v6.10.0

✨ Headline

bmad-loop lands as an installable module, and the automator that came before it steps aside. bmad-loop — the successor project for unattended dev-loop orchestration, adversarial review, and deferred-work sweeps — is now selectable straight from the installer picker, driven by the new bmad-dev-auto skill: a single-iteration unattended worker that clarifies intent, creates or resumes a spec, implements, reviews, and finalizes, all off a spec-frontmatter state machine an orchestrator can poll. bmad-automator, the experimental predecessor, is now deprecated in its favor.

Also in this release: party-mode gets an anti-consensus room and two sync fixes, the code-review/edge-case-hunter pipeline gets sharper severity triage and a named-set generalization pass, and bmad-investigate is retired.

💥 Breaking Changes

• bmad-automator deprecated, replaced by bmad-loop (#2532). New installs no longer show BMad Automator in the picker. Existing installs are untouched and keep showing, with a migration hint pointing to BMad Loop. If you're on Automator, plan the move to bmad-loop.

🎁 Features

• bmad-loop — new marketplace module (#2532). BMad's unattended-dev orchestrator ships as an opt-in installer module (bmad-loop, not selected by default). Its skills live behind a .claude-plugin/marketplace.json rather than a normal module.yaml folder, so the installer gained a marketplace-plugin registry flag that routes it through the existing custom-plugin resolver, and now fails loudly instead of installing an empty module if resolution comes up short. Installing the module only stages files — finish setup by running the bmad-loop-setup skill, which installs the orchestrator and wires up per-project hooks and policy; automation doesn't run until that completes. A new post-install-message registry field surfaces this instruction right after install (blocking on interactive installs so it isn't missed, non-blocking with --yes).
• bmad-dev-auto — new unattended workflow skill (#2500 and nine follow-ups). A Quick Dev sibling built to keep moving without a human in the loop, driven entirely off spec-frontmatter status so an orchestrator like bmad-loop can poll it. Hardened through the release: an append-only review-triage log with loopback tracking (#2505); an end-of-run commit so the worktree stays clean into the next iteration (#2506); a fix to the Blind Hunter reviewer, which was wrongly denied project access ("blind" means blind to intent, not to the codebase) (#2507); re-entry on a completed spec to trigger a fresh follow-up review pass (#2508); a final_revision recorded in frontmatter at exit, the only link back from an out-of-tree spec to its in-tree commits (#2522); a closed gap where Finalize could leave status: draft on an otherwise-done run (#2536); and a hardened contract requiring subagents to be invoked synchronously, since there's no event loop to resume a yielded turn (#2543). Reference doc at docs/reference/dev-auto.md (#2519), retitled "Autonomous Development Loops" (#2521). Some of the same prompt fixes were backported to Quick Dev (#2501).
• party-mode: anti-consensus club (#2530). New built-in persona group (Wildcard, Level, Killjoy, Splinter) for decision rooms that resist fast agreement while keeping a human in control. Launch with --party=anti-concensus-club, and use --mode subagent for best results. These can be configured as defaults if you use bmad customize and specify that.
• Two new elicitation methods: Subtraction and Map Is Not the Territory (#2515). Subtraction counters additive bias; Map Is Not the Territory guards against over-trusting a lossy model.
• Edge Case Hunter: named-set generalization pass (#2524). Catches diffs that special-case some members of a fixed set (enum, status code, sentinel, flag) while leaving the rest as silent unhandled branches. Measured catch-rate improvement of 50% to 100% on a real regression, at a 19% token cost per run.

🐛 Fixes

• party-mode stays interactive and the room stays in sync (#2531). Fixes a bug, observed under Codex, where a runtime treated the opening prompt as one-shot and closed spawned agents once satisfied. Party mode is open-ended by default now, ending only on explicit signal, with an opt-in --non-interactive flag; standing agents are kept alive and resumed rather than dropped.
• party-mode: agent-team sync corrected to point-to-point (#2539). Claude Code Agent Teams communicate mailbox-style, not over a shared broadcast channel, so an idle member doesn't see exchanges it isn't addressed in. Docs updated so the lead relays turns; subagent mode, which is genuinely broadcast, is unchanged.
• Code-review triage severity calibration hardened (#2523). Requires reading surrounding source (call sites, guards) before rating severity instead of judging from the diff hunk alone, fixing over-rated unreachable findings, and drops a "prefer conservative when uncertain" tie-breaker that was inflating severity.
• Deletion audit folded into Edge Case Hunter (#2525). Retires the standalone deletion-contract auditor layer, which added cold-start cost for near-zero yield, in favor of a gated deletion check inside Edge Case Hunter's existing turn.
• Review layer invocation normalized (#2526). Removes stale "no access/control" wording from code-review/quick-dev/dev-auto prompts and normalizes Blind Hunter / Edge Case Hunter invocation phrasing.
• Installer accepts Windows custom module paths (#2511). Local paths like C:\modules\foo, C:/modules/foo, and .\foo no longer fall through to the Git-URL parser and get rejected.
• bmad-help reads central config (#2541). Its config data source now goes through the shared four-layer TOML resolver instead of legacy config.yaml/user-config.yaml, fixing communication_language and project_knowledge not reaching the skill.

📚 Docs

• bmad-forge-idea wording tightened (#2513). Overview, session, persona, and exit language rewritten more directly; no behavior change.
• validate-skills exempts deprecated skills from the trigger-phrase check (#2486). Thin compatibility shims (bmad-create-prd, bmad-edit-prd, bmad-validate-prd, bmad-create-architecture) intentionally omit a trigger phrase to steer users to their replacement.

🗑️ Removed

• bmad-investigate retired. It reached the same conclusions as plain investigation at higher cost; the case-file artifact didn't justify the overhead.