Call 1: Read <agent>.md           (skip if already in context)
Call 2: bash dispatch-job.sh ... && node pack-status.mjs --move N in-progress

SANDBOX RULES (GKE Autopilot):
- REPO_ROOT is /workspace/repo — hardcoded, no discovery needed.
- Each Bash tool call = fresh shell. ALWAYS prefix: cd /workspace/repo && <command>
- Set timeout: 600000 (10 min) on long-running commands (playwright, next build).
- The entrypoint already handled: git clone, branch checkout, pnpm install, git identity.

**Step 1 — Verify environment (quick sanity check):**
cd /workspace/repo && git branch --show-current && node -v && ls package.json 2>/dev/null || ls development/ledger/package.json
If Playwright tests are needed:
  cd /workspace/repo && npx playwright install chromium 2>/dev/null || true
  ln -sf /workspace/repo/development/ledger/node_modules /workspace/repo/quality/node_modules 2>/dev/null || true

**Step 1b — Check for prior work on this branch (UNBREAKABLE):**
cd /workspace/repo && git fetch origin && git log origin/main..HEAD --oneline
If commits exist beyond main:
  - A previous agent session already did work on this branch.
  - Read ALL changed files: cd /workspace/repo && git diff origin/main --name-only
  - Read each changed file before writing a single line of new code.
  - Understand what was already implemented, then continue from where it left off.
  - DO NOT rewrite or redo work that is already committed. Pick up from the current state.
If no commits beyond main: branch is fresh, proceed with full implementation.

TODO TRACKING (UNBREAKABLE):
Use TodoWrite to plan and track ALL work. Create todos at the START of the session
covering every numbered step. Update each todo to in_progress when you start it and
completed when done. This is your progress ledger — if the session dies, the todo
state shows exactly where you stopped.

INCREMENTAL COMMIT + VERIFY LOOP (UNBREAKABLE):
After every logical chunk of implementation work (~5-10 min or 1-3 files changed):
  1. cd /workspace/repo && git add -A && git commit -m 'wip: <what> — issue:<NUMBER>' && git push origin <BRANCH>
  2. cd /workspace/repo/development/ledger && pnpm run verify:tsc
  3. If tsc fails: fix immediately, commit+push, re-run tsc.
  4. Update your todo progress.
Do NOT batch all changes into one commit at the end. Sessions can die at any time —
uncommitted work is lost work.

VERIFY — tsc + build + Vitest (UNBREAKABLE):
All agents run tsc, build, AND the full Vitest suite before handoff.
Do NOT run Playwright E2E tests — Vitest only. E2E runs via CI.
  cd /workspace/repo/development/ledger && pnpm run verify:tsc
  cd /workspace/repo/development/ledger && pnpm run verify:build
  cd /workspace/repo/development/ledger && npx vitest run
**Trust the exit code.** Exit 0 = all tests pass. Non-zero = failures.
Do NOT grep vitest output for FAIL, do NOT parse ANSI escape codes.
Just run the command and check whether it succeeded or failed.
On failure: read the output to see which tests failed, fix them, commit+push, re-run.
Repeat until the command exits 0.
Do NOT proceed to handoff with ANY failing Vitest tests.
**All verify steps MUST run in the foreground (blocking).** NEVER use `run_in_background`
for tsc, build, or Vitest commands. You MUST confirm each step passes before proceeding
to the next step or to merge/handoff. Background verify = unverified merge = bug.

NO MONITOR-UI / ODINS-SPEAR TESTS (UNBREAKABLE):
NEVER write tests for `development/odins-throne-ui/` (Odin's Throne) or `development/odins-spear/`.
These packages have no test infrastructure that agents should use. All tests target
`development/ledger/` only. For odins-throne-ui or odins-spear issues, validate via tsc + build only.

STRICT SCOPE (UNBREAKABLE):
Execute ONLY your numbered steps — nothing more. Do NOT close issues, merge PRs,
declare things "done", or add commentary beyond your handoff step.
If something is ambiguous, comment on the issue asking for clarification — don't guess.

CHAIN CONTINUATION (UNBREAKABLE):
Every agent template includes a final chain continuation step. It is MANDATORY when
conditions are met (success verdict) and FORBIDDEN when they are not (failure/partial).

- **Documentation issues (label: `documentation`):** NO chain continuation. The agent
  merges its own PR and closes the issue directly:
  `gh pr merge <PR_NUMBER> --squash --delete-branch && gh issue close <NUMBER>`
  Do NOT run `/fire-next-up --resume`. There is no Loki QA step for doc-sync work.
- Non-final agents (Luna, FiremanDecko, Heimdall): run `/fire-next-up --resume #<NUMBER>`
  as the final step ONLY when all verify steps pass and the handoff comment is posted.
  This dispatches the next agent in the chain on the same branch.
- Final agent (Loki): merge and close ONLY on a clean PASS verdict:
  `gh pr merge <PR_NUMBER> --squash --delete-branch && gh issue close <NUMBER>`
- If verdict is FAIL or any verify step failed: do NOT run chain continuation.
  Stop immediately and leave the issue open for manual triage.
- Chain continuation is the ONLY authorized merge path — do not merge any other way.
- **GKE sandbox: NEVER fall back to local worktree.** If `/dispatch` fails inside a GKE job
  (e.g. dispatch-job.sh path error, kubectl error), post a comment on the issue with the
  exact error output, then stop. Do NOT spawn an Agent tool with `isolation: "worktree"`.
  The orchestrator (Odin) will re-dispatch from outside the sandbox.

  Luna's wireframes are on this branch. Read EVERY file listed below BEFORE writing any code:
  - `ux/wireframes/chrome/dashboard-tab-headers.html`
  - `ux/wireframes/chrome/dashboard-tab-headers-interaction-spec.md`
  These are the exact wireframe files Luna produced. ...

---
## Additional Context (from orchestrator)
<prompt-extra content>

REPO_ROOT=$(git rev-parse --show-toplevel) && \
SESSION_ID="issue-<N>-step<S>-<agent>-$(uuidgen | cut -c1-8 | tr A-Z a-z)" && \
PROMPT_FILE=$(mktemp /tmp/agent-prompt-XXXXXX.md) && \
cat > "$PROMPT_FILE" <<'AGENT_PROMPT_EOF'
<composed prompt content here>
AGENT_PROMPT_EOF
bash "$REPO_ROOT/infrastructure/k8s/agents/dispatch-job.sh" \
  --session-id "$SESSION_ID" \
  --branch "<BRANCH>" \
  --model "<MODEL>" \
  --prompt-file "$PROMPT_FILE" && \
node "$REPO_ROOT/.claude/skills/fire-next-up/scripts/pack-status.mjs" --move <N> in-progress
rm -f "$PROMPT_FILE"

node "<repo-root>/.claude/skills/fire-next-up/scripts/pack-status.mjs" --move <N> in-progress

**Dispatched #<N>**: <TITLE>
**Agent:** <AgentName> (Step <S>) | **Model:** <MODEL>
**Branch:** `<BRANCH>`
**Session:** `<SESSION_ID>`
**Auto-continue:** enabled — agent will run `--resume #<N>` on completion

**GKE:**
  Cluster: `<kubectl config current-context>`
  Namespace: `fenrir-agents`
  Job: `agent-<SESSION_ID>`
  Image: `<image from dispatch-job.sh output>`

**Links:**
  Logs: `just infra agent-log <SESSION_ID>`
  Logs (verbose): `just infra agent-log-verbose <SESSION_ID>`
  Brandify: `/brandify-agent <SESSION_ID>`
  Pod status: `kubectl get pods -n fenrir-agents -l job-name=agent-<SESSION_ID>`