fix(hide-consumed): parse tag-prefixed compress call text, stub live anchor summaries

[ranxianglei]

Fixes #230.

Bug

rewriteCompressText in src/hide-consumed.ts JSON.parsed the compress tool-call text starting at index 0. Adapters (pi) persist the rendered ref tag in front of the tool-call text, so the parse always threw on tagged calls and the call was treated as unparseable:

1. consumed sibling entries were never dropped from kept calls' args;
2. live compress anchors kept their full summary text in the args forever — duplicating the rendered acp_summary message in every request (~22K tokens of duplicated args measured in session 01a07b3c, billion-context-pi #336).

Reproduced on master before the fix: a tag-prefixed mixed call (live + consumed entries) kept both entries, and a fully-live call with a 5000-char summary kept all 5000 chars in its args.

Fix

• parseCallText: locate the JSON by scanning for the first { before parsing (the closed PR feat: reasoningReplay policy + anchor arg stubs + extraTokens accounting #225 approach), preserving the tag prefix on output.
• Stub entry summaries over 200 chars in kept compress calls (active and newest orphans); the rendered block summary remains the authoritative copy.

Independent of any reasoning feature — shipped on its own per #230.

Relationship to PR #226

PR #226 (open) bundles the same fix as its item 3, using a regex that strips the exact ref-tag prefix and without stubbing orphaned calls. This PR implements the {-scan approach prescribed in #230 (handles any leading metadata; also stubs orphans). If this PR merges first, #226 should drop its hide-consumed hunk; if the regex variant is preferred, this PR can be reworked.

Verification

• npm run typecheck — clean
• npm test — 628/628 pass (4 new tests in tests/hide-consumed-stub.test.ts)
• npm run build — success

[github-actions]

📦 Built Package Artifact

Branch: 2026-09-09_hide-consumed-tag-prefix (cbc78fb)

Option A — Install from npm PR tag (recommended)

In your adapter project:

npm install acp-kernel@pr-232

Each push to this PR publishes a new version under the pr-232 npm tag.

Option B — Download artifact

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

tar xzf acp-kernel-pr232.tgz
npm install ./package

---

This comment is automatically updated on each push.

[ranxianglei]

Linking: fixes #230 (tag-prefixed call text parse failure). Verified locally on 8a9ffbd: 628/628 tests, typecheck + build clean. Note: the separate latent prune sub-id coverage bug is tracked in #231 (no PR yet — no shipped adapter projects sub-ids).

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

Independent review on 8a9ffbd — verified and looks good:

Verification (fresh checkout, not your local run)

• npm run typecheck — clean
• npm test — 628/628 pass (7 suites)
• npm run build — success

Code review

• parseCallText first-{ scan (src/hide-consumed.ts:25-39): safe — the ref-tag prefix (<acp tokens="…" type="…">ref</acp>\n, src/render-refs.ts:39-46) contains no braces, and scanning for { is strictly more general than fix: break the reasoning context floor and live-anchor summary duplication #226's exact-tag regex (survives any future leading metadata).
• rewriteCompressText behavior change: fully-live calls are now re-serialized too (old code bailed when kept.length === content.length). Output is idempotent and byte-stable across turns for compact JSON, so no churn.
• Stubbing: slice(0, 199) + "…" → exactly 200 chars, only applied when summary.length > 200; safe because the rendered acp_summary message remains the authoritative copy. Orphan stubbing via the compactCompressText fallback is correctly scoped to the newest orphans (KEEP_LAST_ORPHANED).
• Pipeline ordering is fine: hide-compress-calls runs before render-refs (src/compress.ts:572); the preserved prefix is stripped and re-rendered with fresh token accounting by render-refs, so no stale tags leak.
• Tests: all 4 new cases are meaningful — including building the tagged input from real renderVisibleRefs output rather than a hand-written prefix, and the mixed-call case asserting the consumed sibling is dropped with the live entry byte-identical. No as any / @ts-ignore.

Process notes

• This counts as the second independent agent review (authoring session + this one) per AGENTS.md §6.
• fix: break the reasoning context floor and live-anchor summary duplication #226 overlap: its item 3 is the same fix via the regex variant. If this PR merges first, fix: break the reasoning context floor and live-anchor summary duplication #226 should drop its hide-consumed hunk — your call.
• prune: isCovered misses base-id coverage when block effective ids use sub-id forms (base#suffix) #231 (prune sub-id coverage) correctly split out, no PR needed until an adapter projects sub-ids.

Ready to merge on your side — I can't merge PRs (human-only).

[ranxianglei]

Extended: forensic replay of the #336 storm session showed ALL 52 compress calls use content as a stringified array (non-strict-tool provider form), 0 tag-prefixed — so the index-0 scan alone would still not have compacted that session. Added: parseCallText accepts the string form; serializeCompacted re-stringifies to preserve the original shape. Verified on the real session's last call: 924 → 278 chars, string shape kept, live range kept. 629/629 tests.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

Reviewed the extension on cbc78fb — verified and good to go:

Verification (fresh checkout)

• npm run typecheck — clean
• npm test — 629/629 pass
• npm run build — success

Diff review

• parseCallText string path (src/hide-consumed.ts:37-52): safe on all edge shapes — inner JSON.parse throwing or yielding a non-array (numbers, null, objects, garbage) leaves content = null → no rewrite; empty inner array bails like the array path.
• serializeCompacted shape preservation: JSON.stringify(compacted) embedded as a JSON string keeps {"content":"[...]" byte-shaped; key order is preserved since content is already a key on the spread obj. Round-trip is idempotent — re-parsing a rewritten call yields stubbed (≤200-char) summaries, so compactCompressText reports no change and the text stops churning.
• New test asserts the full round-trip: string shape preserved, consumed sibling dropped, live entry stubbed to exactly 200 chars ending in ….

On the forensic finding — this is the important one, and it changes which fix actually addresses #336: on master, a stringified-content call parses fine at index 0 but then bails at the Array.isArray(obj.content) check, so the tag-prefix fix alone (8a9ffbd) would indeed not have compacted that session. The #230 repro (tag prefix) and the #336 storm (stringified content) were two distinct parse-failure modes of the same fragile parser; both are now handled. The 924 → 278 char number comes from your replay — I verified the mechanism via the test path rather than the session data, but the stub math is consistent.

Ready to merge on your side — I can't merge PRs (human-only).