Cross-repo delegate (work tickets)

Who runs: the coding agent (Host TUI / OpenClaw), not the human.
Never tell the user to copy-paste run.sh as the primary path.

Single source: this directory (SKILL.md + run.sh + test.sh). Other paths are symlinks from scripts/dev/link-platform-skills.sh.

Skill id: cross-repo-delegate. Short CLI on PATH (optional): delegate.

Three modes (hard rule)

claim 默认 session（owner=human）。Headless worker 才用 claim --as-worker（由 run 内部调用）。
认领 ≠ 派工人。 用户在 TUI 认领后，由当前主会话读票、调研、改代码、challenge / close。

Concrete repo names live only in ~/.agent/tickets/config.yml (and the user’s words), not in this skill’s trigger text.

Routing: ticket vs multi-dir mount

Default for 跨仓委托 / 认领 / 挑战 / 关单: this skill (Modes 1–3).
Product multi-dir mounts (Claude --add-dir, Codex multi-folder, Cursor multi-root, parent-folder workspaces) are a tactical way to keep one semantic edit coherent across checkouts. They do not replace the ticket protocol (ownership, observed/assumed, challenge/reply, allowlist, audit).

More visible files are not better cross-repo context for delegation: the target session must load that repo’s AGENTS.md / CLAUDE.md, keep the single-writer lease, and leave a reviewable contract. Policy detail: docs/agent-scheduling.md#ticket-vs-multi-dir-mount (from repo root: docs/agent-scheduling.md#ticket-vs-multi-dir-mount).

Skip / redirect: long exploratory design → interactive TUI / OpenClaw C2 handoff, not a ticket loop.

Mode 3 only — explicit headless dispatch

"$D" create … --run [--backend claude|codex|grok]
# = research; if verdict=accept → implement immediately
# if challenge/need_input → stops for initiator

"$D" run --id <id> --phase auto|research|implement
"$D" reply --id <id> --decision "…" --continue   # after challenge (worker-owned only)
"$D" watch --id <id> [--once] [--interval 20]    # blocked initiator loop (worker-owned)

Policy for Mode 3: no objection after research → start implement; objection → human reply then continue. Filing with --run means the initiator is blocked — prefer watch so work resumes when the ball returns. Use --mock offline.

If a ticket is already session-claimed (owner=human), run / reply --continue / watch will not steal the lease unless the user explicitly asks to force headless (run --steal). Prefer developing in the claiming TUI.

Prompt injection (token trim): workers embed a phase view, not the full ticket. Omit only high-noise / phase-irrelevant sections — never the hop’s contract. Both phases keep Summary + Assumptions; implement also gets Verification + Decision (+ Implement if present). Thread / events.jsonl stay on disk under _data/<id>/ (and .delegate/ticket.md) for on-demand Read only.

Headless workers call shared scripts/dev/host-agent-invoke/ (write mode), set AGENT_ATTENTION_SKIP=1 / DELEGATE_HEADLESS, and (Claude) disableAllHooks so they must not appear in the human attention badge / Alt+/ list. Scheduling map: docs/agent-scheduling.md.

Resolve TOOL_HOME (first hit wins)

1. $CROSS_REPO_DELEGATE_HOME or $DELEGATE_HOME
2. directory of this SKILL.md if run.sh is co-located
3. $HOME/.agents/skills/cross-repo-delegate   # usual after link
4. $WEZDECK_ROOT/skills/cross-repo-delegate
5. $HOME/github/wezterm-config/skills/cross-repo-delegate

TOOL_HOME=…   # dir that contains run.sh
D="$TOOL_HOME/run.sh"
# If `delegate` is on PATH (install-cli), you may use it; otherwise always $D.

Never assume the skill lives only under the TARGET repo (TOOL ≠ TARGET).

One-time (only if missing)

"$D" init
# optional PATH: "$D" install-cli   # installs short name `delegate`

If ~/.agent/tickets/config.yml lacks the target, add it (path + aliases) before create.

Agent procedure

A) Mode 1 — File a ticket (initiator project cwd)

1. Resolve TOOL_HOME; confirm "$D" is executable.
2. Infer --to <allowlist_key> from the user; use --from . when cwd is the source project (or an explicit source key).
3. Write observed + assumed from the user’s finding (required).
4. Run:

"$D" create \
  --to <target_key> \
  --from . \
  --title "<short>" \
  --observed "<what you saw>" \
  --assumed "<what target should guarantee>" \
  [--snippet-ref "path:line"] \
  [--summary "…"] [--source-pr "#…"]

5. Report the printed id to the user. Do not commit the ticket into any git tree.
   Do not add --run unless the user asked for Mode 3 (委托开发).

B) Mode 2 — Inbox / claim / develop in this session (receiving project cwd)

"$D" inbox --to .
"$D" next --to .
"$D" claim --id <id>    # owner=human, mode=session
"$D" show --id <id>

--to . / --from . resolve the allowlist key from cwd via the primary git worktree root (git rev-parse --git-common-dir), so a linked worktree (…/.worktrees/<repo>/dev-…) counts as the same repo as the primary checkout — never as the worktree slug basename.

Read the full ticket, then edit code in the current worktree/TUI.
If claim returns lease held → stop; do not dual-write.
Do not call "$D" run after a successful session claim.

C) Mode 3 — Create and delegate development

Only when the user explicitly wants a background worker:

"$D" create … --run [--backend claude|codex|grok]
# or later:
"$D" run --id <id> --phase auto
"$D" watch --id <id>   # initiator blocked on worker

D) Challenge (target) → reply (initiator)

"$D" challenge --id <id> \
  --measured "…" --impact "…" --recommended "…" [--needs "…"]
# initiator side:
"$D" reply --id <id> --decision "…" --status waiting_target
# Mode 3 worker-owned only:
"$D" reply --id <id> --decision "…" --continue

Session-claimed tickets: after reply, the claiming TUI continues implement — no --continue.

E) Close

"$D" close --id <id> --doc <topic-doc-in-target-repo>
# or explicit skip:
"$D" close --id <id> --no-doc "<reason>"
# keep Mode 3 worktree (rare):
"$D" close --id <id> --no-doc "<reason>" --keep-worktree

close also reclaims the Mode 3 delegate-<id> worktree when meta.worktree is set (strips .delegate/**, keeps the local branch). Failure to reclaim does not undo close.

F) Inspect Mode 3 worktrees (decision table)

"$D" worktrees                 # JSON + stderr skim (suggest/dirty/ahead/age)
"$D" worktrees --alive         # only trees still on disk
"$D" worktrees --to .          # filter by target allowlist key

G) GC leftover Mode 3 worktrees

"$D" gc --dry-run
"$D" gc --execute --status closed,rejected,failed
"$D" gc --execute --include-shipped --min-age 7d
"$D" gc --execute --force --id <id>   # discard non-.delegate dirt

Update/create the topic solution doc in the target repo yourself; never commit ~/.agent/tickets/** into the project.

Status (authoritative)

submitted · in_progress · waiting_initiator · waiting_target · shipped · closed · rejected · failed

owner: worker | initiator | human

• human = Mode 2 session claim (main TUI develops)
• worker = Mode 3 headless path

Use "$D" next --id <id> when unsure whose turn it is.

Don't

• Don’t ask the human to run CLI as the main path
• Don’t commit tickets into project git / docs/decisions/
• Don’t invent a second OpenClaw-only ticket store — same skill
• Don’t skip --observed / --assumed on create
• Don’t close without --doc or --no-doc
• Don’t start a second claim while a lease is active
• Don’t treat Ticket-headless as C3 ACP (or the reverse); OpenClaw Main picks 执行通道 — see openclaw/docs/agent-interaction.md §6
• Don’t after claim (session) call run / spawn delegate-* worktree / headless worker
• Don’t use --run / run unless the user asked to 委托 / 后台 / 派工人 (Mode 3)
• Don’t replace Modes 1–3 with a standing multi-dir mount from the source repo into the target (“context is better if I edit both”) — tickets own delegation; mounts are tactical same-intent co-edits only

Tests (operators / CI — not the user path)

"$TOOL_HOME/test.sh"    # isolated MOCK fs; no LLM; no ~/.agent pollution

Related

• Runner: run.sh · lib: lib/ (lifecycle.sh = claim/run/reply) · offline: test.sh
• Link: scripts/dev/link-platform-skills.sh
• Sibling: adversarial-review, brainstorm
• Scheduling: docs/agent-scheduling.md (ticket vs multi-dir: #ticket-vs-multi-dir-mount)