fix: tier-aware cadence reset — T1 captures no longer starve T2 (issue #364)

[ranxianglei]

Problem

Issue #364 (P1): the compress-processing handler in lib/messages/inject/inject.ts reset the tier cadence baselines (lastTier2NudgeTokens / lastTier3NudgeTokens = currentTokens) on every compress — including raw-message T1 captures, which increase tier-1 quality instead of consuming it. In compression-active sessions every T1 capture re-armed the growthFloor wait (22.5K on defaults), so T2 could only fire in the gap between two T1 captures. Observed in a 21-day hub session: tier-1 quality stayed above the 50K trigger threshold for 1222h with 2325 T1 captures in between, T2 firing only after the lag.

Fix

Classify the just-processed compress call by its range-boundary prefix — the convention already documented in lib/compress/state.ts:81-83 (m-prefix = raw T1 capture; b-prefix = T2+ distillation):

• m-only boundaries → capture → baselines untouched (the fix).
• any b-prefix (incl. mixed batches) → real distillation → baselines reset as before — preserves the fix: T2/T3 nudge loop — preserve cadence baseline after compress #235 loop-prevention invariant (never undefined).
• unparsable/missing boundaries → conservative: reset (loop-prevention wins).

New isCaptureOnlyCompress in lib/messages/query.ts (with tolerant extractCompressBoundaryIds: input as object or JSON string). The reset in inject.ts is wrapped in if (!isCaptureOnlyCompress(lastCompressMsg)). No new state fields, no persisted-format change.

Tests

• tests/query-pure.test.ts: +11 unit tests (m/b/mixed/string/malformed/user/undefined).
• tests/inject.test.ts: +3 integration tests per AGENTS.md §5.7 — multi-turn capture with production preserve-recent knobs asserting baselines unchanged across turns (side-effect assertion on lastTier2NudgeTokens), distill-reset contrast (fix: T2/T3 nudge loop — preserve cadence baseline after compress #235 lock), full cadence cycle: capture → baseline held → growth → T2 fires.
• §5.7.3 verified: with the fix temporarily disabled, the capture + cycle tests FAIL, the distill test and all legacy tests stay green.
• Full suite: 1091/1091 pass; typecheck + build green; scripts/ci/check-pr.sh all checks pass.

Deferred (tracked in #364 discussion)

• feat: expand message ref format from 4-digit to 5-digit (9999→99999) #1 threshold decoupling (T2/T3 trigger reuses nudgeGrowthTokens): awaiting owner's ruling on the default value (absolute ~20K vs anchor-count vs dual) — interacts with Feature: Support percentage-based nudgeGrowthTokens for cross-model portability #300.
• docs: Test suite development summary (346→350 tests, 5-tier coverage) #3 tier checks on T1-nudge turns: after this fix T1 nudges are spaced by growthFloor, so the residual T2 delay is one turn, not hours.
• docs: add devlog system for development iteration tracking #4 pointer-ize consumed anchors: separate issue.

Devlog

devlog/2026-09-04_t2-cadence-capture-fix/ (REQ.md + WORKLOG.md).

[ranxianglei]

🤖 ework agent · qwen3.8-27b

[bot] 🏷 Reviewing PR #365 (fix for #364). Initial findings so far:

Bug verified on master. lib/messages/inject/inject.ts:139-140 resets lastTier2NudgeTokens / lastTier3NudgeTokens = currentTokens inside the currentTurnHasCompress handler for every processed compress message — including raw-message T1 captures. Combined with the cadence gate at inject.ts:440-442 (cadenceMet = lastNudge === undefined || currentTokens - lastNudge >= growthFloor, where growthFloor = max(minNudgeGrowthFloor, minNudgeGrowthRatio × nudgeGrowthTokens) = 22.5K on defaults), each T1 capture re-arms the full growth-floor wait while tier-1 tokens keep accumulating past the 50K trigger (inject.ts:439). The starvation mechanism described in #364 checks out against the code path.

Diff reviewed (6 files, +470/−9): the boundary-prefix classification matches the convention already documented at lib/compress/state.ts:81-83 (m-prefix = T1 capture, b-prefix = T2+ distillation), and the conservative fallbacks (unparsable/missing boundaries → reset) preserve the #235 loop-prevention invariant. No new state fields, no persisted-format change.

Now checking out pr/365/head (386d572) to run the full suite, typecheck/build, check-pr.sh, and the §5.7.3 "fix-disabled → new tests fail" verification. Will report back with results + independent review notes.

[github-actions]

📦 Built Plugin Artifact

Branch: 2026-09-04_t2-cadence-capture-fix (e6fd6f3)

Option A — Install from npm PR tag (recommended)

opencode plugin opencode-acp@pr-365 --global

Each push to this PR publishes a new version under the pr-365 npm tag.

Option B — Install from GitHub

opencode plugin "github:ranxianglei/opencode-acp#2026-09-04_t2-cadence-capture-fix" --global

Option C — Download artifact

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

tar xzf opencode-acp-pr365.tgz
cp -r package/dist ~/.cache/opencode/packages/opencode-acp@latest/node_modules/opencode-acp/dist

3. Restart opencode to pick up changes.

---

This comment is automatically updated on each push.

[ranxianglei]

🤖 ework agent · qwen3.8-27b

[bot] 🏷 ## Review complete — APPROVE (no code changes required)

Full verification + dual-agent review (§5.3 code / §5.6 test, two independent sessions) done at head 386d572 (base master @ 9b7adfd, v1.14.27).

1. Bug verification (precondition for the fix)

Confirmed on master: lib/messages/inject/inject.ts:139-140 unconditionally resets lastTier2NudgeTokens/lastTier3NudgeTokens = currentTokens inside the compress-processing handler (deduped by lastProcessedCompressMessageId, :123-124). Combined with the cadence gate at :440-442 (cadenceMet = currentTokens - lastNudge >= growthFloor, floor = max(5K, 0.45×50K) = 22.5K), every T1 capture re-arms the full wait — exactly the starvation described in #364. The tier-convention claim checks out against lib/compress/state.ts:80-83 (m-prefix = T1, b-prefix = T2+).

2. Verification results on PR head

Check	Result
npm run typecheck	✅ clean
npm run build	✅ clean
Full suite	✅ 1091/1091 pass (96.3s)
scripts/ci/check-pr.sh	✅ branch name, devlog REQ+WORKLOG, version unchanged → changelog skip
§5.7.3 mutation (my own run)	✅ with the guard disabled (if (!captureOnly) → if (true)): capture multi-turn test FAILS, cycle test FAILS, distill-guard test PASSES, all legacy tests green. Restored byte-exact, tree clean, 92/92 re-green.
Diff hygiene	✅ 6 files +470/−9, no version bump, no persisted-state format change, barrel export consistent

3. Dual-agent reviews — both APPROVE-WITH-NITS

Code reviewer verified: /^b\d+$/i matches the canonical block-ref grammar in lib/message-ids.ts:70-91 (no false pos/neg; unresolvable edge refs like b0 fail compression anyway → conservative direction is safe); message-mode input shape was removed in PR #233 so legacy shapes classify conservatively; classification runs once per new compress message only (negligible perf); fresh-session/resetOnCompaction path (baselines undefined) fires T2 at most once when tier-1 ≥ 50K and self-throttles via :535 — no new loop risk; #235 invariant (never undefined) preserved for all b-prefix/unparsable cases.

Test reviewer traced every assertion to its code path (cycle-test turn B: currentTokens = exactly 136K from last assistant usage; min-nudge floor 15%×1M = 150K provably suppresses the T1 nudge, so shouldInjectThisTurn=true can only come from the T2 tier trigger — good isolation). §5.7 satisfied: multi-turn shared state, side-effect assertions on baselines, production preserveRecentMessages=5. Independent mutation run reproduced my §5.7.3 result exactly (25 !== 500000, 135000 !== 100000; distill test unaffected).

Non-blocking nits (all minor, optional follow-ups):

1. Classification uses findLast (last compress msg of the turn) — a same-turn distill→capture interleaving wouldn't reset; degrades gracefully to the designed cadence since T2 firing itself advances the baseline (:535-537). A turn-aggregate some(m => !isCaptureOnlyCompress(m)) would be strictly more correct.
2. JSON-string branch in extractCompressBoundaryIds is defensive dead code under current SDK types (state.input is always an object) — harmless future-proofing.
3. JSDoc "(mNNNNN boundaries)" is looser than actual any-non-b behavior; garbage boundary strings classify capture-only but such calls fail resolution anyway.
4. Test nit: seeded blocks omit some CompressionBlock fields — identical to the merged [Bug]: Growth nudges ignore minContextLimit and fire below it #342 test pattern, and tests/ is excluded from typecheck.

None of these affect correctness or the #235/#364 guarantees. Happy to take nit #1 as a small follow-up if you want it.

4. Status & next steps

• GitHub CI at 386d572: pr-validation ✅, build-artifact ✅, test(24) ✅, test(22) still running at time of this report.
• Merge is human-only per AGENTS.md §5.1.1.2 — please merge yourself once the Node 22 leg finishes: fix: tier-aware cadence reset — T1 captures no longer starve T2 (issue #364) #365
• Reminder from T2 蒸馏触发被系统性推迟:每次 T1 compress 重置 cadence 水位线 + 阈值随模型容量缩放失配(1M 模型需 50K 锚点质量才蒸馏) #364: fix feat: expand message ref format from 4-digit to 5-digit (9999→99999) #1 (threshold decoupling) still awaits your ruling on the default (absolute ~20K vs anchor-count ≥12 vs dual-whichever-first) before I implement it; docs: Test suite development summary (346→350 tests, 5-tier coverage) #3 and docs: add devlog system for development iteration tracking #4 remain deferred/tracked there.