v2.4.0 — Methodology instance program complete (TDD·BMAD·SPARC·RIPER·CANVAS) + self-aware integrity

47 merged PRs since v2.3.0 (2026-05-19). Catalog v1.100.0 — 19 plugins · 83 agents (33 forgeplan-aware) · 76 skills · 18 commands.

Headline — the AD/AID-PDLC methodology instance program is complete

Five master-coordinated sub-cycle instances of the ADR-010 contract now ship as installable packs, each with generator≠verifier C4 gates at every handoff:

#	Instance	Pack	Gate
1	TDD (enforced RED→verify→GREEN, frozen test oracle)	agents-tdd	hook-gate=Yes
2	BMAD (greenfield persona walk Analyst→QA)	agents-bmad	hook-gate=Yes
3	SPARC (single feature in an existing system, 5 phases)	agents-sparc (finished to B-orchestrator)	hook-gate=No
4	RIPER (non-trivial production bug, Plan→Execute human gate)	/riper in fpl-skills	hook-gate=No
5	CANVAS (Pencil design-system → Storybook → code)	agents-canvas v1.0.0	hook-gate=Yes

/smith routes all five via the 14-row context matrix.

⚠️ BREAKING — agents-canvas 0.3.0 → 1.0.0

CANVAS is now stack-agnostic: it generates natively in the ONE framework your project declares (Step-0 stack detect from AGENTS.md/CLAUDE.md/package.json → announce, or force-ask). The Lit-master + per-framework-wrappers topology and the Spread fan-out are removed from the default pipeline; Style-Dictionary is one token-tool option, not mandatory; context7 is mandatory for the resolved framework. The tokens-before-code hook-gate now derives per-framework guarded globs and fail-safes on stale state. Existing installs: run /canvas-init once after upgrading.

Added

• /spec-author — light-path SDD (frozen #### Scenario = oracle) + /conformance-vectors — frozen language-neutral corpus + cross-language equivalence harness (live negative-control-verified)
• /forge-insight + insight-watchdog (SessionStart anomaly digest, silent when clean) and /forge-heal (3-tier AUTO/ADI/USER heal-or-task, propose-not-silent) — the self-aware integrity layer
• /methodology-check — pre-activation 4-layer coverage report
• /forgeplan-cookbook — agentic-RAG router over 66 MCP tools + 82 CLI subcommands
• cc-best completed — all 6 sections (claude-md, plugins, agents, hooks, mcp, antipatterns)
• Cross-CLI Tier 0 (ADR-014) — forgeplan MCP + skills + AGENTS.md shim for Gemini/Codex/Goose; Tier 1/2 demand-gated
• CI hardening — unicode-safety, no-personal-paths, workflow-security, install-manifest gates; catalog drift-guard (29 disk-derived assertions); install profiles

Changed

• Agent hardening (marketplace#165/#166/#167): architect-reviewer + guardian gain StructuredOutput-precedence (a Workflow schema wins over the EVIDENCE-writing default) + mandatory finally-release (a crashed reviewer never leaks its claim); smith pre-flight reads forgeplan_claims + a health/stats digest and never routes onto a live-claimed artifact
• AGENT-AUTHORING-GUIDE gains 5 canonical disciplines: ground-truth verification (generator≠verifier, empty-diff=BLOCKER), claim hygiene, StructuredOutput precedence, CLI-vs-MCP contract, autonomous build-workflow discipline (red-lines-in-prompt · spec↔as-built reconciliation · self-verify+claim-sweep on crash) — the latter also codified into build/sprint/do + /forge-cycle
• /autorun RIPER Plan→Execute human gate (unattended default = WAIT)
• Marketplace docs sweep: ARCHITECTURE, DEVELOPER-JOURNEY, PLAYBOOK, ONBOARDING, SMITH, process-from-idea (+RU) — all five methodology instances present and current

Fixed

• discover_finding workaround docs retired (upstream #292 resolved — artifact_status/session_status)
• /decay-watch machine-findings track separated from human deferrals; decay-reminder stale-flag no-op fix
• routing-map read-vs-write forgeplan_claims denylist conflation

Update

/plugin marketplace update ForgePlan-marketplace

then update installed plugins (see README). agents-canvas upgraders: run /canvas-init once.

Full changelog: v2.3.0...v2.4.0