Caro Coder Loop

---
description: Continuous v1.2.0 feature development — claims next ready beads task, implements via kraken (TDD), opens PR. Loop until no ready work remains.
---

# Caro Coder Loop (v1.2.0)

You are the **local coder agent** for the caro v1.2.0 milestone (Epic #792, due 2026-04-30). Your job: drain the beads queue by implementing one task per iteration, opening a PR per task, then looping.

## Preflight (run once at loop entry)

```bash
cd /Users/kobik-private/workspace/caro
bd stats | grep -E "Ready to Work|In Progress"
git fetch origin main

If Ready to Work == 0 on v1.2.0 tasks, exit gracefully with: "No ready v1.2.0 work remaining. Backlog-grooming routine will re-populate on next 6h tick."

Per-iteration workflow

1. Claim

# Get next ready v1.2.0 task (highest priority, Phase-1 first by dependency order)
# Skip beads claimed by named delivery agents (`claim_policy:manual_only`) — e.g.
# the caro-waitlist-engineer's path:deep beads filed by caro-merge-review-integrate.
# Skip beads in the rebase queue (`loop:rebase`) — those are drained by the
# rebase loop (follow-up; see policy.yaml + dispatcher SKILL for the contract).
NEXT=$(bd ready --json 2>/dev/null | jq -r '
  .[]
  | select(.labels[]? == "v1.2.0")
  | select((.labels // []) | index("claim_policy:manual_only") | not)
  | select((.labels // []) | index("loop:rebase") | not)
  | .id
' | head -1)
[ -z "$NEXT" ] && { echo "no ready v1.2.0 work"; exit 0; }
bd claim "$NEXT"
bd show "$NEXT"
bin/notify coder "claim $NEXT"

Capture:

• NEXT — beads ID (e.g. caro-xk0.1)
• GH_ISSUE — from external_ref field (e.g. gh-793)
• TITLE, DESCRIPTION, LABELS

2. Create isolated worktree

Follow .claude/rules/git-workflow.md — NEVER work on main.

SLUG=$(echo "$TITLE" | tr '[:upper:]' '[:lower:]' | tr -c 'a-z0-9' '-' | sed 's/--*/-/g' | cut -c1-40)
cd /Users/kobik-private/workspace/caro
bin/sk-new-feature "$SLUG"
cd .worktrees/*"$SLUG"*/

3. Gather context

Before delegating to kraken:

# Read the linked GitHub issue body
gh issue view "${GH_ISSUE#gh-}" > /tmp/issue-body.md
# Read the v1.2.0 tech spec if referenced
ls specs/v1.2-delivering-on-the-promise/ 2>/dev/null && cat specs/v1.2-delivering-on-the-promise/tech-spec.md

4. Pick the right specialist agent

Consult .github/STAKEHOLDERS.yml first. The map pairs codebase paths with the specialist agents that should own changes there. Pick the most specific glob match for the paths the task is likely to touch — derived from the issue body, beads labels (area:safety, area:ml, …), or the title.

# Heuristic: extract candidate paths from the issue body
PATHS=$(grep -oE '(src/[a-z_/]+|tests/[a-z_]+|website/[^ ]+|\.github/[^ ]+|Cargo\.[a-z]+)' /tmp/issue-body.md | sort -u)

# Look up agents per path; the first concrete glob match wins
yq -r --arg p "$PATHS" '
  .areas | to_entries[] | select(.key as $k | $p | test($k))
  | "\(.key) -> \(.value.agents | join(\",\"))"
' .github/STAKEHOLDERS.yml | head -3

The agent named on the most-specific match becomes the primary specialist for the task. Examples:

Path touched	Primary specialist
src/safety/patterns.rs	safety-pattern-developer (TDD discipline required)
src/safety/**	safety-pattern-developer, fall back to safety-pattern-auditor
src/backends/embedded_backend.rs	llm-integration-expert
src/ai/**, src/eval/**	ml-ds-engineer
src/cli/**, src/main.rs	rust-cli-expert
Cargo.toml, .github/workflows/release.yml	caro-release-expert
website/**	technical-writer
website/src/i18n/**	technical-writer + cultural-heritage-expert
specs/**	spec-driven-dev-guide

Fallback rules (when STAKEHOLDERS.yml has no match):

• labels contains documentation → use spark (Sonnet, fast)
• Otherwise → use kraken (Opus, TDD for Rust safety-critical code)

Critical: never bypass safety-pattern-developer for src/safety/** changes — that path requires the TDD-first workflow encoded in the skill.

5. Delegate to the specialist

Spawn via Task tool with this prompt template:

You are implementing beads task $NEXT (GitHub #${GH_ISSUE#gh-}) for caro v1.2.0.

**Working directory**: $(pwd) (a git worktree on branch $(git branch --show-current))

**Task**: $TITLE

**Issue body**: [contents of /tmp/issue-body.md]

**Acceptance criteria**: see the task description and the v1.2.0 tech spec at specs/v1.2-delivering-on-the-promise/tech-spec.md.

**Conventions** (from /Users/kobik-private/workspace/caro/CLAUDE.md):
- Rust edition 2021, MSRV 1.85
- Use `thiserror` for error types, `anyhow::Result` for application errors
- TDD: write failing test FIRST, then make it pass
- All commit messages conventional: `<type>(<scope>): <subject>` + Co-Authored-By
- Run `cargo test` + `cargo clippy` before declaring done
- For safety-related changes (src/safety/), add test cases to patterns.rs tests

**Apply the Good Boy Scout rule** (.claude/rules/good-boy-scout.md) — fix obvious
nearby issues but don't gold-plate.

When done:
1. Commit all changes with conventional message
2. Report: files changed, tests added, clippy status, test pass count

6. Reviewer gate (swarm-forge pattern)

Before validating or opening a PR, spawn an independent Reviewer sub-agent via the Task tool. Borrowed from swarm-forge's Architect → Coder → Reviewer pipeline: the coder cannot self-approve.

Subagent: qa-testing-expert (model: opus)

Reviewer prompt template:

You are the REVIEWER gate for caro v1.2.0 task $NEXT (PR not yet opened).

Working directory: $(pwd)
Branch: $(git branch --show-current)
Diff to review: `git diff main...HEAD`

You MUST run, and your verdict MUST cite the actual output of:
  1. `cargo fmt --all -- --check`
  2. `cargo clippy --all-targets --all-features -- -D warnings`
  3. `cargo test`
  4. Constitution check: any change under `src