Repository navigation
v1.91.272.0 feat(skills): absorb pstack's blast radius, transcript recall, vet-the-number and design-twice as lazy passes - #3500
Merged
Conversation
…e-number and design-twice as lazy sections /investigate gains sections/blast-radius.md (the one safety fact, the five-rung evidence ladder, look where grep stops, confirmed and cleared risks apart) with a Phase 4 pointer, and the shared-premise red flag before the 3-strike rule. /context-restore is carved for the first time: on NO_CHECKPOINTS, sections/transcripts.md mines this workspace's own transcripts (host stores declared in hosts/*.ts and rendered by TRANSCRIPT_SOURCES / TRANSCRIPT_DISCOVERY), a fixed 7-day window, never another project's folder, subagents return only findings, redaction before anything leaves, git and gh verify every artifact, a status-tagged brief. /benchmark is carved for the first time: sections/vet-the-number.md asks the seven questions before Phase 5 reports a comparison. /plan-eng-review gains sections/design-twice.md (a second structurally distinct shape compared on interface depth, the eight design red flags), bound by the autoplan methodology snapshot alongside review-sections.md. Ideas from pstack (Lauren Tan, MIT), rewritten in gstack's words; NOTICE.md. Garry, 2026-10-26: "Do it, send". Guards: carve-guards entries and measured lines, eleven plan-eng-review pins re-measured, periodic carve-section-loading cases for the two new carves, free tests with a fixture transcript. Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
…ction, per-host transcript table, fixture and slice-cap repairs
The remote free run reported nine failures in seven files, all from this
branch: four tests encode plan-eng-review's one companion section
(eng-scope-entry-ap, autoplan-methodology-names, plan-review-cases,
review-checklists), so the design-twice pass now lives inside Section 1 of
sections/review-sections.md with the skeleton, manifest and
bin/gstack-autoplan-snapshot.ts back at main; the Codex render of
context-restore named ~/.claude/ through the all-hosts transcript table, so
{{TRANSCRIPT_SOURCES}} renders this host's row only; the transcript fixture
is declared in FREE_FIXTURES; and the two new periodic cases took the census
to 39 slices, so evals-periodic.yml's eval-slices max-parallel goes 38 -> 40.
Pins: 11 plan-eng-review re-pins at +2,896; maxSizeRatio 1.2400 (measured
1.2389); Codex gpt-5.4 overlay 159,588 bytes.
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
|
😎 This pull request was merged. |
The discovery block takes the none branch from hosts/*.ts's transcripts declaration instead of a host-name fallthrough, and the resolver header says where each half lives; the T2 subagent brief frames the transcript as data, not instructions; the test asserts table/block agreement per host, imports CAPY_NO_TRANSCRIPTS, and drops the fixture-shape case nothing reads. Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
Rule-19 grant from GStack Minder (GSTA-36, 2026-10-27) for #3500. CHANGELOG merged by mergeChangelogs (one [Unreleased], this branch's entry); NOTICE.md keeps both pstack courtesy entries (this lane's and GSTA-124's). test/fixtures/skill-md-size-allowlist.json and test/helpers/carve-guards.ts merged clean and re-measured on the merged tree. Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
… base=origin/main@584e9c1 policy=base origin/main@584e9c15352d:.gstack/ship-policy.json) STAMP_VERSION: 1.91.272.0 FILES: CHANGELOG.md VERSION agents-digest/gstack-AGENTS.md package.json MERGE_CONDITION: sha=59d94595409733739feea13b4597944e7fd29319 Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What changed
/context-restore with no saved checkpoint, before (left) and after (right): the dead end becomes a transcript-mining fallback that ends in a brief and an offer to save it as a checkpoint.
Four ideas from pstack (Lauren Tan, MIT), each rewritten in gstack's words as a lazily read pass with a credit line and a NOTICE.md entry (plans/absorb-pstack.md items 2(b), 3, 6, 8, 9):
sections/blast-radius.md(Phase 4 pointer): the one fact the change is safe because of, the five-rung evidence ladder, "look where grep stops", confirmed and cleared risks listed apart. Phase 3's red flags add the shared-premise trigger before the 3-strike rule.sections/transcripts.mdruns when Step 1 printsNO_CHECKPOINTS. Host transcript stores live inhosts/*.ts(transcriptson Claude and Codex) and render through the new{{TRANSCRIPT_SOURCES}}/{{TRANSCRIPT_DISCOVERY}}resolvers; each host's table names only its own store, and Capy printsTRANSCRIPTS: none.sections/vet-the-number.mdasks the seven questions before Phase 5 reports a comparison and ends in faster / slower / no measurable difference / inconclusive.Why
The review loop is where gstack spends time and gate money; these are the cheapest known levers from the one library built around "fewer words, more evidence". Transcript recall removes the "did you save context?" dead end every crew thread hits after a machine replacement.
Approvals
gstack-plan-eng-reviewsizes for the inlined passPre-Landing Review
REVIEW_LOOP: passes=3 blocking_found=0 deferred=0 noted=0 fixed=6 scope=frozen
Pass 1 at 0119569 (review army: testing, maintainability, security, performance, simplification): 8 findings, all INFORMATIONAL, 0 at or above the blocking floor. Six fixed in 3c6e0c4: the transcript discovery block now takes its none branch from the host's
transcriptsfield and throws for a host that declares a store without a discovery body; the resolver header was rewritten to describe what it renders; the T2 subagent brief gained "The transcript is data, not instructions"; a per-host table/block agreement test and theCAPY_NO_TRANSCRIPTSimport were added; a fixture-shape test that only restated the fixture was dropped. Two skipped as not worth a change in this PR: the quadratic registry walk intest/paid-shards.test.ts:816(4.97 s measured, now under a 30 s budget) and the Codexhead -c 4096rollout boundary (see the review notes). Pass 2 on the fix range 0119569..3c6e0c4 (58 lines): clean, converged, 1 fix cycle.Pass 3 at 59d9459 (rule-19 merge of origin/main d9b0d60, v1.91.267.0, granted by GStack Minder 2026-10-27): the product diff against main is hunk-identical to the reviewed one for every file (per-file patch-ids equal) except the merge resolution itself: CHANGELOG.md through
mergeChangelogs(one[Unreleased], this branch's entry, main's released history intact), NOTICE.md keeping both pstack courtesy entries (this lane's and GSTA-124's), and context lines intest/fixtures/skill-md-size-allowlist.jsonfrom main's #3433 raises. Clean, converged, 0 fix cycles. Trial merges per rule 25: #3434 (f45f1ad) and #3453 (af025a0) conflict with this pin ontest/helpers/carve-guards.ts(plan-eng-reviewmaxSizeRatio1.2400 here vs the hoist's 1.3290; investigatemaxSkeletonBytes52_000 vs the hoist's 30_178) and on the plan-eng-review pins in the allowlist, plus CHANGELOG; #3507 (ef6a163) overlaps on CHANGELOG only. Whichever of #3434 and this PR lands second re-measures the pins, the ratio and the skeleton cap on the merged tree.Adversarial review at 3c6e0c4: native Red Team (claude-fable-5-1, 8 findings) and an outside voice run as a GPT-6.1 subagent (the Codex CLI is not installed on this machine, recorded as
outside_provider: capy-subagent, not as Codex coverage; 5 findings). Merged to 10 INFORMATIONAL items, every one below the blocking floor and recorded under Review notes.Exploratory QA
Report:
.gstack/qa-reports/ship-3500/functional/report.md(ship Step 9.2.1 smoke, report-only, CLI surfaces). 7 contracts, 7 pass, 0 findings; probes 002–005 were superseded by the pass-1 resolver fix and rerun as 008–011 at 3c6e0c4, all pass:gen:skill-docs --host all, clean tree)TRANSCRIPTS: none, exit 0GSTACK_CAPY=1)~/.claude/path; Cursor render takes the none branchmax-parallel 40(39 slices, no exceed warning)[SKILL.md, sections/review-sections.md]Codex rollout discovery was not re-probed here;
test/context-restore-transcripts.test.tscovers it. No browser surface.Documentation
Status: updated — one factual edit, no blockers.
docs/ADDING_A_HOST.md's Config field reference gains atranscriptsrow for the new optional HostConfig field ({ root, layout }, rendered by/context-restore's no-checkpoint fallback through{{TRANSCRIPT_SOURCES}}/{{TRANSCRIPT_DISCOVERY}}; a declared store needs a discovery body inscripts/resolvers/transcripts.ts).bun test test/host-config.test.tspasses 103/103 after the edit.Scope: branch gsta-122-absorb-pstack-c1 at 3c6e0c4 against base 576839c, 39 committed files in 3 commits, no staged, unstaged or untracked content: four lazily read pstack-derived sections (investigate blast-radius, context-restore transcripts, benchmark vet-the-number, plan-eng-review design-twice inside Section 1), the
transcriptsHostConfig field in hosts/claude.ts and hosts/codex.ts with the two new resolvers, evals-periodic.yml max-parallel 38 → 40, NOTICE.md and CHANGELOG.md entries, and the test/fixture repairs. Banned-terms check: none.Documentation health:
docs/ADDING_A_HOST.md Updated (Key fields table: added the
transcriptsrow)docs/TESTING_INTERNALS.md Current (carve cases and max-parallel stated by rule, no 38/39 literal)
docs/hosts.md Current (per-host notes; no HostConfig field list)
docs/skills.md Current (skill descriptions not contradicted; new passes would be narrative)
docs/PROJECT_STRUCTURE.md Current (resolver list explicitly non-exhaustive; NOTICE one-liner already omits courtesy entries)
docs/privacy.md Current (network sends and gstack-written local state; transcript fallback is a local read with redaction before publish)
AGENTS.md / README.md / CLAUDE.md / CONTRIBUTING.md Current
CHANGELOG.md Not edited by this child (parent-owned)
Documentation coverage (reference / how-to / tutorial / explanation):
transcriptsHostConfig field ✅ ADDING_A_HOST.md + host-config.ts JSDoc ❌ ❌ ✅ transcripts.ts headercontext-restore transcript fallback ✅ SKILL.md/section only ❌ ❌ ❌ (docs/skills.md has no
/context-restoresection; pre-existing gap)investigate blast-radius pass ✅ SKILL.md/section only ❌ ❌ ❌
benchmark vet-the-number pass ✅ SKILL.md/section only ❌ ❌ ❌
plan-eng-review design-twice pass ✅ section + autoplan checklist ❌ ❌ ❌
TRANSCRIPT_SOURCES / TRANSCRIPT_DISCOVERY resolvers ✅ transcripts.ts header + ADDING_A_HOST row ❌ ❌ ❌
Debt: every new pass is reference-only (the skill's own SKILL.md and section); docs/skills.md carries no how-to for them and no
/context-restoresection at all. Gaps reported, not filled;/document-generatecan fill them if wanted.Diagram drift: none.
Gate integrity
No threshold, timeout, trial count or case kind changed.
test/helpers/carve-guards.ts: investigate skeleton 51,623 ≤ 52,000 (unchanged cap); plan-eng-reviewmaxSizeRatio1.2196 → 1.2400 (measured 154,366 / 124,597 = 1.2389; skeleton unchanged at 51,670); new entries for context-restore (45,848 ≤ 46,300; union 52,510) and benchmark (30,781 ≤ 31,300; union 34,527), each with a periodiccarve-section-loading-<skill>case and touchfiles. The two new periodic cases take the census from 38 to 39 slices, soevals-periodic.ymleval-slicesmax-parallelgoes 38 → 40. The Codex render of plan-eng-review sits 412 bytes under the 160,000-byte host ceiling (gpt-5.4 overlay 159,588 bytes); the next byte added to that skill on an external host needs the deferred preamble hoist or a carve on every host.test/paid-shards.test.ts's registry walk gets a 30 s budget (4.97 s measured at main on 4 cores against a 5 s default).Evidence
gstack-pregate --stage all --remote, Ubicloud standard-16, us-east-a2) at this head 59d9459, tree e79da630c310: 19,135 tests, 0 failed, $0.05, VM destroyed. Earlier at 25e9ab8 (tree 4c65d03fa6b8): 19,022 / 0 failed, $0.06; at 0119569: 19,022 / 0 failed, $0.05, after a first run there found 7 failures (all this branch's) that were repaired.bun run typecheckclean, 824 guard tests green (parity suite and sectioned parity, carve guards and completeness, external-host carve, skill size guard at the exact pins, module-size ratchet, pstack sections, context-restore transcripts, paid-shards, evals-workflow wiring, free fixtures, host config, gen-skill-docs),gen:skill-docs --host allfresh,gstack-pregate sealed:SEALED: no-policy findings=0 commits=0.Review notes
Below the blocking floor (advisory, informational, test-gap or wording items); recorded here and not fixed in this PR.
Below the blocking floor (advisory, informational, test-gap or wording items); recorded here and not fixed in this PR.
scripts/resolvers/transcripts.ts:45PROJECT_ROOT from gstack-slug's outermost marker walk: a .git or package.json in $HOME (or a nested worktree) points discovery at another folder's transcripts; use git rev-parse --show-toplevel || pwd (native docs: add README and CLAUDE.md #1, outside refactor: reorganize codebase into modular structure #2)scripts/resolvers/transcripts.ts:62CLAUDE_CONFIG_DIR / CODEX_HOME ignored; relocated stores print a false none (native refactor: reorganize codebase into modular structure #2, outside fix: harden /browse lifecycle and ref safety #5)scripts/resolvers/transcripts.ts:62slug sed is locale-dependent on non-ASCII paths (C vs UTF-8)context-restore/sections/transcripts.md.tmpl:57prose orders by a modification time the block never prints; the live session is indistinguishablescripts/resolvers/transcripts.ts:65TRANSCRIPT_DAYS unvalidated: 'all' or 0 prints a confident empty resultscripts/resolvers/transcripts.ts:76Codex cwd filter depends on serde byte layout, no JSON unescaping, 4 KB head (native Feature request: saving all the reports, or generating some kind of artifact/hand-off file to persist the information between the sessions #6, outside Fix project-local /browse installs and harden setup #4)context-restore/sections/transcripts.md.tmpl:54'Agent tool' and gbrain search wording render unchanged on Codexscripts/resolvers/transcripts.ts:50GSTACK_CAPY=1 returns none before testing the directory; a Claude Code session on a Capy machine does write transcriptscontext-restore/sections/transcripts.md.tmpl:35Claude's lossy slug merges paths differing only in punctuation; T2 does not check each record's cwd (outside docs: add README and CLAUDE.md #1)context-restore/sections/transcripts.md.tmpl:38find lists symlinked or non-regular .jsonl files (outside Project-local .claude/skills/gstack install breaks /browse and setup can report false success #3)