workbuddy-guide

name

workbuddy-guide

description

Explain and operate workbuddy: what it is, which command to reach for, and where to find deeper references. **Trigger whenever the user mentions 'workbuddy' in any context** — including operational requests (register/onboard a repo, add/bind a worker, 注册仓库, 加仓库, 接入新仓库, onboard repo, set up worker), explanatory questions (how to use workbuddy, workbuddy guide, teach me workbuddy, 怎么用workbuddy, 使用指南), deployment/running/operating/debugging workbuddy, and questions about how issues/tasks are scheduled or how dependencies between issues work (depends_on, blocked, dependency gate, 依赖, 调度, 排队, '为什么没派发'), and questions about parallel-development workflows (rollouts:N fan-out, synthesize/synthesis reduce step, cherry-pick, '并行开发', '多个方案', 'rollout'). Invoke this skill before running workbuddy CLI commands so the documented onboarding flow (coordinator register + worker repos add) and the dependency/scheduling model aren't re-derived from scratch.

Workbuddy Guide

A map of what workbuddy is and which command solves which problem. This skill is intentionally terse — for command details always run workbuddy <cmd> --help. The CLI help is the authoritative reference and stays in sync with the code; duplicating it here would only drift.

Bundled references (read when relevant)

• references/new-repo-onboarding.md — step-by-step checklist for adding a new repo to a workbuddy deployment. Read when setting up a repo.
• references/known-pitfalls.md — real-world failure modes and fixes. Read when something misbehaves.
• references/writing-good-issues.md — how to write issues that agents can actually process. Read when issues keep failing or getting blocked.

What workbuddy is

GitHub Issue-driven agent orchestration. Humans file issues with ## Acceptance Criteria and the workbuddy+status:developing labels; workbuddy polls GitHub, dispatches agents (Claude Code or Codex), and reacts to label changes the agents make as they hand off work.

Only two agent roles exist: dev-agent and review-agent. Runtime (claude-code | codex) is a config field, not a separate agent.

developing ⇄ reviewing → done
   ↕
 blocked  (missing Acceptance Criteria — human fixes, flips back)

Dev writes code + PR → flips to reviewing. Review evaluates each AC → done on pass, developing on fail (max 3 retries via REQ-055 cap).

Deployment modes — pick one

Mode	Command	When to use
Bundle (default since v0.5: supervisor + coordinator + worker)	workbuddy supervisor + workbuddy coordinator + workbuddy worker	Production, multi-host, anywhere agent runs must survive worker restarts
Serve (single process, legacy)	workbuddy serve	Local dev, single-host quick-start; does not preserve agents across restart
Managed install (systemd)	workbuddy deploy install	Long-lived production; defaults to the bundle layout; deploy redeploy/deploy upgrade/deploy rollback/deploy watch for updates

All three share the same state machine, SQLite store, and agent configs. Run workbuddy <mode> --help for flags, workbuddy deploy install --help for the systemd wrapper examples. See the /deploy skill for the bundle topology rationale and the cutover from serve.

Command map — what to reach for

Run workbuddy --help for the full list. Grouped by intent:

Setup a repo

• workbuddy init — scaffold .github/workbuddy/ in a new repo
• workbuddy validate — sanity-check config before running
• workbuddy repo register — attach a repo to a running coordinator (no restart)

Run workbuddy

• workbuddy supervisor + workbuddy coordinator + workbuddy worker — bundle mode (default)
• workbuddy serve — single-process legacy mode (local dev only)
• workbuddy deploy install|redeploy|upgrade|rollback|watch — managed systemd install

Observe

• workbuddy status — issues, tasks, events, stuck issues, or --watch until done
• workbuddy status --events --since 10m — audit trail with real RFC3339 timestamps (v0.6.24+ fix); useful event types include the recovery-loop signals periodic_recovery_tick, task_reaped, dispatch_skipped_inflight (with source discriminator: statemachine / router / resync), plus the original state_entry, dispatch, completed, dispatch_blocked_by_dependency
• workbuddy logs <issue> — per-attempt session logs (stdout/stderr/tool calls)
• workbuddy diagnose — surfaces stuck issues, 3-retry caps, stale claims, and (v0.6.24+) worker_tunnel_down findings per registered repo; --fix for safe auto-remediation
• workbuddy hooks list|status|test|reload — operator-owned event hooks (Slack/Feishu/etc.)

Recover

• workbuddy cache invalidate — force re-poll after manual label edits (cache-invalidate remains as a deprecated alias)
• workbuddy issue restart — clear poller cache + claim lease for one issue (was admin restart-issue, now deprecated)
• workbuddy recover — clean stale processes/worktrees/claims after a crash (--kill-zombies --prune-worktrees --prune-remote-branches --reset-db)
• workbuddy operator-watch — tail the operator incident inbox and auto-dispatch Claude per incident

Worker runtime ops

• workbuddy worker repos add|list|remove — change a running worker's repo bindings without restart
• workbuddy coordinator token create|list|revoke — per-worker auth tokens

Retry and failure semantics (worth knowing)

The state machine has several guardrails that change how failures look — read the relevant decision docs for depth; the short version:

• 3-retry cap (REQ-055): after 3 consecutive agent failures on the same issue, dispatch stops. workbuddy diagnose surfaces this explicitly.
• Infra vs verdict (REQ-056): launcher/runtime crashes are reported with an "Infra Error" comment header and do not burn the retry budget. Only agent-decided FAIL verdicts count toward the cap.
• Issue-claim with TTL (REQ-057, hardened in REQ-059): a SQLite claim prevents two workers from dispatching the same issue concurrently. Stale claims self-heal after TTL expiry.
• Worktree isolation (REQ-058): every task runs in its own .workbuddy/worktrees/issue-N/. If worktree setup fails, the worker reports loudly instead of falling back to CWD.
• Snapshot resync (REQ-150, v0.6.24): every 5 minutes the poller re-emits EventIssueResynced for every cached open issue carrying a workflow trigger label. The state machine consults IsInflight + HasAnyActiveTask; if both guards pass, the issue is re-dispatched. Closes the post-restart silent-stall pattern.
• task_queue reaper (REQ-151, v0.6.24): every 60 seconds the coordinator flips status='running' rows with stale heartbeat (default 5 min grace) to failed/exit_code=-2, emitting task_reaped. Unblocks HasAnyActiveTask after worker host death.
• Periodic recovery sweep (REQ-152, v0.6.24): every 60 seconds PollerManager.recoverAllPeriodically walks every active repo, re-enters recoverOrphanedActiveStates, and emits a periodic_recovery_tick event with {repos_swept, issues_redispatched}.
• Fault-injection hooks (REQ-148/153, build tag faultinject): six named failpoints (poller.list_issues.before/after, wstunnel.send.drop, worker.claim_task.before, worker.agent_exec.die_mid, store.insert_event.busy) for reproducing silent-stall scenarios. Zero-cost in production builds.

Impact for operators: if diagnose says "failed 3 times in a row", check the issue's comments — if they're "Infra Error", the bug is infrastructure (usually runtime/launcher), not the acceptance criteria. Fix infra and workbuddy cache invalidate to restart; don't rewrite the issue.

Dependencies & scheduling

Workbuddy has no separate "scheduler" or DAG runner — scheduling falls out of three pieces working together. Knowing them avoids re-deriving this every time:

1. Polling. The Coordinator polls each registered repo at poll_interval (config.yaml). GitHub label changes drive the state machine; nothing happens between polls. Force one with workbuddy cache invalidate --repo owner/name --issue N.

2. Task queue + claim lease. Eligible issues go into SQLite task_queue. Workers long-poll the Coordinator and claim by claim_token with a TTL, so two workers can never run the same task concurrently (REQ-057/059). Parallelism is per-worker per-repo — bind multiple workers (or --repos owner/A=...,owner/B=...) to fan out across repos.

3. Dependency gate. Declared in the issue body, not via labels:

   ## Dependencies
   
   ```yaml
   workbuddy:
     depends_on:
       - "#26"
       - "Lincyaw/workbuddy#27"
     ```
   

   The Coordinator parses this every poll, computes a verdict (ready / blocked / override / needs_human), and refuses dispatch when blocked, logging a dispatch_blocked_by_dependency event. UX signal on GitHub: a single 😕 reaction added when blocked, removed on unblock. No managed comment, no resolver agent — the gate lives entirely in Coordinator Go.

   • Same-repo shorthand #26 is normalized to owner/repo#26.
   • owner/repo#N cross-repo refs are accepted but unsupported in v0.1 (verdict goes to needs_human).
   • Cycles, malformed refs, or "trigger label present + depends_on declared but no status:* label" all surface in workbuddy diagnose.
   • Human override: add the override:force-unblock label — verdict flips to override (treated as ready). Don't toggle status:blocked by hand; the Coordinator owns it.

   To see the verdict for one issue:

   workbuddy status --repo owner/name | grep "  N "      # CYCLES + DEPENDENCY columns
   workbuddy status --events --type dispatch_blocked_by_dependency --since 1h
   

   To re-evaluate after editing a dependency declaration:

   workbuddy cache invalidate --repo owner/name --issue N
   

There is no priority queue, no cron, no time-based scheduling. If you want issue B to wait for issue A, declare it. If you want N tasks in parallel, run N workers (or one worker bound to N repos).

Parallel rollouts + synthesize (per-issue fan-out)

Beyond cross-issue parallelism, a single issue can also be developed N ways in parallel and then reduced to one PR. This is opt-in per issue.

Mental model:

  status:developing  ──fan-out──►  N parallel dev-agent runs
       (rollouts:N label)            each on its own branch
                                     workbuddy/issue-<N>/rollout-K
                                     each opens its own PR labeled
                                     rollout:K with "[rollout K/N]" title
                                              │
                          (≥ join.min_successes succeed)
                                              ▼
                                    status:synthesizing
                                              │
                                  review-agent runs in mode: synthesize
                                  → reads all sibling PRs/branches
                                  → emits a structured synthesis_decision:
                                      • pick one rollout as-is, OR
                                      • cherry-pick / merge the best
                                        pieces across siblings into a
                                        single result branch
                                              ▼
                                    status:reviewing
                                  (normal AC review on the chosen artifact)
                                              ▼
                                       status:done

Key points worth knowing:

• Triggered by adding a literal rollouts:N label (N in 2..5) to the issue. No label, or rollouts:1, takes the legacy single-dev fast path that skips synthesizing entirely.
• Each rollout is a fully isolated dev run — its own worktree, branch, PR, and session — so failures in one rollout don't block siblings. join.min_successes (in the workflow YAML) decides when the join fires.
• The synthesize step is a reduce, not a vote. It runs the review-agent in mode: synthesize; the agent must return a structured decision (chosen_rollout_index, plus optional cherry-pick instructions). If the structured output is malformed, the Coordinator stops the reduce step rather than silently auto-picking — surfaces in the issue as a synthesize failure for human attention.
• Audit trail: synthesis_decision events plus per-rollout sessions visible via the audit/web UI (rollout_index/rollouts_total/rollout_group_id).
• The state graph is in .github/workbuddy/workflows/default.md; the rollout fan-out and mode: synthesize are first-class in the workflow YAML, not bolted on.

Use this when the AC admits multiple plausible implementations and you want the agents to explore them in parallel, then keep the best one. For straightforward issues, leave rollouts:N off and stick with the single-dev path.

Common workflows

Create a workbuddy issue

gh issue create -R owner/name \
  --title "…" \
  --body '## Description
…

## Acceptance Criteria
- [ ] …' \
  --label "workbuddy,status:developing"

Unstick an issue

# Check what's wrong
workbuddy diagnose --repo owner/name

# Manually nudge state and force re-poll
gh issue edit N -R owner/name --remove-label status:blocked --add-label status:developing
workbuddy cache invalidate --repo owner/name --issue N

# Or, if the poller cache + claim lease both need clearing (stuck after a crash)
workbuddy issue restart --repo owner/name --issue N \
  --coordinator http://127.0.0.1:8081 \
  --token-file /home/$USER/.config/workbuddy/auth-token

Add a repo to a running deployment (no restart)

cd /path/to/new-repo
workbuddy repo register --coordinator http://coord:8081
# For a worker already running:
workbuddy worker repos add owner/new-repo=/path/to/new-repo

Environment variables

workbuddy <cmd> --help lists the flags each command honors. Globals that matter across commands:

Variable	Purpose
WORKBUDDY_AUTH_TOKEN	Bearer token for coordinator auth (used with --auth)
WORKBUDDY_SLACK_WEBHOOK_URL / WORKBUDDY_FEISHU_WEBHOOK_URL / WORKBUDDY_TELEGRAM_*	Optional notification sinks
WORKBUDDY_SMTP_*	Optional email notifications

When you need more detail

1. workbuddy <cmd> --help — flags + examples (authoritative)
2. references/ in this skill — setup, pitfalls, issue-writing
3. docs/decisions/ and docs/implemented/ in the repo — architecture rationale
4. project-index.yaml — requirements map (REQ-XXX IDs referenced above)

Related skills

• /setup-repo — fully onboard a new repo (labels + config + validation)
• /pipeline-monitor — watch a running pipeline, diagnose stuck issues
• /merge-flow — batch-merge approved workbuddy PRs with conflict handling