f6

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".