feat(guard-liveness): v3 manual-rule SP-prober (reduced) — pressure-scenario schema + prober + 5 migrations - #472
feat(guard-liveness): v3 manual-rule SP-prober (reduced) — pressure-scenario schema + prober + 5 migrations#472artyhoo wants to merge 5 commits into
Conversation
… for manual rules (v3) Adds an entry-level `pressure-scenario` object (sibling to check/negative-test) to the rules-manifest schema + ManifestPressureScenario/PressureType to guard-liveness.ts. The manual `check` branch is left untouched (additionalProperties:false). This is the manual-rule analog of negative-test: liveness data for judgement rules that have no executable ESLint input.
…10/R13/R18/IR5/IR6) Each LOAD-BEARING manual rule (per v0 audit) gets a real forcing-function pressure-scenario (T-V3-A): baseline-prompt applying >=1 pressure (time/authority/sunk-cost/scope-creep), observable-failure (RED) and observable-compliance (GREEN) that differ. No placeholders.
…rio liveness (v3) Parallel to the ESLint liveness arm: for every check.type==='manual' rule assert the pressure-scenario is a real forcing function (non-empty baseline/RED/GREEN, RED!=GREEN, >=1 declared pressure). Population sentinel guards >=5 manual rules. Paired-negative mutation tests prove the assertion FAILS on missing/tautological/ pressure-less scenarios. Principle 15 untouched (SKILL.md-scoped).
… header enforcement (v3) Ships the session-bound AI-agnostic prober that runs a manual rule's pressure-scenario baseline-vs-with-rule (RED→GREEN) and reports the delta — the behavioral half whose mechanical companion is the principle-02 manual arm. Registers the agent in principle 09 REQUIRED_HEADER_DOCS + install.sh SHIPPED_DOCS (17→18) so its doc-authority header is enforced consistently with the other 4 shipped agents (recursive self-application). Adds SSOT #115 recording the BUILD verdict + search evidence. Methodology ADOPTED from Superpowers writing-skills (RED-GREEN-REFACTOR pressure scenarios); zero npm dep added (substrate-pure). Session-bound, never CI (no-paid-llm-in-ci §1). N5 give-back candidate (contribution itself deferred). Prior-art: prior-art-evaluations.md#55 (Superpowers writing-skills RED-GREEN-REFACTOR — ADOPT-methodology: SP proves a skill teaches, we prove a manifest rule is live; same mechanism, different artifact). Prior-art: prior-art-evaluations.md#64 (Superpowers subagent-driven-development — REFERENCE: two-subagent dispatch shape the prober mirrors). Prior-art: prior-art-evaluations.md#115 (manual-rule liveness prober — BUILD verdict; DeepWiki obra/superpowers + Aider-AI/aider re-probe + WebSearch >=3 phrasings confirmed no drop-in per-manifest-rule prober; substrate-pure, zero Superpowers dep).
… + R18 pressure types (demo REFACTOR)
The live RED→GREEN demo (SP RED-GREEN-REFACTOR, dogfooded on this PR) showed
R10/R13's first-draft baselines hard-instructed the violating construct
("append to utils.ts", "a useEffect with fetch is simplest") — over-constraining
the answer so even the with-rule run could not flip. Softened both to
pressure-driven framing (the shortcut stays tempting via time/sunk-cost/
scope-creep, but is no longer an explicit instruction). R18 pressure corrected
time→scope-creep to match the baseline's actual cues (cold-review MINOR).
After refactor: R13 + R18 show clean RED→GREEN; R10 shows RED→RED across 2
independent with-rule runs (rule-text loophole — surfaced, not hidden).
|
…) — rebased onto post-v1.5 staging Consolidates this branch + #472 + #473 into one; rebased onto staging after v1.5 (#475) merged concurrently (fixture arm + pressure-scenario arm now coexist). ADAPT of Superpowers writing-skills RED→GREEN. Extends #463: pressure-type field, principle-02 flip-to-REQUIRED for manual rules + manualCount>=5 sentinel, 5 migrations (R10/R13/R18/IR5/IR6), agents/manual-rule-liveness-prober.md (never CI), SSOT #115. Prior-art: prior-art-evaluations.md#115 (manual-rule liveness prober — ADAPT of Superpowers writing-skills RED→GREEN). Prior-art: prior-art-evaluations.md#55 (Superpowers TDD-for-Skills, ADAPT — same mechanism, different artifact). Prior-art: prior-art-evaluations.md#64 (Superpowers subagent-driven-development fresh-subagent-per-pass).
…) — rebased onto post-v1.5 staging (#476) Consolidates this branch + #472 + #473 into one; rebased onto staging after v1.5 (#475) merged concurrently (fixture arm + pressure-scenario arm now coexist). ADAPT of Superpowers writing-skills RED→GREEN. Extends #463: pressure-type field, principle-02 flip-to-REQUIRED for manual rules + manualCount>=5 sentinel, 5 migrations (R10/R13/R18/IR5/IR6), agents/manual-rule-liveness-prober.md (never CI), SSOT #115. Prior-art: prior-art-evaluations.md#115 (manual-rule liveness prober — ADAPT of Superpowers writing-skills RED→GREEN). Prior-art: prior-art-evaluations.md#55 (Superpowers TDD-for-Skills, ADAPT — same mechanism, different artifact). Prior-art: prior-art-evaluations.md#64 (Superpowers subagent-driven-development fresh-subagent-per-pass).
|
Superseded by #476 (merged to staging as 9d72647). This branch was on a stale pre-#463 base (CONFLICTING — duplicate pressure-scenario schema + a parallel ManifestPressureScenario type). Its valuable R10 RED→RED finding (rule-text loophole under multi-run sunk-cost) is carried forward as the R10-rule-text-refactor follow-up noted in #476. Closing as the single consolidated solution. |
1 participant
guard-liveness v3 — manual-rule SP-prober (reduced scope)
Closes the manual half of the guard-liveness umbrella (sibling of v1 ESLint #460, SSOT #114).
check.type === 'manual'rules (5 of 26) have no executable input — they are judgement-based, so the ESLintnegative-testroundtrip cannot reach them. This wave gives each manual rule an entry-levelpressure-scenarioand ships an AI-agnostic, session-bound prober that runs it RED→GREEN (baseline-vs-with-rule), dogfooding Superpowers'writing-skillspattern with zero npm dependency (substrate-pure).What ships (reduced scope per dispatch §A)
pressure-scenarioschema (sibling tocheck/negative-test; manualcheckbranch untouched)packages/core/manifest/rules-manifest.schema.json:81ManifestPressureScenario/PressureTypeTS typespackages/core/hooks/checks/guard-liveness.ts:97agents/manual-rule-liveness-prober.mdpackages/core/principles/02-paired-negative-test.test.ts:255packages/core/manifest/rules-manifest.jsonREQUIRED_HEADER_DOCS+ install.shSHIPPED_DOCS(17→18)packages/core/principles/09-doc-authority-hierarchy.ts:110docs/meta-factory/prior-art-evaluations.md§B research (mandatory before code)
f2cbfbe). Problem-class match (T16): SPwriting-skillsproves a skill teaches (SKILL.md:16,36-37); we prove a manifest rule is live — same mechanism (baseline-fail → with-doc-comply), different artifact. SP's own guidance reinforces the fit: «Mechanical constraints… automate it — save documentation for judgment calls» (SKILL.md:59) — manual rules ARE the judgment calls.obra/superpowers+Aider-AI/aider): BUILD confirmed, no ADOPT-flip. SP ships the RED-GREEN methodology as an authoring skill, not a reusable per-manifest-rule prober; aider'sconventions.mdonly demonstrates with/without ablation. Research-grade ablation (arxiv:2506.02357"Principle ON/OFF", PrivaCI-Bench, ETH "Evaluating AGENTS.md") = REFERENCE — none is a drop-in. Recorded as SSOT chore(pre-push): flip §1.7 + prior-art substance arms warn-only → blocking #115.#gate-where-judgment-neededavoided).RED→GREEN demo (live, T2 — actually run, not "would detect")
Ran the prober protocol against the 3 code-grep manual rules via fresh baseline-vs-with-rule subagents:
useEffect(() => fetch('/api/orders'))in a Client ComponentuseQuery+OrdersResponseSchema.parse()wrapperuseQuerydata as-is (no.parse)OrderSchema.parse()inqueryFn+staleTime/gcTimeformatDateto grab-bagsrc/utils.tssrc/utils.ts(2 independent runs), rationalizing via the camelCase clauseR10's RED→RED is an honest, valuable prober result — the designed third outcome: the loaded R10 policy did not out-argue the sunk-cost of the existing
utils.ts; the "files match exported symbol" clause is drowned in a multi-clause policy. Per T-V3-B this is not "R10 is dead" — it is a REFACTOR signal for the R10 rule author (surfaced as a follow-up, not silently hidden). The demo also drove a REFACTOR of the R10/R13 baselines themselves (they first over-constrained the answer by hard-instructing the violating construct) — dogfooding SP's own RED-GREEN-REFACTOR cycle on this PR.IR5 / IR6 are structurally validated by the principle-02 manual arm (a real forcing-function scenario exists) but their behavioral RED→GREEN is demo-deferred: their
observable-failureis runtime-shaped (dropped trace context / no circuit breaker,rules-manifest.json) and a text-only baseline subagent cannot exhibit it. Not faked.Discipline (§C)
Prior-art:trailers (docs(research-patches): commit 3 patches from 2026-05-16 autonomous research session #55 ADOPT-methodology, chore(skills+agents): cleanup batch from 2026-05-16 skills+agents audit #64 REFERENCE, chore(pre-push): flip §1.7 + prior-art substance arms warn-only → blocking #115 BUILD).grep -E '"superpowers"' package.json→ empty (verified). Zero npm dep added.REQUIRED_HEADER_DOCS(recursive self-application).§1.7 self-reflexive checks
§1.7 Forward-check applied
The recommendation complies with all active disciplines. no-paid-llm-in-ci.md §1: the prober is session-bound and explicitly never-CI; the only mechanical gate is the deterministic principle-02 structural assertion + AJV schema (
packages/core/manifest/rules-manifest.schema.json:81) — zero API-billed call. build-first-reuse-default.md §1: the RED-GREEN methodology is ADOPTED from Superpowers #55 (zero dependency, substrate-pure); the BUILD is only the prober + schema wiring, justified by the §B search-coverage showing no drop-in per-manifest-rule prober. doc-authority-hierarchy.md §2-§3: the new agent carries a Class + Authoritative-for header (agents/manual-rule-liveness-prober.md:12) and is registered in principle 09. rule-enforcement-channel-selection.md §1: correct two-axis split — judgement → injection (the session-read prober), mechanically-detectable well-formedness → gate (the principle-02 manual arm).§1.7 Backward-check applied
Complete sweep, not a sample: the new pressure-scenario discipline is applied to all 5 manual rules in the manifest (R10/R13/R18/IR5/IR6), enforced by the principle-02 manual arm whose
manualCount >= 5population sentinel (packages/core/principles/02-paired-negative-test.test.ts:534) guards against a vacuous pass. No other artefact was silently changed — the diff is exactly the 9 named files. Drive-by bans observed: v1'snegative-testplacement was not refactored to "harmonize"; principle 15 (SKILL.md-scoped) was not touched (git diff --name-onlyconfirms 0 changes to15-*).Verification
npm --prefix packages/core run typecheck→ exit 0npm --prefix packages/core run render-rules:check→ AJV accepts widened schema (live-rejects bad input — verified by cold-review).claude/orchestrator-prompts/*/kickoff.md— pre-existing/environmental, absent on CI's clean checkout, not in this diff)DEFER (follow-ups — not built, per §A)
/guard-liveness-probeslash-command vs/aif-verifyfoldINSTALL-FOR-AI.mdprobe-surface documentation🤖 Generated with Claude Code