feat(compaction): enforce 384k active budget ceiling and sticky generation latch for request-shape stability

[codeg-dev]

Summary

Separates physical context window from operational working-set budget (ContextBudgetPolicy) in core compaction and builtin compaction extensions, enforcing a 384k active ceiling for 1M models and a sticky generation latch to prevent request prefix shape oscillations.

Key Changes

1. ContextBudgetPolicy (384k Active Ceiling & 288k Warmup):

   • Sets initial operational active ceiling maxActiveContextTokens = 384_000 for 1M-class models (contextWindow >= 500_000), regardless of adaptive ratio resulting in 650k.
   • Speculative warmup begins at 384_000 * 0.75 ≈ 288_000 tokens.
   • 1M model default keepRecentTokens = 35_000 (explicit user settings take precedence).
   • Preserves standard behavior for 200k/smaller models without forced 384k ceiling.

2. Request-Shape Sticky Generation Latch:

   • Once context reduction engages (e.g. at 50% usage), the transform stays latched across the current compaction generation.
   • Usage oscillations (e.g. 501k → 499k → 505k) do not alternate the historical request prefix shape, protecting prompt cache hits.
   • Latch releases only on accepted & persisted session_compact or session tree reload/reset.

3. Context Pressure Metric Interface:

   • Introduces typed ContextPressure interface with activeTokens, requestBodyBytes, and generation for non-token metric tracking.

Verification

• packages/coding-agent/test/compaction/policy.test.ts: 5/5 PASS (200k compatibility, 1M 384k/288k matrix, user overrides).
• packages/coding-agent/test/compaction/context-reduction.test.ts: 9/9 PASS (collapse, shrink, clear, sticky generation latch & prefix stability).
• packages/coding-agent/test/compaction.test.ts: 43/43 PASS.
• Full compaction test suite: 57/57 files PASS.
• npm run check: 0 errors (biome, shrinkwrap, install-lock, tsc --noEmit, browser-smoke).
• Full monorepo build: npm run build PASS.

---

Summary by cubic

Enforces a 384k active-context ceiling for 1M-class models and adds a branch-scoped sticky context-reduction latch to stabilize request prefix shape and preserve cache hits. Previously we used a 0.65 ratio (~650k on 1M) and reduction could oscillate; now blocking caps at 384k, speculative warmup starts at 288k, and reduction stays latched until compaction is accepted or the model or session branch changes.

• Policy: ContextBudgetPolicy decouples physical window and active budget. 1M defaults to maxActiveContextTokens=384k, keepRecentTokens=35k, warmup at 75%, reserveTokens=16,384; standard models keep keepRecentTokens=20k. Users can override via settings compaction.maxContextTokens and compaction.keepRecentTokens; explicit overrides (including zero) are preserved.
• Threshold plumbing and final admission: core and harness shouldCompact() honor the active ceiling; computeEffectiveBlockingThresholdTokens() caps adaptive thresholds; the final provider admission path now enforces the same ceiling before dispatch; speculative warmup derives from the policy.
• Sticky latch: builtin compaction engages once over the gate and remains active; resets on accepted session_compact, session_tree, and model_select. Provider-native compaction stays excluded.
• Migration: to keep 20k retention on 1M models, set compaction.keepRecentTokens: 20000. To lower the active ceiling, set compaction.maxContextTokens.

^{Written for commit 2cf8c81. Summary will update on new commits.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 31f29200ca

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".

[codeg-dev]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 31f29200ca

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".

[codeg-dev]

@codex review

[codeg-dev]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c9837c5acd

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".