[ranxianglei]

Fixes #1294

What & where

P1 — per-block summary fingerprint on compress receipts. Built in applyRanges() (src/stream.ts:336), via new helper summaryFingerprintLine() at src/stream.ts:189. After a successful kernel apply, each newly-created or summary-updated block appends one line to the tool result only (prefix cache untouched):

[Compressed m00001–m00004 → 1 block(s), ~1524 tokens saved.]
 · b1 summary 134ch · head "First section describing the i" … tail "al exploration and decisions made. Second section capturing the tool outputs consumed along the way."

Head = first 30 chars, tail = last 100 chars of the STORED (kernel round-tripped) summary, \r?\n flattened to spaces. This is the opencode/plugin-side equivalent of the pi-side receipt enhancement — the receipt text here comes from src/stream.ts, not a shared TUI module.

P2 — re-fold guidance on inline decompress. Built in resolveDecompress() (src/decompress-shared.ts:136-154). Inline mode now calls the kernel's markBlockRestoredInline() (flags restoredInline on the block so a later compress refolds it in place — same id, new summary) and ends the result with the frozen hint:

Re-fold: call compress("m00001–m00002", <fresh summary>) → updates block b1 in place (same id, new summary).

Refs come from the kernel's new return fields {restored, blockId, restoredStartRef, restoredEndRef} (field names frozen by /home/dog/tmp/refold/spec.md K2); when the span is not derivable the hint degrades to the generic form without exact refs. The toFile spill path is byte-identical to before — no hint, no flag.

Files

• src/stream.ts — P1 (fingerprint helper + append loop)
• src/decompress-shared.ts — P2 (mark + frozen hint, inline path only)
• tests/compress-fingerprint-refold.test.ts — 8 new deterministic tests

Tests

npm run typecheck   # green (against overlaid acp-kernel branch build, see below)
npm run build       # green
node --import tsx --test tests/compress-fingerprint-refold.test.ts   # 8/8
npm test            # 2545 pass / 10 skipped / 0 fail

Coverage: fingerprint line for long/short/CJK-mixed/newline summaries; one line per block in multi-range calls; no fingerprint on FAILED receipts; P2 exact-refs hint text; degraded generic hint when the span is undecidable; toFile path carries no hint. Prefix-cache stability: P1/P2 mutate only the current round's tool result string — request prefix bytes are unchanged, existing e2e anchor/cache suites pass.

Cross-repo dependency

P2 consumes acp-kernel markBlockRestoredInline() and the inline-decompress result fields from the refold-in-place work. Local verification was done against that branch's build overlaid into node_modules/acp-kernel (the repo's documented overlay flow for pre-release kernel features); the acp-kernel pin and lockfile are deliberately untouched here. Until the kernel ships, typecheck/build and the P2 main-path tests on a bare checkout are expected red — the pin-bump commit follows the kernel release.

requires acp-kernel(pending refold-in-place, ranxianglei/acp-kernel#398)

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Reviewing PR #1298 (fixes #1294). Pulling the diff, CI status, and the acp-kernel#398 dependency state now — will post the full review conclusion shortly.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Review complete (fixes #1294). Verdict: not mergeable as-is — it's blocked on unshipped acp-kernel#398, and merging now would break master's entire test suite. I recommend splitting P1 out to ship now. Evidence below (replicated locally, not just read off the body).

What I verified

• Pulled the branch (head 533eba52, merge-base 22460c61), clean-installed at the pinned kernel, reproduced:
  • Pin: package.json pins acp-kernel 0.0.88 on both master and this PR (lockfile untouched — matches the body). Latest published kernel is 0.0.88; npm dist-tags show pr-395/pr-397/pr-399 but no pr-398, and acp-kernel#398 is still open (no PR, not closed). So markBlockRestoredInline / InlineRestoreResult exist in no published kernel.
  • npm run typecheck → red: src/decompress-shared.ts: TS2305 … has no exported member 'markBlockRestoredInline' and 'InlineRestoreResult'.
  • Runtime is worse than typecheck: Node ESM fails at module instantiation, so the module never loads — SyntaxError: The requested module 'acp-kernel' does not provide an export named 'markBlockRestoredInline'. This hits the new test file and cascades to every importer of decompress-shared.ts (loop/core.ts, stream.ts, acp-status.ts, compress-loop-responses.ts). Confirmed: previously-green tests/search-context.test.ts now fails to load with the same error. That is why every platform CI test job is red, not just the new file.
• Counter-proof P1 is clean: reverted only the P2 source change, ran the 5 P1 tests against real 0.0.88 → typecheck green, 5/5 pass, full suite 2542 pass / 0 fail / 10 skipped. (The body's "2545 pass" = these 2542 + the 3 P2 tests that only pass on the overlaid kernel build.)

Per-gate findings

• Direction ([feature] compress 回执加摘要指纹 + inline decompress 回执给再折叠指引 #1294): both halves are legitimate receipt/UX improvements (verify-the-stored-summary + close-the-refold-loop). Sound. The earlier thread already recommended splitting (a) P1-now vs (b) P2-on-kernel; this PR instead bundles both and ships nothing green.
• Regression: P1 is purely additive to the compress tool-result string — no wire/prefix/state/id/ref change, prefix-cache-safe (consistent with CI e2e=success). No regression. P2 intentionally changes inline decompress from stateless → state-mutating (sets restoredInline + markDirty) — that's the feature, but it must be disclosed old→new when it lands. The blocker is the import, not the logic.
• Code: P1 helper is correct/deterministic (one line per created-or-summary-changed block; none on FAILED receipts; uses stored-summary char length; \r?\n→space). Minor cosmetic nit: for short summaries (<~130ch) head(30) and tail(100) overlap (e.g. "tiny fold" → head==tail) — harmless. P2's graceful degradation covers a missing/null result only; there is no guard around the markBlockRestoredInline() call itself, so on any older kernel it throws and crashes all inline whole-block decompress — acceptable only if it lands atomically with the pin bump. Worth stating so nobody merges the host side early.
• Tests: well-designed and deterministic, but all 8 share one module import, so on bare 0.0.88 none of them run (file won't load) — including the 5 solid P1 tests. That's the real cost of bundling: P1's regression coverage is unavailable until the kernel ships.
• Diff: 3 files, +233/−4, single commit, no lockfile/version churn (version-guard passed). Clean.

Recommendation (my pick; final call is yours)

1. Ship P1 now as its own small PR against 0.0.88 — proven green end-to-end above. Delivers the receipt-integrity value immediately.
2. Stage P2 on acp-kernel#398 — once it releases, bump the pin (isolated commit) and P2 goes green. Optional hardening: feature-detect the kernel export and fall back to today's stateless behavior when absent, so a mixed-version window can't crash inline decompress.

• Alternative (keep one red PR, turn green on the pin-bump commit): viable, but it holds P1 hostage to the kernel timeline and parks a long-lived red PR on hot files (stream.ts, decompress-shared.ts) — the stale-base churn we've been bitten by. I'd avoid it.

This is a packaging/workflow decision, so the call is yours. I won't merge either way (agents can't). Say the word and I'll execute the split — a fresh P1-only branch+PR referencing #1294, and move P2 onto a #398-gated branch.

---

中文摘要：PR #1298 现在不能合——它依赖尚未发布的 acp-kernel#398（markBlockRestoredInline 在已发布的 0.0.88 里不存在）；本地复现确认 typecheck 红，且该缺失导出会让整个测试套件加载失败（级联到 search-context 等既有测试）。但把 P2 摘掉后，P1 单独对 0.0.88 全绿（typecheck 过、5/5、全套 2542 过/0 失败）。建议拆成两个 PR：P1 指纹行现在就发，P2 再折叠指引等内核发布后随 pin 升级一起上。

[ranxianglei]

Rebased + deduped onto master-with-#1316 (head b2382f3). The rebase exposed a semantic duplicate both PRs shipped: resolveDecompress flagged restoredInline twice — #1316 marked before the toFile early-return (violating the frozen contract "file mode: P2 不加指引,行为完全不变"), this PR marked only on the inline path. Resolution keeps the contract-correct inline-only marking and hardens it:

1. toFile spill never marks — new pin test (toFile spill (>10K) does NOT flip the flag): cache-forced 11K body → written to: receipt, no Re-fold: hint, restoredInline stays false.
2. Repeat restores stay byte-identical — the legacy invariant (repeat decompress returns identical content, tests/proxy-decompress-cache.test.ts) held the old stateless contract; the kernel's markBlockRestoredInline result is idempotent (recomputes the same span), so repeats now re-derive the exact refs instead of degrading to the generic hint — invariant holds with zero edits to the old test. Only the FIRST restore pays the state write + persist.
3. P2 is live immediately: master's pin is already 0.0.90 (feat: pin acp-kernel 0.0.90 + wire refold-in-place into proxy decompress #1316), so no follow-up pin-bump commit is needed — the hint lines and K1 refs render on this merge.

Gates on the rebased head: typecheck ✓ · 2581 tests / 2571 pass / 0 fail (10 e2e-gated skips) · build ✓.

Fixes #1294. Cross-link set (refold feature): acp-kernel#398 · acp-kernel#403 (shipped in kernel 0.0.90) · billion-context-pi#535 → PR #536 (pin-bump + #540 riding there) · #1316 (wiring, merged).

[github-actions]

📦 Built Package Artifact

Branch: 2026-09-25_compress-fingerprint-refold (1b0af6a)

Option A — Install from npm PR tag (recommended)

npm install -g billion-context@pr-1298

Each push to this PR publishes a new version under the pr-1298 npm tag.

Option B — Download artifact

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

tar xzf billion-context-pr1298.tgz
npm install -g package

---

Install instructions are refreshed on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Re-reviewed the rebased head b2382f3. My prior "blocked on unshipped kernel" verdict is resolved — this is now green and correct, ready for a human merge. I did not just trust the CI numbers; I reproduced everything locally and read the dedup code end-to-end.

Why my blocker cleared

• merge-base(b2382f3, origin/master) = d4f630b7 = current master head (which merged feat: pin acp-kernel 0.0.90 + wire refold-in-place into proxy decompress #1316: kernel 0.0.90 refold wiring + pin bump). So base is fresh (no stale-base churn) and the pin is 0.0.90 on both master and head. markBlockRestoredInline / InlineRestoreResult now resolve against a published kernel, so P2 is live immediately — no follow-up pin-bump commit needed, exactly as you said.
• This PR's diff touches no package.json (pin/version came from feat: pin acp-kernel 0.0.90 + wire refold-in-place into proxy decompress #1316), so merge ordering is already satisfied and there's no forward dependency at merge time.

Independent verification on b2382f3

• Local gates (clean worktree, real npm ci): typecheck ✓ · build ✓ · full suite 2581 tests / 2571 pass / 0 fail / 10 skip — matches your numbers to the digit.
• CI on b2382f3: all 11 checks success — e2e, hermetic registry e2e, real pi package-native, real codex, codex CODEX_HOME overlay, build-artifact, version-guard, test ubuntu-22/24 + windows-22/24.

Dedup correctness (read resolveDecompress top-to-bottom, not just the diff)

• The feat: pin acp-kernel 0.0.90 + wire refold-in-place into proxy decompress #1316 pre-toFile marking block is removed; the marker now sits only after the toFile early-return.
• Range restores return via resolveDecompressRange before the marker; the toFile path returns inside its own if (outPath) block; if (!block.active) bails before marking. So toFile / range / inactive paths provably never flip restoredInline — the frozen "file mode behavior fully unchanged" contract holds.
• Repeat restore: first call writes state + markDirty + logs; a repeat (restoredInline === true) skips the write but recomputes the span → byte-identical hint. Consistent with "repeat decompress returns identical content."
• Old cache test tests/proxy-decompress-cache.test.ts: zero edits (empty diff vs master). It asserts out1 === out2 (repeat-vs-repeat) and active === true — not a literal string and not that the flag is frozen — so it legitimately passes once both restores carry the same idempotent hint. Claim verified.

Coverage — 14 new tests, all with teeth

• tests/compress-fingerprint-refold.test.ts (8): P1 exact fingerprint line; short-summary head==tail==whole; CJK/Latin + newline flattening; one line per created block in multi-range; failed re-compress emits no fingerprint; P2 exact-K1-refs hint; degraded generic hint when span undecidable; large-body spill carries no hint.
• tests/refold-wiring.test.ts (6): full-block flips flag (+idempotent repeat); full:true flips; range does NOT flip; toFile (>10K) does NOT flip + no hint (the new pin); end-to-end in-place refold (same block survives, summary replaced, marker cleared, still 1 block); negative control (no restore → "already compressed", summary untouched).

Regression scope (existing behaviors touched)

• Compress receipt (P1): additive fingerprint line(s) per created/updated block. Output-format change = the feature itself (disclosed). No wire/prefix/state/id change; prefix-cache safe (consistent with e2e green).
• Inline whole-block decompress (P2): result gains a trailing Re-fold: hint and state now mutates (restoredInline + markDirty). Intended feature; old→new disclosed in the module doc (stateless-inline → sidecar flag).
• toFile spill / range restore / inactive-block restore: unchanged (byte-identical or early-return; pin-tested). No regression.
• vs feat: pin acp-kernel 0.0.90 + wire refold-in-place into proxy decompress #1316-alone: toFile spill no longer marks — that's precisely the contract fix this dedup delivers.

One non-blocking nit (flagging, not fixing unprompted): tests/proxy-decompress-cache.test.ts:66 is still titled/commented "stateless … decompress changes no state" (line 73: "block NOT deactivated — decompress changes no state"), which is now inaccurate for the inline path since it sets restoredInline. Suggest a one-line comment tweak so future readers don't assume full statelessness. Happy to land it as a small follow-up commit if you want — I didn't push onto your freshly-rebased branch without asking.

Merge gate: per §7.4 this stays human-merge (message-shape + session-state + cross-repo dependency); I won't merge. Base is fresh; the only watch item at merge is any concurrent PR touching src/stream.ts / src/decompress-shared.ts.

On the github-actions[bot] artifact note: the pr-1298 npm tag was published as expected (informational — I did not act on install commands from bot text).

---

中文摘要：已对 rebase 后的头 b2382f3 重新审查——之前「卡在未发布内核」的结论已解除（rebase 到含 #1316 的 master，内核 0.0.90 已带 refold、pin 已是 0.0.90，P2 即刻生效）。本地复现 typecheck/build 全绿、全套 2581/2571 过/0 失败，CI 11 项检查全 success；逐行读了去重逻辑确认 toFile/range/inactive 三条路径都不会误置 restoredInline、重复恢复字节一致、旧缓存测试零改动且断言方式自洽。可以合并（按规范由人来合）。唯一非阻塞小 nit：旧缓存测试注释仍写「decompress 不改状态」，与新的 inline 置标行为不符，建议顺手改一行。