Delegate To Other Repo

Parent Claude (you, in the current session) sets up an isolated worktree in a different target repo, constructs a self-contained brief, and dispatches a subagent with a fresh context to do the actual work. The subagent reads the target repo's conventions, opens a PR, and returns a structured final message. You relay the PR URL and a short summary.

Core principle: parent does infrastructure, subagent does content. Parent never cds — everything uses git -C <target>.

Announce at start: "I'm using the delegate-to-other-repo skill to set up a subagent for cross-repo work in <target>."

When to use

• •User asks for a change in another repo mid-conversation ("also fix the typo in the blog", "while you're at it, bump the version in the other service")
• •Task is self-contained enough for a subagent to handle end-to-end
• •You want to preserve the current session's context — cross-repo work would otherwise load another repo's CLAUDE.md, file reads, and diff into your working memory
• •Target repo already exists locally (this skill does NOT clone)

Do NOT use when:

• •The task needs back-and-forth with the user during the work
• •The user picks "branch" in the same-repo prompt (see Phase 1c-bis below) — in that case tell them to create a feature branch in-session and exit
• •The user explicitly said "just do it in this session"
• •The target isn't under ~/gits/ yet — tell the user to run gh repo clone owner/repo ~/gits/repo first

Flow at a glance

Parent (you)                             Subagent (fresh context)
──────────────────                       ─────────────────────────
1. Resolve target repo
2. git -C <T> fetch origin
   (plus upstream if present)
3. Create worktree off
   upstream/<default> if
   upstream exists, else
   origin/<default>
4. Build brief (template +
   substitutions)
5. Agent tool dispatch  ─────────────►   1. cd <worktree>
                                         2. Read CLAUDE.md greedily
                                         3. Enumerate skills/
                                         4. Do the task
                                         5. Detect fork vs direct
                                         6. Commit code, push, PR
                                         7. Commit reasoning doc
                                            separately
                                         8. Reflect on lessons
6. Receive final message ◄───────────    9. Return PR: / Summary: /
7. Relay to user                            Notes: / (optional) Lessons:
8. Offer lesson follow-up
   (if present)

Note: if the resolved target is the SAME repo as the current session, the flow branches at Phase 1c-bis — the user picks worktree (continue) or branch (abort and proceed in-session).

Phase 1: Resolve the target repo

Follow this checklist in order:

1a. Explicit argument

If the user passed a target:

• •Absolute path (/home/user/gits/blog) → use it
• •Relative path (../blog) → resolve against current pwd
• •Bare name (blog) → resolve to ~/gits/blog
• •owner/repo slug → STOP and tell the user:

  "I don't clone repos. Run gh repo clone <owner>/<repo> ~/gits/<repo> first, then retry."

1b. Inferred from conversation

If no argument, scan the recent conversation for repo references ("the blog", "chop-conventions", "that other repo") and match them against ~/gits/ entries.

Inference is never final. Always propose the match to the user and wait for confirmation before dispatching:

"You mentioned 'the blog' — I'm reading that as ~/gits/idvorkin.github.io. Proceed?"

If multiple candidates or no match, fall through to 1c.

1c. Ask

Run /bin/ls ~/gits/ and present the list. Let the user pick.

1c-bis. Same-repo handling

Do this BEFORE calling prepare_dispatch.py. The helper has no awareness of the parent session's cwd and will happily set up a worktree in the same repo you're sitting in — so the parent must gate the call on this check.

Compare the resolved target's toplevel against the current session's toplevel:

target_top=$(git -C "$T" rev-parse --show-toplevel)
session_top=$(git rev-parse --show-toplevel)

If they match, the target is the current repo. Delegation still has value (fresh subagent context, isolation from in-session uncommitted work), but the user may prefer to just branch in-session. Ask:

"Target resolves to the current repo. Two options:

1. •worktree — create a fresh worktree off upstream/main and dispatch a subagent in isolation (normal flow)
2. •branch — skip delegation; create a feature branch in the current session and proceed here Which?"

On "worktree" → continue to Phase 1d unchanged. The worktree isolation still applies; the subagent will run in a fresh context.

On "branch" → abort this skill. Print:

"Not delegating. Create a feature branch with git checkout -b <slug> and proceed in-session."

Do NOT create a worktree, do NOT dispatch a subagent. End the skill cleanly.

1d. Validate the resolved target

All validation runs via git -C <path> — never cd into the target:

1. •Path exists
2. •Is a git repo: git -C "$T" rev-parse --is-inside-work-tree
3. •Has an origin remote that resolves: git -C "$T" remote get-url origin
4. •Default branch is resolvable (see Phase 2 recipe)
5. •$base_ref (upstream/<default> if upstream exists, else origin/<default>) is reachable after git fetch

Not required to be clean. Worktrees off the remote base ref are safe even when the target's working tree is dirty.

Phase 2: Create the worktree (and finalize Phase 1 validation)

prepare_dispatch.py does Phase 1d validation AND Phase 2 worktree setup in one call. Phase 1a–1c (resolving the user's intent) and Phase 1c-bis (same-repo guard) stay in the parent because they need conversational context the helper can't see.

DO NOT delegate to superpowers:using-git-worktrees. That skill branches off current HEAD, auto-runs npm install / cargo build, and runs baseline tests — none of which are correct for delegating a change off a fresh canonical base ref that may be a doc-only edit.

Preferred: prepare_dispatch.py

Run the helper — it performs Phase 1d validation + Phase 2 worktree setup in a single call and returns the inputs Phase 3 needs as JSON:

skills/delegate-to-other-repo/prepare_dispatch.py \
  --target <name|abs-path|rel-path> \
  --slug <task-slug> \
  --task "<task description>" \
  --pretty

Add --dry-run to validate and inspect the JSON without creating the worktree or writing .git/info/exclude.

Output shape (stable JSON contract — the parent reads these keys only):

{
  "target": "/abs/path/to/target/repo",
  "worktree_path": "/abs/path/to/.worktrees/delegated-<slug>",
  "branch": "delegated/<slug>",
  "slug": "<final-slug-after-collision-resolution>",
  "base_ref": "upstream/main",
  "base_remote": "upstream",
  "default_branch": "main",
  "target_repo_slug": "owner/repo",
  "session_log": "/path/to/jsonl or null",
  "task": "<the --task string verbatim>",
  "dry_run": false,
  "errors": [],
  "warnings": []
}

Exit code and error handling

• •Exit 0 iff errors is empty. Exit 1 otherwise. The helper always emits the JSON on stdout, even on failure — do NOT gate JSON parsing on the exit code.
• •Non-empty errors ⇒ helper stopped before creating the worktree. worktree_path is null in this case. Relay the first error to the user and abort — do NOT dispatch the subagent. Typical causes: target path missing, not a git repo, owner/repo slug passed (helper does not clone), base ref unreachable after fetch, git worktree add collision.
• •Non-empty warnings ⇒ helper recovered and proceeded. The worktree exists and dispatch is safe. Current producers: upstream fetch failure (fell back to origin/<default>). Surface warnings to the user in the dispatch summary so they notice drift from the fork-workflow assumption.

Additional keys like task and dry_run are echoed for log-reading — the parent does not need to consume them. Future fields will be added as optional; keys are never renamed or removed without a V2 bump.

Steps the helper performs (all git -C <target>, never cd) are documented in prepare_dispatch.py's module docstring and mirrored in worktree-recipe.md for the manual fallback. Do NOT re-list them here — the two sources drift and this skill becomes the rotting one. When reviewing a helper change, diff the docstring and the fallback recipe together.

Testing

cd skills/delegate-to-other-repo && python3 -m unittest test_prepare_dispatch

Pure functions (slug sanitization, default-branch chain, base-ref selection, session-log hash, remote-URL parse) are unit-tested in isolation — no subprocess mocking required for those.

V1 limitation

This skill hardcodes .worktrees/delegated-<slug> and does NOT honor a target repo's CLAUDE.md worktree.*director preference. using-git-worktrees does honor that — revisit if it becomes a problem.

Phase 3: Construct the brief

The brief is the single most important artifact this skill produces. It must be fully self-contained — the subagent sees none of the current conversation.

Template lives at brief-template.md. Read it, substitute the slot placeholders, and pass the result as the prompt parameter to the Agent tool.

Slots to substitute

Session log resolution

Use the value from prepare_dispatch.py's session_log field. The helper handles the physical-path + [/.]-hash quirks and falls back to the repo toplevel hash. If session_log is null, omit the entire "Historical context" section from the brief — do NOT guess a path.

The shell-path reconstruction used to live here and is preserved inline in worktree-recipe.md §7 as the manual-fallback equivalent. Do not re-derive the hash rule in the parent — it's load-bearing enough that a single source of truth matters. Parallel sessions in the same cwd resolve to "whichever jsonl was most recently written"; this is an accepted v1 ambiguity since the log is an escape hatch, not a required input.

Phase 4: Dispatch the subagent

Agent tool:
  description: "Delegated work in <target-repo>"
  subagent_type: "general-purpose"
  prompt: <the substituted brief from Phase 3>
  run_in_background: true

Async by default. Delegated work is usually long-running (minutes). Blocking the parent on it wastes the user's time and burns the parent's context budget while it sits idle. The harness sends a <task-notification> when the subagent completes — that notification is your trigger for Phase 5. No polling required.

After dispatch

1. •Summarize what you dispatched to the user in one short message — worktree path, branch, key checkpoints from the brief. This gives the user a chance to course-correct before the subagent burns minutes in the wrong direction.
2. •End the turn. The parent is now free to accept other unrelated work while the subagent runs.
3. •When the <task-notification> arrives, resume at Phase 5 automatically (relay the result, offer follow-ups).

Monitoring

• •Default: trust the completion notification. Simple, reliable, no overhead. Do NOT poll, do NOT sleep, do NOT Read the agent's output JSONL — the tool explicitly warns that reading the transcript will overflow the parent's context.
• •Heartbeat (opt-in): for long-running or risky delegations where the user wants progress checks, run /loop 2m "status check on delegation to <target-repo>". The loop wakes the parent every 2 minutes to summarize state from memory (what was dispatched, how long ago, what's expected). The parent answers from its recollection of the brief — NOT by reading the output file. When the completion notification arrives, the parent processes the real result and the loop self-terminates on the next tick.
• •Never: tail the output file, sleep in a bash loop, call the Agent tool again with the same prompt, or claim "done" before the notification arrives.

If the user explicitly asked for sync

If the user said "wait for it" or "block until done", pass run_in_background: false instead — but note that the harness may still dispatch async regardless. Either way, the parent's Phase 5 trigger is the final message, wherever it arrives from.

Never retry automatically. If the subagent fails, escalate to the user (see Phase 5).

If dispatch fails after the worktree exists

The helper is mutating (worktree created, .git/info/exclude appended) before dispatch. If the Agent tool call itself fails (permission denied, transport error), the worktree is orphaned — neither the subagent nor Phase 5 will reach it. Surface this to the user with the exact cleanup command and let them decide:

"Dispatch failed before the subagent started. Orphan worktree at <path> on branch <branch>. Clean up with git -C <target> worktree remove <path> && git -C <target> branch -D <branch>, or re-run the skill (the helper resolves slug collisions, so a retry will pick a fresh name)."

Do NOT auto-remove. The user may want to inspect the failed state.

Phase 5: Relay the result and offer follow-ups

The subagent returns a final message with:

1. •PR: URL — on its own line
2. •Summary: 3–5 bullets — what changed and why
3. •Notes: pointer — <hostname>:/tmp/agent-notes/YYYY-MM-DD-<slug>.md on the parent's machine. Same string appears as a Reasoning: trailer in the work commit. Required.
4. •Lessons: block (optional) — draft CLAUDE.md insertions the subagent thinks are worth capturing

Happy path

Relay the sections the subagent returned verbatim to the user (PR:, Summary:, Notes:, and Lessons: if present). Add a note with the worktree path and the cleanup command:

"Worktree preserved at <path> on branch <branch>. Delete both with git -C <target> worktree remove <path> && git -C <target> branch -D <branch> once the PR is merged (or you're done iterating). The branch is safe to -D after squash-merge — delegated/<slug> is never fast-forwarded onto main."

Leaving the worktree is intentional: lessons follow-up and take-over recovery both need it. Do NOT auto-remove.

If Lessons: is present

Show the block verbatim, then offer two follow-up paths:

1. •Quick path — "Open a second PR in the same worktree with just this CLAUDE.md addition?" If accepted, run the commit and gh pr create in the existing .worktrees/delegated-<slug> worktree (still exists, since cleanup is manual). Branch name: delegated/<slug>-lessons.
2. •Full path — "Run /learn-from-session on the target repo for multi-file routing?" For lessons that span multiple CLAUDE.md files or need deeper routing.

If the user declines both ("skip it"), that's a normal terminal state — omit the follow-up and consider the delegated run complete. Rejected lessons are NOT a failure.

If the subagent's final message doesn't match the contract

Missing PR: line OR missing Notes: line → treat as failure. Show the user the subagent's last message and ask:

• •Retry with the same brief?
• •Abandon — delete the worktree and stop?
• •Take over in-session — you (parent) cd into the worktree and continue the work manually?

If the code pushed but /tmp/agent-notes/<file> is missing or unwritten, re-write the file locally and amend the Reasoning: trailer into the commit message. No auto-retry.

Fork workflow detection (reference)

The subagent handles this itself — the brief walks it through the decision tree. For the full 4-case tree with ASCII diagram and the diagnose.py reuse strategy, see fork-detection.md.

TL;DR: the subagent runs gh auth status, git remote -v, and (for single-remote cases) gh repo view <slug> --json isFork,parent, then classifies into:

• •Case A — two remotes (origin + upstream): fork workflow, push to origin, PR --repo <canonical>
• •Case B — one remote, canonical origin matching auth: direct push, PR with no --repo
• •Case C — one remote, fork-as-origin matching auth (chop-conventions' real pattern): push to origin, PR --repo <parent-from-gh-json>
• •Case D — one remote, canonical origin NOT matching auth: look for sibling fork remote; if none, STOP

Reasoning audit trail (local-only, referenced from commit)

Every delegated run writes /tmp/agent-notes/YYYY-MM-DD-<slug>.md on the parent's machine. The file stays local — it is NOT committed to the target repo. The commit that ships the work includes a Reasoning: trailer pointing back to it:

Reasoning: <hostname>:/tmp/agent-notes/YYYY-MM-DD-<slug>.md

<hostname> comes from hostname on the parent's machine. <slug> strips the delegated/ prefix from the branch name. Future-Igor (or future-Claude) can ssh <hostname> and cat the file to recover context, without the audit trail polluting public repo history or leaking intent text.

Six level-2 sections in the file, in order:

1. •User request — brief intent summary plus a pointer to the source (session jsonl path, Telegram msg id, PR review comment). Verbatim ONLY for chore-style asks with no private content.
2. •Parent's interpretation — scope decision + why delegated.
3. •Subagent's plan — pre-execution, unchanged after the fact.
4. •Decisions — deliberate forks taken during execution.
5. •Outcomes — commit SHAs, files touched, verification run, PR URL.
6. •Deferred — what was explicitly NOT done.

Ephemerality is a feature. /tmp/ survives for the session (and typically until reboot); beyond that it may disappear. The whole point is that reasoning docs are working-memory artifacts, not repo history. If a specific reasoning doc matters beyond a session, upgrade it — copy to ~/agent-notes/ or into a longer-lived location, on purpose.

brief-template.md carries the authoritative spec the subagent follows.

Integration with learn-from-session

Reflection happens in the subagent, not the parent. The parent has no visibility into what tripped up the subagent during the work (missing docs, hook reformats, unclear conventions) — only the subagent lived it. So the subagent runs through learn-from-session's reflection prompts on its own work, applies the durability filter, and drafts any surviving lessons inline in its final message.

The parent's job is to relay, not to reflect:

1. •Subagent drafts — never commits CLAUDE.md edits to the work PR
2. •Parent relays the Lessons: block verbatim
3. •Parent offers quick-path or full-path follow-ups (see Phase 5)
4. •User owns the approval gate

Hard prohibitions

These apply to both parent and subagent:

• •No git push --force on any branch
• •No --no-verify on commits
• •No commits on any branch of the target's primary checkout — the parent never mutates the target's branches. The only thing the parent writes is .git/info/exclude (local-only, untracked). All subagent work happens on the delegated branch inside the worktree and is pushed via normal PR flow.
• •No rm -rf or destructive ops without explicit user confirmation
• •No gh pr merge — opening the PR is the terminal action
• •No committing CLAUDE.md edits derived from lessons reflection to the work PR. Lessons are draft material in the subagent's final message only; the user owns the approval gate.
• •No cd in the parent. Parent always uses git -C <target>. Only the subagent cds, into the worktree, as its first action. Take-over recovery may enter the worktree briefly and cd - out.
• •No skipping the target's root CLAUDE.md read in the subagent. Missing file means STOP and ask the user, not proceed on defaults.
• •No committing the reasoning doc. It lives on the parent's machine at /tmp/agent-notes/; the Reasoning: trailer in the commit message is the only durable breadcrumb.
• •When fixing PR review comments, commit the fix, reply to the thread with the commit SHA + a short how/why, then resolve the thread.
• •When updating this skill (or any skill in this repo), code-review your own change before opening the PR. Skill files are contracts — drift is invisible until it bites.

Concurrency — parallel delegations to the same target

Two parent sessions (or one parent dispatching twice back-to-back) can point at the same target repo at once. The helper is safe under concurrency because:

• •Slug collision resolution checks refs/heads/delegated/<slug> AND refs/remotes/origin/delegated/<slug> — distinct tasks get distinct branches. Identical task slugs get -2, -3, ...
• •git worktree add is atomic from git's side — one of the two will fail the <path> collision and the helper's errors will surface the git error verbatim.
• •.git/info/exclude append is NOT atomic. Two racing helpers can both miss the .worktrees/ line and both append — benign (git tolerates duplicate exclude lines). File is never rewritten, only appended, so the old-line survives if a write interleaves.

What the helper does NOT prevent:

• •Two parents dispatching subagents that work on overlapping files. Scoping is the user's job — if you delegate two changes to the same file, they will conflict at merge time. Warn the user if the task description mentions a file you already saw in a recent delegation summary.
• •A third delegation starting while the first two are pending <task-notification>s. The parent is free to dispatch during that window, but each dispatch consumes a subagent slot.

Failure handling

Parent-side

Subagent-side

Common mistakes

Using cd <target> instead of git -C <target>

Parent ends up stranded in the target's cwd, polluting shell state for any post-dispatch commands. Fix: always git -C "$T" in the parent. Only the subagent cds.

Hardcoding origin/main

Breaks on repos with master or trunk as the default branch, and on two-remote fork workflows where origin is the fork and can lag upstream. Fix: resolve both the default branch name AND the base remote via the chain in worktree-recipe.md — upstream preferred when it exists, origin as fallback.

Skipping remote set-head --auto after fetch

Plain git fetch origin does NOT refresh refs/remotes/origin/HEAD. A target whose default branch was renamed (e.g. master → main) since clone yields a stale value from symbolic-ref. Fix: always run git -C "$T" remote set-head origin --auto between fetch and the symbolic-ref lookup.

Committing .worktrees/ to a branch's .gitignore

Committing on the target's current branch pollutes arbitrary branches and vanishes on next checkout. Committing on the default branch requires branch-switching (destructive) and breaks on protected defaults. Fix: write to .git/info/exclude instead — local-only, untracked, branch-independent, per-repo (shared across worktrees).

Editing this skill through the ~/.claude/skills/ symlink

~/.claude/skills/delegate-to-other-repo/ is usually a symlink into the chop-conventions primary checkout at ~/gits/chop-conventions/skills/delegate-to-other-repo/. Editing SKILL.md or any supporting file directly through the symlink mutates whatever branch the primary checkout is currently on — silently mixing skill-fix edits with whatever unrelated work is checked out there. Verified this session: the primary checkout was on a different feature branch with uncommitted Python edits when a well-intentioned edit would have polluted both.

Fix: when modifying this skill (or any skill in the chop-conventions skills/ tree), always create a worktree off upstream/main first and edit there. Same recipe as the worktrees this skill creates for its own delegations — the self-referential case doesn't get a free pass. Double-check with realpath before editing any file under ~/.claude/skills/ to confirm where it actually lives.

Trusting diagnose.py's is_fork_workflow: false

It misclassifies chop-conventions' real pattern (single fork-as-origin with no upstream). Fix: for the single-remote case, always run gh repo view <slug> --json isFork,parent manually — see fork-detection.md.

Nested fences in the brief

Embedding ``` inside the brief's outer ```markdown fence terminates the fence early when the brief is stored in SKILL.md. Fix: the brief uses plain-text formatting for all embedded structure (decision trees, output contract, Lessons: block). See brief-template.md — it's pre-formatted to avoid this.

Red flags — STOP and reconsider

• •You're about to cd into the target repo in the parent
• •You're about to call using-git-worktrees to create the worktree
• •You're about to retry a failed subagent dispatch automatically
• •You're about to commit a drafted lesson to the work PR
• •You're about to dispatch without confirming an inferred target
• •The target slug is non-ASCII and you're about to use it unfiltered
• •You're about to dispatch a subagent to the same repo without first asking worktree-or-branch (see Phase 1c-bis)
• •The subagent's final message has a PR: line but no Notes: line — treat as contract failure

Related

• •REQUIRED SUB-SKILL: superpowers:using-git-worktrees — read it so you understand why this skill deliberately does NOT call it
• •learn-from-session — the reflection flow the subagent runs inside itself before returning
• •up-to-date — its diagnose.py helper is what the subagent uses as a fork-detection shortcut (Cases A and simple B only)

Manual fallback

If prepare_dispatch.py is broken or unavailable (stale checkout, Python not on PATH, uv bootstrap failure), drop to the bash recipe at worktree-recipe.md and follow it verbatim. The Python helper was extracted FROM that recipe and the pure functions are unit-tested, but the two implementations are NOT byte-for-byte equivalent — they can drift. When touching either, diff against the other as part of the change. Prefer the Python path whenever both work; the shell recipe exists only as a break-glass.

Supplementary files

• •prepare_dispatch.py — preferred Phase 1-3 helper (resolves target, fetches, picks base ref, slug-collision resolves, creates worktree, emits JSON)
• •test_prepare_dispatch.py — unit tests for the helper's pure functions and dry-run orchestration
• •worktree-recipe.md — manual-fallback shell recipe with full commentary on every step
• •brief-template.md — the full self-contained brief the parent substitutes slots into and passes to Agent
• •fork-detection.md — the 4-case decision tree with ASCII diagram and diagnose.py interaction rules