research(dn-decisions): R-phase DN-1 permission-boundary + DN-4 ship-vs-gate (best-practices + companion survey) - #262
Merged
Conversation
…undary + DN-4 ship-vs-gate
DN-1 verdict: Option C — Bash(bash helpers/*.sh *) glob pattern in allowed-tools, backed
by CC permissions docs (glob supported), OMO per-command granularity precedent, Cline
deny-override-allow pattern, and NVIDIA/AWS/OWASP least-privilege consensus. Critical caveat:
CC Issue #14956 (open bug: skill allowed-tools may not auto-approve) — settings.json fallback
required. Also surfaces DECISION-NEEDED on ${CLAUDE_SKILL_DIR} expansion in frontmatter fields.
DN-4 verdict: REVISE-TO-DEFER — defer F6 entirely, consistent with maintainer 2026-05-27 stance.
Backed by Superpowers (no speculative fixes), OMO (direct implementation over premature abstraction),
dual-implementation-discipline.md §3 (meta-orchestrator = Internal = CC-native only default), and
§1.5c finding that ${umbrella:-} defensive bash coding already mitigates the functional gap.
Prior-art: skipped — R-phase research-patch, not a capability commit (markdown only, no new dependency or new ≥80 LOC code file)
… --> annotation Principle 10 (research-patch annotation invariant) requires every research patch to carry <!-- scope:<slug> --> on its first line. Initial patch (bf95b60) shipped without it; pre-push hook caught the gap. Prior-art: skipped — fixup for principle-10 annotation; not a capability commit.
…IONAL + verbatim Issue #14956 body excerpt Addresses Phase -1 cold-review verdict (1 BLOCKER + 1 MAJOR) on PR #262. BLOCKER (#verdict-without-load-bearing-probe) — §1.5a verdict was claimed `VERIFIED-SYNTAX-SUPPORTED` based on CC docs extrapolation + "SHOULD match" pattern analysis, but the §1.5a kickoff mandate was a runtime probe (invoke synthetic skill, capture verbatim classifier output: auto-approve/prompt/deny). Probe setup was created but NEVER invoked through CC's slash-command interface. Fix: downgraded verdict to `PROVISIONAL-SYNTAX-EXTRAPOLATED — probe-deferred- to-I-phase-Stage-0`. Added explicit probe-execution-disclosure paragraph above the verdict noting (a) what was actually done (setup), (b) what was extrapolated (SHOULD match), (c) what was deferred (runtime classifier observation). Added "MANDATORY I-phase Stage 0 task" — if probe contradicts SHOULD-match extrapolation, fall back to Option A or A+settings-fallback. MAJOR (#claim-without-verbatim) — Issue #14956 citation at §1.3 Search 6 was Worker's paraphrase/summary, not a verbatim body excerpt. Reviewer caught that title + summary != actual issue text passage; kickoff §1.2 mandated verbatim excerpts for cited sources. Fix: replaced paraphrase with verbatim excerpt from issue body (Description + Actual Behavior + Workaround paragraphs), explicitly marking the title as title-verbatim. State verified via `gh api repos/anthropics/claude-code/issues/14956` (state=open, 2026-05-28). Bug confirmed against CC v2.0.75 per original Worker context. Net change: ~15 LOC across two edits. DN-1 final verdict (Option C narrow glob + settings.json fallback) UNCHANGED — only the claim-strength language is corrected (overclaim → probe-deferred PROVISIONAL). DN-4 verdict unchanged. 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."
15 tasks
artyhoo
added a commit
that referenced
this pull request
May 28, 2026
…delta-write-from-state helpers (DN-1..DN-4) (#263) F.3 I-phase implementation per kickoff `.claude/orchestrator-prompts/meta-orch-f3-iphase/kickoff.md`. Closes the F.3 helper-collapse arc started by PR #261 (R-phase verdict) + PR #262 (DN-1..DN-4 best-practices research). Changes: - NEW `.claude/skills/meta-orchestrator/helpers/dispatch-from-state.sh` — emits dispatch state (winner_id, sub_wave_state) + head-120 kickoff body. Replaces the §3 inline `cat .../kickoff.md` block per DN-3 A. - NEW `.claude/skills/meta-orchestrator/helpers/delta-write-from-state.sh` — sibling helper to `update-delta.sh`; writes the two arrays (`untracked_seen` + `closed_since_last`) in `_master-backlog-delta.json`. Replaces the §10 step 5b inline jq block per DN-2 B. Existing `update-delta.sh` UNCHANGED — preserves paired-negative test contract. - NEW paired-negative tests: `packages/core/hooks/dispatch-from-state.test.ts` (6 cases) + `packages/core/hooks/delta-write-from-state.test.ts` (7 cases). All 13 pass. - SKILL.md: added `Bash(bash \${CLAUDE_SKILL_DIR}/helpers/*.sh *)` to `allowed-tools` frontmatter (DN-1 Option C narrow glob). Added Issue #14956 caveat note in §0 — settings.json fallback is load-bearing while #14956 stays open. §3 launch-table now invokes `dispatch-from-state.sh`. §10 step 5b now invokes `delta-write-from-state.sh` then `update-delta.sh`. Added `<!-- @dual-pair: meta-orchestrator-dispatch-from-state -->` marker. - wave-sequencing-plan.md §0: single-line row update — F.3 helper-collapse I-phase done (this PR) + R-phase refs #261/#262. Stage 0 probe (DN-1 Option C): Per kickoff §1 Stage 0, the runtime probe via synthetic `tmp-glob-probe/` skill could not be executed autonomously — `/glob-probe` slash-command requires user-input action not available from this session. Substituted with authoritative ruling via `claude-code-guide`: Issue #14956 confirmed OPEN as of 2026-05-27 (labels `bug` + `has repro` + `area:tools` + `area:security`); skill-scoped `allowed-tools` with specific `Bash(<pattern>)` patterns does NOT auto-approve matching commands in current CC version. Verdict matches kickoff §1 Stage 0 Decision-branch row 3: «Issue #14956 still bites in current CC version → settings.json fallback becomes load-bearing; document in SKILL.md header.» No `tmp-glob-probe/` artefacts created (probe done via authoritative ruling, no runtime invocation); zero traces in `git status --short`. Stage 1d (settings.json fallback) — recipe-for-maintainer: `.claude/settings.json` is in the project deny-list (`Edit(.claude/settings.json)` + `Write(.claude/settings.json)` at settings.json:56-57) — agent-uncommittable by design (per maintainer convention; agent preserves snapshot + hands recipe). Required entry to add manually to `permissions.allow`: "Bash(bash *helpers/*.sh *)" Rationale: skill-agnostic fallback path for Issue #14956 — remove this line if/when #14956 closes. ### Removed §3 inline !shell block (was at SKILL.md:242-244) \`\`\`bash \`\`\`! cat ".claude/orchestrator-prompts/\${umbrella:-}/kickoff.md" 2>/dev/null | head -120 || echo "MISSING kickoff" \`\`\` \`\`\` Reason: superseded by `\${CLAUDE_SKILL_DIR}/helpers/dispatch-from-state.sh` per F.3 helper-collapse verdict (PR #261). The inline block was classifier-incompatible under no-arg mode (`\${umbrella:-}` substituted before §2 picks winner, so §3 always emitted "MISSING kickoff" in arg-mode and contributed nothing in no-arg mode). Helper reads winner-id from state-file populated by §2. Test results: - 13/13 new paired-negative tests pass (`packages/core/hooks/dispatch-from-state.test.ts` 6 cases + `packages/core/hooks/delta-write-from-state.test.ts` 7 cases). - `packages/core/hooks/update-delta.test.ts` UNCHANGED, all 4 tests pass — DN-2 B contract preserved. - `packages/core/principles/12-ai-laziness-traps.test.ts` passes (6/6) — §3 inline-block removal does not break principle 12. - 1 pre-existing test failure on `hooks/pre-push.test.ts` (staging baseline, unrelated — references `s17.test.sh` + `hook-stub-completeness.test.sh` that do not exist in current `packages/core/audit-self/`; not introduced by this PR — verified via `git diff origin/staging -- packages/core/hooks/pre-push.test.ts` = zero diff). Prior-art: PR #261 (F.3 R-phase verdict) + PR #262 (DN-1..DN-4 best-practices research-patch) — both already merged to staging.
artyhoo
added a commit
that referenced
this pull request
May 28, 2026
…nvocations + dup-detect empty-arg = --all (#266) Stage 4 of the `meta-orch-no-arg-overview` umbrella. Two coupled wire-level fixes that unblock the skill itself: P4-a closes the form-mismatch between SKILL.md `!`-block helper invocations and the maintainer-side allow-list shape, and P4-b removes the compound `||` chain that no single allow-rule pattern matched. (a) SKILL.md: rewrite 7 of 14 `!`-blocks from direct-path form `${CLAUDE_SKILL_DIR}/helpers/<file>.sh "$arg"` to bash-quoted form `bash "${CLAUDE_SKILL_DIR}/helpers/<file>.sh" "$arg"`. Scope = the 7 blocks that ARE direct-path helper invocations (lines 72, 104, 172, 180, 188, 239, 243). The other 7 blocks (`cat`/`git`/`gh`/`head`/`jq` at 56, 60, 64, 68, 162, 285, 360) are not direct-path helper invocations — the transformation rule does not apply, current form preserved. Semantics preserved: same helper, same args, same exit codes. (b) dup-detect.sh: empty arg now silently falls through to the `--all` iteration (was: emit `Usage:` and exit 1). Lets the SKILL.md §2.5 Step 2 invocation use a single allow-rule pattern instead of the compound `<arg> || --all` chain. (c) packages/core/hooks/dup-detect-empty-arg.test.ts: paired-negative contract — empty-arg-equals-all (positive), explicit-all regression, valid-umbrella-name preserves named-path, missing-umbrella preserves the original `MISSING:` error path (not the old `Usage:`), absent prompts-dir handled gracefully. Pattern mirrors packages/core/hooks/delta-write-from-state.test.ts (F.3 sibling helpers PR #261). The maintainer-side companion — the `.claude/settings.json` allow-rule diff that complements the bash-quoted form — ships in the PR body as a HEREDOC recipe per [[feedback_settings_json_agent_uncommittable]]: the agent cannot Write/Edit `.claude/settings.json` (self-protected deny entry). Maintainer applies the diff manually before the next `/meta-orchestrator` invocation can succeed. ### §1.7 Forward-check applied `build-first-reuse-default.md §3` — six-layer mechanism not required: this is a wire-level rewrite of an existing helper (no new capability, no new substrate). The paired-negative test pattern is ADAPTED from SSOT #55 (Superpowers TDD-for-Skills, prior-art-evaluations.md#55) and the helper-testing variant is established at `packages/core/hooks/delta-write-from-state.test.ts:30-99` (F.3 helper-collapse PR #261, 2026-05-28). `no-paid-llm-in-ci.md §1` — all work is session-bound; the new test is pure-bash spawnSync + vitest, no API-billed call. `doc-authority-hierarchy.md` — SKILL.md Authoritative-for header preserved; no goal-redefinition. `parallel-subwave-isolation.md §1` — work executed in worktree `/Users/art/code/rules-as-tests-aif-overview-s4/`, branch `feat/meta-orch-no-arg-overview-s4`, not primary workdir. ### §1.7 Backward-check applied No artefact silently superseded. SKILL.md frontmatter `allowed-tools: Bash(bash ${CLAUDE_SKILL_DIR}/helpers/*.sh *)` already covers the bash-quoted form — this commit makes the helper invocation sites *use* that form rather than the direct-path form. SKILL.md §0 caveat about CC issue #14956 still applies — the maintainer-side `.claude/settings.json` allow-rule is the load-bearing fallback per DN-1 Option C (PR #262 §3). The compound `|| --all` chain at SKILL.md §2.5 Step 2 line 172-174 collapses into a single invocation; the collapsed form preserves all existing helper output paths (POTENTIAL_DUPE / MISSING / OK / `(no orchestrator-prompts dir)`). The named-path branch in dup-detect.sh:74 is preserved verbatim — only the empty-arg edge behaviour changes. Reference: kickoff §5 backward-check at `.claude/orchestrator-prompts/meta-orch-no-arg-overview/kickoff.md:156`. Prior-art: prior-art-evaluations.md#55 (Superpowers TDD-for-Skills, ADAPT — extends paired-negative discipline to helper testing; helper-testing variant established at packages/core/hooks/delta-write-from-state.test.ts shipped under F.3 helper-collapse PR #261).
artyhoo
added a commit
that referenced
this pull request
Aug 18, 2026
… (#1475) * docs(pre-merge-carrier): S0 design spec + prior-art SSOT rows #259-#263 Design stage S0 of the pre-merge-carrier umbrella (kickoff merged as PR #1469). Deliverables: (1) prior-art consult — SSOT rows #259 (git merge-tree, DEFER), #260 (GitHub merge queue, REJECT), #261 (nektos/act, REJECT), #262 (jjq, DEFER), #263 (timeliner pre-merge-local.sh, ADAPT) + #176 Last-reviewed touch; (2) design spec with sections (a)-(g), measured four-lane interlock inventory, both #1465 riders landing in build stage B1 (same change as the carrier ship, per the ratified record), falsifiers carried verbatim; (3) three DECISION-NEEDED forks (delivery depth, ancestor case, NOT-COVERED policy) left open. Prior-art: prior-art-evaluations.md#259 #260 #261 #262 #263 (this commit IS the consult record; docs-only, no capability shipped) * docs(pre-merge-carrier): T19 round-1 fixes — UI-preset lanes in §b/§a.4/§g, append-only SSOT restored Cold review round 1 (REVISE) findings applied: - MAJOR-1: three UI preset lanes (react-next/react-spa/react-native, delivered as consumer ci.yml via setup.d/40-configs.sh:399/414/437) added to the §b inventory (measured: 0-hit shared-resource grep; new interlocks = fixed TCP ports + playwright browsers), §a.4 derivation, §g B2 scope; population corrected to seven lanes. - MAJOR-2: prior-art-evaluations.md restored to origin/staging content + clean append of rows #259-#263 and the #176 Last-reviewed touch (the whole-file prettier reformat mutated historical row #152 'en/*' -> 'en/_' and rewrote ~80 rows it never touched — reverted). - MINOR-1/2/3: code-span spacing in new SSOT rows; VACUITY included in the §a.2 terminal-verdict enumeration (three shas + ledger on every outcome); Russian citation removed from §b.2 trap-5 row. - Reviewer note (a): the part-1 landing-site choice (new probe script vs docs-only recipe) now stated explicitly in §d as a resolved design choice. * docs(pre-merge-carrier): fix the react-next census claim (ls-blind negative) Cold fidelity audit round 1 (GO) notes-lane MINOR: the spec said packages/core/templates/react-next/ is EMPTY on the strength of a bare `ls`, which hides dot-entries — `find … -type f` returns 2 (.storybook config). The design conclusion is unchanged (the directory carries no CI template, and the repo-wide census `find . -name 'github-actions-ci*.yml'` returns exactly the 7 lanes §b inventories); only the measurement claim is corrected, and the ls-blindness is named so a later reader does not repeat it. Scope-neutral: no deliverable, constraint, or decomposition row changes. --------- Co-authored-by: Test <test@example.com>
artyhoo
added a commit
that referenced
this pull request
Aug 18, 2026
…s-server lane) + both #1465 riders (#1479) * feat(audit-self): opt-in pre-merge carrier (ts-server lane) + ci-available-probe B1 of the pre-merge-carrier-build umbrella (#1466/#1465; spec 2026-08-18-pre-merge-carrier-design.md §a/§d). pre-merge-local.sh — gates the MERGE RESULT, never the head (#1466 class): throwaway worktree at $(git rev-parse --git-path getff/), git merge --no-ff of base into head, npm ci + npm run validate on the result. Three-sha report (head/base/merge) on every verdict; exit table 0 PASS / 1 FAIL / 2 CONFLICT (distinct outcome + GitHub-runs-nothing warning) / 3 CANNOT-RUN (named) / 90 VACUITY (a declared gate that never reported — npm's per-gate header line is the signal, not the command echo); F2: base already contained -> proceed, merge = head, shas intact. NDJSON run ledger + retained PASS logs live under --git-path getff/, outside the worktree. Atomic mkdir lock (exit 75). Honest framing per §e: "LOCAL PRE-MERGE PASS" (never "CI green"), NOT-COVERED list on every verdict, §e.4 copy-paste PR-body citation block. Bash 3.2-compatible. ci-available-probe.sh — #1465 part 1: classifies GitHub CI state with the Actions-quota third state. A CI waiter reading only red/green misreads a DEAD CI (GitHub Free private repos share 2 000 Actions minutes/month per ACCOUNT) as red: every first-party (app.id 15368) check fails ~2 s with zero executed steps while third-party checks stay green. The probe detects that signature -> "CI UNAVAILABLE" exit 4, distinct from RED, and fetches the check-run annotation naming the true cause. gh absence degrades to a named CANNOT-RUN (exit 3), never a hard dependency. Delivery is file-landing only (W-3): no husky/validate/CI wiring rides along — promotion to default-on is B3's trigger-gated decision. Prior-art: prior-art-evaluations.md#263 (ADAPT — artyhoo/timeliner PR 229 reference: worktree+merge --no-ff+same-validate+vacuity control transfer; the six timeliner interlocks re-derived per lane, not vendored — Turbo/ pnpm/Postgres absent from all seven shipped lanes). Prior-art: prior-art-evaluations.md#259 (DEFER — git merge-tree lacks the runnable checkout the lane gates need), #260 (REJECT — merge queue: Enterprise-only + spends the same metered minutes), #261 (REJECT — act simulates the merge, does not construct it; Docker dep), #262 (DEFER — jjq owns the landing and is jj-only) — the S0 consult this build cites. * test(audit-self): live-fire self-tests — arms 1-10 + probe P1/P2 B1 Task 3 (spec §a.6 + kickoff §3). T-PMC-D contract: every arm RUNS the real carrier/probe against a seeded fixture repo and asserts OUTPUT + exit code — nothing greps the scripts' source text. Carrier arms: three-sha block asserted on EVERY arm (W-1/T-PMC-B); seeded failing gate -> 1; seeded conflict -> 2 (+ the GitHub-runs-nothing warning + conflicted file named); vacuity -> 90 via a shim that skips a declared gate while the aggregate exits 0 (the #1466 item-1 shape); PASS run retains its log OUTSIDE the worktree and cleans the worktree up; F2 containment (base ancestor of head) proceeds with merge = head; real-merge sharp edge — verified sha must differ from head sha when base advanced; .nvmrc major mismatch -> 3 with the pin named; unresolvable base -> 64 with NO ledger line; held lock -> 75. Probe arms: gh absent from PATH -> 3 named (invoked via /bin/bash under an emptied PATH with only a git symlink); a gh shim (emitting post---jq output) drives the four classifications: quota signature (zero steps, sub-5s first-party failure + green third-party) -> 4 with the annotation's cause named; real red with steps -> 1; all green -> 0; no check-runs -> 2 with the merge-conflict-state hint. Fixtures are throwaway git repos with a zero-dep package-lock so npm ci works offline; the npm-run-all2 shim loops gates via `npm run` so npm prints the SAME per-gate report headers the production aggregate produces. Prior-art: prior-art-evaluations.md#263 (ADAPT — live-fire self-tests for the carrier capability of commit 8d63f23; same SSOT row, arms re-derived per T-PMC-D, not a new capability) * feat(install): deliver carrier scripts to all profiles + CI wiring + the two #1465 docs lines B1 Tasks 4-7 (kickoff §2 B1 row + §3 delivery-mechanism constraint). Delivery (F1 — ALL profiles, core included; the delivery is NOT profile-gated): setup.d/40-configs.sh copy_safe + chmod_safe +x pairs for pre-merge-local.sh and ci-available-probe.sh in the unconditional scripts section; install.sh do_refresh "Scripts -> scripts/" pairs for both (refresh_safe honours <dst>.override.md — Layer-3 escape preserved). W-3: file-landing only — nothing wires the carrier into husky/validate/CI. Enforcement wiring the new test files force (principle 41's population is every tracked *.test.sh, allowlist empty): a step in the EXISTING audit-self.yml mechanical job runs packages/core/audit-self/pre-merge-local.test.sh; a literal `run: bash tests/install-sh/pre-merge-local-delivery.test.sh` step in the EXISTING install-sh-c job (meta-all-wired greps the literal); the install-sh-c shellcheck step now also sweeps the two new shipped scripts, making the shellcheck-clean constraint CI-enforced; a run-local-ci-sweep.sh gate_table row (rank 5, trigger packages/core/audit-self/) keeps the local sweep predicting CI (sweep-coverage gate: the delivery-test step is battery-covered by the install-sh-suite row; the shellcheck step is a multi-line run: block, outside the single-line population). tests/install-sh/pre-merge-local-delivery.test.sh — installs a fixture consumer at `--profile core`, asserts BOTH scripts land executable and byte-identical to the framework source, asserts W-3 negatively (no package.json/husky wiring), then proves --refresh updates a stale planted carrier copy (paired-negative control included). The two #1465 part-2 docs lines: AI-USAGE-GUIDE.md §3 item 5 names scripts/ci-available-probe.sh (+ the opt-in carrier) for the quota-dead-CI state; INSTALL-FOR-AI.md gains the Actions-minutes fact (2 000/month per account on GitHub Free private repos) + both script pointers — 11 lines, 594/600 against the markdown wall. Prior-art: skipped — delivery wiring, CI steps, docs lines, and a test for the capability committed earlier (cited there as #263 ADAPT); no new capability. * chore(install): snapshot regen for carrier delivery Deliberate baseline regen after the B1 shipped-file edits (kickoff §3): every npm stack gains scripts/pre-merge-local.sh + scripts/ ci-available-probe.sh and the refreshed .ai-factory/AI-USAGE-GUIDE.md hash; python lanes carry the guide-hash change only (the agent surface, not 40-configs scripts/); cargo/go are untouched (toolchain artefacts only). Diff reviewed — no unexpected drift. Prior-art: skipped — snapshot regen after a delivery edit, no new capability. * fix(audit-self): shellcheck disables for SC1003/SC2329 false positives The B1-widened CI shellcheck step (audit-self.yml, pinned-idiom 0.9.0 on the runner) reds on pre-merge-local.sh: SC1003 (info) fires on the tr -d literal set at the NDJSON ledger remote-sanitize line on every shellcheck version; SC2329 fires on cleanup() from 0.10.0 up (trap-invoked, not dead). Both are false positives — annotated with rationale, no behaviour change. Reproduced red at v0.9.0 (exit 1) and green after (exit 0) on both v0.9.0 and v0.11.0. Prior-art: skipped — two-line lint-annotation fix, no new capability * chore(install): snapshot regen after shellcheck-annotation fix Prior-art: skipped — snapshot regen after lint-annotation edit, no capability * fix(ci): quote the B1 self-test step name — unquoted inner colon broke YAML actionlint at pre-push (host harvest gate): audit-self.yml:99:70 'mapping values are not allowed in this context' — the step name contains 'B1: live-fire', and an unquoted scalar with an inner ': ' is invalid YAML, breaking the whole workflow. The sibling install-sh step names added in the same change were already quoted; this one was not. Verified: full-file js-yaml load now passes. Prior-art: skipped — one-line YAML quoting fix, no new capability --------- Co-authored-by: Test <test@example.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
R-phase research-patch backing two inline-recommendation classes that recurred in the
meta-orch-no-arg-lazinessumbrella (PR #261 follow-up) — moves them from session-recall to evidence-cited.allowed-toolsscope for skill helpers: verdict Option C —Bash(bash ${CLAUDE_SKILL_DIR}/helpers/*.sh *)narrow glob, with~/.claude/settings.jsonpermissions.allowfallback while CC Issue #14956 (verified open) remains unresolved. Evidence: CC permissions docs (glob syntax confirmed), OMO per-command granularity ({ git: allow, rm: deny }), Cline deny-overrides-allow, NVIDIA/AWS/OWASP PoLP consensus. Two open implementation questions surface as DECISION-NEEDED-Q1/Q2 for the eventual I-phase.dual-implementation-discipline.md §3(~/.claude/skills/= Internal tooling = CC-native default), and §1.5c probe showing\${umbrella:-}defensive bash in the existing skill already mitigates the functional gap.Deliverable:
docs/meta-factory/research-patches/2026-05-28-dn-decisions-best-practices.md(453 LOC + 1-line scope annotation).Scope: R-phase research-patch only. No code changes; no skill / hook / rule modifications. F.3 and F.6 implementation decisions are out of scope regardless of this R-phase's DN-4 verdict — they trigger separate future umbrellas.
Method (compliance with
.claude/rules/)phase-research-coverage.md §1— all 6 search items + §1.5 probes ran (6 DeepWiki queries, 6 WebSearches, 84-row SSOT sweep, CC primitive probes with safety-mandated cleanup, repo+PR history sweep).recommendation-laziness-discipline.md §3— every verdict in patch §4 cites file:line / URL / DeepWiki excerpt evidence from §1-§3.build-first-reuse-default.md §3— 5 of 6 layers applied (layer 1 = Prior-art trailer = inapplicable per CLAUDE.md Build-vs-reuse invariant: this is R-phase markdown, not a capability commit).no-paid-llm-in-ci.md §1— research via DeepWiki MCP + WebSearch + WebFetch + Bash; zero API-billed LLM calls.parallel-subwave-isolation.md §1— Worker executed inisolation: \"worktree\"ephemeral worktree.ai-laziness-traps.md §3— T1/T3/T4/T7/T11/T12/T13/T15/T16/T19/T20 + domain-specific T-N5/T-N6/T-N7 enumerated in kickoff; T15 self-application addressed in patch §6.§1.7 Forward-check applied
This patch complies with existing disciplines:
phase-research-coverage.md §1— 6-item checklist scaffold executed (patch §1.1-§1.6); see patch §6 forward-check.doc-authority-hierarchy.md §3— patch header carries Class C + Authoritative-for + NOT authoritative-for (patch lines 3-5).recommendation-laziness-discipline.md §3— every §4 verdict backed by ≥1 evidence-bearing tool result cited inline; e.g. DN-1 falsifier patch line 316, DN-4 evidence patch lines 331-336.dual-implementation-discipline.md §3— meta-orchestrator skill classified as Internal tooling = CC-native default; informs both DN-1 and DN-4 verdicts (patch §4 + §3 lines 282-284).no-paid-llm-in-ci.md §1— all research via free-tier tools (patch §6 forward-check, line 365).§1.7 Backward-check applied
New verdicts applied to existing artefacts:
allowed-toolsfield (no other in-repo skill carries bash-helper auto-approval entries — verified via patch §1.6 repo sweep, lines 210-218).prior-art-evaluations.md— no new SSOT rows needed (patch identifies no capability commits; markdown-only output)..claude/rules/*.mdis superseded; the patch extends PR research(meta-orchestrator): no-arg-laziness R-phase — F.3 + F.6 verdict #261 R-phase findings with backed evidence per patch §6 backward-check lines 372-378.DECISION-NEEDED (4 items for maintainer)
Surfaced per
reviewer-discipline.md §2— these are picks for maintainer, not orchestrator/reviewer:\${CLAUDE_SKILL_DIR}expansion inallowed-toolsfrontmatter: CC docs describe\${CLAUDE_SKILL_DIR}as a skill content substitution. Whether it expands in the frontmatterallowed-toolsfield is not confirmed by docs or probes. Determines exact Option C implementation form (\${CLAUDE_SKILL_DIR}/helpers/*.shvs relativehelpers/*.shvs absolute path from settings.json). Patch §8 Q1 lines 411-415.allowed-toolsalone suffices. If still open:~/.claude/settings.jsonfallback required. Patch §8 Q2 lines 417-422.Bash(rm *)/Bash(curl *)deny rules insettings.jsonalongside the Option C allowlist (belt-and-suspenders). Patch §8 Q4 lines 431-436.Out of scope (do NOT cascade out of this PR)
.claude/skills/,.claude/hooks/,settings.json, or rule files (R-phase boundary).Test plan / verify-trace
tmp-glob-probe/+tmp-arg-probe/absent from commit; verified viagit status --short+git diff --stat --cachedpre-commit (Worker REPORT verify-trace).gh api repos/anthropics/claude-code/issues/14956returnedstate: open,created_at: 2025-12-21, title matches Worker citation.research/dn-decisions-best-practicesbased onorigin/staging @ fd93ed5; ahead of staging by 2 commits (initial research + principle-10 annotation fixup).<!-- scope:dn-decisions-best-practices -->present on patch line 1 (fixup commit9488731).medium(3 companion + 3 industry citations on DN-1 direction; DN-4 backed by 2 companion + 1 project rule; medium-not-high due to Q1/Q2 open).Phase -1 audit trail
Cold-reviewer dispatched on kickoff (1× Opus general-purpose subagent) before Worker dispatch. Findings + amendments: