Dynamic workflows

The newest harness primitive, adopted day one. Dynamic workflows are a Claude Code research preview (requires v2.1.154+; on Pro, enable in /config → Dynamic workflows). If your build doesn't have them, the rest of compass is unaffected — these are additive and clearly gated.

A workflow is a small JavaScript script the Claude Code runtime executes in the background. Unlike a subagent or a skill — where Claude decides turn-by-turn what runs next and every result lands in the chat context — a workflow moves the plan into code: the loop, the branching, and the intermediate results live in script variables, so your context holds only the final answer, and it can fan out tens to hundreds of subagents (16 concurrent, 1,000 per run) that the session never has to hold in memory.

That shift is what lets a workflow apply a quality pattern, not just run more agents: independent agents can adversarially review each other's findings before anything is reported, or draft a plan from several angles and weigh them — so the result is more trustworthy, not merely bigger.

Why compass ships its own

The everyday review/audit/plan loops are exactly the cases worth running the same way every time, and they benefit most from parallelism + verification. compass's three workflows also do something a generic one can't: each routes its stages to compass's own cost-tiered subagents (agentType), so the security dimension runs on the opus security-auditor while the rest run on the sonnet code-reviewer — cost follows risk, the same principle as the rest of the repo.

Command	Pattern	The lever
/compass-review	parallel dimensions → adversarial verify → synthesize	Everyday. Reviews the branch diff on five dimensions (correctness, security, performance, tests, conventions) concurrently; a skeptic tries to refute each finding before it's reported; returns one Blocking/Should-fix/Nit verdict. Faster and less noisy than a single-pass review.
/compass-audit	multi-modal finders → loop-until-dry → 3-lens vote	Coverage. Six blind finders hunt different failure classes (authz, injection, error-handling, concurrency, resource leaks, secrets), repeated until two dry rounds; each fresh finding ships only if 2 of 3 perspective-diverse lenses confirm it. Scope with args.path.
/compass-plan	N angles → judge panel → grafted synthesis	Confidence. Drafts a plan MVP-first / risk-first / simplicity-first, scores each on one rubric, then synthesizes a single plan from the winner while grafting the best ideas of the runners-up. For ambiguous changes where the approach isn't obvious.

Run them

After make install (or ./quickstart.sh) the scripts are symlinked into ~/.claude/workflows/, so they're available as slash commands in every repo:

/compass-review
/compass-audit                                   # whole repo; pass a path to narrow
/compass-plan add a token-bucket rate limiter to the login endpoint

A run goes to the background; watch it with /workflows (arrow-keys to a phase, Enter to drill into an agent, x to stop, p to pause/resume, s to save your own). The subagents always run in acceptEdits and inherit your tool allowlist, so add any commands they need to permissions.allow first or a long run may stall on a prompt.

You don't have to install these to get workflows at all: put the word workflow in any prompt and Claude writes one for the task; /effort ultracode makes it do that automatically for every substantive task in the session. compass ships these three because they're the ones worth codifying — and because they're readable, you can audit exactly what fans out before you trust it.

Cost & control

A workflow spawns many agents, so one run uses meaningfully more tokens than the same task in conversation — it buys thoroughness, spend accordingly. Every agent uses your session model unless a stage routes elsewhere (compass's do, via agentType). Check /model before a large /compass-audit. Stop any run from /workflows without losing completed work.

Off-switch (any of): toggle Dynamic workflows off in /config; set "disableWorkflows": true in ~/.claude/settings.json; or CLAUDE_CODE_DISABLE_WORKFLOWS=1.

Write your own

Each file in claude/workflows/ is export const meta = { name, description, phases } (the name becomes the /command) followed by a body using agent(), parallel(), and pipeline(). Read compass-review.js first — it's the canonical shape:

• pipeline(items, …stages) is the default: each item flows through every stage independently, no barrier — item B's findings verify while item A is still being reviewed.
• parallel(thunks) is a barrier — use it only when a stage genuinely needs all prior results at once (e.g. dedup across every finding before verifying).
• agent(prompt, { schema, agentType, label, phase }) spawns one subagent; with a schema it returns a validated object. Route to a compass subagent with agentType.

Validate the shape (and JS syntax, when node is present) with scripts/check-workflows.sh — it runs in CI, so a malformed workflow fails the build like everything else here.

Honest limits

• Research preview. The feature, the save path (/workflows → s), and the exact runtime are still moving. compass pins the scripts and validates their shape; treat anything your build doesn't support as aspirational, per the repo's grounded-over-impressive rule.
• Plugin install doesn't carry workflows. They ship via the full install (symlinked ~/.claude/workflows/), not the core plugin — plugins bundle agents/commands/skills/hooks/MCP, not workflows (yet).
• The human still merges. A workflow reviews, audits, or plans; it doesn't push or deploy. Same boundary as the rest of compass.