[ranxianglei]

What

Fixes #333. codex's native remote-compaction request (compaction_trigger form) used to flow through prepareResponses (kernel transform + payload rebuild) and the preflight overflow-compress before the compact detection/gate ran. Even when the result was a passthrough, the upstream received a mutated compaction request — diverging from codex's local history, and 400-looping on non-OpenAI backends that reject the compaction_trigger input item (real-device evidence in the issue: 44354 tok > 32000 window → preflight compressed + rebuilt → relay 400 ×3).

How

Detection is moved to the top of the per-session pipeline, before runPrepare()/preflight, with a three-way dispatch exactly as proposed in the issue:

1. E2 on + gate preconditions pass → prepareResponses runs (transform + existing forge logic) → forged 2-frame SSE served locally, upstream not contacted.
2. E2 off (pass mode) or gate preconditions fail → verbatim passthrough: the raw request bytes are forwarded (prepared = null path, same as the fix: stitched-stream usage double-count after compress re-request (adapter-anthropic >0 guard) #300 hop-marker passthrough). No preflight, no rebuild, no window clamp, no session-state folding as a side effect.
3. Non-codex client / non-compact request → normal pipeline, unchanged.

Changes:

• src/codex-compact.ts — split the safety valve: codexCompactGatePre(session, limit) (limit > 0, steady-state < 90% of window, ≥1 active block — everything knowable before the transform) and codexCompactGate = transformOk && gatePre (identical semantics to before).
• src/server.ts (pipeline) — detect isCodexClient && last input item === compaction_trigger on the raw parsed body; dispatch as above. When not intercepted, logs the reason (BILI_CODEX_COMPACT=pass / gate preconditions not met / transform/forge failed).
• src/server.ts (prepareResponsesCompact, endpoint form) — gate preconditions checked before processTurn, so a compact that will not be intercepted no longer mutates session state as a side effect (wire behavior unchanged: raw body passthrough).

Notes / scope:

• The x-codex-beta-features: remote_compaction_v2 header stripping in buildForwardTarget is left as-is (pre-existing deliberate behavior for all codex requests, with its own aggregator-rejection rationale) — flagging it here in case pass-mode native compaction on a backend that requires the beta header turns out to need it preserved.
• The triggerFinal nudge-skip inside prepareResponses is now dead code for the trigger form (the rebuilt body is never forwarded for it) but left in place — it belongs to the feat: intercept + forge codex native compaction when ACP is healthy (#321 PR-E2) #325 change set.
• This branch is stacked on 2026-08-28_codex-compact-intercept (PR feat: intercept + forge codex native compaction when ACP is healthy (#321 PR-E2) #325, the E2 intercept/forge feature this ordering fix completes). Merge feat: intercept + forge codex native compaction when ACP is healthy (#321 PR-E2) #325 first, or both together.

Tests

4 new e2e regression tests in tests/codex-compact-e2e.test.ts using an oversized payload (14 msgs ≈ 15.4k tok > 10k window — the condition that made the old ordering fire preflight):

• pass mode + oversized compact → forwarded byte-identical to the raw request, session state untouched
• intercept + gate preconditions fail (≥90%) + oversized compact → forwarded byte-identical (the issue's exact log scenario)
• intercept + healthy ACP + oversized compact → forged, upstream untouched (preflight no longer interferes with the forge)
• endpoint form + gate preconditions fail → raw body forwarded, processTurn skipped (no state mutation)

All 4 fail on the pre-fix code (verified by stashing src/server.ts and re-running) and pass with the fix.

Pre-flight

• npm run typecheck — clean
• npm test — 761/761 pass
• npm run build — clean

[github-actions]

📦 Built Package Artifact

Branch: 2026-08-29_compaction-preflight-order (2eb7759)

Option A — Install from npm PR tag (recommended)

npm install -g billion-context@pr-335

Each push to this PR publishes a new version under the pr-335 npm tag.

Option B — Download artifact

1. Download the artifact from the Actions run
2. Extract the tarball and install:

tar xzf billion-context-pr335.tgz
npm install -g package

---

This comment is automatically updated on each push.

[ranxianglei]

🤖 ework agent · vllm-qwen/qwen3.8-27b

[bot] On it — reviewing PR #332 (branch 2026-08-29_compaction-preflight-order, stacked on 2026-08-28_codex-compact-intercept / PR #325). I'll verify the claims: check out the branch, run typecheck/tests/build, review the diff (gate split in src/codex-compact.ts, pipeline dispatch in src/server.ts, the 4 new e2e regression tests), and confirm the tests fail on pre-fix code. Full review reply to follow shortly.

[ranxianglei]

🤖 ework agent · vllm-qwen/qwen3.8-27b

[bot] Review of PR #332 complete — approved, all claims verified. Details below.

Verification (all run on 2026-08-29_compaction-preflight-order @ 3b34dde)

Check	Result
npm run typecheck	✅ clean
npm test	✅ 761/761 pass (matches the claim)
npm run build	✅ clean (tsup, dist/index.js 2.53 MB)
4 new tests fail pre-fix	✅ Verified: checked out 3b34dde^:src/server.ts and re-ran tests/codex-compact-e2e.test.ts → exactly the 4 #332 tests fail, the 5 pre-existing E2 tests still pass. Restored afterwards.
Merge vs master	✅ git merge-tree against origin/master (caca703, PR #329) → no conflicts (disjoint files: branch touches src/codex-compact.ts, src/server.ts, tests/codex-compact*.test.ts; master added tests/e2e/, AGENTS.md, ci-e2e.yml, test:e2e script)

Code review

The implementation matches the description exactly:

• src/codex-compact.ts — codexCompactGatePre(session, effectiveLimit) = limit > 0 && lastInputTokens < 0.9·limit && ≥1 active block; codexCompactGate = transformOk && gatePre. Semantics identical to the old combined gate.
• src/server.ts:1064-1082 — trigger detection on the raw parsed body inside withSessionLock, before runPrepare()/preflight. Three-way dispatch as proposed: intercept+gatePre → runPrepare() → forge served locally via forward()'s codexForge branch (upstream untouched, rememberPluginMessages called); otherwise raw bodyBuffer forwarded with prepared = null (so no #300 chain-marker stamp — correct). The why log (BILI_CODEX_COMPACT=pass / gate preconditions not met / transform/forge failed) is accurate for each branch.
• src/server.ts:1665-1671 (prepareResponsesCompact) — gatePre checked before processTurn, so a non-intercepted endpoint compact returns base (raw body) with no state mutation.
• Consistency: the dispatch and prepareResponses both use the same reqConfig.modelContextLimit (the dispatch's gatePre at src/server.ts:1071 and the forge gate at src/server.ts:1531 see the identical limit), so there's no window where the dispatch says "intercept" but prepareResponses' full gate disagrees except via transformOk — which is exactly the "transform/forge failed" branch.
• No regression to the E2 e2e suite: the forge log codex compact intercepted (trigger) (src/server.ts:1539) that tests/e2e/e2e-codex.test.ts asserts on is unchanged.
• The two flagged out-of-scope notes check out: x-codex-beta-features stripping in buildForwardTarget is untouched, and the triggerFinal nudge-skip (src/server.ts:1499-1502) is now dead for the trigger form but correctly left for the feat: intercept + forge codex native compaction when ACP is healthy (#321 PR-E2) #325 change set.

Minor observations (non-blocking)

1. "transform/forge failed" branch still touches a little state: when mode=intercept + gatePre passes but the kernel transform throws, runPrepare() has already run ++session.stats.requests, reconcileNativeCompactionBoundary, and stripBiliCompactionItems before the failure is discovered. processTurn/session.state is not mutated on the throw path (state is only assigned after success), so this is the same pre-existing behavior as any transform-failure request — not a regression from this PR, just noting it for completeness.
2. Redundant (harmless) check: src/server.ts:1684 re-checks the full codexCompactGate after processTurn in the endpoint form, but at that point gatePre already passed and transformOk === true, so it can never fail. Fine as defense-in-depth.
3. Test quality is good: the new tests assert byte-identical forwarding (h.bodies[last] === rawBody), session-state immutability via JSON.stringify snapshot, and upstream non-contact on the forge path — with the oversized 14-msg payload that specifically reproduces the old preflight-firing condition.

Merge ordering

⚠️ One dependency to keep in mind (already noted in the PR body): PR #325 (2026-08-28_codex-compact-intercept) is not in master yet — none of its commits (008bb95, c507e0c, c959942) are ancestors of origin/master. This branch is stacked on it, so #325 must merge first (or both land together). No conflicts either way.

LGTM — ready to merge once #325 lands.