v1.1.0

What's new

• Direct cursor-agent dispatch — no runner middleware needed. Scouts invoke cursor-agent --model composer-2.5-fast --mode ask directly. Install drops from 3 pieces to 2.
• Composer 2.5 Fast model pin — model: composer-2.5-fast in all scout frontmatter; --model enforced at CLI level.
• Severity calibration — observed (cited path:line) vs inferred ("implies...") vs hypothesis (needs evidence). Scouts can't say "broken" without citing the behavior.
• Task restatement — every scout echoes a Task: line restating what it understood. Drift = bad brief, caught early.
• Convergence framing — weight findings by independent rediscovery across scouts, not raw count.
• Hypotheses bucket — speculation separated from findings in a dedicated output section.
• PR comment filter — only cited + convergent + actionable findings earn a PR comment.
• Natural adoption triggers — skill description matches task shape ("find every file that touches X"), not product name.
• Compact descriptions — skill-cleaner principles applied to all frontmatter descriptions.
• Adjacent surprises tightened — must be plausibly tied to the task; unrelated oddities filtered out.

Composer Swarm v0.7.0

Three-scout delegation pack. Composer Swarm collapses into a small set of read-only-by-default Cursor Composer scouts the main agent fans out marginal-value work to.

The three scouts

• composer-wide-search — coverage-disciplined map of a subsystem
• composer-deep-search — depth-disciplined trace of one behavior end-to-end
• composer-runner — execution of one named command, structured summary returned

Each scout takes a budget knob (quick / thorough / exhaustive) and ends its output with an Adjacent surprises footer that flags 1-3 things the main agent did not explicitly ask about — each tied to path:line. This is the scout's blindspot-hunting move: it sees the surroundings the main agent doesn't.

What changed vs v0.5.0

• 6 reviewer agents and 5 playbooks removed
• 17 fewer tracked files net (down to 21)
• Sync infrastructure collapsed — plugins/composer-swarm/agents/ is the single source of truth, no more .agents/composer-*.md duplicates, no more scripts/sync-bundle.mjs
• Tests strengthened: bundle no-extras check no longer regex-filtered (catches any junk file), legacy-ban scan includes CLAUDE.md + CONTRIBUTING.md, cross-host plugin.json defaultPrompt parity asserted

Install

Claude Code:

```text
/plugin marketplace add jxucoder/composer-swarm
/plugin install composer-swarm@jxucoder-composer-swarm
```

Codex:

```text
codex plugin marketplace add jxucoder/composer-swarm
```

Requires Cursor CLI and the shinpr/sub-agents-skills Runner plugin. See README for the full flow.

Docs

• Design rationale
• Main-agent dispatch protocol (SKILL.md)
• Edit workflow & release checklist (CONTRIBUTING.md)

Composer Swarm v0.2.0

Highlights

• Treat read-only review/research as scout signal, not reviewer-of-record output.
• Cleaner result output for review and research tasks, with final worker reports preferred over noisy stream traces.
• Dirty and untracked worktree snapshots for read-only review/research.
• inspect and logs commands for local task state and worker transcripts.
• setup --init no longer guesses a verifier; the host agent/user configures workers.verifier only when the correct repo check is known.
• Composer workers remain pinned to composer-2.5-fast; Codex/Claude stay in control of verification and apply.

Verification

• CI runs syntax checks, tests, CLI smoke, and package dry-run on Node 20 and Node 22.

Composer Swarm v0.1.0

Composer Swarm v0.1.0

Initial repo-only release for running local Cursor/Composer worker teams from Claude Code, Codex, or a shell.

Highlights

• Local Node CLI for planner, builder, reviewer, verifier, apply, cleanup, and status/result workflows.
• Isolated git worktrees for Composer workers.
• Claude Code local plugin commands under /composer:*.
• Codex plugin metadata and bundled composer-swarm skill.
• Cursor workers pinned to composer-2.5-fast.
• Manual apply only, with candidate comparison and verification support.
• MIT licensed.

Verification

• npm test: 29 pass, 0 fail.
• npm pack --dry-run --json: clean 31-file package preview, excluding local state and reference checkouts.
• CLI syntax and smoke checks passed.