chore(meta-factory): Phase 3 Step 0 retrofit + 3 MAJOR fixes by artyhoo · Pull Request #3 · artyhoo/getff

artyhoo · 2026-05-07T23:48:55Z

Summary

Apply EXECUTION-PLAN §5.5 «Existing solutions research» retroactively к Phase 3 (monorepo split, merged 2026-05-08) + close 3 MAJOR findings из reviewer verdict 2026-05-08.

Verdict: KEEP merged Phase 3 as-is. Все 5 capability decisions (workspaces, linking, build, tests, version mgmt) validate ex-post через context7 research; нет capability change worth revert+redo cost.

Commits

3acf37a — Phase 3 Step 0 retrofit (Task 2 + bonus)
- phase-3-research.md (136 lines) — context7 matrix против nx/turborepo/pnpm/lerna/changesets, KEEP verdict
- PHASE-3-PROMPT-v0.2.0.md (108 lines) — counterfactual delta prompt
- EXECUTION-PLAN.md §5.5 — closure pointer
29acb8d — 3 MAJOR fixes (Task 1)
- MAJOR-1: terminology two-AI review → via AIF review-sidecar (PROPOSAL §1.2, architecture L0, risks, EXECUTION-PLAN Phase 5)
- MAJOR-2: 4 AIF coupling risks added к risks.md (version, adoption, API contract, identity dilution)
- MAJOR-3: Phase 4-9 caveat pointer на §5.5 Step 0 gate
a5e8f56 — paper trail commit PHASE-3-RETROFIT-PROMPT.md

Reuse-vs-build matrix (from `phase-3-research.md`)

#	Capability	Decision	Forward trigger
1	Workspace declaration	KEEP npm	watch pnpm catalog at ≥5 packages × ≥8 shared dev-deps
2	Cross-package linking	KEEP `*` peerDeps	switch to `workspace:*` if pnpm migration
3	Build orchestration	DEFER (no Nx/Turbo)	re-eval at ≥30s CI runtime OR cross-package compile chain
4	Cross-package tests	KEEP `npm --workspaces`	same threshold
5	Version management	DEFER → Phase 8/11 (Changesets)	pre-publish trigger

Test plan

Pre-commit hook (bash/JSON/YAML/markdown ≤500): green
Pre-push hook (actionlint, zizmor, principles 24/24, render-rules drift, spec validation): green
make self-audit: 5 pass / 0 fail
All 4 spec'd terminology edits applied (PROPOSAL:48, architecture:29, risks:18, EXECUTION-PLAN:500)
4 AIF coupling risks count = 4
Phase 4-9 caveat at line 212
phase-3-research.md ≤200 lines (136)
CI on push: audit-self.yml green
Reviewer GO/REVISE on merge

References

Three actions after independent validator pass 2026-05-08: 1. EXECUTION-PLAN §5.5 — Phase entry gate "context7-first research" mandatory for Phase 3 retrofit + Phase 4-9. Hard constraint: no git clone for external libraries, only context7 MCP queries. 2. aif-comparison.md corrections: - §4 split: 4 confirmed differentiators (mutation, manifest-SSOT, paired examples, recursive self-application) vs 2 conventions atop AIF infrastructure (two-AI tautology, stack scoping). - §9 reuse-vs-build matrix with per-Phase impact (3-9). - §10 confirmed differentiators after validator + alternatives research (Reflexion, LangGraph, AutoGen, CrewAI, Aider, etc). 3. PROPOSAL §1.4 — strategic positioning: project is plug-in for AIF runtime, not parallel framework. ~30-40% capability via AIF reuse; uniqueness is logical self-application layer. Estimated reuse savings: 30-40% original Phase 4-9 scope. Verification: all claims via context7 MCP, no source clones.

Apply EXECUTION-PLAN §5.5 «Existing solutions research» retroactively to Phase 3 (monorepo split, merged 2026-05-08). Verdict: KEEP merged work as-is; all 5 capability decisions validate ex-post. Capability matrix (context7 queries against nx_dev, turborepo_dev, pnpm.io, lerna, changesets — no source clones): | Capability | Decision | |---------------------------------|------------------------| | Workspace declaration | KEEP npm; watch pnpm | | Cross-package linking | KEEP `*` peerDeps | | Build orchestration | DEFER (no Nx/Turbo) | | Cross-package test execution | KEEP `npm --workspaces`| | Version management | DEFER → Phase 8/11 | Acceptance per §5.5: 4 KEEP/DEFER + 1 forward-add (Changesets). Files added: - docs/meta-factory/phase-3-research.md — Step 0 matrix (136 lines, under §5.5 transient ≤200 cap) - docs/meta-factory/PHASE-3-PROMPT-v0.2.0.md — counterfactual delta prompt; v0.1 body Blocks 2-6 unchanged (108 lines) EXECUTION-PLAN §5.5 updated with closure pointer (2 lines). Cost-benefit of revert+redo: discard ≥40h Phase 3+3.1 effort for zero capability change at current scale (3 packages, <3s test runtime, no cross-package compile chain). Verdict: KEEP. Forward Step 0 triggers tracked: Phase 4 entry, Phase 8 (Changesets pre-publish), Phase 9+ (pnpm catalog / Nx / Turbo re-eval thresholds).

…ixes) After reviewer verdict 2026-05-08: - MAJOR-1: terminology fix 'two-AI review' → 'via AIF review-sidecar' in PROPOSAL §1.2, architecture.md L0, risks.md, EXECUTION-PLAN Phase 5 - MAJOR-2: add 4 AIF coupling risks to risks.md (version, adoption, API, identity) - MAJOR-3: Phase 4-9 caveat pointer to §5.5 Step 0 gate before drafts Reference: aif-comparison.md §9 reuse matrix, §10 confirmed differentiators.

Self-contained delegation prompt that drove commits: - 3acf37a (Task 2 retrofit research) - 29acb8d (Task 1 — 3 MAJOR fixes) Tracked alongside existing PHASE-N-PROMPT.md artifacts for orchestrator audit trail consistency.

…(a)) Independent review (Opus 4.7) flagged L1 (a) idempotence test as vacuous on this repo: husky cp/chmod blocks at setup.sh:339,345 gate on a templates path that doesn't exist (templates moved in Phase 3.1), so dry-run idempotence proves only non-husky steps. Lowering self-app score 9→7, promoting Open question #3 from out-of-scope cleanup to Phase 5 entry blocker. M2 (arch.md §2.3 drift) and m1 (aif-comparison.md §5 sync) deferred to Phase 5 entry hygiene. Probes 1–10 still green; verdict remains GO.

… 7→8 Reflects state after 1555fb1 (setup.sh templates path fix). L1 (a) idempotence criterion no longer vacuous: dry-run output includes 4 husky [dry-run] would: cp/chmod lines, proving the install path is exercised. Score evolution now visible in retro: 9/10 (initial Phase 4 close) 7/10 (post-review calibration, reviewer M1) 8/10 (post-fix, reversal-condition met) Open question #3 closed; new Open question #6 logged for wider Phase 3.1 fallout (react-next templates path rot in setup.sh:230-244, stack-gated so doesn't affect this repo's L1 self-app, but will break react-next consumers). Probes 1-10 still green; verdict remains GO.

Open #2 (Tailwind v3/v4 version-aware) closed by ea3e006 via patterns[] field including tailwind-v3-config + tailwind-v4-css-tokens. Mantine/Chakra additions remain additive scope. New versioning entry covers Bundle 4-partial-extended completion: - missing[] + patterns[] schema additions (v1-active partial) - arch.md §2.3 v1.1 subset annotation (Q2/M2 partial closure) - detector-v0 deprecation declared - Test count semantics clarified per prior reviewer note Phase 5 entry closures total: 5 of 6 (Open #2, #3, #5, #6 + m1). Q2/M2: partial via annotation; full v2 deferred to Phase 7+.

Phase 8.8.1 addendum push-based health check on 12 non-cascade §13.x triggers (excluding §13.2 narrative, §13.3/§13.6/§13.7/§13.9 closed, §13.11 cascade, §13.10 #5 cascade, §13.15 unused). For each entry: decompose trigger into observable signals → run verification probe → classify FIRED / STILL ARMED / CASCADE-DEPENDENT. Result: 12/12 STILL ARMED or PENDING T7. Zero FIRED. No per-fire trigger-fire-§<N>.md patches needed. Probes consulted: - §13.1: research-store file count (12 at 1-pattern-per-file granularity); - §13.4 / §13.5 / §13.10 #4 / §13.12 / §13.13: consumer evidence (none); - §13.8: self-application.md §3 row count (9, no 10th proposal); - §13.10 #1: Phase 8 closure status (no curated-store gap); - §13.10 #2: Phase 9 entry coverage gap recorded f92f60b — DEFER carries forward; - §13.10 #3: Phase 9 implementation scope (housekeeping only — no new pattern); - §13.14: rules-lock.json schemaVersion = 1 (no bump); - §13.16: PENDING T7 self-review. Sweep adds value most when external signals (consumer issues, version bumps) accumulate between sessions OR a phase fires a trigger the originating session didn't notice. Both vectors quiet at 2026-05-08; unsurprising outcome. Re-sweep cadence: next phase entry research session. T6 distillation folds the sweep methodology into rule §1 as 6th checklist item. Prior-art: skipped — docs-only sweep report companion to T5.5 audit, no new capability surface or explicit dependency added (per CLAUDE.md capability-commit definition).

…-1, MAJOR-2, MINOR-1, MINOR-2) MAJOR-1: §3 ATTN + REPORT ATTN — correct shipped-artefact paths (skills/, agents/ at repo root, not .claude/skills/, .claude/agents/); narrow Claude-first bias claim to .claude/skills/self-reflection/ + proposed harness-hook layer. MAJOR-2: §0 row + §1 row 1 + §8 entry #16 + §11 sub-wave 7.1 + §12.1 — lint-staged is NOT installed and root ESLint config does NOT exist; reframe as PRECONDITION for sub-wave 7.1 (not baseline). MINOR-1: cascade-table footnote — STILL ARMED includes deferred-with-armed-trigger. MINOR-2: §13.23 trigger #3 (pre-push surface widening) satisfied by Wave 7 sub-waves 7.1.b/7.1.c — explicit note in §0.1 + §6. Prior-art: skipped — text corrections to research patch, no new capability shipped

…yer pre-push enforcement) Adds §1.7 discipline trailer check as section 9 of .husky/pre-push, closing §13.23 4th-layer enforcement per Wave 7 sub-wave 7.6 research (research patch 2026-05-11-§13.23-4th-layer-research.md) and review verdict GO Path F1 (2026-05-11-§13.23-4th-layer-review-verdicts.md). Atomically lands SSOT entries #23-#26 in prior-art-evaluations.md per CLAUDE.md «add a new SSOT entry … in the same commit as the capability artifact». Precedent: sub-wave 7.1.d landed #16 atomically with D-3 BUILD shipment; Round 1 cold-start review (2026-05-11) §3 adjudication confirmed 7.1.d as exemplary same-commit pattern. - #23 commitlint — ADOPT VOCABULARY (push-time vs commit-time gap; grep-pattern borrowed by s17_check_trailer()) - #24 semantic-release — DEFER (release automation, orthogonal) - #25 commitizen — DEFER (Python dep + FORMAT gap) - #26 git interpret-trailers — ADOPT VOCABULARY (bash case-match parallel; two-parser inconsistency avoided) Synthesis (from §11 Decision 1): - C4 scope predicate: file-glob (.claude/rules/, packages/core/principles/, .claude/skills/) ∧ content-marker (## § heading or top-level export const) ∧ NOT in D3 allow-list (docs(research-patches): / chore(snapshot-regen): / chore(prior-art-update):) - B1 bootstrap exemption: §1.7 Bootstrap: <≥20-char rationale> satisfies §1.7 at introduction-time commits (resolves chicken-and-egg) - T1 trailer format: separate stanza after Prior-art, ≥40-char payload, same placeholder-detection as pa_check_trailer() - D1 warn-only default: S17_WARN_ONLY=true for 30 days post-ship (TODO flip 2026-06-10); operator can flip locally for testing - UTF-8 locale guard: LC_ALL=en_US.UTF-8 wraps every grep matching § (fixup #3 from 7.6.b §4) Adversarial test: 50-commit sweep against main HEAD pre-push. FP count: 0/50 (0%) — below threshold → 30-day calibration default retained. Predicate verified fires on branch commit 4ec69e5 (reviewer skill section add — confirmed true positive: touches .claude/skills/reviewer/SKILL.md + adds new ## § heading content). Predicate is sound. Bootstrap: this commit introduces the §1.7 check; the rule cannot apply to its own introduction without circular dependence. Prior-art: prior-art-evaluations.md#23 (commitlint, ADOPT VOCABULARY — push-time-vs-commit-time framing borrowed; grep-pattern reused in s17_check_trailer()); prior-art-evaluations.md#26 (git interpret-trailers, ADOPT VOCABULARY — RFC 822 trailer convention, bash case-match parallel). Entries #24 (semantic-release DEFER) + #25 (commitizen DEFER) land in this commit per CLAUDE.md same-commit rule but are not positively cited — they document candidates considered + rejected. §1.7 Bootstrap: introducing §1.7 pre-push enforcement section in .husky/pre-push — self-application would be circular at introduction time (the rule cannot apply to its own introducing commit; B1 exemption per 7.6.a research §4 + 7.6.b §2 Problem 2 verdict).

Picks up EXEMPT_DIRS for research-patches/ from PR #57 (d189814) to unblock CI overweight check on 786-line patch #3. No file conflicts expected — PR #57 only touched .github/workflows/ while PR #55 only adds files under docs/meta-factory/research-patches/.

… §0 snapshot) (#155) Plan statuses had drifted (#stale-claim). Reconciled against merged PRs + per-wave patches: - new §0 verified status snapshot (N1✅ N2🟡 N3✅ N4a✅ N4b🟡 N5🔲 N6a✅ N6b🔲 N7🟡 N8🟡) - Status line: PLANNING ONLY → ACTIVE (committed #151) - Track I.1 staging-trunk migration: awaits-execute → DONE (#144/#150) - §5 decision #3 (migration execute/hold): struck → DONE Companion=C (Track 0.2 / §5.1) already landed via #153; this builds on it.

… consolidation + F.6 self-contained rationale Two of three MINOR findings from PR #261 Phase -1 cold-review applied (the third — SSOT #20 line-citation gap — was a false-positive; #20 already cites :L88 at patch:79 identical to peer rows). MINOR #1 (#ai-traps-section-fragmented) — added new §4.5 "AI-traps applied (canonical + domain-specific)" consolidating the T-trap enumeration that was previously fragmented inline at §4 §1.7 line 333. Each T-number now names the specific action in this R-phase that honours it (T1 sampling floor=5 → §1.4 enumerated all 84 SSOT rows; T3 → file:line + verbatim excerpts at lines 140-143/87/93-95/101-105/109-113; etc.). T-N1/T-N2/T-N3/T-N4 listed with their honouring action too. The prior inline mention at §4 §1.7 line 333 now points forward to §4.5. MINOR #3 (#f6-supplementary-rationale-relies-on-matrix-jump) — added "Supplementary, not subsumed by F.3 (self-contained statement)" paragraph at end of F.6 section. Now self-contained explanation: F.3-without-F.6 = Bug #3 fixed by side-effect but latent landmines remain; F.6-without-F.3 = Bugs #2/#4 unfixed → REJECT; combined F.3+F.6 = defence in depth on Bug #3. Reader no longer needs to jump to coverage matrix at lines 297-307 to understand why F.6 is additive not replacement. MINOR #2 dropped — reviewer misread: #20 citation has explicit ":L88" line ref identical to peer rows (#9/#43/#68/#77). Verified at patch:79. Prior-art: skipped — markdown polish on a research-patch; no new capability; per CLAUDE.md "Refactors, doc edits, test additions for existing capabilities ... NOT capability commits."

Wire a SubagentStart hook (.claude/hooks/inject-subagent-digest.sh) that injects the session-bootstrap digest into every dispatched junior at spawn — $0, zero per-prompt boilerplate. Mechanizes the orchestrator's manual "anchor the junior" discipline (SSOT #108). STEP-1 dual-channel verification (WebFetch code.claude.com/docs/en/hooks + DeepWiki anthropics/claude-code) confirmed the SubagentStart contract: NON-blocking, context via JSON hookSpecificOutput.additionalContext — plain stdout is a silent no-op here (unlike UserPromptSubmit), so the hook emits jq-wrapped JSON (mirrors inject-matching-rule.sh). Single digest source: reuses inject-session-bootstrap.sh (no #two-prompts-drift). Scope: only the verified candidate ships. ConfigChange settings-self-protection (census #2) DROPPED — its matcher scopes on settings-tier (source), not actor, so it cannot distinguish agent vs maintainer (T16); settings.json is already deny-listed for agents, so a block would only hit the maintainer. SubagentStop/TaskCompleted REPORT-gate (census #3) deferred as DECISION-NEEDED (block-vs-warn is a reviewer-discipline §2 fork). Paired self-test packages/core/hooks/inject-subagent-digest.test.ts (4 tests, green): asserts JSON contract (not plain stdout, T-108-A), digest content, verbatim single-source equality, exit-0 non-blocking. settings.json wiring is maintainer-landed (self-protected) — hook is INERT until then. Prior-art: prior-art-evaluations.md#108 (CC orchestrator-gate hooks, ADOPT operator — SubagentStart digest-injection subset; ConfigChange dropped on T16, SubagentStop/TaskCompleted deferred).

…330) Wire a SubagentStart hook (.claude/hooks/inject-subagent-digest.sh) that injects the session-bootstrap digest into every dispatched junior at spawn — $0, zero per-prompt boilerplate. Mechanizes the orchestrator's manual "anchor the junior" discipline (SSOT #108). STEP-1 dual-channel verification (WebFetch code.claude.com/docs/en/hooks + DeepWiki anthropics/claude-code) confirmed the SubagentStart contract: NON-blocking, context via JSON hookSpecificOutput.additionalContext — plain stdout is a silent no-op here (unlike UserPromptSubmit), so the hook emits jq-wrapped JSON (mirrors inject-matching-rule.sh). Single digest source: reuses inject-session-bootstrap.sh (no #two-prompts-drift). Scope: only the verified candidate ships. ConfigChange settings-self-protection (census #2) DROPPED — its matcher scopes on settings-tier (source), not actor, so it cannot distinguish agent vs maintainer (T16); settings.json is already deny-listed for agents, so a block would only hit the maintainer. SubagentStop/TaskCompleted REPORT-gate (census #3) deferred as DECISION-NEEDED (block-vs-warn is a reviewer-discipline §2 fork). Paired self-test packages/core/hooks/inject-subagent-digest.test.ts (4 tests, green): asserts JSON contract (not plain stdout, T-108-A), digest content, verbatim single-source equality, exit-0 non-blocking. settings.json wiring is maintainer-landed (self-protected) — hook is INERT until then. Prior-art: prior-art-evaluations.md#108 (CC orchestrator-gate hooks, ADOPT operator — SubagentStart digest-injection subset; ConfigChange dropped on T16, SubagentStop/TaskCompleted deferred).

…r; the ping is two halves §7d.3 as first written told the worker to route §7a #3's «one real minimal model call» through `POST /runtime-profiles/validate` and left a falsifier for the case where that endpoint is only a reachability check. The dispatcher fired that falsifier itself rather than shipping the trap: - for `transport: "api"`, `validateClaudeConnection` returns ok after checking only that `apiKey` and `baseUrl` are non-empty — no network call (aif-handoff packages/runtime/src/adapters/claude/index.ts:468-479). Live 2026-08-09: `/validate` against the Qwen3.8-Max-Preview profile returned ok:true sub-second. - `POST /runtime-profiles/models` is not a substitute — it returned a static Claude catalogue (Sonnet 4.6, Opus 4.6) for a Qwen profile, so it is not querying the provider. So §7a #3 and §7c #3 cannot both be satisfied by one aif REST call. §7d.3 now binds both halves — `/validate` for the route proof, plus one 1-token-scale completion against the `baseUrl` READ BACK from that response (not a hardcoded `$GLM_BASE_URL`, which was run 2's W-3 defect). Provenance is stated as dispatcher-authored and operator-overridable, matching §7b. §7d.4 gains an exact, value-free verifier for §7b #3: `hasApiKey` is `Boolean(resolved.apiKey)` and `resolved.apiKey` is `normalizeString(env[envVarName])` off the aif runtime's own process.env (packages/runtime/src/resolution.ts:426, :217-219), so `hasApiKey:true` in the §7d.3(1) response IS proof that §7b #1's outcome was achieved — without dereferencing the key, without argv exposure. Prior-art: skipped — kickoff prose correction only, no new capability; markdown is exempt from the capability-commit LOC triggers per CLAUDE.md.

#1318 landed a §7d on the same file while this branch was in flight, so the block authored here is renumbered §7d → §7e and rebuilt on top of #1318 rather than over it. #1318's content is kept intact. Two corrections to §7d, both evidence-backed: §7e.0 — §7d step 1 orders `git merge feature/beta-delivery-ux-92bf00`, which is run 2 (task 92bf0019, 2026-08-08T01:04Z). Run 3 (task e65989fa, commit 53fce45, 14:51Z) is newer and already carries every §7c correction. Measured: getff-glm-onebutton.sh is 212 lines on 92bf00, still holding `PATCH "$AIF_URL/project"` (:135) and the vendor-direct ping (:171), versus 361 lines at 53fce45 with the profile-routed path in place. Basing on 92bf00 re-does ~150 accepted lines and re-opens two closed defects. §7e.3 — §7d.1 states `POST /runtime-profiles/validate` «exercises the route the flow just built» and is «exactly what §7a #3 / §7c #3 ask for». For `transport: "api"` it makes no network call at all — it checks only that apiKey and baseUrl are non-empty (aif-handoff packages/runtime/src/adapters/claude/index.ts:468-479); live, `/validate` against the Qwen3.8-Max-Preview profile returned ok:true sub-second, and /runtime-profiles/models returned a static Claude catalogue for that same Qwen profile. The ping is therefore split in two: /validate for the route proof, plus one 1-token-scale completion against the baseUrl read back from that response. Carried over unchanged in substance: §7e.1 (host-path retraction + the container base-URL trap — localhost:3009 → 000, api:3009 → 200), §7e.2 (runtimeId/providerId required), §7e.4 (§7b #1 undelivered; hasApiKey as its exact value-free verifier), §7e.5 (companion install in a dead field), §7e.6 (fail-closed mock, it.fails-as-delivery banned). §6 gains T-BDU-D and T-BDU-E. Prior-art: skipped — kickoff prose revision only, no new capability; markdown is exempt from the capability-commit LOC triggers per CLAUDE.md.

… path, bind run-3 blockers (#1319) * docs(beta-delivery-ux): S2 kickoff §8a — operator resolutions of the six #1284 parks (re-dispatch rev) Records the operator's 2026-08-08 batch: Park-1 schema = flat Candidate A + description field; Park-2 format = JSON (F-B' shell-sourced rejection recorded with the WHY); Park-3 economy review = whole line on executor tier, aif cap 1 iteration, external cold fidelity mandatory; Park-4 CC detection = CLAUDE_CODE_SESSION_ID env-presence capability check; Park-5 conditional park did not fire (live runtime-profiles probe, no substring collision); Park-6 ship the full functional set (create-worktree.sh + worktree-node-modules.sh + link-coordination.sh) under the same $1j gate, fresh-consumer smoke AC added. Unblocks the S2 re-dispatch (A4 presets + A5 status + A9 parts 2-3). * docs(beta-delivery-ux): S4 kickoff §7b/§7c — close run-2 KICKOFF-AMBIGUOUS + host-verify contract Run 2 (aif task 92bf0019) returned FIDELITY STOP with KICKOFF-AMBIGUOUS (PR #1300). The kickoff's §7a never states how the aif runtime acquires the key VALUE: #1 stores only the env-var NAME in the profile, #4(ii) writes the value to ~/.config/getff/glm.env, and #3 requires the validation ping to run through the created profile. Verified against source: aif resolves the key from its own runtime process.env by that name (aif-handoff packages/runtime/src/resolution.ts:217-219, :247), and that env comes from the compose env-file (docker-compose.yml:15,59,94) — so glm.env is a dead end as written and #3 is unreachable without dereferencing the value, which #1 forbids. §7b closes it by binding the OUTCOME (value reachable in the aif runtime process env under the §7a #1 name, verified before the ping, honest objective-3 MISS on failure) while leaving the wiring mechanism to the worker as a §7 technical fork. It is marked as a dispatcher-authored completion, not an operator decision — the operator can override it. §7c carries run 2's settled corrections: PUT /projects/:id with a full body (aifHttp.ts:96) instead of the invented PATCH /project, no GET /projects/:id (:90), both defaults halves, ping through the profile, 10-skills.sh untouched. §4.1 adds the missing host-verify contract — the stage ships a shell helper, an install-time gate and a tests/install-sh suite, i.e. exactly the container-vs-host surface destination-environment-verification.md §1 exists for; the umbrella had no contract on any of its six kickoffs (exit 2). Prior-art: skipped — dispatch-input documentation only, no new capability; resolves an ambiguity in an existing stage kickoff against in-repo source citations. * docs(beta-delivery-ux): S4 kickoff §7d — retract the unreachable host path, bind run-3 blockers §7d.0 retracts the §1 + §7a #1 instruction to re-read `~/code/aif-handoff/packages/api/src/routes/runtimeProfiles.ts` at stage entry. That path is host-only: the agent container mounts `$PROJECTS_DIR` (= `$PROJECTS_HOST_ROOT/projects`), so the repo root holding `packages/api/` sits one level above the mount. The instruction was unexecutable and run 3 guessed instead of parking — same class as state.md §2 decision 14. Replacement is an oracle the container can actually reach: the live aif API. §7d.1 pins two probes (required-field discovery + route-existence discovery) with the dispatcher's own 2026-08-09 outputs recorded for reproduction, not for trust. Run-3 blockers bound as corrections: §7d.2 `runtimeId`+`providerId` required in the create body; §7d.3 the ping route is `POST /runtime-profiles/validate` (the guessed `/<id>/v1/messages` → 404); §7d.4 §7b #1 still undelivered (the helper verifies reachability but never establishes it); §7d.5 companion install is prose in a field `engine.sh:18` early-returns past; §7d.6 (MAJOR) the suite mocks the non-existent endpoint and a body-blind 201, so §4.1 passed over §7d.2/§7d.3 — the stub becomes fail-closed with a paired-negative that fails on the run-3 helper, and `it.fails()`-as-delivery is banned. Adds T-BDU-D: unreachable instructed source → probe, else PARK; never substitute a plausible shape. Prior-art: skipped — kickoff prose revision only, no new capability; markdown is exempt from the capability-commit LOC triggers per CLAUDE.md. * docs(beta-delivery-ux): S4 kickoff §7d.1 — name the real base-URL var, and the container trap under it Own cold-QA of #1319 caught two defects in §7d.1. (1) It invented `$AIF`; the helper actually resolves `AIF_URL="${RUNTIME_BRIDGE_AIF_URL:-http://localhost:3009}"` (scripts/getff-glm-onebutton.sh:94), so the probe block now uses that name. (2) The more load-bearing half: that shipped default is unreachable from the agent container. Measured 2026-08-09 from aif-handoff-agent-1 — `http://localhost:3009/runtime-profiles` → curl exit 7 (000), `http://api:3009/runtime-profiles` → 200. `RUNTIME_BRIDGE_AIF_URL` is unset in the container, so an in-container run of the run-3 helper could not have reached aif at all, which is consistent with the suite mocking every call (§7d.6). §7d.1 now instructs the export for probing, explicitly forbids changing the shipped consumer default, and §4 item 1 must quote the base URL alongside any live end-to-end transcript. Prior-art: skipped — kickoff prose correction only, no new capability; markdown is exempt from the capability-commit LOC triggers per CLAUDE.md. * docs(beta-delivery-ux): S4 kickoff §7d.3/§7d.4 — fire my own falsifier; the ping is two halves §7d.3 as first written told the worker to route §7a #3's «one real minimal model call» through `POST /runtime-profiles/validate` and left a falsifier for the case where that endpoint is only a reachability check. The dispatcher fired that falsifier itself rather than shipping the trap: - for `transport: "api"`, `validateClaudeConnection` returns ok after checking only that `apiKey` and `baseUrl` are non-empty — no network call (aif-handoff packages/runtime/src/adapters/claude/index.ts:468-479). Live 2026-08-09: `/validate` against the Qwen3.8-Max-Preview profile returned ok:true sub-second. - `POST /runtime-profiles/models` is not a substitute — it returned a static Claude catalogue (Sonnet 4.6, Opus 4.6) for a Qwen profile, so it is not querying the provider. So §7a #3 and §7c #3 cannot both be satisfied by one aif REST call. §7d.3 now binds both halves — `/validate` for the route proof, plus one 1-token-scale completion against the `baseUrl` READ BACK from that response (not a hardcoded `$GLM_BASE_URL`, which was run 2's W-3 defect). Provenance is stated as dispatcher-authored and operator-overridable, matching §7b. §7d.4 gains an exact, value-free verifier for §7b #3: `hasApiKey` is `Boolean(resolved.apiKey)` and `resolved.apiKey` is `normalizeString(env[envVarName])` off the aif runtime's own process.env (packages/runtime/src/resolution.ts:426, :217-219), so `hasApiKey:true` in the §7d.3(1) response IS proof that §7b #1's outcome was achieved — without dereferencing the key, without argv exposure. Prior-art: skipped — kickoff prose correction only, no new capability; markdown is exempt from the capability-commit LOC triggers per CLAUDE.md. --------- Co-authored-by: Test <test@example.com>

@Profile

…with the three false greens fixed (#1323) * feat(beta-delivery-ux-s4): GLM one-button helper + aif-handoff guided-install helper + entry verification S4 (spec §4 A2 + A1) — the GLM executor tier connects with ONE human-entered key, and the factory profile UPGRADES the aif-handoff companion row from S1's detect+instruct declaration to a consented guided INSTALL. Three artefacts: - scripts/getff-glm-onebutton.sh — the one-button automation (detect | explain | provision). Executor of the flow = the consumer's in-session AI agent reading an INSTALL-FOR-AI step (the aider pattern). detect probes /runtime-profiles for a Z.ai-shape profile; explain prints the one explanation block (z.ai Coding Plan, env-file path); provision runs REST create + per-mode defaults + validation ping (real model call per §7a #3). KEY-HANDLING INVARIANT (§4 item 2 — binding): the helper references ONLY the env-var NAME ANTHROPIC_AUTH_TOKEN; the value lives only in the untracked env file the human creates. - setup.d/aif-handoff-guided-install.sh — the consented guided INSTALL for the factory profile. Sources bridge-guided.sh for bridge_diagnose (SSOT per dual-implementation-discipline.md §7); states: up (no-op, detect-first) | docker (consented git clone + docker compose up -d + 30s health wait) | native (instruct start) | absent (env-level degrade). Decline → env-level degradation is a DESIGNED SUCCESS PATH (kickoff §6 T-BDU-B). - docs/meta-factory/research-patches/2026-08-08-s4-glm-onebutton-entry-verification.md — the entry verification record. Bridge unreachable + schema inaccessible at build time (both explicit §7 park triggers); §7a resolutions implemented on operator authority (binding) with falsifiers UNRESOLVED at build time. engine.sh does NOT gate on @Profile: (comment-only convention — the helper is invoked from install.sh under PROFILE=factory). Prior-art: skipped — bash helpers under scripts/ and setup.d/, not packages/; not a capability-commit per CLAUDE.md definition (no new package.json dependency, no new file ≥50 LOC under packages/core/<new-dir>/, no new file ≥80 LOC under packages/). * feat(beta-delivery-ux-s4): wire S4 automation into install.sh factory gate + INSTALL-FOR-AI one-button step + tests Four wiring changes completing S4's scope: - setup.d/companions.manifest — aif-handoff row install_cmd upgraded from the S1 placeholder to the real pointer: setup.d/aif-handoff-guided-install.sh under --profile factory. The @Profile: factory marker (S1) is preserved; engine.sh contract unchanged (kind=external-service → print + return 0; actual handling routes through the helper invoked from install.sh). - install.sh — factory gate added after the setup.d layer loop (after line 1115). Gating matches setup.d/10-skills.sh:95 exactly: PROFILE=factory OR WITH_AIF_SUITE set. The helper is invoked via `bash ... || true` (decline → env-level degrade is a designed-success path). - INSTALL-FOR-AI.md — new "GLM executor one-button (factory profile ONLY)" subsection under the depth-profiles section. Documents the detect→explain→provision flow the consumer's in-session AI agent runs; carries the key-handling invariant sentence. - tests/install-sh/glm-onebutton.test.sh — 20-assertion firing test covering: (a) detect cases (present|missing|bridge-unreachable), (b) explain output (env-file path + env-var name + cost line + parent dir), (c) provision cases (happy path, step-A failure, preflight env-file missing), (d) key-handling invariant grep (no echo/printf/log/_log/_warn references the env-var name; indirect expansion present in curl header). Snapshot baselines byte-identical (companions.manifest is not part of the installed-tree fingerprint). Stage boundary respected: setup.d/10-skills.sh and .zcode/ are NOT touched (S5's sole-editor scope for skill-set wiring). packages/runtime-bridge is NOT touched (S5 vendoring scope). Prior-art: skipped — wiring edits + bash test; not a capability-commit per CLAUDE.md definition (no new package.json dependency, no new file ≥50 LOC under packages/core/<new-dir>/, no new file ≥80 LOC under packages/). * fix(beta-delivery-ux-s4): repair one-button helper per §7b/§7c/§7d — PUT /projects/:id, profile-routed validate, key-reachability preflight Round 3 rework of scripts/getff-glm-onebutton.sh — fixes the four §7c dispatcher defects from run-2 plus the §7b key-reachability gap (resolved 2026-08-09): - §7c #1: replace invented PATCH /project (singular, no id) with PUT /projects/:id using a full createProjectSchema body (jq-mutated from GET /projects response). Aligns with packages/runtime-bridge/src/cli/aifHttp.ts:96 which documents PUT /projects/:id as the only parallelEnabled write path. - §7c #2: write BOTH Task+Review halves of per-mode defaults, with distinct preserve-vs-park branches for the Plan half based on defaultPlanRuntimeProfileId read from GET /projects. Conflating them silently overwrites a consumer's existing top-tier or silently writes a null. - §7c #3: route the validation ping through POST $AIF_URL/runtime-profiles/validate with payload {profileId:$id} — proving key + model + route together via aif. Run-2 pinged the Z.ai vendor URL directly, bypassing the runtime the flow built. - §7c (implicit): drop the invented x-api-key header attribution. - §7b: print docker-compose env_file wiring instruction for the consumer/AI agent to apply (the aif runtime resolves ANTHROPIC_AUTH_TOKEN from its own process.env by NAME; a file at $GLM_ENV_FILE is invisible unless deployment loads it). Honor §7b #2 design-call: print, do NOT auto-patch deployment files. Key-handling invariant (kickoff §4 item 2) holds: helper references ONLY the env-var NAME ANTHROPIC_AUTH_TOKEN (line 31); value lives only in untracked env file (line 29); post-source reachability check at line 249 uses ${!GLM_ENV_VAR:-} indirection — never expands the value into a log/echo line. Pre-flight grep returns empty (rc=1). INSTALL-FOR-AI.md updated to reference the §7b key-reachability wiring the helper prints. Tests updated with happy-path stubs for GET /projects, POST /runtime-profiles, PUT /projects/:id, POST /runtime-profiles/validate, plus regression guards for all four §7c defects. 25/25 PASS. Hook: worker-container mirror at /tmp/fix-hook/pre-commit (per memory project_handoff_root_owned_pre_commit_err path 3 — mktemp substitutes for root-owned /tmp/pre-commit-err; substantive probes unchanged). markdownlint-cli2 + prettier skipped (npm install blocked in container); maintainer-side pre-commit + CI on PR branch carries those gates. Prior-art: skipped — refactor only, no new capability (round 3 repairs run-2's existing payload per dispatcher §7c/§7d resolutions; no new dep, no new module). * fix(beta-delivery-ux-s4): point aif-handoff clone URL at lee-to upstream + drop park language (§7d.2) Two changes per §7d.2 dispatcher resolution (host-verified 2026-08-09): setup.d/aif-handoff-guided-install.sh: - Default AIF_HANDOFF_REPO_URL → https://github.com/lee-to/aif-handoff.git (gh api confirmed 2026-08-09: sst-aif/aif-handoff returns 404; lee-to/aif-handoff returns 200). The sst-aif reference is retained only in the comment documenting the host-side verification rationale. - Park language ("PARKED — repo URL unverified") dropped — the §7d.2 resolution closes the run-1 park with host-verified evidence. - # shellcheck source= directive restored to the canonical setup.d/bridge-guided.sh form (run-2 had drifted it). docs/meta-factory/research-patches/2026-08-08-s4-glm-onebutton-entry-verification.md: - Prepended scope slug () as the first line per principle 10's research-patch annotation requirement. - Item 5 verdict updated from "PARKED — repo URL unverified" to "RESOLVED — default upstream is lee-to/aif-handoff" matching the §7d.2 host-verified resolution. Hook: worker-container mirror at /tmp/fix-hook/pre-commit (per memory project_handoff_root_owned_pre_commit_err path 3). Prior-art: skipped — refactor only, no new capability. * chore(beta-delivery-ux-s4): MINOR-4 manifest field + wire glm-onebutton.test.sh into audit-self shard A (§7d.3 #2) Two cleanup/wire changes per §7d dispatcher resolutions: setup.d/companions.manifest: - Added MINOR-4 resolution comment above the aif-handoff row recording that kind=external-service rows have an informational-only `install` field — setup.d/engine.sh companion_step returns 0 for this kind without consulting it. Actual install routes via install.sh → setup.d/aif-handoff-guided-install.sh (mirrors the existing runtime-bridge row precedent). Closes the §7d.4 MINOR-4 run-2 finding that the manifest carried placeholder install semantics. .github/workflows/audit-self.yml: - Added `bash tests/install-sh/glm-onebutton.test.sh` step to the shard A sequence (after the bridge-guided test). Closes §7d.3 #2 — the new test existed but was not wired into CI. Indentation 6/8 spaces matches adjacent entries per ci-tool-pinning §1 (no version-pinned tool install needed — bash is system-provided). Hook: worker-container mirror at /tmp/fix-hook/pre-commit (per memory project_handoff_root_owned_pre_commit_err path 3). Prior-art: skipped — refactor only, no new capability. * fix(beta-delivery-ux-s4): the one-button flow could not complete a single live run — three false greens Round-4 review found the shipped flow 400s on its first live call while its own suite reported 25/25 green. Three independent false greens, each verified against the live aif on 2026-08-09. 1. Create body omitted the two schema-REQUIRED fields. `createRuntimeProfileSchema` requires `runtimeId` and `providerId` (no .optional(), no .nullable()). Firing the helper's exact body: POST /runtime-profiles -d '{"name":…,"defaultModel":…,"apiKeyEnvVar":…,"baseUrl":…}' → HTTP 400 ZodError, path ["runtimeId"] and ["providerId"] With `curl -sf`, step A therefore always returned non-zero and the flow always fell back to guided-manual — an objective-3 MISS on every run, by the kickoff's own §4 item 5. 2. Transport was unset, which voided step C entirely. Omitting `transport` resolves the profile to SDK, and for SDK transport validateClaudeConnection returns ok unconditionally ("using session auth"). Measured against a non-persisted inline profile: no transport → {"ok":true, "transport":"sdk", "hasApiKey":false} ← passes with NO key transport=api → {"ok":false, "transport":"api", "hasApiKey":false, "message":"Missing API key (expected env var: …)"} ← real gate One missing field was the difference between a ping that cannot fail and a ping that gates. 3. Step C read the verdict from the exit code, but /runtime-profiles/validate answers HTTP 200 even when validation fails — the verdict is in `.ok`. Measured: `HTTP 200 {"ok":false,"message": "Missing API key …"}` and `curl -sf` exits 0 on it. Step C now parses `.ok`. Also corrected: comments claiming the validate call performs "profile resolution + key lookup + model call" and that it "returns auth-error" — it does neither. §7a #3's real model call is now an explicit §7 PARK with both options and the binding each one breaks (argv exposure of the key vs. no aif endpoint that completes through a stored profile), instead of being reported as delivered. §7b #1 is stated honestly rather than claimed closed: the helper prints wiring instructions and does not wire. What changed is that an un-applied instruction is now a hard MISS at step C instead of a green run — detection, not wiring, and the comment says so. Test suite made fail-closed (25 → 33 checks). The stub had two defects that hid the above: - the /validate arm was DEAD CODE — `*"/runtime-profiles"*` matched first, so step C was tested against the CREATE response. shellcheck flags this automatically (SC2221/SC2222) and fires on the original file; it simply is not pointed at tests/install-sh/** today. - the create arm answered success regardless of body, hiding defect 1. Now: ordered arms, an explicit endpoint allowlist (each entry dated by live probe), a body rule mirroring the live 400, and paired negatives N1/N1b/N2/N2b/N3 that fail on run 3's invented ping path and run 4's create body. Gates shown to discriminate, not merely to pass — each fix reverted in turn: remove transport → PASS=32 FAIL=1 remove runtimeId/providerId → PASS=30 FAIL=3 drop the .ok parse → PASS=30 FAIL=3 restored control → PASS=33 FAIL=0 refresh-covers-full-delivery 6/6 PASS; shellcheck -S warning clean. Prior-art: skipped — bug fix to an existing capability plus test hardening, no new capability. --------- Co-authored-by: Test <test@example.com>

…harvested, 995e9c dead-end) (#1324) §7e was written before two things happened: run 3 got harvested to origin, and a duplicate dispatch (995e9c) landed a dead-end branch in the container. §7f records both so run 4 cannot repeat the base-selection mistake that produced the duplicate. - §7f.0 — run 3 is now PR #1322 on origin (`feature/beta-delivery-ux-e65989`, `53fce45f51`); its body carries the round-3 watch-list W-1..W-6. - §7f.1 — `feature/beta-delivery-ux-995e9c` is a dead end: do not base on it, do not delete it (T18). Its one salvageable item is the audit-self shard-A wiring, and only after the §7e.6 stub fix lands. - §7f.2 — measured from `aif-handoff-agent-1`: `POST /runtime-profiles/:id/v1/messages` → 404, the exact path `scripts/getff-glm-onebutton.sh:53` targets. §7e.3's two-half split is the only reachable form of §7a #3, not a preference. - §7f.3 — the §2.4 2-consecutive-REVISE cap counts unchanged scope; §7d/§7e are each a scope change, so run 4 is legitimate. Prior-art: skipped — kickoff dispatch-facts addendum, markdown only, no new capability Co-authored-by: Test <test@example.com>

…, name frozen against the live registry (#1336) * docs(beta-delivery-ux): S2 kickoff §8a — operator resolutions of the six #1284 parks (re-dispatch rev) Records the operator's 2026-08-08 batch: Park-1 schema = flat Candidate A + description field; Park-2 format = JSON (F-B' shell-sourced rejection recorded with the WHY); Park-3 economy review = whole line on executor tier, aif cap 1 iteration, external cold fidelity mandatory; Park-4 CC detection = CLAUDE_CODE_SESSION_ID env-presence capability check; Park-5 conditional park did not fire (live runtime-profiles probe, no substring collision); Park-6 ship the full functional set (create-worktree.sh + worktree-node-modules.sh + link-coordination.sh) under the same $1j gate, fresh-consumer smoke AC added. Unblocks the S2 re-dispatch (A4 presets + A5 status + A9 parts 2-3). * docs(beta-delivery-ux): S4 kickoff §7b/§7c — close run-2 KICKOFF-AMBIGUOUS + host-verify contract Run 2 (aif task 92bf0019) returned FIDELITY STOP with KICKOFF-AMBIGUOUS (PR #1300). The kickoff's §7a never states how the aif runtime acquires the key VALUE: #1 stores only the env-var NAME in the profile, #4(ii) writes the value to ~/.config/getff/glm.env, and #3 requires the validation ping to run through the created profile. Verified against source: aif resolves the key from its own runtime process.env by that name (aif-handoff packages/runtime/src/resolution.ts:217-219, :247), and that env comes from the compose env-file (docker-compose.yml:15,59,94) — so glm.env is a dead end as written and #3 is unreachable without dereferencing the value, which #1 forbids. §7b closes it by binding the OUTCOME (value reachable in the aif runtime process env under the §7a #1 name, verified before the ping, honest objective-3 MISS on failure) while leaving the wiring mechanism to the worker as a §7 technical fork. It is marked as a dispatcher-authored completion, not an operator decision — the operator can override it. §7c carries run 2's settled corrections: PUT /projects/:id with a full body (aifHttp.ts:96) instead of the invented PATCH /project, no GET /projects/:id (:90), both defaults halves, ping through the profile, 10-skills.sh untouched. §4.1 adds the missing host-verify contract — the stage ships a shell helper, an install-time gate and a tests/install-sh suite, i.e. exactly the container-vs-host surface destination-environment-verification.md §1 exists for; the umbrella had no contract on any of its six kickoffs (exit 2). Prior-art: skipped — dispatch-input documentation only, no new capability; resolves an ambiguity in an existing stage kickoff against in-repo source citations. * docs(beta-delivery-ux): S4 kickoff §7d — retract the unreachable host path, bind run-3 blockers §7d.0 retracts the §1 + §7a #1 instruction to re-read `~/code/aif-handoff/packages/api/src/routes/runtimeProfiles.ts` at stage entry. That path is host-only: the agent container mounts `$PROJECTS_DIR` (= `$PROJECTS_HOST_ROOT/projects`), so the repo root holding `packages/api/` sits one level above the mount. The instruction was unexecutable and run 3 guessed instead of parking — same class as state.md §2 decision 14. Replacement is an oracle the container can actually reach: the live aif API. §7d.1 pins two probes (required-field discovery + route-existence discovery) with the dispatcher's own 2026-08-09 outputs recorded for reproduction, not for trust. Run-3 blockers bound as corrections: §7d.2 `runtimeId`+`providerId` required in the create body; §7d.3 the ping route is `POST /runtime-profiles/validate` (the guessed `/<id>/v1/messages` → 404); §7d.4 §7b #1 still undelivered (the helper verifies reachability but never establishes it); §7d.5 companion install is prose in a field `engine.sh:18` early-returns past; §7d.6 (MAJOR) the suite mocks the non-existent endpoint and a body-blind 201, so §4.1 passed over §7d.2/§7d.3 — the stub becomes fail-closed with a paired-negative that fails on the run-3 helper, and `it.fails()`-as-delivery is banned. Adds T-BDU-D: unreachable instructed source → probe, else PARK; never substitute a plausible shape. Prior-art: skipped — kickoff prose revision only, no new capability; markdown is exempt from the capability-commit LOC triggers per CLAUDE.md. * docs(beta-delivery-ux): S4 kickoff §7d.1 — name the real base-URL var, and the container trap under it Own cold-QA of #1319 caught two defects in §7d.1. (1) It invented `$AIF`; the helper actually resolves `AIF_URL="${RUNTIME_BRIDGE_AIF_URL:-http://localhost:3009}"` (scripts/getff-glm-onebutton.sh:94), so the probe block now uses that name. (2) The more load-bearing half: that shipped default is unreachable from the agent container. Measured 2026-08-09 from aif-handoff-agent-1 — `http://localhost:3009/runtime-profiles` → curl exit 7 (000), `http://api:3009/runtime-profiles` → 200. `RUNTIME_BRIDGE_AIF_URL` is unset in the container, so an in-container run of the run-3 helper could not have reached aif at all, which is consistent with the suite mocking every call (§7d.6). §7d.1 now instructs the export for probing, explicitly forbids changing the shipped consumer default, and §4 item 1 must quote the base URL alongside any live end-to-end transcript. Prior-art: skipped — kickoff prose correction only, no new capability; markdown is exempt from the capability-commit LOC triggers per CLAUDE.md. * docs(beta-delivery-ux): S4 kickoff §7d.3/§7d.4 — fire my own falsifier; the ping is two halves §7d.3 as first written told the worker to route §7a #3's «one real minimal model call» through `POST /runtime-profiles/validate` and left a falsifier for the case where that endpoint is only a reachability check. The dispatcher fired that falsifier itself rather than shipping the trap: - for `transport: "api"`, `validateClaudeConnection` returns ok after checking only that `apiKey` and `baseUrl` are non-empty — no network call (aif-handoff packages/runtime/src/adapters/claude/index.ts:468-479). Live 2026-08-09: `/validate` against the Qwen3.8-Max-Preview profile returned ok:true sub-second. - `POST /runtime-profiles/models` is not a substitute — it returned a static Claude catalogue (Sonnet 4.6, Opus 4.6) for a Qwen profile, so it is not querying the provider. So §7a #3 and §7c #3 cannot both be satisfied by one aif REST call. §7d.3 now binds both halves — `/validate` for the route proof, plus one 1-token-scale completion against the `baseUrl` READ BACK from that response (not a hardcoded `$GLM_BASE_URL`, which was run 2's W-3 defect). Provenance is stated as dispatcher-authored and operator-overridable, matching §7b. §7d.4 gains an exact, value-free verifier for §7b #3: `hasApiKey` is `Boolean(resolved.apiKey)` and `resolved.apiKey` is `normalizeString(env[envVarName])` off the aif runtime's own process.env (packages/runtime/src/resolution.ts:426, :217-219), so `hasApiKey:true` in the §7d.3(1) response IS proof that §7b #1's outcome was achieved — without dereferencing the key, without argv exposure. Prior-art: skipped — kickoff prose correction only, no new capability; markdown is exempt from the capability-commit LOC triggers per CLAUDE.md. * docs(beta-delivery-ux): R1 stage kickoff — npm release mechanics (A6) R1 is the last unstarted stage of the umbrella (S1-S5 all merged). This is its dispatch input, scoped from the spec (§4 A6 `:278-284`, §11 F-C' `:467-468`) and the binding input `launch-preannounce-track/s6-u10-handoff.md`. Three things the kickoff decides, because a worker cannot: 1. Dispatchability (§0.1) — R1 IS aif-dispatchable. `private: true` STAYS on `packages/core/package.json` for the whole stage; npm refuses to publish a private package, so the publish guard is mechanical rather than a reminder (attention-is-not-a-mechanism §1). Verified `npm pack --dry-run` works on a private package, so the tarball loop runs with the guard armed. 2. Name architecture (§0.2) — probed the live registry, not the spec's prose: unscoped `getff` ALREADY EXISTS at 0.0.1, published 2026-06-23 by the maintainer as a reservation stub. So the registry-side freeze is done, 0.0.1 is spent as a version, and `@getff` SCOPE ownership stays unproven (a 404 on `@getff/core` proves the package is free, not the scope) — that is an entry probe with a STOP attached. 3. Scope boundary (§0.3) — R1 does NOT execute the `@rules-as-tests/*` rename. Measured blast radius 33 tracked files / 55 occurrences; umbrella §3 puts the rename in U9 (post-announce), and the acceptance gate is name-independent because `npm i <tarball>` installs by path. Also recorded: two gates of the binding input's §0 are SUPERSEDED by the 2026-07-23 amendment (U9 no longer the rename window; the private-drop ordering), and the honest gap R1 does not close — U10's `npx getff init` gate needs a package with `bin: getff`, which does not exist in this repo and which A6 does not ask R1 to build. Host-verification contract declared (no opt-out): the new tarball cell plus the existing pnpm start cell, which is the file-copy fallback the §4 rollback doctrine requires to keep working. `scripts/host-verify.sh --list` parses both, exit 0. Prior-art: skipped — dispatch-input authoring only; a markdown kickoff adds no capability, no dependency, and no code module (CLAUDE.md doc-file carve-out). --------- Co-authored-by: Test <test@example.com>

…cted, model proof measured (#1340) The Round-4 fidelity audit emitted KICKOFF-AMBIGUOUS on §7e.3 and a DECISION-NEEDED fork (weaken §7a #3's completion proof, or weaken §2 constraint 1's key-handling invariant). Both options were unnecessary: the fork rested on §7f.2's claim that every aif-side completion route is closed, and that claim came from a single probe. Corrected here, each against a measurement rather than a re-reading: - §7f.0 — «base on run 3» pointed at PR #1322, which is still OPEN and red. The S4 work is on `staging` as `6ea6011560` (PR #1323); `git merge-base --is-ancestor 6ea6011 49eaafe` → YES, so a staging-based run 4 satisfies §7e.0 with a superset, not a regression. - §7f.1 — `…-995e9c` was labelled a dead end. The BRANCH is; its CONTENT was repaired dispatcher-side, pushed as `harvest/s4-995e9c` and merged as #1323, shard-A wiring included. - §7f.2 — `POST /runtime-profiles/<id>/v1/messages` → 404 was measured and stands, but the conclusion drawn from it («the only reachable form of §7a #3») generalised one probe into a negative-existence claim, missing project invariant #3. `POST /chat/sessions` → 400: the route exists. `chat.ts:923-937` project-scope-validates `runtimeProfileId`; `chat.ts:1275` is the completion endpoint; §7e.4 already established the key is read off aif's own `process.env`, so aif makes the call and the helper never handles the value. Both constraints hold. Proven end-to-end with one operator-approved billed call: `assistantMessage` returned, `costUsd:0.117219`, `runtime.profileId:53eca24c-…`, and that profile's `lastUsageAt` moved `09:15:11.263Z` → `09:51:42.017Z`. Cost is recorded as ~$0.12/run, not token-scale, because aif injects project context — the suite must stub it and only host-verify may fire it live. What is still unproven is stated as run 5's task rather than implied closed: the probe used an already-created profile on a warm aif, not a cold install immediately after the helper's own `POST /runtime-profiles`. The file lands at exactly 600 lines. `.husky/pre-commit` carries no kickoff exemption (only `md-line-gate.sh` does, and its glob is `kickoff.md`, which never matches `kickoff-s4.md`), so the next correction to this file needs a split, not a trim. Prior-art: skipped — documentation only; corrects three falsified facts in an existing kickoff section and adds no capability, no dependency, and no new module. Co-authored-by: Test <test@example.com>

…un did not decide night-mode delta item 1 says an unattended session resolves technical forks autonomously and LOGS genuine owner forks instead of deciding them. Two blocked S4 from closing complete: 1. Spend authorisation for the §7a #3 model proof. The route is measured and proven; step D is ordinary work, but proving it on a cold install needs another real billed call, and the one the operator authorised is spent. Cost is ~$0.12 per run, not token-scale. 2. The terminal-token contract when a binding objective is undelivered — DONE vs FAILED vs a new PARTIAL token. Changing it touches what INSTALL-FOR-AI.md:184 tells consumer agents to report. Also records what was NOT escalated and why: the round-4 KICKOFF-AMBIGUOUS dissolved under measurement rather than needing an operator amendment, so it was resolved autonomously. Prior-art: skipped — a decisions log for an existing umbrella; no capability, dependency, or module.

…and correct how it pins Operator authorised the spend, so the round-5 BLOCKER is closed by building the thing rather than by re-parking it. Step D makes one real minimal completion through the profile the run just created, which is what turns `GLM_PROVISION: DONE` from a claim about resolution into a claim about the vendor accepting the key. Fork 2 dissolved rather than being decided: `DONE` became true, so no new terminal token was needed. The first implementation was WRONG, and the way it was wrong is the point. Step D was written as a single `POST /chat` carrying `runtimeProfileId` — `chatRequestSchema` accepts that field, and the call appeared to pass because the echoed `runtime.profileId` came back matching. It matched by coincidence: the pinned profile was already the project's default. Fired against a freshly created profile, the completion ran on the project default and echoed THAT back. `chat.ts:1336` is why — `POST /chat` resolves the profile from the chat SESSION (`existingSession?.runtimeProfileId ?? null`) and never reads the chat body's field when opening a new conversation. Delivered form is therefore two calls: `POST /chat/sessions` pins the profile, `POST /chat` sends the completion to that session. Without the cold-profile probe this would have shipped as a proof that silently validated whatever the project already defaulted to. Failure modes are covered rather than assumed: N7 (call unreachable), N8 (completion billed to a DIFFERENT profile), N9 (zero-token "completion"), plus a §2 constraint 1 assertion that the request body carries the profile id and never the key value. Removing the session pin from the helper turns 3 assertions red, so the wiring is exercised, not decorated. Suite 53 → 60, green on the host. The park block is replaced by a record of why it was wrong: it generalised a handful of 404s into "every aif-side completion route is closed", which is a negative-existence claim from a partial probe. Option A's rejection stands (T16 — `curl --header @-` moves argv exposure but still needs the helper to read the value); Option B was rejected on a false premise and is the delivery path. HONEST GAP, stated in the code and in the decisions log: the full chain against a profile THIS helper created could not be exercised on the verifying host. That aif has `ZAI_API_KEY` in its process env and no `ANTHROPIC_AUTH_TOKEN` (measured — names only, never values), so a helper-created profile returns CHAT_AUTH_ERROR there regardless of correctness. On a consumer machine the §7b wiring supplies that var, which is the case the helper targets — but the first consumer run is still the first end-to-end execution, and a `FAILED step-D` there means the wiring did not land, not that the proof is broken. INSTALL-FOR-AI.md now says DONE means the vendor accepted the key, that the call is billed, and what a `FAILED step-D` line means. Prior-art: skipped — completes an existing helper's declared objective using the aif REST API it already speaks; no new capability, no dependency, no new module.

…ertion that could not fail Round-6 cold audit on `34ccc8cece` confirmed both round-5 carry-overs genuinely closed, and found three MAJORs of its own. All three are closed here. MAJOR-1 — the §2 constraint 1 assertion on the step-D surface was VACUOUS, and it is the third time this branch has applied a principle to one branch and not its neighbour. It accumulated request argv into the shell variable `_ND_SEEN`; the helper calls curl inside `$( )`, so every stub call ran in its own subshell and the variable was empty at assertion time. The suite reported the key never reaches the request while a deliberately injected key value sailed straight past. The file-based fix for exactly this subshell trap was already documented and applied twelve lines away, for the session pin. Now: capture goes to a file, a guard asserts the capture is NON-EMPTY (so the check cannot pass on nothing), and a paired negative proves the detector fires on a request that does carry the value. Verified by injecting a leak into step D's body — the suite goes red, and green again when reverted. Same block, second defect: the create arm answered 2xx with no body inspection, unlike every sibling stub — W-4's tell verbatim; it now calls `_stub_create_body_ok`. MAJOR-2 — `INSTALL-FOR-AI.md` promised the model call costs "a few cents at most". This branch's own measurement is `costUsd:0.117219`, and its decisions log says "not token-scale". A consumer-facing ceiling falsified by the author's own number in the same commit range is exactly the false green this stage keeps producing. Corrected to the measured figure, and the sentence now also states the two things it silently omitted: the helper runs `docker compose up -d` against the consumer's aif (restarting their services), and every run leaves a chat session titled "getff GLM provisioning proof" in their project. MAJOR-3 — the research patch still carried "§7e.3 — model proof: PARKED", "the fork is genuine: two binding constraints in conflict", and a run-5 §1.7 note calling the DONE gap "NOT fixed" — all false at HEAD, and the newest dated section was among them, so a reader following the file's own ordering would land on withdrawn claims. The folder is append-only, so a superseding addendum now closes each of the three by name at the end of the file. MINOR — narrowed, but not the way the audit proposed. The audit read "session pinning measured live" as broader than the evidence, having seen only completions that ran on the project default. The discriminating observation exists and is now named in the code: a session pinned to a NON-default, freshly created profile returned CHAT_AUTH_ERROR. Had pinning fallen back to the default, that call would have SUCCEEDED, since the default is a working profile — so the error is positive evidence the pinned profile was used. NOT closed — the audit's KICKOFF-AMBIGUOUS stands and is an owner question: §7a #3 binds a "1-token-scale" call, but the only reachable route injects project context, so every consumer provision bills ~$0.12. Recurring per-consumer spend was never decided by the kickoff. Suite 60 → 62, green on the host. Prior-art: skipped — test-material repair plus doc corrections to an existing helper; no new capability, no dependency, no new module.

… on step D The round-6 audit flagged KICKOFF-AMBIGUOUS: §7a #3 binds a '1-token-scale' model call, but the only reachable completion route injects project context — measured 39058 input tokens and costUsd:0.117219 for a one-word prompt. Step D runs on every provision, so that is a recurring charge on each consumer's plan rather than a one-off verification cost. Logged, not decided: nothing in the code settles it, and whether ~$0.12 per provision is acceptable is a product call about other people's money. INSTALL-FOR-AI.md states the measured figure to the consumer either way, so nobody is billed by surprise while it is open. Prior-art: skipped — decisions log entry for an existing umbrella; no capability, no dependency.

…ree false greens closed (#1341) * feat(beta-delivery-ux): S4 run-4 — §7e.4 wiring + §7e.3 refreshed park + cold-QA revisions Run 4 closes the run-3 #warning-nobody-reads shape on the GLM one-button helper. The helper now MAKES the key reachable in the aif runtime's process env (best-effort W1) and the §7e.4 verifier (`.profile.hasApiKey` from POST /runtime-profiles/validate) is a binding gate, not a printed warning. Three substantive deliverables (T1–T5 of the dispatched plan): 1. §7e.4 wiring (W1 + W2). W1: when $AIF_HANDOFF_CHECKOUT/docker-compose.yml exists and no unmarked override collides, the helper writes a marker-bearing docker-compose.override.yml that adds $GLM_ENV_FILE to each detected service's env_file list. Idempotent (marker grep) + collision-back-off (unmarked override → rc=2 → W2 fallback). W2: instruction + the §7e.4 binding gate (hasApiKey:false → FAILED step-C key-unreachable, objective-3 MISS). Live-fired evidence in the PR body. 2. §7e.3 model-proof — PARK refreshed. Re-probed every plausible aif completion route (404 across /runtime-profiles/<id>/{v1/messages,chat/completions,...}, root-level /chat /prompt /infer, /openapi.json empty). Tested third options for the helper-direct route: `curl --header @-` and `curl --config -` both avoid argv exposure AND temp file, but require the helper to expand the env var — i.e. read the value — which §2 constraint 1 forbids ("automation reads the ENV VAR NAME (never the value) when calling the model"). Genuine fork between two binding constraints; parked per §7 with T16 verdict on the curl pattern (solves argv but presupposes the script reads the value). 3. .profile.hasApiKey parsing fix. Run 3 assumed top-level; live aif (probed 2026-08-09) nests under .profile: {"ok":true,"profile":{"hasApiKey":...}}. Helper now tries .profile.hasApiKey first (live-accurate) then top-level .hasApiKey (older-aif fallback) then .ok-only (oldest aif). Three paired-negatives cover each branch: N1 happy-path (live nested), N4 hasApiKey:false (FAILED step-C key-unreachable), N5 defensive-fallthrough (older aif without the field anywhere). Cold-QA revisions (T19 MAJOR + 2 MINOR, addressed before commit): - M1: stubs now use the LIVE nested .profile.hasApiKey shape, not top-level. Previously the primary parse path (the one that fires against live aif) was untested — only the elif fallback was. The T-BDU-E shape the §7e.6 rule exists to prevent. Fixed. - m1: N2b description corrected (it tests arm-ordering, not defensive parse); N5 added for the actual defensive-fallthrough branch. - m2: W1-write test comment corrected — `docker() { return 127; }` is a function stub, `command -v docker` finds it (returns 0), so the helper takes the "docker compose up -d failed" branch, not the "docker not in PATH" branch. Comment now describes the actual path. §4.1 host-verify contract CANNOT be exercised in-container (destination-environment-verification rule's whole point is container≠host). The contract command `bash tests/install-sh/glm-onebutton.test.sh` runs green (47/0) in-container; the host re-run is pending at PR merge. §1.7 Forward-check: complies with §7 park-don't-guess (model proof parked, not guessed); §7e.4 (verifier wired, not warning); §7e.6 (fail-closed + paired-negatives N1-N5); §2 constraint 1 (helper never expands the value); no-paid-llm-in-ci.md (no LLM in CI); doc-authority-hierarchy.md §2-§3 (research-patch is folder-level authority, no per-file header needed); language-discipline.md (all English machinery). §1.7 Backward-check: Class of this change = "shell helper that auto-writes a docker-compose.override.yml with an env_file path, gated by a binding validate-response field". Surfaces where class-X occurs, per cold-agent sweep (backward-sweep-auditor, PR-blind): - scripts/getff-glm-onebutton.sh — IN SCOPE (the changed file) - setup.d/aif-handoff-guided-install.sh + setup.d/bridge-guided.sh + packages/runtime-bridge/scripts/* — SWEPT-CLEAN (no parallel docker-compose.override.yml writer exists; grep `docker-compose|env_file` across all deployment-writing helpers returns zero writers) - POST /runtime-profiles/validate consumers — SWEPT-CLEAN (single consumer: the changed file) - tests/install-sh/* fail-closed stub pattern — SWEPT-CLEAN (mirror of bridge-guided.test.sh; no other /validate consumer) - INSTALL-FOR-AI.md + docs/runtime-bridge-setup.md consumer wiring instructions — SWEPT-CLEAN (INSTALL-FOR-AI.md:180 updated for W1; docs/runtime-bridge-setup.md:190 uses the same filename for a different purpose (operator-axis volume bind-mounts, not env_file injection); W1 collision-back-off handles the filename-sharing correctly). Overall backward-sweep verdict: GO. T-BDU-C/D/E tells countered: every aif API claim in the diff is established by a command output in this PR body or the research-patch's run-4 update. The suite's green run is disambiguated by the live end-to-end probe (T4) — both the W1 wire path (override written with marker covering agent+api services) and the §7e.4 verifier firing on the throwaway key (validate returned ok:false "Missing API key" because the env file was visible to the helper but not to aif's process env — exactly the gap the verifier exists to detect). Prior-art: skipped — refactor only, no new capability (no new dependency, no new file ≥80 LOC under packages/, no new packages/core/<new-dir>/ entry; the helper script + test suite + research-patch + INSTALL-FOR-AI.md edits are all additions to existing surfaces). Refs: .ai-factory/plans/feature-beta-delivery-ux-5cfe83.md (run-4 dispatched plan) * fix(beta-delivery-ux): S4 run-5 — close both round-4 MAJORs, and fix what caused one of them Round-4 cold fidelity audit returned REVISE on `91f99d06be` with two MAJORs and no BLOCKER. Both are closed here, and the second turned out to have a fixable root cause rather than an unavoidable degradation. MAJOR-1 (W-4 REINTRODUCED) — the N5 block shipped a fail-OPEN `curl()` stub: its create arm answered 200 with no body inspection and its catch-all returned rc=0 for any path, so it would have green-lit run 3's invented `/v1/messages` ping and a create body missing the required ids. The base stub and the round's own N4 stub both fail closed; N5 was a copy that drifted. Fixed in place, and — because a copy that drifts is the actual defect class — a structural gate now asserts the invariant over this file's own source: every case-based `curl()` stub must fail closed on an unallowlisted path. A prose reminder rots under exactly the fatigue that causes the drift (attention-is-not-a-mechanism.md §1). Its paired negative proves non-vacuity, and the gate earned its keep immediately by flagging its own first fixture, which is why the fixture is now assembled programmatically instead of written as a heredoc. MAJOR-2 — on a failed per-mode-default PUT the helper warned an objective-3 MISS and continued, then printed `GLM_PROVISION: DONE`. `INSTALL-FOR-AI.md:184` tells the consumer's agent to report that line verbatim, so a missed binding objective surfaced as success. §2 constraint 4 is explicit that a degrade to manual steps is a MISS, not a neutral fallback: the branch now emits `GLM_PROVISION: FAILED step-B per-mode-defaults` and returns non-zero. N6 is the paired negative; reverting the fix turns it red on 3 assertions. ROOT CAUSE of that MISS, measured rather than reasoned about. The research patch attributed the `rc=22` to a schema-shape incompatibility. One layer down: PUT /projects/<id>, GET body verbatim -> 400 expected number, received null path ["plannerMaxBudgetUsd"] same body, null *MaxBudgetUsd keys omitted -> 200 `createProjectSchema` declares the four budget fields `z.number().positive().optional()` — optional but NOT nullable — while `GET /projects` returns them as `null`. The helper mutated the GET body and fed those nulls back. So objective 3 was missing on every consumer for a reason unrelated to the fields it was setting. One jq filter drops null-valued `*MaxBudgetUsd` keys; the 200 above was read back with `autoQueueMode`, both profile defaults and every other value unchanged. Objective 3 is met, not degraded. Suite: 47 → 53 assertions, green on the host (§4.1). The §7e.3 model proof stays parked: the route is now proven reachable (kickoff §7f.2), but demonstrating it on a cold install costs a real billed call and that spend is the operator's to authorise. Prior-art: skipped — bug fix to an existing helper plus test material for it; no new capability, no dependency, no new module. * fix(beta-delivery-ux): S4 run-5 — round-5 audit MINORs, and the withdrawn park stated as undelivered Round-5 cold fidelity audit returned STOP on `86b01848ed`. This commit closes the two MINORs and stops the park from resting on a rationale the kickoff has withdrawn. The BLOCKER and the MAJOR are deliberately NOT closed here — see below. MINOR-1 — the §7e.6 meta-scanner caught only `return 0`. A case arm with no explicit `return` inherits the exit status of its last command, so `*) printf 'x' ;;` is exactly as fail-open while reading as if it were safe. Both shapes are now caught. The scanner is also scoped to `curl()` blocks, which is its declared population: without that, the `_wrong_order` hazard reproducer — not a stub — would be graded against a stub invariant and flagged falsely. Verified: the widened rule flags the new shape (`FLAGGED line 3` on a fixture) and the real file stays clean at 53/53. MINOR-2 — the research patch's §1.7 note was labelled `(run-4)` and the run-5 addendum carried none. Added, including the part that does not flatter this branch: the backward-check enumerates the false-green shape across `do_provision` and records that the terminal `DONE` printf is a GAP-FOUND that is NOT fixed. Park text — the block claimed «Root-level completion routes also 404: /chat, …» and concluded «Option B is a dead-end at the aif API surface as it stands today». Both are false, as §7f.2 records: `POST /chat/sessions` → 400 (exists), `POST /chat` returns a real billed completion, and `chat.ts:923-937` project-scope-validates `runtimeProfileId`. The word «PARKED» is withdrawn with the rationale: a park is an open question, and settled-but-unbuilt work called a park is the shape §7d.2 already ruled against. It now reads as undelivered work, named as such. NOT closed, and why: - **BLOCKER (§7a #3 model proof absent).** Step D is buildable now that the route is measured. Verifying it on a cold install costs a real billed call (~$0.12/run), and that spend was authorised once, for the probe already spent. Logged as an owner fork. - **MAJOR (`GLM_PROVISION: DONE` over an undelivered objective).** §7e.3 makes either half of the proof failing an objective-3 MISS, so the honest token is not `DONE` — but changing it changes the consumer-facing contract `INSTALL-FOR-AI.md:184` tells the consumer's agent to report verbatim, and the alternatives (always-FAILED, or a third `PARTIAL` token) are an owner call, not an implementation detail. Recorded in code and in the patch instead of being papered over; the inconsistency with the step-B fix twelve lines above is stated, not hidden. The night-mode session-local fidelity cap of 2 rounds is reached (round 4 REVISE, round 5 STOP), so this branch is pushed and NOT merged. Prior-art: skipped — test-material widening plus comment corrections to an existing helper; no new capability, no dependency, no new module. * docs(beta-delivery-ux): S4 — log the two owner forks the unattended run did not decide night-mode delta item 1 says an unattended session resolves technical forks autonomously and LOGS genuine owner forks instead of deciding them. Two blocked S4 from closing complete: 1. Spend authorisation for the §7a #3 model proof. The route is measured and proven; step D is ordinary work, but proving it on a cold install needs another real billed call, and the one the operator authorised is spent. Cost is ~$0.12 per run, not token-scale. 2. The terminal-token contract when a binding objective is undelivered — DONE vs FAILED vs a new PARTIAL token. Changing it touches what INSTALL-FOR-AI.md:184 tells consumer agents to report. Also records what was NOT escalated and why: the round-4 KICKOFF-AMBIGUOUS dissolved under measurement rather than needing an operator amendment, so it was resolved autonomously. Prior-art: skipped — a decisions log for an existing umbrella; no capability, dependency, or module. * feat(beta-delivery-ux): S4 — deliver §7a #3's model proof as step D, and correct how it pins Operator authorised the spend, so the round-5 BLOCKER is closed by building the thing rather than by re-parking it. Step D makes one real minimal completion through the profile the run just created, which is what turns `GLM_PROVISION: DONE` from a claim about resolution into a claim about the vendor accepting the key. Fork 2 dissolved rather than being decided: `DONE` became true, so no new terminal token was needed. The first implementation was WRONG, and the way it was wrong is the point. Step D was written as a single `POST /chat` carrying `runtimeProfileId` — `chatRequestSchema` accepts that field, and the call appeared to pass because the echoed `runtime.profileId` came back matching. It matched by coincidence: the pinned profile was already the project's default. Fired against a freshly created profile, the completion ran on the project default and echoed THAT back. `chat.ts:1336` is why — `POST /chat` resolves the profile from the chat SESSION (`existingSession?.runtimeProfileId ?? null`) and never reads the chat body's field when opening a new conversation. Delivered form is therefore two calls: `POST /chat/sessions` pins the profile, `POST /chat` sends the completion to that session. Without the cold-profile probe this would have shipped as a proof that silently validated whatever the project already defaulted to. Failure modes are covered rather than assumed: N7 (call unreachable), N8 (completion billed to a DIFFERENT profile), N9 (zero-token "completion"), plus a §2 constraint 1 assertion that the request body carries the profile id and never the key value. Removing the session pin from the helper turns 3 assertions red, so the wiring is exercised, not decorated. Suite 53 → 60, green on the host. The park block is replaced by a record of why it was wrong: it generalised a handful of 404s into "every aif-side completion route is closed", which is a negative-existence claim from a partial probe. Option A's rejection stands (T16 — `curl --header @-` moves argv exposure but still needs the helper to read the value); Option B was rejected on a false premise and is the delivery path. HONEST GAP, stated in the code and in the decisions log: the full chain against a profile THIS helper created could not be exercised on the verifying host. That aif has `ZAI_API_KEY` in its process env and no `ANTHROPIC_AUTH_TOKEN` (measured — names only, never values), so a helper-created profile returns CHAT_AUTH_ERROR there regardless of correctness. On a consumer machine the §7b wiring supplies that var, which is the case the helper targets — but the first consumer run is still the first end-to-end execution, and a `FAILED step-D` there means the wiring did not land, not that the proof is broken. INSTALL-FOR-AI.md now says DONE means the vendor accepted the key, that the call is billed, and what a `FAILED step-D` line means. Prior-art: skipped — completes an existing helper's declared objective using the aif REST API it already speaks; no new capability, no dependency, no new module. * fix(beta-delivery-ux): S4 — round-6 REVISE, incl. a key-invariant assertion that could not fail Round-6 cold audit on `34ccc8cece` confirmed both round-5 carry-overs genuinely closed, and found three MAJORs of its own. All three are closed here. MAJOR-1 — the §2 constraint 1 assertion on the step-D surface was VACUOUS, and it is the third time this branch has applied a principle to one branch and not its neighbour. It accumulated request argv into the shell variable `_ND_SEEN`; the helper calls curl inside `$( )`, so every stub call ran in its own subshell and the variable was empty at assertion time. The suite reported the key never reaches the request while a deliberately injected key value sailed straight past. The file-based fix for exactly this subshell trap was already documented and applied twelve lines away, for the session pin. Now: capture goes to a file, a guard asserts the capture is NON-EMPTY (so the check cannot pass on nothing), and a paired negative proves the detector fires on a request that does carry the value. Verified by injecting a leak into step D's body — the suite goes red, and green again when reverted. Same block, second defect: the create arm answered 2xx with no body inspection, unlike every sibling stub — W-4's tell verbatim; it now calls `_stub_create_body_ok`. MAJOR-2 — `INSTALL-FOR-AI.md` promised the model call costs "a few cents at most". This branch's own measurement is `costUsd:0.117219`, and its decisions log says "not token-scale". A consumer-facing ceiling falsified by the author's own number in the same commit range is exactly the false green this stage keeps producing. Corrected to the measured figure, and the sentence now also states the two things it silently omitted: the helper runs `docker compose up -d` against the consumer's aif (restarting their services), and every run leaves a chat session titled "getff GLM provisioning proof" in their project. MAJOR-3 — the research patch still carried "§7e.3 — model proof: PARKED", "the fork is genuine: two binding constraints in conflict", and a run-5 §1.7 note calling the DONE gap "NOT fixed" — all false at HEAD, and the newest dated section was among them, so a reader following the file's own ordering would land on withdrawn claims. The folder is append-only, so a superseding addendum now closes each of the three by name at the end of the file. MINOR — narrowed, but not the way the audit proposed. The audit read "session pinning measured live" as broader than the evidence, having seen only completions that ran on the project default. The discriminating observation exists and is now named in the code: a session pinned to a NON-default, freshly created profile returned CHAT_AUTH_ERROR. Had pinning fallen back to the default, that call would have SUCCEEDED, since the default is a working profile — so the error is positive evidence the pinned profile was used. NOT closed — the audit's KICKOFF-AMBIGUOUS stands and is an owner question: §7a #3 binds a "1-token-scale" call, but the only reachable route injects project context, so every consumer provision bills ~$0.12. Recurring per-consumer spend was never decided by the kickoff. Suite 60 → 62, green on the host. Prior-art: skipped — test-material repair plus doc corrections to an existing helper; no new capability, no dependency, no new module. * docs(beta-delivery-ux): S4 — log fork 3, recurring per-consumer spend on step D The round-6 audit flagged KICKOFF-AMBIGUOUS: §7a #3 binds a '1-token-scale' model call, but the only reachable completion route injects project context — measured 39058 input tokens and costUsd:0.117219 for a one-word prompt. Step D runs on every provision, so that is a recurring charge on each consumer's plan rather than a one-off verification cost. Logged, not decided: nothing in the code settles it, and whether ~$0.12 per provision is acceptable is a product call about other people's money. INSTALL-FOR-AI.md states the measured figure to the consumer either way, so nobody is billed by surprise while it is open. Prior-art: skipped — decisions log entry for an existing umbrella; no capability, no dependency. --------- Co-authored-by: Test <test@example.com>

…age gate that fell open (#1350) ## Summary The round-7 cold fidelity audit returned `FIDELITY: GO` against `e3e705eb26` and graded five findings MINOR, deliberately leaving them unfixed so the merged artefact would not differ from the audited one. This PR closes all five as follow-up work on `staging`. Four are in the S4 GLM one-button helper and its suite — one real fail-open branch in the helper, and three assertions that were weaker than the properties they protect. The fifth is a stale provenance label in the S4 research patch. ## Changes - **`scripts/getff-glm-onebutton.sh` — step D's zero-usage gate fell OPEN on a non-numeric `totalTokens`.** `[ "$chat_tokens" -le 0 ] 2>/dev/null` errors on a non-numeric value, so the `if` is false and control reached `GLM_PROVISION: DONE`. `null` and an absent `usage` were always safe (jq's `// 0` → `0`) and aif types `RuntimeUsage.totalTokens` as `number`, so the window was narrow — but this was a fail-open branch inside the one step whose entire job is to fail closed. Replaced with an explicit numeric `case` test emitting a distinct terminal token, `FAILED step-D model-proof-unusable-usage`. Paired negatives **N10** (string `"abc"`) and **N10b** (non-integer `12.5` — valid JSON, and `[ -le ]` errors on it identically), plus paired positive **N10c** so the widened gate is not reject-all. - **`tests/install-sh/glm-onebutton.test.sh` — the §2-constraint-1 guard-the-guard was weaker than the property it protects.** It asserted the capture file was non-empty, not that step D's completion request was in it. Now matches a whitespace-delimited token ending in `/chat` — `/chat/sessions` must not satisfy it alone. - **Same file — the shipped paired-negative for the leak detector was a tautology.** It wrote a leak-shaped string to a temp file and grepped it back, proving `grep` works rather than that the capture wiring carries a real leak. It now builds a scratch copy of the helper with the key value injected into step D's chat body (`--arg LEAKV "${!GLM_ENV_VAR}"`), runs it against the same stub, and asserts the value lands in the capture — with a companion assertion that the injection actually applied, so a drifted `sed` anchor cannot make it vacuous. - **Same file — the §7e.6 meta-scanner's declared population excluded one fail-open shape.** It scoped itself to `case`-based stubs, so a caseless stub — which answers 2xx-equivalent for *every* path — was invisible, and the file carried one (the preflight stub, harmless only by accident of call order). Declaring a population and scanning a subset of it is the same can't-fail shape this gate exists to reject, so the population was **widened** rather than narrowed in prose. Comment lines are now skipped, and one-line `case`-based stubs are graded on their catch-all. Preflight stub is now `return 1`. - **`docs/.../2026-08-08-s4-glm-onebutton-entry-verification.md`** — the `SUPERSEDING ADDENDUM` was headed «(run-5, commit `34ccc8cece`)» but was added by `3bf1d18ce9`, and its body describes round-6 corrections postdating the commit named in its own heading. It now names its own commit, in the section whose entire subject is provenance honesty. The round-7 follow-ups are recorded in the same file's append-only tail. ## Prior-art consult - [x] Capability commits in this PR carry a `Prior-art:` trailer (or `skipped — <≥20 chars rationale>` for hook-flagged commits that aren't real capability additions). Not a capability commit: no new dependency, no new file, edits only to an existing helper, its own suite, and one research patch. - [x] If a new capability area surfaced during this PR: n/a — none surfaced. - [x] If existing entries matched: n/a — no SSOT entry is cited or touched. - [x] context7 queries (≥3 phrasings) for new capability areas were run and the results cited: n/a — no new capability area. ## Test plan - [x] `bash tests/install-sh/glm-onebutton.test.sh` → **73/0** on the host (62/0 at the audited SHA) - [x] `bash tests/install-sh/bridge-guided.test.sh` → **11/0** (sibling suite, unchanged) - [x] `npm run --prefix packages/core test:principles` → **353 passed, 1 skipped, 37 files** - [x] `bash -n` clean on both edited shell files; `shellcheck -S warning` on the helper shows only the two pre-existing `SC1090` hits (identical count on `staging`) - [x] `markdownlint-cli2` clean on the edited research patch - [x] Manual smoke — each fix watched FAIL before the fix, and each falsified after: - N10 pre-fix reported «non-numeric totalTokens fell through to DONE»; green post-fix. - Guard-the-guard: a scratch helper with an injected `return 1` after step A produced `✓ capture is non-empty (9 requests recorded)` under the old guard with **zero** `/chat` requests made; the new guard goes red on the same scratch copy. - Leak paired-negative: neutering the injection turns exactly those two assertions red and nothing else (66/2). - Meta-scanner: reintroducing `curl() { return 0; }` in the real file turns the meta assertion red, naming the line. ## Provenance n/a — non-stage follow-up PR closing MINORs carried by the merged S4 stage PR #1341. ## Review findings n/a — non-stage PR. Origin of the five items is #1341's round-7 audit (`FIDELITY: GO` against `e3e705eb26`), which graded them MINOR and left them for follow-up. ## Fidelity verdict FIDELITY: skipped — non-stage follow-up bugfix PR; no kickoff or stage contract applies, and the five items originate from the already-merged #1341 round-7 verdict. ## Parked questions The `§7a #3` per-consumer spend fork (≈$0.12 per `provision` run) is **not** re-opened here — it is recorded in `.claude/orchestrator-prompts/beta-delivery-ux/kickoff-s4.decisions.md` and is the operator's call. One item is deliberately **not** closed: the §7e.6 gate remains scoped to its own file's source (`$0`), so the `bridge-guided.test.sh` sweep below is a one-off measurement, not a standing gate. Promoting the scanner to a shared checker over `tests/install-sh/**` is a capability commit needing its own prior-art consult, out of scope for a follow-up bugfix. ## §1.7 Self-discipline check (REQUIRED if PR touches discipline-bearing files) ### §1.7 Skipped: mechanical maintenance — this PR fixes a fail-open branch and three weak assertions in an existing helper plus its own test suite, and corrects one stale provenance label; it introduces and extends no discipline rule, and touches no file under .claude/rules/, packages/core/principles/, EXECUTION-PLAN.md, prior-art-evaluations.md, CLAUDE.md or any shipped template.

…h open forks (#1351) Both questions the unattended run logged rather than decided are now answered, and the answers are written where a later reader can judge them rather than inherit them. §4 item 1 (live end-to-end on a clean machine) — ACCEPTED AS MET on block-level evidence. The record states exactly what was measured (each step live against a running aif; step D's four failure modes each with a paired negative observed RED before GREEN; 62/62 on the host) and exactly what was not: those steps executing in sequence against a helper-created profile on a machine where the §7b wiring supplied the key. The accepted risk is named — the first consumer run is the first full-chain execution — together with its falsifier and the distinguishing signal (`CHAT_AUTH_ERROR` from `POST /chat`, which the helper already hints at). §7a #3 recurring per-consumer spend — ACCEPTED AS-IS. Step D bills ~$0.117 per provision because aif prepends project context on the only reachable completion route. The kickoff's «1-token-scale» wording is left standing and marked inaccurate-as-shipped rather than quietly amended, since the measured figure is what governs and `INSTALL-FOR-AI.md:184` states it to the consumer before the step runs. The two rejected alternatives are recorded so the choice is legible. Also corrects a framing error this session made twice in chat: `done.md` was described as blocked by these forks. It is not. Per operational-conventions.md §1 it marks the UMBRELLA complete and is written at the LAST stage's merge only — and `beta-delivery-ux` still has R1 in flight (aif task `7a12c5bc`) plus S2's follow-up unharvested (`a8337963`), among 8 done-but-unharvested tasks. Writing it now would make priority-score.sh Layer C3 score the umbrella DONE and stop dispatching that work. S4 as a STAGE is closed; the umbrella is not. Prior-art: skipped — decisions record for an existing umbrella; no capability, no dependency, no new module. Co-authored-by: Test <test@example.com>

…ws, routing bindings (#1462) * feat(arch): frontier pacing delta in §1 — ADAPT of grill-me/grilling (SSOT #253) Two deltas over the wrapped brainstorming loop: batch prerequisite-settled questions per round (dependent ones stay serial), and enumerate-before-done — the dialogue closes only when every design decision is answered or an explicit operator fork. Superpowers 6.2.0 verified to lack the mechanism (grep over the whole installed plugin: 0 hits; brainstorming pins one-question-per-message). Prior-art: prior-art-evaluations.md#253 (grill-me/grilling, ADAPT — only the tree/frontier mechanic transfers; recommendation-per-question and probe-don't-ask already exist as H1 + T8/T20). * feat(arch): grilling becomes the questioning engine — SSOT #253 lifted ADAPT→ADOPT Operator-ratified design session (D1-D4): the compressed frontier-pacing paraphrase measurably lost upstream's non-blocking probe rule (same-day cold review vs the raw upstream text), so /arch §1 now consumes the grilling skill AS IS via the mattpocock-skills companion plugin (MIT, versioned, precedent #64 brainstorming) and keeps only a thin binding: brainstorming collision resolution, probe routing (T20/§1.5), AskUserQuestion as the round carrier (added to allowed-tools), and the spec's live decision register as the tree surface. Register format lands in the spec-template obligation; SSOT #253 revisit triggers gain a named recording surface + a vendor-copy fallback arm. Prior-art: prior-art-evaluations.md#253 (grill-me/grilling, ADOPT — companion plugin consumed AS IS; paraphrase channel measured lossy, hence the lift from ADAPT). * feat(arch): §2 no-rerank rule — the two altitudes are never merged into one list Adopted from mattpocock code-review's two-axis separation (one axis must not mask the other) during the 2026-08-17 plugin sweep; the §2 seats already report independently, this pins that their findings are presented side by side and never reranked across altitudes. Prior-art: prior-art-evaluations.md#253 (mattpocock-skills plugin sweep; doc-only edit, no new capability). * docs(arch-prep): three-stack skill harmonization — collision map + raw ownership idea Prep-doc for a future /arch design session: enumerates all three skill populations (ours 16, superpowers 6.2.0 14, mattpocock-skills 1.2.3 35 — all 35 read in full), maps collisions per capability area (sharpest: Matt tdd vs SP TDD contradict on refactor placement and seam scoping; diagnosing-bugs vs systematic-debugging claim the same trigger space), lists the six available resolution mechanisms with two unknowns (per-skill disable, routing precedence) as probes, and drafts a one-owner-per-area map plus a live decision register the design session starts from. Prior-art: prior-art-evaluations.md#253 (grilling ADOPT — this prep extends the same three-stack comparison to the full plugin; doc-only, no capability). * docs(arch-prep): §1.5 dependency edges — collision risk weighted by our hard references Measured map of what our machinery hard-references upstream (grep over skills/rules/agents/CLAUDE.md/templates): SDD is the most-referenced upstream and crosses the shipped axis (tier-home.md); requesting-code-review is the highest-risk collision zone because dispatcher/harvest contracts name it while Matt's code-review claims the same trigger space; TDD/debugging collisions carry routing risk only (zero hard edges from us). Also: D-H4 recorded as answered (parallel commit 09569a3 landed mid-session), §7 gains the re-probe-before-edit note. Prior-art: prior-art-evaluations.md#253 (same three-stack comparison; doc-only edit). * docs(arch-prep): DeepWiki pass on both satellites + our-side thinning audit Per-repo DeepWiki interrogation folded in: (a) the measured routable surface is exactly 11 mattpocock skills (user-invoked ones never enter the router — his collision policy is the user/model-invoked split, confirmed live in this session's skill listing); (b) Matt's refactor-out-of-loop is a June-2026 behavioral measurement («agents essentially never performed it»), not doctrine — D-H2 needs our own corpus check; (c) TDD edge CORRECTED: SDD's implementer-prompt.md:36 says bare «TDD», so the collision is transitive-contract grade, not routing-only; (d) superpowers documents Project > Personal > Plugin per-skill shadowing — new mechanism 7, P1 narrowed. New §4.5: our 16 skills audited — nothing deletable, orchestrator is the one THIN candidate (D-H9); D-H10 TDD shadow, D-H11 domain-modeling pairing added to the register. Prior-art: prior-art-evaluations.md#253 (same three-stack comparison; doc-only edit). * docs(arch-prep): slash-only planning skills evaluated as adoption candidates + 4 raw ideas Operator correction folded in: «no collision» ≠ «no value» — the user-invoked planning skills get per-skill adopt/adapt verdicts (wayfinder ADAPT strongest; to-tickets ADAPT mechanizable; to-spec one section; implement REJECT; triage two residues). New §4.6 carries four raw ideas for the design session: (1) the decision map as the multi-session layer over /arch — D4's register lifted to wayfinder shape, map-location sub-fork included; (2) kickoff Blocked-by edges with a pipeline-computed frontier; (3) seams-first Testing-seams slot in the spec template, unlocking the seams half of D-H2; (4) glossary SSOT as a term-ownership generated index — the CONTEXT.md-free adaptation that makes the grilling+domain-modeling pairing adoptable (D-H11 re-opened from defer). Register grows D-H12-D-H14. Prior-art: prior-art-evaluations.md#253 (same three-stack comparison; doc-only edit). * docs(arch): three-stack skill harmonization — design spec + continuation handoff Interview phase complete (frontier empty): P-1..P-6 operator premises, 15-area ownership map ratified (D-H1), decision register D-H0..D-H16 with falsifiers, mechanism set (prune script, CONTEXT.md rule+test, claim reorder, Blocked-by frontier, seams slot, aif plugin), probe register P1/P2a-c/P5-pending/P6, routed-work inventory for §3 exit routing. Awaiting §2 cold two-altitude review (this session's next step). * docs(arch): harmonization spec v2 — round-1 cold-review dispositions landed Both §2 seats returned REVISE (9 + 8 findings). All round-triggering findings repaired in place: §1 restated as two declared lanes (TD-F1); prune radius narrowed to 2 machine-globally-justified items per the operator's F7 answer + --check pre-push drift detector (TD-F2/F7); D-H17 completes the ownership map to all 11 model-invocable skills (TD-F3); setup run re-bucketed attended (TD-F4); D-H5 claim mechanics specified with real machinery + P4 restored (TD-F5, B-M1/M2); §5.6 non-target named (B-M3); /vitest transfer dissolved (B-M4); D-H7/D-H8 counter statuses corrected (B-M5); D-H13 adopts incumbent 'Depends on' spelling (B-M6). New: P-7 premise + D-H18 consumer-axis contour routed out via chip. Full dispositions: §9 v2 entry. * docs(arch): harmonization spec v3 + SSOT #253 counter arm + REJECT rows #254-257 Round-2 delta review (both seats REVISE; all round-1 closures confirmed): - --check channel corrected: owner:'maintainer' section in the pre-push.ts section registry (the file ships to consumers but maintainer sections never compose on a consumer layout, fail-closed) — .husky/pre-push is an exec dispatcher with no sections (convergent TD/B finding). - D-H16 build item DISSOLVED: aif container mounts the host ~/.claude/plugins read-only (docker-compose.override.yml), so the plugin is already visible in-container and the prune/--check cover it by construction (measured round-2). - 'counter armed' made true instead of re-worded: D-H7/D-H8 arm + observation No.0 appended to SSOT #253; REJECT rows #254-257 added (Matt implement, ADR dir, severity-less review model, total-sweep pruning). Spec SS8 item 5 DONE in-session. Dispositions: spec SS9 v3 entry. * docs(arch): harmonization contour GO — round-3 record + routed in-session edits Round 3 (targeted delta): both cold seats GO. Spec header → REVIEWED-GO; §9 round-3 entry (one TD MINOR accepted as recorded limit: container premise rests on untracked local docker-compose.override.yml — covered by D-H16 falsifier). Routed §8 item 2 small edits, per spec: - arch/SKILL.md §1: Testing seams slot added to the spec-template obligation (D-H14; seams-first adopted WITHOUT Matt's refactor placement) - ai-doc/SKILL.md: skill-authoring ownership note (standard=ours, process=SP writing-skills, writing-for-agents=REFERENCE) - rule-tests/SKILL.md: tautological-test anti-pattern REFERENCE note (D-H2 transfer (b)) * docs(arch): close harmonization contour handoff — full tail executed Review GO (3 rounds), exit routing done (3 chips + in-session edits), SSOT appends landed. Handoff retained as closure record; residue = operator actions (spec SS8 item 1) + chip-routed umbrellas. * docs(arch): consumer-axis satellite harmonization — design v1 + round-3 handoff Round-2 /arch contour (D-H18): interview closed, D-C1..D-C8 ratified with falsifiers; three-class collision model (factory CI / install-time census / informed consent); detect+declare+prescribe mechanism recorded. Cold review and exit routing DEFERRED behind the operator-mandated round-3 top-down creative re-examination (P-C3) — handoff written for the fresh session. * docs(arch): harmonization round 3 — registers amended, injected-context bindings land Round 3 (D-C8, operator-mandated P-C3) executed per the handoff's membrane phase order. Operator-axis spec v4: D-H15 SUPERSEDED — the prune apparatus (script / wizard / --check pre-push section / gate P5) dissolved, replaced by CLAUDE.md routing bindings (repo section + ~/.claude/CLAUDE.md machine-global half, written in-session with live operator approval) + meta-kickoff.template.md binding line (D-H10 fallback promoted to primary); D-H8 gains a frontmatter-neutering ladder step. Round-2 spec v2: D-C1 re-cut to the thin form (static census prose + known-pair presence check; inventory-join engine not built), D-C9 fourth-stack admission boundary added (knowledge-work trio stays on SSOT #235). Round-3 handoff closed with the continuation-state staleness correction; keen-shannon merged in (3ae6981) so both specs live on one branch. * docs(arch): P7 recorded — fresh-session bindings probe 2/2 vs P2 baseline Both P2-class triggers flip with the CLAUDE.md bindings in context (headless claude -p, fresh sessions reading the worktree CLAUDE.md from disk). Method finding recorded: in-session subagent probes are invalid for mid-session binding edits — subagents inherit the parent's session-start CLAUDE.md snapshot (measured via a failed in-session probe plus its diagnostic follow-up). * docs(arch): round-3 review R1 — both seats REVISE, dispositions landed Convergent BLOCKER fixed: the meta-kickoff.template.md binding line REMOVED — .claude/skills/pipeline/ ships to consumers via GETFF_SKILLS_ENV (setup.d/lib.sh:59) at the default env profile, so carrier #3 breached the operator-axis membrane while buying no coverage; its removal restores all 8 install fingerprints to the baseline blob. §5.1's «no mechanical channel at all» premise corrected (config layer only; frontmatter + a possible Skill-matched PreToolUse hook priced — P8 records the hook UNVERIFIED: guide claims no Skill matcher, live harness observation contradicts). P7 restated honestly (1 measured flip + 1 post-only confirmation). SSOT #253/ #257 got dated supersession notes (no prune ever executed). Five residual prune assertions re-cut. Consumer spec: population corrected — TWO shipped cc-plugin rows (superpowers + ast-grep, the latter disabled on the operator's own machine); presence check re-keyed on installed_plugins.json + enabledPlugins; D-C5/D-C6 aligned; class-2 own-skills half recorded as prose-only limit. ESCALATED to operator: ast-grep shipping fate (ESC-1) + the detection-wire fork (TD-M2/P8). ~/.claude/CLAUDE.md section relocated to file end (orphaned AIF bullet restored to its heading). * docs(arch): round-3 review R2 — residuals closed, operator answers landed Both R2 seats REVISE with a convergent root cause: R1 edited the surfaces findings argued FROM, not every surface repeating the claim. Closed: §1 premise re-cut to config-layer wording; §1 scope guard now names the pre-round-3 routed edits as verified degrade-safe REFERENCEs; sixth prune assertion re-cut (D-H16); handoff header unmerged label; D-H15 exclusivity hedge; consumer D-C1/§7 re-keyed on installed_plugins.json + enabledPlugins; both §8 inventories carry the escalations. Operator answers recorded live: ESC-1 → retro-census BOTH shipped rows, keep ast-grep on a clean census; P8 → VERIFY via the settings.json hand-off (§8 item 7). Review round cap (2 REVISE) reached — residual state surfaced in §9 instead of a third cold round. * docs(kickoffs): round-3 exit routing — two build umbrellas authored consumer-satellite-contract (thin form: retro-census of BOTH manifest rows per the answered ESC-1, D-C2 principle test, AGENTS.md.template section + parity line, install-registry-keyed presence check) and skill-harmonization-mechanisms (CONTEXT.md pointer-rule test, four-part claim machinery closing probe P4, Depends-on frontier). Both carry host-verify contracts and the PR-pause note: they become dispatchable only when the spec branch merges to staging. * docs(kickoffs): declare the effort-worthiness L0 rigor label on both round-3 kickoffs Principle 40 (`packages/core/principles/40-kickoff-rigor-label.test.ts`) requires every post-cutoff kickoff to carry a `Rigor label … L0 …` line with a legal value. Both round-3 kickoffs were authored without it and failed the gate at push time. - skill-harmonization-mechanisms → `build-and-verify`: all three surviving stages are factory-internal and reversible, each with a live RED/GREEN seam proof. - consumer-satellite-contract → `research-grade`: S3/S4 touch consumer-shipped surfaces (AGENTS.md.template, ./setup), which effort-worthiness §1 reserves for the research-grade contour. Prior-art: skipped — mechanical gate compliance on two doc files, no new capability * docs(arch): P8 CONFIRMED live — detection wire v0 declared; kickoff L0 labels The operator-registered log-only PreToolUse Skill hook fired on a forced model-invoked skill in a fresh headless session (JSON with tool_name=Skill + the skill name in tool_input). The guide-agent's 'skill loading bypasses the tool pipeline' claim is falsified — the P6 failure class again. Measured boundary: user-typed slash commands bypass the Skill tool (invisible to the wire, irrelevant: misroutes are model-invocations). TD-M2 closes — the log IS the v0 misroute detection wire feeding the D-H7 counter; spec §6 P8 + D-H8 + §8 item 7 updated. Both kickoffs gained effort-worthiness L0 rigor labels (parallel session's edit kept as-is). * test(install-sh): regenerate baselines for the three edited shipped skills The branch edits `.claude/skills/{ai-doc,arch,rule-tests}/SKILL.md` — all three are shipped artefacts, so their fingerprints move in every stack baseline that carries them. Captured with `SNAPSHOT_MODE=capture bash tests/install-sh/snapshot.sh`. Diff reviewed before committing (round-3 R1 precedent: an unreviewed template edit turned 8 fingerprints stale): exactly three payload paths changed hash — rule-tests (22 occurrences), arch (16), ai-doc (16) — and every fingerprint file is 1:1 on line count, so no payload entered or left any stack. Prior-art: skipped — snapshot regeneration after a shipped-file edit, no new capability --------- Co-authored-by: Test <test@example.com>

… carve-outs dropped (#1506 parked closure) (#1511) * docs(corpus): legacy lint+link repair — orchestrator-prompts to markdownlint-zero and lychee-zero Fires both re-entry triggers recorded in the host-verify retrofit (PR #1506): the A/M carve-outs in the markdownlint arm and pre-push §8 can now be dropped (next commit). Measured before: 401 markdownlint violations (MD040 ×326 bare fences, MD007 ×44, MD034 ×19+3 bare emails, MD009 ×7, MD001 ×5) and 340 lychee offline errors across the dir. After: 0 / 956 files, 0 / 3670 links. All repairs are machine-verified and content-preserving: - MD040: bare fences (plain, indented, blockquoted) gain a 'text' info-string — fence content byte-identical - MD007: dedent to the linter's own Expected value (CommonMark parses ≤3 spaces as top-level; no AST change) - MD009/MD034: trailing-space strips; bare URLs/emails wrapped in angle-bracket autolinks - MD001: five h1→h3 skips re-levelled ### → ## - links: 226 relinked to verified live targets (repair rewrites: renamed skill meta-orchestrator→orchestrator, doubled .claude/.claude, wrong relative roots; suffix search over git-tracked files, unique matches only, fragments preserved); 102 de-linked to code spans where the target is machine-local (~, file://, /private/tmp, .zcode), gitignored scratch (drafts/), or gone with no unique survivor (e.g. agents/docs-auditor.md, master-backlog-delta.md, meta-launch kickoffs never committed) - two multi-line links (state.md 248, top-down-r1.md 331) hand-fixed to the correct relative depth - lychee.toml: one unanchored exclude '/\.\.\.$' for the literal ellipsis placeholder quoted inside an escaped code span (migration-ast stage-1) — a parser phantom in the placeholder class, not a link Stance note: lychee.toml's exclude_path keeps research-patches/retros as 'expected archaeology'; the orchestrator-prompts corpus instead gets repaired-to-zero (PR #1506's shipped decision) so the LIVE kickoffs keep full link checking with no exclusions at all. * fix(hooks): drop both legacy-corpus carve-outs — re-entry triggers fired The orchestrator-prompts corpus is now markdownlint-zero (956 files) and lychee-zero (3670 links, offline + --root-dir) after the repair commit — the re-entry conditions both carve-outs recorded. Dropping them: - .husky/pre-commit markdownlint arm: back to the plain staged-*.md set; the A/M awk filter and its exclusion note replaced by a history note. - packages/core/hooks/pre-push.ts §8: same removal; the lychee invocation gains --root-dir REPO_ROOT — root-relative links (leading /) now resolve against the repo instead of erroring unresolved (the class the repair surfaced; without a root dir lychee cannot form a URL for such links, so neither checking nor exclude patterns can see them). Consumer-layout narrowing (S2 Part 1) untouched — suite 32/32. - packages/core/hooks/utils/git.ts: getAddedFiles() removed with its last caller (dead export; the A/M split was its only consumer). - baselines regenerated (shipped bytes changed): exactly the pre-push.ts, git.ts and transitive refresh-baseline manifest hash lines ×8 npm cells. Sweep: 35/35 gates, FULL mode (byte-identical 15/15, typecheck, all vitest lanes, render/rule-index, format — clean). * fix(review): cold-review round 1 — collapse 3 duplicate de-link code spans + on-disk precision in gate comments Reviewer CLEAN verdict (0 blocker/major); folded the three MINORs: - swarm-tools-research/kickoff.md:76 and wave-9 post-9 ×2 — the de-link transform appended the href code span even when the label already WAS that span ('label (label)'); collapsed to the single span. The foreign coordination mirror (repo-hygiene-cleanup) carries the same shape but is another session's artifact — left untouched. - '956 files'/'3670 links' comments now say on-disk/checked — the corpus dir includes gitignored sidecars beyond the tracked set. - stray state.md.tmp from the aborted first fixer attempt removed. * fix(review): fidelity round 1 — relink the 3 live targets the de-link pass over-applied Auditor MINOR: think-time-s17-gate/kickoff.md:32,:92,:172 were de-linked although their targets are live and tracked (research-patch §13.24, agents/compliance-verifier.md, README.md) — first-pass target matching missed them (percent-encoded § in one report form; ambiguous bare basenames in the others). Relinked to correct relative paths; both zero states re-verified (markdownlint 0/956, lychee 0/3670 exit 0). The ESCALATED finding is a wording correction only (recorded in the PR body): the parked-#3 premise 'cli2 does not auto-read .markdownlint.json' is empirically false — cli2 DOES auto-discover the v1 config from cwd (auditor A/B: with the json present the corpus reported 1 error class; with it removed, 3 — MD013/MD041 surface). In-repo the hook therefore enforces the project's declared 6-rule set (shared with editors); consumers, where install.sh ships no config, get cli2 defaults — the strict side lands there. No wiring needed; the retrofit PR's parked wording was wrong and this PR's body records the corrected decision. * chore(lychee): exclude explainx.ai — 200 locally, deterministic TIMEOUT from GH runners The live link audit (pre-existing link, docs/meta-factory/prior-art- evaluations.md:167 + the companion-capability survey — NOT part of this PR's diff) timed out twice from the CI runner on the same URL while a local curl -I returns HTTP/2 200. Same unreachable-from-runner class as the documented npmjs/gptlint entries: runner-path problem, not a dead link. Documented with the measurement. --------- Co-authored-by: Test <test@example.com>

Promote staging → main #3 — beta candidate (682 commits since 2026-07-10)

… economy → GLM-5.3 SDK (#1608) ## Summary Two `/pipeline` launch presets named runtime profiles that no longer resolve on the live aif instance: `presets/aif.json` marker `Claude Opus (plan+review)` is a DISABLED profile (matches by exact name in `AifHandoffBackend._resolveProfileId`, then blocks at runtime auth — the `cf894220` shape), and `presets/economy.json` marker `Z.AI GLM-5.2 SDK` matches no profile at all (the instance runs 5.3; dispatch throws). Both were recorded as a live defect in `.claude/orchestrator-prompts/beta-docs-showcase/kickoff.decisions.md:53-54` and left for the pipeline skill owner; the operator invited the fix on 2026-09-05. Data fix, one concern. ## Changes - `.claude/skills/pipeline/references/presets/aif.json` — `marker: null` (the preset dispatches on the project's per-mode default profiles — the ratified GLM-only posture of PR #1446 / kickoff.decisions.md Decision 1 — instead of forcing a disabled Claude profile); description says so. - `.claude/skills/pipeline/references/presets/economy.json` — `marker: "Z.AI GLM-5.3 SDK"` (unique under the resolver's match; live list quoted below). - `.claude/skills/pipeline/SKILL.md:56` — the Seam #3 sentence names the presets that now carry a marker (economy) and the null-marker set (aif/night/sdd). Live profile list at authoring (`curl -s "$RUNTIME_BRIDGE_AIF_URL/runtime-profiles"`, 2026-09-05): ```text 41315ef6 | Claude Opus (plan+review) | enabled=False 53eca24c | Z.AI GLM-5.3 SDK | enabled=True b6f8d94e | Qwen3.8-Max-Preview | enabled=True 088182b8 | Z.AI GLM-5.3 Flash (implementer) | enabled=True ``` `Z.AI GLM-5.3 SDK` is a strict prefix of nothing and a substring of exactly one row — the marker value rule (`packages/core/templates/shared/tier-home.md:63`) holds. ## Prior-art consult - [x] Not a capability commit (two JSON values + one sentence; `packages/core/hooks/checks/prior-art.ts:142`). - [x] No new capability area. - [x] No SSOT entry matched. - [x] context7 not applicable. ## Test plan - [x] `bash .claude/skills/pipeline/helpers/list-presets.sh` — `aif` renders without a marker (null → no `marker=` suffix, same path as night/sdd, `list-presets.sh:45-47`), `economy` renders `marker=Z.AI GLM-5.3 SDK`. - [x] `bash .claude/skills/pipeline/helpers/resolve-preset.sh aif` emits no `PRESET_MARKER` line (`resolve-preset.sh:74-75` handles null); `... economy` emits `PRESET_MARKER=Z.AI GLM-5.3 SDK`. - [x] `grep -rn 'GLM-5.2 SDK\|plan+review' .claude/skills/pipeline` → only the two edited files before, zero stale hits after; historical kickoffs under `.claude/orchestrator-prompts/` keep their markers as records and are out of scope. ## Provenance n/a ## Review findings n/a ## Fidelity verdict FIDELITY: skipped — data fix of two preset marker values plus one SKILL.md sentence; no stage kickoff applies. ## Parked questions n/a ## §1.7 Self-discipline check (REQUIRED if PR touches discipline-bearing files) ### §1.7 Skipped: data fix of two preset marker values plus one descriptive SKILL.md sentence — no rule, principle or convention introduced; the marker value rule itself is unchanged ## Follow-ups landed in this PR - Regenerated the 8 npm-stack + 3 python snapshot baselines (the pipeline skill payload and the re-rendered AI-USAGE-GUIDE §6a ship at every depth; cargo baselines untouched — green in CI). - `packages/core/hooks/parse-preset.test.ts` follows the new markers (aif: no marker line; economy: GLM-5.3 SDK). - `AI-USAGE-GUIDE.md` §6a re-rendered via `scripts/render-presets.mjs --write` (the Manifest → RULES.md drift job checks it).

artyhoo added 4 commits May 8, 2026 02:19

artyhoo merged commit 12b4213 into main May 7, 2026
18 checks passed

artyhoo mentioned this pull request May 8, 2026

Phase 4: Stack Detector v1 (read+write AIF bridge, CI gate, reviewer cycle complete) #4

Merged

6 tasks

This was referenced May 8, 2026

docs: Phase 9 entry research — capability matrix + verdict for Phase 9 implementation #13

Merged

docs: Phase 9 implementation prompt draft (PHASE-9-PROMPT.md) #15

Merged

This was referenced May 16, 2026

docs(research-patches): commit 3 patches from 2026-05-16 autonomous research session #55

Merged

docs(research-patches): B+C session + parallel + errata + 2 mechanical fixes #56

Merged

artyhoo deleted the chore/phase-3-aif-retrofit branch May 22, 2026 18:08

This was referenced May 25, 2026

research(bundle-autonomous): prior-art for 3 sub-mechanisms (B1/B2/B3) #237

Merged

research(meta-orchestrator): no-arg-laziness R-phase — F.3 + F.6 verdict #261

Merged

artyhoo mentioned this pull request Jun 1, 2026

feat(hooks): #108 orchestrator-gate hooks (SubagentStart digest-injection) #330

Merged

artyhoo mentioned this pull request Jun 14, 2026

fix(install): R2 reaches brownfield/monorepo consumers — #507 reopen #513

Merged

5 tasks

This was referenced Aug 9, 2026

feat(beta-delivery-ux): S4 run 3 — GLM one-button per kickoff §7b/§7c (A2) [FIDELITY STOP — base for run 4] #1322

Closed

feat(beta-delivery-ux): S4 — GLM one-button + aif companion install, with the three false greens fixed #1323

Merged

artyhoo mentioned this pull request Aug 9, 2026

docs(beta-delivery-ux): S4 kickoff §7f — run-4 dispatch facts (run 3 harvested, 995e9c dead-end) #1324

Merged

artyhoo mentioned this pull request Aug 9, 2026

docs(beta-delivery-ux): S4 §7f — three falsified dispatch facts corrected, model proof measured #1340

Merged

artyhoo mentioned this pull request Aug 9, 2026

feat(beta-delivery-ux): S4 — GLM one-button model proof (step D) + three false greens closed #1341

Merged

artyhoo mentioned this pull request Aug 9, 2026

docs(kickoffs): correct three falsified container-reachability claims, back a fourth with its probe #1345

Merged

artyhoo mentioned this pull request Aug 9, 2026

docs(beta-delivery-ux): S4 — operator resolutions on both open forks #1351

Merged

This was referenced Aug 16, 2026

triage-kernel-v2 #1397

Merged

fix(triage-kernel-v2): close cold-review defects in the S4b kickoff before dispatch #1398

Merged

This was referenced Aug 20, 2026

docs(orchestrator): FRS S3 task-2 — Final PR backfill for the 2 real metadata gaps (#1308, #1289) #1494

Merged

legacy corpus repair: orchestrator-prompts to lint-zero + link-zero — carve-outs dropped (#1506 parked closure) #1511

Merged

artyhoo added a commit that referenced this pull request Sep 4, 2026

Merge pull request #1597 from artyhoo/staging

9b768ca

Promote staging → main #3 — beta candidate (682 commits since 2026-07-10)

artyhoo mentioned this pull request Sep 5, 2026

fix(pipeline): repoint stale preset markers — aif → project defaults, economy → GLM-5.3 SDK #1608

Merged

7 tasks

Navigation Menu

chore(meta-factory): Phase 3 Step 0 retrofit + 3 MAJOR fixes - #3

chore(meta-factory): Phase 3 Step 0 retrofit + 3 MAJOR fixes#3
artyhoo merged 4 commits into
mainfrom
chore/phase-3-aif-retrofit

artyhoo commented May 7, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

artyhoo commented May 7, 2026

Summary

Commits

Reuse-vs-build matrix (from phase-3-research.md)

Test plan

References

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Reuse-vs-build matrix (from `phase-3-research.md`)