crew v0.1.1 — skill-aware agents

crew agents are now skill-aware. Each agent that can act on one carries the Skill tool and invokes the skills its role turns on:

• coder, quick — run the ponytail:ponytail laziness ladder before editing; pull in a design skill (make-interfaces-feel-better, emil-design-eng, apple-design, animate, dataviz) on UI, motion or chart work.
• planner — holds the plan to the ponytail ladder and the same design skills.
• architect — reaches for dataviz and the design skills when a cross-subsystem plan lands in the UI.
• reviewer — runs code-review, security-review and break-ui.
• scout — left alone; a read-only locator has no skill to lean on.

governor 0.5.0

A PreToolUse hook that watches what a Claude Code session is spending and stops it running away. Zero always-on cost for the gate itself — hooks run outside the model's context.

Gates

gate	default	effect
burn rate	25k output tokens in 5 min	warn
soft	60% of budget	warn
hard	100% of budget (200k)	deny every tool call

The model keeps its voice when denied, so it reports what is done and what is left rather than going silent.

The budget, and the ceiling

A denied session may still edit the project config — otherwise the gate strands it with no way to ask for more. That doubles as the self-extension: it can raise budget up to ceiling. Past the ceiling nothing it can reach moves the line; enabled, ceiling and both ratios live in ~/.claude/governor.json.

New in 0.5.0 — weighted

Off by default. The gate counted output tokens because they drive cost — true for a session that writes, false for one that reads. Measured on a real agent turn: 168 output tokens against 322,512 cache reads, costing $0.3562, of which output was $0.0042 — 1.2%. Across four tool-heavy runs, output was 5% of the bill.

"weighted": true counts each class at its price relative to output (input ×0.2, cache write ×0.25, cache read ×0.02), so the budget tracks dollars while keeping its unit. Weighted totals run far higher on the same work — raise budget in the same edit.

0.4.0 — security

Three ways out of the gate, closed. The escape hatch let a denied session write the project config, and the project config could set any field — so {"enabled": false} or {"hardRatio": 100} turned the gate off from inside and the ceiling was decoration. The project file may now set budget, burnTokens, burnMinutes and rewarnMinutes only.

Also: the state file was trusted on the way in and a non-array window threw on every tool call for the rest of the session; state moved out of world-writable /tmp to ~/.claude/governor-state; and find-plugin now allowlists repo slugs before printing them inside a paste-able command and strips control characters that \s did not cover.

Measured

See BENCHMARK.md — the gate replayed over 395 real session transcripts: 6.1% would have been hard-stopped, $40.34 of spend sat past the line, and the published median/p90/p99 reproduce from raw data.

Install

claude plugin marketplace add tanvoid0/claude-code-plugins
claude plugin install governor@tanvoid0

crew 0.1.0

Six subagents, each pinned to the cheapest model that holds for its job.

agent	model	when
quick	haiku	mechanical, target already named
scout	sonnet	where does X live, what would this touch — read-only
coder	sonnet	implementation against a settled plan
planner	opus	single-subsystem plan, bug-fix strategy
reviewer	opus	review a diff with real logic in it
architect	fable	spans subsystems, migrations, wire contracts

The routing table is the boring part. What makes it work is the refusals: an agent handed work below its tier names the right agent and hands back instead of doing it anyway.

Measured

From BENCHMARK.md — four identical tasks run with and without routing:

• 68.3% cheaper ($1.2792 → $0.4049), every task correct in both arms, verified by executing the code rather than trusting the agents' reports.
• A typo fix cost 33× more at the default tier than routed.
• Blind dispatch picked the right agent 18/18 from the descriptions alone.

Honest caveats are in the doc: the cheap tier reported three fewer real defects on the review task, one of two hand-back probes did not fire, and on same-model work roughly half the saving is prompt discipline rather than routing.

Install

claude plugin marketplace add tanvoid0/claude-code-plugins
claude plugin install crew@tanvoid0