Skip to content

v1.12.0 — accessibility-agent patterns + audit-FP prevention + self-audit fixes

Choose a tag to compare

View all tags
@lasswellt lasswellt released this 16 May 18:18
· 179 commits to main since this release
v1.12.0
96966e8

[1.12.0] — 2026-05-16

Minor release across three themes: (1) incorporating six specialist-agent patterns from GitHub's May-15-2026 accessibility-agent post-mortem; (2) closing two broken pipeline handoffs (sprint-review and ship) plus a self-audit-driven compactness + consistency sweep; (3) a four-iteration recursive cycle that hardened audit-agent false-positive prevention, validated against Anthropic's shipped Code Review Plugin pattern, and converged to a stable rule.

Added — GitHub accessibility-agent pattern incorporations

Six patterns from docs/_research/2026-05-16_github-accessibility-agent-patterns.md (commit 317c47d). The blog reported 3,535 PRs reviewed at 68% resolution rate; we mapped its 8 reusable patterns + 5 failure modes against blitz state and adopted six:

  • KNOWLEDGE.md slice injection into sprint-dev worker prompts (skills/sprint-dev/SKILL.md Phase 0.5 + Dev Agent Prompt Specification item 14). Counters training-data bias per knowledge-protocol.md — workers see project-specific gotchas before generation, not after type-check fails post-hoc. Opt-out: BLITZ_SKIP_KNOWLEDGE_INJECTION=1.
  • Pre-flight complexity gate in sprint-dev/SKILL.md Phase 1.4: complexity_score = story_count * 2 + est_loc / 100. Warn at >40, hard-stop at >80 (escape: BLITZ_SPRINT_COMPLEXITY_OVERRIDE=1). Prevents token-explosion that the ratchet only catches retrospectively.
  • Sequential review fallback in sprint-review/SKILL.md Phase 2.2.0: triggered by BLITZ_REVIEW_SEQUENTIAL=1 or diff >2000 LOC. Passes prior reviewer findings as ## Prior Reviewer Findings context to the next spawn. Default parallel behavior unchanged.
  • Reviewer "Instruction Gaps" field in sprint-review/references/main.md: non-empty entries route to .cc-sessions/KNOWLEDGE.md under ## Skill Instruction Drift — <reviewer-role>, creating an automated feedback loop from reviewer agents to skill authors.
  • Phase 3.7 Automation Coverage in sprint-review/SKILL.md: declares deterministic-gates-passed vs human-judgment-required boundary. Sets REVIEW_RECOMMENDATION to auto-merge-safe (all gates + Phase 2.5 full + zero critical/major) or needs-human-review otherwise. Mitigates F4 over-confidence by making the gap explicit in the report.
  • Mandatory Playwright gate in sprint-review/SKILL.md Phase 2.5: when MCP available, skipping smoke test counts as phase_2_5_coverage: partial and surfaces in Phase 4 Recommendations. When unavailable, gate still skips silently (gap, not failure).

Added — Audit-Agent False-Positive Prevention

Self-Falsification + Confidence pattern (commit 6199a65), mirrored on Anthropic's shipped Code Review Plugin (github.com/anthropics/claude-code/plugins/code-review, 129K+ installs, <1% FP rate). Driven by 3 false positives in the 2026-05-16 self-audit + literature review (CHIIR 2026, EMNLP 2025, arxiv 2309.11495 CoVe).

  • skills/codebase-audit/references/main.md Rules block: new rule 2 (Falsify before recording — count/negative/duplication artifacts) + rule 3 (Confidence: 0-100 on every finding). Pillar agents inherit immediately. Refined across 3 iterations to add: routing for Confidence < 50 to ## Discarded Drafts, ## Verified Clean section for no-violation reports, and count-discipline rule disambiguating grep -l | wc -l (files) from grep -rn | wc -l (hits).
  • skills/_shared/agent-prompt-boilerplate.md §Self-Falsification: inheritance target for all future audit-style skills via Pattern A author-time reference. Same 5 clauses as above plus a ${VAR} output-path resolution clause to prevent agents from taking placeholder text literally.
  • skills/_shared/shortcut-taxonomy.md detector #20 (Unverified pattern-match claim): P3 advisory tier; canonical grep pattern in §3 that flags Evidence blocks with count-only claims or missing Confidence scores.
  • agents/critic.md §2.9: audit-finding integrity check that fires detector #20 against audit findings files in the sprint diff. Advisory only — adds findings to critic issues[] with severity: advisory, signaling re-run with Self-Falsification rule.
  • skills/codebase-audit/SKILL.md Phase 2.1.5: confidence threshold filter (default 80, tunable via BLITZ_AUDIT_CONFIDENCE_THRESHOLD). Filters findings below threshold before deduplication. Findings missing Confidence trigger detector #20 advisory.

The four-iteration cycle (initial implementation → generous test → blind retest #1 → blind retest #2) reached a fixed point: blind retest #2 surfaced zero new rule gaps and only two Confidence-65 housekeeping items, both verified real on independent falsification.

Added — Shared Protocols

  • skills/_shared/skill-cross-references.md (new) — canonical source-of-truth for the 2-line "Additional Resources" block (spawn-protocol + terse-output refs) shared by 7 SKILL.md files. Each file carries an <!-- import: from _shared/skill-cross-references.md ... --> marker. No runtime line reclaim — Claude Code's skill loader needs each SKILL.md to declare its own resources — but provides a canonical-wording target plus a drift-detection grep snippet at the file's bottom.
  • skills/_shared/project-context.md (new) — canonical source-of-truth for the ## Project Context heading + detect-stack.sh invocation shared by 29 SKILL.md files. Same <!-- import: --> marker pattern. design-extract is intentionally excluded (has bespoke body). The independent-falsification process that produced this file disambiguated three counts (30 files with heading, 29 with full block, 30 grep-hits) — the kind of confusion the new count-discipline rule prevents.

Fixed — Broken Pipeline Handoffs

  • skills/sprint-review/SKILL.md Phase 0.0 — hard-fail input gate on sprint-registry.json, ${SPRINT_DIR}/manifest.json, and ${SPRINT_DIR}/stories/S*.md. Override (not recommended): BLITZ_REVIEW_NO_MANIFEST=1. State-handoff.md declared sprint-review consumed manifest.json but the skill never gated on its existence — would proceed silently on missing input. Bash block lives in references/main.md §Phase 0.0 Input Gate.
  • skills/ship/SKILL.md Phase 0.1 — added [ -s "${SPRINT_DIR}/review-report.md" ] || exit 1 check when sprint context exists. Ship would previously cut a release without a passing review when the review-report file was missing.

Changed — Self-Audit Driven Sweep

Three waves of fixes from docs/_research/2026-05-16_blitz-self-audit.md (commits b06fa08, 4182ea0, cbada5b):

  • 18 SKILL.md compatibility floors bumped from >=2.1.50 to >=2.1.71 (CLAUDE.md-declared project floor). Affects ask, bootstrap, browse, completeness-gate, dep-health, fix-issue, health, migrate, next, perf-profile, quick, refactor, release, retrospective, test-gen, todo, ui-audit, ui-build. design-extract retains >=2.1.117 (holistic-machine orchestrator dependency, now annotated with an HTML comment in the file).
  • agents/orchestrator.md §2 routing matrix expanded from 19 to 38 skills (full coverage of the user-invocable skill catalog). Reorganized into 5 intent groups: greenfield/setup, sprint pipeline, research/audit/quality, dev/maintenance, diagnostics/meta. The two greenfield-pipeline entry points (bootstrap and roadmap) were previously unrouted from natural-language input.
  • skills/_shared/verbose-progress.md Sprint Selection Verbosity section: 300→247 lines (−53). Replaced 85-line per-format ASCII-art dashboard examples with a 4-row trigger/format/content table plus three condensed canonical examples. Box-drawing dashboards still acceptable but the inline form is now also valid.
  • skills/_shared/spawn-protocol.md §7 trimmed: ~21 lines reclaimed via Output Style prose tightening; Historical Reference section (v1.4.0 merge notes) deleted.
  • 7 SKILL.md persona preambles trimmed (bootstrap, completeness-gate, dep-health, retrospective, migrate, release, perf-profile): dropped "You are a X. You..." framing while keeping the imperative-mood description. The frontmatter description already conveys the role. ui-audit and browse preserved as-is — their preambles contain meaningful read-only constraints and loop-mode behavior worth keeping inline.
  • v1.4.0 SendMessage tombstone dedup (research, sprint-plan, sprint-review, codebase-audit): five identical historical paragraphs collapsed to single-line synthesized by orchestrator (not peer-to-peer, per spawn-protocol.md) references. Exact duplicate at research/SKILL.md:155 deleted entirely.
  • skills/research/SKILL.md:92 dangling reference: relative ../_shared/token-budget.md → canonical absolute /_shared/token-budget.md.
  • skills/health/SKILL.md description: "activity feed" → "activity-feed" (hyphen form when referring to the .cc-sessions/activity-feed.jsonl file, matching retrospective's description).
  • skills/design-extract/SKILL.md: HTML comment after frontmatter explains why compatibility floor is >=2.1.117 (requires holistic-machine orchestrator for the DESIGN.md handoff to ui-build, frontend-design, and design-critic).

Compatibility

No breaking changes. Drop-in upgrade from v1.11.2.

  • New env vars introduced (all opt-out / tunable): BLITZ_SKIP_KNOWLEDGE_INJECTION, BLITZ_SPRINT_COMPLEXITY_OVERRIDE, BLITZ_REVIEW_SEQUENTIAL, BLITZ_REVIEW_NO_MANIFEST, BLITZ_AUDIT_CONFIDENCE_THRESHOLD.
  • New shared protocols (5 total now at 23 files: was 21): skill-cross-references.md, project-context.md.
  • New detector (#20 in shortcut-taxonomy.md): P3 advisory tier, does not block sprint-review.
Assets 2
Loading