Skip to content

daily-triage.md

Latest commit

 

History

History
142 lines (104 loc) · 5.89 KB

daily-triage.md

File metadata and controls

142 lines (104 loc) · 5.89 KB
Raw

Daily Triage — Opencode

Same pattern as Grok and Claude Code; scheduling runs from cron/systemd and each tick invokes opencode run instead of a TUI /loop.

Prerequisites

  1. Opencode CLI installed and authenticated against your preferred provider.
  2. STATE.md at the repo root — copy from starters/minimal-loop-opencode/STATE.md.example.
  3. loop-triage skill — copy templates/SKILL.md.loop-triage to skills/loop-triage/SKILL.md.
mkdir -p skills/loop-triage
cp templates/SKILL.md.loop-triage skills/loop-triage/SKILL.md
cp starters/minimal-loop-opencode/STATE.md.example STATE.md

Report-Only (Week 1)

Use cron or a systemd timer to run a fresh opencode session each morning. The prompt forces the run to read STATE.md first so state carries across sessions.

opencode run \
  "Run the loop-triage skill. Read STATE.md first. Append high-priority items under High Priority and Watch List. Update Last run timestamp. Do not edit source code. End with a 5-line summary." \
  --title "Daily triage — repo:${PWD##*/}"

The report-only stance is enforced by AGENTS.md and the prompt: no source-code edits in week one.

Faster cadence during active periods:

0 */2 * * * cd /repo && opencode run "Run loop-triage. Report obvious small wins only. Update STATE.md. No code changes."

With Small Auto-Fixes (Week 3+)

Create named implementer and verifier agents in opencode.json, isolate each implementer in a git worktree, and pass the worktree path with --dir:

FIX_ID="$(date +%Y%m%d%H%M%S)"
WORKTREE="../wt-small-fix-$FIX_ID"
git worktree add "$WORKTREE" -b "loop/small-fix-$FIX_ID"
opencode run \
  "Run loop-triage. For one high-priority single-file bugfix: implement the minimal fix, run tests, and write a summary plus diff path. Escalate ambiguous or denylisted paths." \
  --agent implementer \
  --dir "$WORKTREE"
DIFF_FILE="$(mktemp /tmp/loop-diff.XXXXXX.patch)"
git -C "$WORKTREE" diff > "$DIFF_FILE"
opencode run "Review this diff against project rules and tests. APPROVE or REJECT only." \
  --agent verifier \
  --file "$DIFF_FILE"

The verifier sees only the diff; the implementer works only inside the worktree. This preserves the same maker/checker split Claude Code expresses with isolation: worktree.

Goal Mode Alternative

For a one-shot "get main green" session, drive opencode with the Goal Engineering /goal skill. opencode is goal-friendly: hand it the goal, the stop condition, and the verifier prompt and let the verifier decide when to stop.

opencode run "Goal: all tests on main pass and lint is clean. Stop when tests pass and write the evidence."
DIFF_FILE="$(mktemp /tmp/goal-diff.XXXXXX.patch)"
git diff > "$DIFF_FILE"
opencode run "Verify the goal is complete. APPROVE only if tests pass and the diff is minimal." \
  --agent verifier \
  --file "$DIFF_FILE"

Skills Setup

Copy from templates/SKILL.md.loop-triage to skills/loop-triage/SKILL.md (the skills/ dir is auto-discovered by opencode at the repo root):

mkdir -p skills/loop-triage
cp templates/SKILL.md.loop-triage skills/loop-triage/SKILL.md

Create named verifier/implementer agents when you graduate to L2. The starter's opencode.json.example already includes both names; copy it to opencode.json or add this shape manually:

{
  "agent": {
    "implementer": {
      "description": "Implement minimal scoped fixes in an isolated worktree.",
      "mode": "subagent",
      "prompt": "Implement only the requested minimal fix. Stay within the supplied worktree and run documented tests."
    },
    "verifier": {
      "description": "Review diffs and test evidence; APPROVE or REJECT only.",
      "mode": "subagent",
      "prompt": "Review the supplied diff against project rules and tests. Do not edit files. APPROVE or REJECT only."
    }
  }
}

Sub-agents

Opencode exposes named agents through opencode agent. Keep each agent's role text aligned with the same maker/checker contract used by the other examples:

Verifier: review code changes against project skills and tests. Always REJECT or APPROVE.

Invoke one role per run with --agent verifier or --agent implementer (see Week 3+ example). Use explicit git worktrees for any implementer that will edit code.

State Conventions

The default spine is STATE.md at the repo root. Use STATE.md.example for the template (gitignored live state). For pattern-specific state, follow the state conventions in primitives-matrix.md.

Safety (L1 defaults)

  • Week one: report-only prompt + AGENTS.md safety rules; no source-code edits.
  • One worktree per fix attempt; discard after verifier REJECT.
  • Never push without a draft PR and explicit human review on denylisted paths (docs/safety.md).
  • Token budget: set loop-pause-all in STATE.md if a run exceeds the daily cap.

Operations

crontab -l
systemctl --user list-timers
opencode session list
opencode export <sessionID> > loop-session.json

Pause: disable the cron/systemd timer or set loop-pause-all in STATE.md and teach the skill to stop acting.

Audit readiness: npx @cobusgreyling/loop-audit . --suggest.

When to pick opencode over a TUI agent

  • You want scheduling without a TUI — cron/systemd can call opencode run in CI, in tmux, and on a headless box.
  • You want to keep your state and skills in plain files (STATE.md, skills/) that survive a host change.
  • You're already running systemd timers or cron for unrelated work — fold opencode loops into the same scheduler.

References