squad

Every "tests pass" / "build green" / "feature works" claim needs the actual command output pasted into the conversation. Bare assertions are worth zero.

The operating loop every session follows — register, tick, pick, work, checkpoint, test, review, commit, close. Use whenever you start or resume work on a squad-managed repo.

Post often, post small, post honestly. The team needs visibility into your non-obvious state — not a change log. Use the typed verbs (thinking / milestone / stuck / fyi / ask) so peers can route attention.

Tick the squad team — surface mentions and file conflicts. Address mentions BEFORE continuing other work.

Before signing off — whether you finished an item or stopped mid-flight — post a 3-bullet summary. Shipped, in flight / queued, surprised by. Three bullets, no ceremony.

Run a structured intake interview with the user — open a session, ask one focused question per turn, draft the bundle, confirm with the user, then commit. Supports green-field (mode=new) and refine-mode (mode=refine) flows.

Evidence-gated close-out for an item. Requires test output paste, code review, and quality-bar pass before squad done runs.

Decompose a squad spec into 3-7 captured items for triage. Takes a spec name; reads the spec; drafts items via squad_new; lands them in the inbox linked to the parent spec.

Capture a quick work item from free text. Defaults to type-heuristic; asks if ambiguous. Lands in the inbox (status=captured) for later triage.

Mark an item blocked and inject the standard

File a new item. Runs squad new with type and title, then injects the standard item template body for you to fill in.

End-of-session 3-bullet brief plus claim release. Invokes the squad-handoff skill.

Atomically claim an item by ID. Verifies references, applies RED-first if AC names concrete failures.

Spawn superpowers:code-reviewer on an item with a self-contained briefing. Includes premise-validation latitude and working-tree hygiene clauses.

Onboard or resume in one step — squad go does init/register/claim/AC/mailbox.

Every item — even a one-line fix — gets reviewed by superpowers:code-reviewer before commit. No exceptions. The agent is cheap; production bugs are not.

Default to parallel when items are independent. Children run scoped tests; the parent runs the full suite as the integration gate. Sequencing independent work out of habit wastes wall-clock time.

Before fixing a claimed bug, prove the bug exists by writing a RED test against the current code. If the test passes unmodified, the bug does not reproduce — reclassify, do not "fix" a non-bug.

Tests passing is necessary but not sufficient. Before committing, run the self-review checklist — no commented-out code, no TODOs, no PM traces, no defensive checks for impossible cases, no half-finished implementations.

When acting as a reviewer agent, attempt to empirically disprove each finding before reporting it. Run the test, read the file, check the assertion. False-positive findings poison the team's trust in the ledger.

Exploratory work has a 2-hour cap. At 2h, stop and write up what you know — hypotheses tried, ruled-out causes, evidence collected, what is still unknown. Then escalate, parallelize, or block. Long unsuccessful sessions are a signal, not a setback.