agent-delegate — Headless Delegation and Review Between Agents

Hand a task to the other agent, or get an independent adversarial review, by calling its CLI headlessly. This is the cmux-free counterpart to cmux-delegate and cmux-second-opinion: it needs only a git repository and the peer CLI installed, and it returns a machine-readable report.json.

The runnable interface is references/scripts/agent-delegate.sh. Its full argument and output contract is in references/contract.md (日本語: contract.ja.md). Upper skills call the script directly; this SKILL.md is the interactive entry point for humans.

Language Rules

1. Auto-detect input language → output in the same language
2. Japanese input → Japanese output, use references/*.ja.md
3. English input → English output, use references/*.md
4. Explicit override takes priority (e.g., "in English", "日本語で")

Prerequisites

• Run inside a git repository (touchedFiles measurement degrades outside one).
• The peer CLI must be installed: codex when running under Claude Code, claude when running under Codex.
• Codex direction only: the workspace must be trusted in ~/.codex/config.toml. The script stops with exit 2 and instructions if it is not.
• node (18 or later) must be on PATH. The script reads and writes its JSON records through references/scripts/json-tool.mjs; no other JSON tool is needed, so it also runs in Git Bash on Windows.

The script self-detects the host agent. If it cannot (neither CLAUDECODE nor a Codex runtime marker is set), pass --target <codex|claude> explicitly.

Request Type Detection

Determine what the user wants from their phrasing:

When ambiguous, ask with bilingual options:

• question: "What should the other agent do?" / "もう一方のエージェントに何をさせますか？"
• options:
  • "Delegate a task" / "タスクを委譲" — hand off work and collect the result
  • "Adversarial review" / "敵対的レビュー" — read-only review, structured findings

Execution Flow

Step 1: Summarize the request

Restate the task or review scope in one or two sentences so the user can confirm what will be sent. Identify the direction (which peer will run).

Step 2: Build the prompt file

Write the full instruction for the peer to a file (e.g. under the out-dir), and pass it with --prompt-file. Never inline a long prompt as an argument.

• For delegate: describe the task, the acceptance criteria, and any constraints.

• For review: describe the review target (diff, spec paths, focus areas). The script prepends the adversarial template automatically; your file supplies only the context. If a review_rules.md exists in the project, include its contents as review criteria and tell the reviewer to also raise issues beyond those rules — matching the cmux-second-opinion behavior. Detect it with:

  find . -name "review_rules.md" -maxdepth 3 2>/dev/null

  If none is found, ask the reviewer to review freely.

Step 3: Choose synchronous or detached

A synchronous call blocks until report.json is written. --detach returns immediately and prints the run id followed by the future report.json path. Save both as the expected run. Poll every 15 seconds by default, never less often than every 30 seconds, and follow the expected-run state machine in references/contract.md: validate the report first, then inspect owner, pid, heartbeat, and process state. A missing report while the run is alive is not a failure. Use the expected-run owner's worker_pid before the first heartbeat, and select states in this order: terminal report, different owner, process disappearance, invalid report, then heartbeat. Re-evaluate every detached run at 30 minutes and every 30 minutes thereafter. At 2 hours, apply the contract's controlled stop: recheck the report, send TERM only to the verified expected monitor, wait up to 90 seconds for its terminal report, then stop waiting and escalate with diagnostics. Never use --force for this timeout path.

Before launch, register each task artifact's exact path, correlation value, freshness baseline, and task-specific validator. For review, also record a pre-launch content fingerprint of tracked worktree and staged diffs plus every non-ignored untracked path and its content, excluding the declared out-dir. If the expected-run terminal report is blocked with blocker_category: env_error, run the contract's fail-closed artifact recovery before reporting task failure. Monitor loss or a missing idle signal alone does not decide completion. On hosts that reap detached monitors, the durable waiter may be a synchronous bounded until loop over the artifact validator, using the original launch deadline. The loop does not relax recovery eligibility: without a valid expected-run env_error report, it rejects recovery and escalates diagnostics at the deadline.

Step 4: Run the script

# Synchronous read-only delegate with a concrete <=5-minute basis
report="$(skills/agent-delegate/references/scripts/agent-delegate.sh \
  --mode delegate --prompt-file <prompt> --out-dir <out> --label <slug> | tail -1)"

# Adversarial review with a concrete <=5-minute basis (always read-only)
report="$(skills/agent-delegate/references/scripts/agent-delegate.sh \
  --mode review --prompt-file <context> --out-dir <out> --label <slug> | tail -1)"

# Writing or otherwise unbounded task, detached
launch="$(skills/agent-delegate/references/scripts/agent-delegate.sh \
  --mode delegate --prompt-file <prompt> --out-dir <out> --label <slug> --detach)"
expected_run_id="$(printf '%s\n' "$launch" | sed -n 's/^run_id: //p')"
report="$(printf '%s\n' "$launch" | tail -1)"
# Register a durable 15-second watcher that applies the contract state machine.

The last line of stdout is always the report.json path. A successful launch also prints run_id: <uuid> immediately before it.

Step 5: Read the report and present the result

Read report.json and summarize for the user in their language:

• status — done or blocked.
• summary — the peer's final message headline.
• touchedFiles — files the peer changed (script-measured, authoritative).
• For review: read artifacts.review_file and validate it in this order — the script verifies structure only, so never adopt the Gate line at face value (see references/adversarial-review-prompt.md):
  1. Every Critical / Improvement finding carries a fix_before tag whose value is in the stage list in effect — the four default values, or the ordered list the review context supplied instead (the prompt template tells the reviewer to use that list). A finding with a missing or out-of-list tag is malformed output (treat as blocked; re-run or inspect) — do not compute a gate from it, or an untagged Critical would silently pass.
  2. Recompute the Gate from the tags — FAIL iff at least one finding carries the gate-blocking stage: the first stage of the list in effect (implementation by default). A Gate line that contradicts this tally is also malformed output.
  3. Report the recomputed Gate, the Critical / Improvement / Minor counts, and the gate-blocking-stage count.
• If blocked: report blocker and blocker_category, and suggest a next step (e.g. resume, fix the trust setting, retry). For the expected run's env_error only, first apply artifact recovery from references/contract.md: prove freshness and correlation, then run the review checks above or the task-specific delegate validator registered before launch. Review recovery must also compare the pre-launch and post-run content fingerprints after excluding the declared out-dir. Record an adopted artifact separately and retain the blocked report as a runtime diagnostic.

To continue a session, re-run with --resume <thread_id> from the prior report (same sandbox stage; see contract.md for the constraints).

Error Handling

Notes

• This skill provides the mechanism. Result quality depends on the peer agent.
• For cmux-based delegation and review, use cmux-delegate / cmux-second-opinion. Use agent-delegate when cmux is not available or a parseable report is needed.
• The report.json schema, sandbox mapping, resume rules, and the read-only guarantee-level difference between directions are documented in references/contract.md.
• Codex under the workspace-write sandbox may refuse to write agent-config dot directories (.agents/, .claude/), reporting writing outside of the project; this can occur with repo layouts that include symlinks. When a task's edits target those directories, prefer assigning it to claude.