Repository navigation
feat: second harvest of real-run cascade learnings into the kit - #5
Merged
Merged
Conversation
Records the second reconciliation pass with the reference instance: the revised framework-not-tooling constraint (per-file merit test replaces the blanket no-new-rule-files decision), the five operator-settled decisions (reviewer memory, deliberate phase invocation, light CI items, one-PR sequencing, tooling-complex templates), the ten cluster summaries, and the verification battery. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…backend write ask-gate Three new PreToolUse guards, each the enforcement companion of a discipline the kit already documents in prose: protect-main-branch.sh hard-denies git commit on main/master (branch-first per cbk-conventions § Branch naming); guard-pr-state.sh ask-gates gh pr ready/merge/close/reopen (PR-state changes are the operator's calls — the permission prompt is the per-action OK; push and pr-create stay unmatched); require-knowledge-backend-ok.sh ask-gates knowledge-backend MCP writes, making the HITL write discipline allowlist-proof (Notion matcher ships as the v1 reference). settings.json rewires the hook registry around the two-tier vocabulary — hard-deny for never-legitimate actions vs ask-gate for operator-only one-way doors — and all guards fail open with warnings that name the surviving backstop. knowledge-backend.md documents the enforcement layer. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… plan (Step 5 split) Step 5 conflated research with plan mode — plan mode is read-only, so research could never run a spike, a feasibility probe, or the test suite to inform the plan. Split into 5a research (executable, operator's current permission mode: read-only probes, step-0 probes for probe-pending assumptions, scratchpad-scoped throwaway spikes; one hard rule — no repo edits) and 5b plan + gate (entering plan mode is the self-disarm; assumption enumeration, guardrail respect, and the non-negotiable approval gate carry over; approval restores the prior mode so Step 6 runs under the same oversight). Adds the deferred-hardening note pattern (why structural enforcement of the no-repo-edits rule is shelved, the residual-gap severity, the revisit trigger), the spikes-are-throwaway rule, the permission-mode expectation with graceful degradation, and the fan-out grounding pointer. Bundled template re-synced byte-parallel with its prose updated to match. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…egative claims Adds the grounding discipline to both research references (canonical three-rule statement in rough-in's; framing's carries the phase-flavored form and points at it): existence/absence claims are verified with a repo-root-wide search covering every source tree, never a single-package grep; fan-out drafter prompts embed the rule verbatim and verifier passes preferentially spot-check negative claims (a wrong absence-claim ships duplicate machinery, a wrong line number is cosmetic); drafts cite the run's existing grounding corpus instead of re-deriving facts it already covers. New rough-in failure mode 13 (narrow-scope existence grepping) names the signal, the cost asymmetry, and the recovery path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… roster, dropped-coverage rule pr-review.md gains the orchestrated-sweep option (find-then-verify with a refute-by-default adversarial stage before triage; refuted findings need no triage; direct dispatch stays the documented fallback and the hand-off records which path ran) and two invariants that hold either way: triage judgment stays with the caller, never delegated to dispatched agents; and a failed review agent is dropped coverage, not zero findings — failed dimensions and unverified findings are tracked and retried before the pass is complete. § Project-local agents becomes the single authoritative dispatch roster (cross-cutting reviewers unconditional with cheap clean verdicts, domain reviewers path-matched), and the pre-filtered changed-path list is computed once and shared. /finish Step 9 consults the roster instead of keeping its own copy; template re-synced byte-parallel. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The three shipped reviewers gain memory: project frontmatter — a persistent per-project precedent surface that ships empty. pr-review.md's authoring note documents the discipline: four memory genres (out-of-scope precedents with the surface split, clean-review calibration baselines that record the method, de-facto-convention prior art that always names its self-expiry condition, both-polarity conforming-pattern records) and four craft rules — verify by artifact never by the diff's self-description, pair red-flag checklists with maintained do-not-flag guard lists, a clean verdict on out-of-surface diffs is correct output, and no unsourced claims about external-platform behavior. When a rule clause changes or retires, the reviewers that restate it are swept in the same change. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…pattern Suites excluded from the offline CI gate run only when someone runs them, so they get an explicit cadence contract: event-triggered (any merge that re-points a live surface) plus periodic; the sweep runs alone against shared backing services; the sweep task carries its own optional dependency groups or suites skip cleanly. New anti-pattern: frozen-state pins over append-only stores — pin the earliest row byte-exactly, constrain later rows to same-grain content copies, stay direction-agnostic on lifecycle flags the suite doesn't own. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…s, body refresh, milestone-scoped re-frame Three lifecycle grains between full supersession and doing nothing, all operator-triggered and ledger-recorded: an additive completeness pass on an already-roughed-in milestone (classify each instinct present / genuine gap / deferred / category-error-with-reason; append-only, nothing closes); an in-place body refresh for un-executed issues carrying factually wrong mechanisms (provenance comment + ledger row; executed issues never retro-edited; shape changes still close-and-recreate); and framing pattern E, milestone-scoped re-framing (supersede one milestone's shape while the prior frame stays Active with an annotated status and the retired AC set recorded as retired-un-executed — new conventions mutation row). The frame template's ledger guidance broadens into a milestone-lifecycle event vocabulary (completeness passes, body refreshes, capstone-close rows with verdict + demonstrables + downstream consumer, re-affirmations, count deviations with operator approval), and its milestone block gains the inline [F<N>.AC<M>] trace-ID examples the conventions and exit checklist already required — turning the kit's own trace-ID verification grep green. The conventions' trace-ID section records the two-level anchor observed in practice (tests trace to R-issue-local AC numbering; R-issue ACs cite the F-level IDs). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…s, scoped standing authorizations The charter pattern gains a pre-registration flavor for wide comparative milestones: rules frozen in an immutable record before any candidate runs (closed candidate set, named targets, evaluation protocol, expected-outcome pre-commitments including likely-nulls, inherited limitations), with three honesty companions — post-freeze arrivals join a surface-only exploratory tier, deviations land as refining records with honest disclosure, and a standing gate's no-change verdict is recorded as a re-affirmation event. Spikes gating risky adoptions pre-declare promote / named fallback / documented drop. adr-new documents the clause-level Refines header narrative, Refines targeting non-decision clauses, and the honest-disclosure refine. The conventions' HITL heuristics gain the mechanize-absolute-gates guidance (two-tier hooks + the deferred-hardening note) and the scoped standing-authorization rule; verify-against-reality gains the event-grain companion shape and citation-tally verification. The frame template's IC guidance names the authored hand-off-spec discharge shape with its audience-assumption header. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…irical rails scaffold/blueprint/framing/rough-in gain disable-model-invocation: true and rewritten trigger sentences — each is an expensive HITL one-way-door workflow, so an accidental auto-trigger mid-conversation costs more than a missed one. Consultation (the funnel entry, where users don't know the cascade vocabulary) stays model-invocable. CLAUDE.md's invocation section rewritten in lockstep to the split posture. Conventions gain the dated-empirical-rails principle (record operational constraints as dated measurements with failure signature + re-check trigger, note retirements inline; contractual limits re-verify instead) and the tooling-conventions row names model/effort orchestration conventions as a recordable per-project choice. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Pin the kit's own workflow action to a full commit SHA with a version comment, and add the policy to the settle-window section: tag refs are mutable and open to tag-retag compromise, so every uses: pins a SHA, initial pins resolve through the settle window, and the update bot's CI-actions ecosystem maintains them. Two CI-trap paragraphs join the conventions: an unattended CI agent's deliverable is the posted artifact, not its exit code (assert the side-effect landed); and linter-excluded surfaces get their own CI gate or the exclusion quietly becomes a standing exemption. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…lates + review-sweep workflow + Explore exemplar Under the revised framework-not-tooling constraint (per-file merit test; harvest-2 spec D11), the tooling complex ships in the kit's template register. workflows.md: the plan-mode/task-tracking/subagent triad, when-to-plan-mode, plan-revision signals, TDD-where-merited pointers, verification-before-completing, dispatch and tracking judgment, narrate-during-iteration, the workflow anti-patterns table, and a rules index. tooling.md: the built-ins-first principle and the per-surface decision-rule table skeleton with bracketed stack sections (distinct from blueprint's task-runner tooling template — noted inline). orchestration.md: the ceiling rule with the one-top-tier-agent allowance, the role ladder, the effort-vs-model heuristic, the dispatch-mechanism decision, and fan-out discipline — fast-aging platform claims carried only as dated observations with re-verify pointers, cost tables deliberately omitted. review-sweep.js is the shipped find-then-adversarially-verify review orchestration (kit reviewers always-on, bracketed domain-trigger map, coverage accounting, triage stays with the caller) that pr-review.md § When to invoke names as the orchestrated path. Explore.md is the mechanical-tier search exemplar of the role ladder. CLAUDE.md's repo layout and contracts list updated. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ehaviors Framing gains Test 5 (additive increment — pattern D, closing the harvest-1 editorial follow-up) and Test 6 (milestone-scoped re-frame — pattern E, incl. the trace-ID template requirement). Rough-in gains Test 6 (additive completeness pass), Test 7 (in-place body-refresh boundaries), and Test 8 (grounding existence claims in fanned-out research, pairing failure mode 13). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ases Rough-in's intro and cross-test-invariants line still said five tests after Tests 6-8 landed; framing's intro still said three after Tests 4-6. Flagged by four independent review dimensions and confirmed by adversarial verification. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two adversarially-confirmed gaps in protect-main-branch.sh: the regex required commit to immediately follow git, so any global option between them (git -c k=v commit, git --no-pager commit) silently defeated a hard-deny guard; and the branch check queried a fixed project dir instead of the call's working directory, judging worktree/nested-repo commits against the wrong repo's branch. The match is now token-anchored and flag-tolerant (over-matching costs a loud wrongly-blocked call the operator can override; under-matching was a silent bypass — the wrong direction for a deny-tier guard), and the branch is read from the payload's cwd with the project-dir fallback. Behavioral smoke: both former bypasses block; reads, feature-branch commits, and non-git commands pass. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…elete verbs The ask-gate matcher missed notion-convert-page-to-skill — a confirmed page-mutating tool — so that write would have skipped the forced HITL prompt. Add convert and (defensively) delete to the verb alternation and note the widen-on-new-verb obligation in the hook header. Reads stay unmatched. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… the memory version-control call Three adversarially-confirmed findings on pr-review.md: the three reviewer bullets still framed dispatch as conditional (runs when the diff touches X) after the new paragraph declared cross-cutting reviewers unconditional — bullets now state scope, with the unconditional + cheap-clean-verdict rule in the section lead; the memory: project platform claim carried no citation despite the sourced-claims rule added in the same diff — now cites the sub-agents documentation with a re-verify note; and the memory directory's version-control question was undecided — commit it (the exercised default) or gitignore it, and record the choice. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Owner
Author
Review-sweep triage auditRan the PR's own
Refuted (no action, per the refuted-findings-need-no-triage rule) — 11, including: review-sweep.js "fails syntax check" (misread the workflow-script execution model as ESM), TDD-coverage demands on the workflow script and hooks (speculative/no testable-unit path), Sweep stats: 30 agents, 0 errors. Verification battery re-run post-fixes: portability greps unchanged, denylist clean, hooks 🤖 Generated with Claude Code |
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.
Second reconciliation of the kit with the reference instance's continued real-world cascade run (delta since the 2026-07-04 harvest snapshot). Design spec with the settled decisions:
docs/superpowers/specs/2026-08-09-cascade-kit-harvest-2-design.md— including the revised framework-not-tooling constraint (per-file merit test replaces the blanket no-new-rule-files decision).The comparison ran as a 14-agent orchestration (7 surface readers, each adversarially verified against current main and the harvest-1 decisions): 97 classified candidates, ~51 landed after dedup. Thirteen atomic commits, one per cluster:
What lands
/finishStep-5 split — research runs executably (current permission mode; read-only probes, step-0 probes for probe-pending assumptions, scratchpad-scoped throwaway spikes; one hard rule: no repo edits) before plan mode, which is entered as the self-disarm gating the plan. Deferred-hardening note pattern included. Both copies byte-parallel..claude/workflows/review-sweep.js) with direct dispatch as the recorded fallback; triage stays with the caller; a failed review agent is dropped coverage, not zero findings; single-source dispatch roster inpr-review.md.protect-main-branch; ask-gatesguard-pr-stateandrequire-knowledge-backend-ok— each the enforcement companion of a discipline the kit already documents.memory: projecton the three shipped reviewers + the four-genre precedent discipline and craft rules (verify-by-artifact, guard lists, clean-verdict-valid, sourced-claims-only) in the authoring note.[F<N>.AC<M>]examples in the frame template (turns the kit's own trace-ID verification grep green).adr-new.disable-model-invocation: true); consultation stays the auto-triggerable funnel entry; CLAUDE.md lockstep.workflows.md,tooling.md,orchestration.mdin the cbk-conventions template register — bracketed stack sections, dated observations for fast-aging platform claims, cost tables omitted. Plus theExplorecheap-tier search exemplar.Verification
bash -n(+ behavioral smoke on the ask-gate and branch guard);settings.jsonvalid/finishcopies byte-parallel;review-sweep.jsparses🤖 Generated with Claude Code