iteration-complete

<!-- allowed-tools intentionally omitted — inherits Bash() from .claude/settings.json. Subcommand-level restriction silently blocked fleet agents (flag #62/#63; devex dispatch #171). This skill composes dispatch, diff-hash, receipt-sign, git-safe, git-safe-commit, and Skill invocations — too broad to enumerate at subcommand level without regression risk. Inherit Bash(). -->

iteration-complete

Run this after completing each iteration of work. Invokes /quality-gate for the review+fix cycle, auto-commits on green (no principal approval at iteration boundaries), and emits a structured dispatch to captain so the small-batch-cadence daemon can pick up for PR landing.

Why this exists

Iteration boundaries are the smallest formal commit-and-review unit in the-agency's valueflow work stream. Without a dedicated skill, agents skip the quality gate, forget the receipt-sign step, or let the commit drift from the plan file. iteration-complete bundles all five responsibilities (QG run → QGR receipt → auto-commit → plan update → handoff update + dispatch to captain) so iteration-end is single-invocation and auditable.

Auto-commit (no principal approval) is specifically for iteration boundaries — they're small, bounded units where the QG + receipt is sufficient attestation. Phase and plan boundaries use sister skills with principal approval gates.

Required reading

Before proceeding, Read the files listed in required_reading: frontmatter. REFERENCE-QUALITY-GATE.md is the protocol the invoked /quality-gate skill runs. REFERENCE-RECEIPT-INFRASTRUCTURE.md explains the five-hash chain that signs the iteration's QGR receipt.

Usage

/iteration-complete <phase.iter>: <description> [--base <ref>]

Examples:

• /iteration-complete 1.2: parser edge cases
• /iteration-complete 1.3: error handling --base abc1234

• <phase.iter>: numeric identifier (phase.iteration), e.g., 1.2
• <description>: one-line summary
• --base <ref> (optional): override baseline for QG diff-hash; resolved automatically if omitted

If invoked with empty arguments, ask what was completed before proceeding.

Preconditions

• Changed files present (git status non-empty).
• On a worktree branch (not master — this skill is worktree-only).
• Prior iteration's commit identifiable (or explicit --base passed). First iteration in a phase uses the phase-start commit.
• A plan file should exist under docs/plans/ or agency/workstreams/*/. If absent, skill proceeds with a "no plan file" note in the commit; handoff captures the omission.

Flow / Steps

Step 1: Preflight

1. If args empty, ask what was completed.
2. git status + git diff --stat HEAD. If empty, report "Nothing to gate" and stop.
3. Identify plan file.

Step 2: Determine prior-iteration base ref

QG's Hash A / Hash E diff is computed against the prior iteration commit (or phase-start commit for first iteration). Resolution order:

1. Read plan file for iteration history / status table; use the last iteration's commit SHA.
2. Grep git log for the prior-iteration commit: git log --oneline --grep="Phase <P>\\.". If this is X.1, use phase-start tag/commit.
3. Fallback: HEAD~1. Note fallback usage in handoff.

Capture as $BASE_REF.

Step 3: Run the quality gate

/quality-gate iteration-complete <phase.iter>: <description> --base <BASE_REF>

Leading iteration-complete <phase.iter> tells QG the boundary type (used in receipt filename). --base tells QG the baseline for Hash A/E via diff-hash --base.

Full QG protocol runs: parallel review → consolidate → bug-exposing tests → fix → coverage tests → confirm clean → present QGR → sign receipt (five-hash chain at agency/workstreams/{W}/qgr/). Iteration-complete is auto-approved — Hash D = Hash C.

Wait for receipt signed before proceeding.

Step 4: Commit (auto)

No approval. Commit automatically after clean QGR.

Use /git-safe-commit with the full structured commit message from the QGR's "Proposed Commit" section.

Step 5: Update plan file

Append to the plan under "Quality Gate Reports":

• What was done
• What QG found + fixed
• Plan changes (scope, reordering, new findings)
• Iteration status-table row

Append the full QGR (all tables + summary). The plan is the living record.

Step 6: Update handoff

Update the worktree handoff file (usr/*/*/handoff.md or usr/*/captain/handoff.md):

• Current phase + iteration status
• What just committed (summary, not full QGR)
• What's next
• Context for fresh-session continuation

Step 7: Emit iteration-complete dispatch to captain

Structured dispatch for captain's auto-ship daemon. Must run AFTER the commit (Step 4) so commit_hash is current.

Capture:

• ITERATION_SLUG (e.g., "1.2")
• PHASE_NUM
• BRANCH = git branch --show-current
• COMMIT_HASH = the Step 4 commit
• SUMMARY = first line of commit message
• RECEIPT_PATH = QGR receipt path from Step 3

Emit:

bash $CLAUDE_PROJECT_DIR/agency/tools/dispatch create \
  --to {repo}/{principal}/captain \
  --type iteration-complete \
  --subject "Iteration {ITERATION_SLUG} complete on {BRANCH}" \
  --body "<yaml body — see below>"

Body:

event: iteration-complete
iteration: {ITERATION_SLUG}
phase: {PHASE_NUM}
branch: {BRANCH}
commit_hash: {COMMIT_HASH}
summary: {SUMMARY}
qgr_receipt: {RECEIPT_PATH}
emitted_at: {ISO-8601 timestamp}

Cascade isolation: export AGENCY_SKILL_BYPASS_CASCADE=1 at skill start and unset at end if running multiple commits within the skill.

Failure modes

• No changes to gate (Step 1): skill stops cleanly with "Nothing to gate" message. Not an error.
• QG fails (Step 3): fix-and-retry per QG protocol — the gate itself handles findings, no special iteration-complete handling needed.
• Commit fails (Step 4): usually a pre-commit hook (lint-staged, oxfmt). Fix the blocking issue; re-run from Step 4 (QG already passed; no need to re-run Step 3).
• Plan file missing (Step 5): skill notes "no plan file" and continues. Handoff records the gap.
• Dispatch emission fails (Step 7): skill warns but exits 0 — the commit itself is the authoritative record; dispatch is notification. Captain catches the omission via other mechanisms (handoff read, manual check).

What this does NOT do

• Does not land on master — iteration-complete is a worktree-side boundary. Phase boundary + pr-submit + captain's pr-captain-land land on master.
• Does not require principal approval — iteration boundaries are auto-approved; the QG + receipt is sufficient attestation.
• Does not run deep QG — the QG here is iteration-scoped (diff against prior iteration). Phase and plan boundaries run deeper variants.
• Does not squash prior iteration commits — those stay individually visible in history; squash happens (if at all) at phase boundary.

Status

active (v2, body retrofit from Arguments/Steps pattern to 9-section structure 2026-04-19).

Related

• /quality-gate — the core QG protocol this skill invokes in Step 3
• /phase-complete — sibling skill for phase boundaries (deep QG + principal approval)
• /plan-complete — sibling skill for plan delivery (final deep QG + A&D)
• /pr-prep — pre-PR gating for accumulated work
• /pr-submit — agent's hand-off to captain for PR landing (runs AFTER iteration-complete when ready to ship)
• agency/tools/receipt-sign — signs the five-hash receipt
• agency/tools/dispatch — emits the Step 7 notification
• agency/workstreams/captain/small-batch-cadence-*.md — design for the auto-ship daemon that consumes Step 7's dispatch
• the-agency#298 — skill refactor recommendation
• the-agency#315 — V1→V2 migration

OFFENDERS WILL BE FED TO THE — CUTE — ATTACK KITTENS!