v2.2.0 - Install into every harness, automatically

The installer now sets Muster up across all your Agent Skills harnesses in one command.

What's new

install.sh clones Muster once to ~/.muster and symlinks the seven skills into every harness it detects:

• Claude Code (~/.claude), Mistral / Vibe (~/.vibe), Codex (~/.codex), Gemini (~/.gemini), Cursor (~/.cursor).

Because they're symlinks to the one clone, git -C ~/.muster pull updates every harness at once.

curl -fsSL https://raw.githubusercontent.com/markstent/muster/main/install.sh | bash

• Target specific harnesses: bash install.sh claude vibe (creates the dir if needed).
• Copy instead of symlink: MUSTER_COPY=1 bash install.sh.
• bash 3.2-compatible (macOS default).

README updated: Option B is now a one-command, all-harness install.

v2.1.0 - Plain-language feedback

The commands now read plainly for users of any technical level, while the rigorous process underneath is unchanged. This is an editorial pass on user-facing output only.

What changed

• Two-layer output: every gate and verdict leads with a plain sentence (what just happened) and a clear Your move: / Next: call-to-action; the technical tables stay underneath.
• Jargon translated on the way out: internal terms (two-track, vertical slice, tracer bullet, burnout cap, state machine, inner review, worktree) no longer appear in printed output - they remain only in the instructions the agent follows. Labels and risk are glossed wherever shown.
• Friendlier slash-menu descriptions for all seven commands.
• README: new "In plain words" quickstart + a Glossary.
• CONTRIBUTING: the two-layer principle is now part of the Output discipline.

Unchanged

Process behaviour, GitHub label names, risk thresholds, the 3-task / 3-builder caps, the TDD and safety rules, and the internal sub-agent prompts.

v2.0.0 - Harness-agnostic Agent Skills

Muster is now distributed as plain Agent Skills, so it runs in Claude Code, Cursor, Codex, Gemini CLI, Mistral, pi, and any other harness that reads the spec - not just Claude Code.

Highlights

• commands/*.md → skills/<name>/SKILL.md as a single source of truth. The seven command bodies are unchanged; the only frontmatter addition is disable-model-invocation: true, keeping each command slash-only (the model never auto-fires /build or /triage). Other harnesses ignore the unknown field.
• install.sh now symlinks skills into ~/.claude/skills/.
• README + CONTRIBUTING rewritten harness-agnostic, with a "Runs in" badge row and an "Other harnesses" install section.

Upgrading (breaking)

• Plugin users: /plugin marketplace update muster then /plugin update muster@muster. Invocation is unchanged (/muster:think, …).
• Symlink users: the install target moved from ~/.claude/commands/ to ~/.claude/skills/. Re-run install.sh (it pulls, then relinks); you can remove the old dangling ~/.claude/commands/<muster>.md symlinks.

Notes

• The previous command-based line is preserved on the legacy/commands branch.

v1.4.0 - independent test re-run

Enforcement

• /build re-runs the test suite itself against each task branch and gates on its own result, rather than trusting the Worker's pasted test output (now only cross-checked for fabrication).
• CONTEXT.md requires a Test command: field. /build stops at startup without it; /triage won't mark a task agent-ready without it.

Wording

• "seam" -> "test point" and review "axes" -> "tracks" across all commands and docs.

v1.3.0 - unified command output

Consistent terminal output across all seven commands

Every command now shares one output style instead of each inventing its own:

• Rendered markdown, not code-fenced ASCII. /build and /status reports are now markdown tables; code fences are reserved for copyable content (commands, diffs, branches).
• One status-glyph set everywhere: ✅ pass/ready/approved · ❌ fail/blocked · ⚠️ warning/judgement/medium-risk · ⏭️ deferred/skipped/on-hold.
• A **Next:** footer on every command.
• House style documented in CONTRIBUTING.md so future commands stay consistent.

No behavioural change - output formatting only.

Full changelog: v1.2.0...v1.3.0

v1.1.1 - punctuation cleanup

Replace all em-dashes with hyphens repo-wide. No behaviour change.

Update: /plugin marketplace update muster then /plugin update muster@muster

v1.1.0 — file-based /spec & token-lean output

What's new

• File-based /spec. Drafts the spec and all task slices into one committed markdown file under docs/specs/<date>-<slug>.md. You edit it in your editor, then issues are generated from the file on create. Out-of-band editing keeps the review loop token-free, and the file is a durable source of truth.
• Token-lean /review. Sub-agents now return terse structured findings instead of ~400-word prose, and the verdict is a compact, verdict-first synthesis rather than two reports dumped verbatim.
• Cleaner /triage summary. Compact verdict-first markdown instead of a raw ASCII block.
• Text status markers only (PASS/FAIL, [x]/[!]/[ok]) — no emojis.
• Next-step pointers. Every command ends telling you which command comes next.

Updating

/plugin marketplace update muster
/plugin update muster@muster

v1.2.0 - richer /spec, clearer deferrals, tutorial

/spec output enrichment (from Matt Pocock's to-prd / to-issues)

• Focused User stories section ("As an X, I want Y, so that Z"), used by the inner-review Spec axis and /triage.
• Per-task AFK / HITL hint (advisory; /triage makes the binding call) and Covers user-story traceability.
• Active quiz at the review gate: the slice breakdown is shown and sanity-checked (granularity, split/merge, tags) before any issue is created.
• Prior-art pointer for test seams and the prototype-snippet exception.

Kept deliberately: the no-dependency design (no Blocked-by) and the Worker's Scope: touch only guardrail.

Clearer handling of non-independent slices

• /spec warns at cut time when two slices touch the same file.
• /triage gives an actionable "merge PR #N first" reason instead of a vague stall.
• /build defers same-file tasks deterministically rather than building against a stale main.

Docs

• README: rendered Mermaid pipeline diagram and a granular step-by-step "idea to merged PR" tutorial.

Full changelog: v1.1.1...v1.2.0