agpair

name: agpair description: "Use when Claude Code handles non-trivial coding, refactor, test-fix, research, review, or multi-file work where AGPair external CLI agents can produce implementation candidates, alternative analysis, or verification evidence."

AGPair for Claude Code

IRON LAW: Non-trivial work requires an explicit routing budget. Use external AGPair lanes by default, but fan out only when each lane has a distinct role, scope, hypothesis, or verification value. Multiple grok-cli lanes are valid when they are not duplicate prompts.

Claude Code is the controller and verifier. AGPair is the durable external-agent control plane.

Default Route

Use AGPair before Claude Code native subagents for non-trivial implementation, refactor, test-fix, research, review, or multi-file work.

Actively outsource low-value, repetitive, time-consuming, or easily verifiable work through AGPair: repo scans, alternative reviews, focused test-fix attempts, multi-file mechanical edits, smoke checks, and implementation slices with clear acceptance criteria. For ordinary non-trivial work, start with a strong external lane: grok-cli and antigravity-cli are peer first choices. Use both when independent evidence, implementation contrast, or latency hedging is useful. It is acceptable to run several attempts from the same executor at once for parallel review, competing implementation approaches, or file-sliced work, but each attempt needs a distinct task id, role, prompt, or scope.

Prefer these external lanes:

1. grok-cli and antigravity-cli as peer first choices for review, research, and implementation attempts.
2. Additional grok-cli or antigravity-cli lanes when prompts, scopes, or acceptance criteria are distinct.
3. codex when an AGPair-managed external Codex CLI worker is useful as a fallback or challenger executor.

codex is the AGPair-managed external Codex CLI worker for Claude Code controllers. It is the cross-controller fallback lane, not a Claude Code native subagent.

Do not request the AGPair-managed external claude-code executor by default; Claude Code already has native subagents and claude-code is suppressed for Claude Code controllers unless --allow-self-executor is explicitly justified.

Only route new work to active registered executor ids. Historical task records may remain inspectable for compatibility, but they are not default dispatch targets.

Use Claude Code native subagents as narrow controller-side reviewers/helpers when they add clear value, and as fallback when AGPair is unavailable, unsuitable for the task, or an external result is not good enough. If skipping external AGPair entirely for non-trivial work, state the skip reason before proceeding. Native subagents are not the default primary execution lane, but they can run in parallel when they materially improve verification, hidden-context reasoning, or integration of external outcomes.

Default executor environments are managed-natural for all active external CLI executors: AGPair manages state and evidence, while the external CLI keeps its normal skills, MCP, memory, plugins, and provider config.

AGPair external-first routing applies to controller sessions. AGPair-started executor, probe, smoke, and retry processes suppress AGPair client hooks to avoid recursive delegation, but external workers still inherit their normal CLI capabilities, skills, MCP, plugins, memory, and provider config unless an explicit diagnostic mode says otherwise.

Use agpair policy list --controller claude-code --json to inspect the effective executor order, suppression, and lifecycle state. Use agpair policy disable/enable/priority/reset for pluggable runtime changes instead of editing source.

Routing Budget And Fanout

Before doing non-trivial work directly or using native subagents, answer:

1. Which external lane is most likely to produce an adoptable result?
2. What is the routing budget: one lane, several role-based lanes, direct work, or a native helper?
3. For every extra lane, what distinct role does it play: implementation, review, test-fix, research, adversarial critique, or file-sliced work?
4. If the task mutates files, can every mutating lane use an isolated worktree or a disjoint scope?

Default to one high-likelihood external lane for ordinary work. Dispatch additional lanes when they are role-distinct and the added evidence is likely to reduce rework, risk, or uncertainty. Do not cap useful fanout by habit, but also do not add lanes as ceremony. Stop when an added lane would only duplicate another lane's prompt, scope, or expected evidence.

Skip AGPair entirely only when one of these is true, and state the reason:

• the task is tiny or mostly mechanical;
• the task is sensitive or depends heavily on current controller context;
• external executors are unhealthy, unavailable, or already produced low-quality output for this task;
• safe isolation is unavailable for mutating work and a report-only external lane would not help;
• the best path is a narrow controller-side check or native helper.

Recommended shapes:

Give each external lane the same goal, explicit scope, and comparable exit criteria. For concurrent lanes, use --no-wait, then task wait / task watch to collect evidence without burning controller turns. For one lane, normal task start waiting is fine.

For code-writing work, fanout is still useful, but every mutating lane must use --isolated-worktree or a disjoint repo/worktree. Do not run multiple mutating executors in the controller worktree. Safe patterns:

• one primary implementation lane plus one external review/test lane;
• two alternative implementation candidates in separate isolated worktrees;
• multiple grok-cli or same-executor instances with distinct task ids, prompts, file slices, or acceptance criteria; mutating instances need disjoint scopes or separate isolated worktrees.

Claude Code native subagents remain available as review or narrow helper lanes when they add useful controller-side verification, and as fallback when external output is unavailable, unsuitable, or not good enough. Prefer external AGPair lanes for primary execution, but do not avoid a native helper when it can run in parallel and materially improve verification. Native helpers are especially appropriate for hidden-context reasoning, quick local sanity checks, integrating multiple external outcomes, or reviewing an external diff before adoption.

Anti-patterns:

• Do not use fanout as ceremony when added lanes no longer improve the result.
• Do not run duplicate external lanes with the same prompt, scope, and exit criteria.
• Do not run multiple mutating lanes in the controller worktree.
• Do not keep waiting on a silent or low-quality lane after another lane has produced adoptable evidence.
• Do not treat task count as success; success is usable evidence that reduces controller rework.

Pre-delivery check:

• Routing budget made: single external lane, role-based fanout, direct work, or native helper.
• Every external brief is self-contained and not dependent on hidden chat context.
• AGPair skip reason stated if no external lane was used for non-trivial work.
• Every mutating external lane is isolated or disjoint.
• task status --json inspected for each lane.
• Useful evidence was accepted/adopted or explicitly rejected.
• Final answer distinguishes external evidence from controller judgment.

Fusion-Style Workflow Fanout

For high-value research, review, design, implementation choices, or competing candidate work, prefer a fanout-synthesis workflow instead of manually starting unrelated tasks. The useful pattern is parallel lane tasks, one synthesis node, and a controller gate:

agpair workflow fanout \
  --controller claude-code \
  --mode review \
  --topic "$TOPIC" \
  --lane grok-cli:primary \
  --lane grok-cli:adversarial \
  --lane antigravity-cli:second-opinion \
  --repo-path "$REPO" \
  --wait --json

For implementation or test-fix panels, mutating lanes are isolated by the fanout preset. Keep the scope explicit:

agpair workflow fanout \
  --controller claude-code \
  --mode implementation \
  --topic "$TOPIC" \
  --scope "$ALLOWED_SCOPE" \
  --lane grok-cli:candidate-a \
  --lane antigravity-cli:candidate-b \
  --lane codex:reviewer \
  --isolated-worktree \
  --repo-path "$REPO" \
  --dry-run --json

Read panel_result, lane_cards, synthesis_result, and evidence_path before answering. The synthesis result is evidence, not final truth. Controller verification still decides whether to use, apply, retry, switch executor, or fall back to a native helper.

Fanout presets populate coordination_role and role_coverage so Claude Code can see whether lanes covered thinker/worker/verifier roles. Treat those fields as prompt/status metadata only; they do not make a result adoptable and missing expected roles are advisory, not a hard failure.

Dispatch

For ordinary tasks, send a clear natural brief. AGPair normalizes useful briefs and should not reject work merely because a section heading is missing. Do not pass placeholders like <brief>, todo, or fix this.

External lanes do not share Claude Code's hidden conversation state. The brief must stand alone: include the goal, relevant paths, constraints, forbidden scope, expected output, and validation or evidence requirements. If the worker would need unstated chat context to succeed, either add that context to the brief or keep the work in Claude Code/native helper lanes.

For complex mutating work, prefer this structured shape because it gives the external executor tighter scope and gives the controller better evidence:

Goal:
State the concrete outcome.

Scope:
Allowed files/areas:
Forbidden files/areas:

Required changes:
Describe the expected edit, or say: None. This is report-only. Do not edit files.

Exit criteria:
List required verification, report format, and expected AGPair evidence.

For a single external lane, let task start wait by default:

agpair task start \
  --repo-path "$REPO" \
  --controller claude-code \
  --executor grok-cli \
  --task-kind quick_review \
  --wait-policy lease \
  --authorization-profile local_readonly \
  --completion-policy report \
  --body "$BRIEF"

For non-trivial implementation, refactor, or test-fix work, dispatch one bounded isolated implementation slice first unless AGPair is unavailable, unsafe, or already low quality for this task. Add a second external implementation or review/test lane only when risk, uncertainty, or verification value justifies the extra coordination:

agpair task start \
  --repo-path "$REPO" \
  --controller claude-code \
  --executor grok-cli \
  --task-kind implementation \
  --wait-policy lease \
  --authorization-profile local_mutating \
  --completion-policy evidence \
  --isolated-worktree \
  --body "$BRIEF"

Use a brief with explicit allowed files, forbidden files, required changes, validation command, and exit criteria. The external worker returns changed_files, validation or validation_not_run, scope_violations, report text, and raw evidence paths. Claude Code integrates or rejects the result in the main worktree after verification.

For isolated mutating evidence/commit tasks, AGPair defaults to --dirty-snapshot tracked: tracked staged/unstaged controller changes are copied into the executor worktree before launch. Ignored and untracked files are not copied; use --dirty-snapshot off when the worker should start from committed HEAD only.

When --wait-policy lease expires and the task is still alive, detach and continue or run a native reviewer in parallel. Do not abandon a complex external task solely because it has not produced a quick final report.

For parallel or background work, dispatch asynchronously and attach a low-noise watch:

agpair task start \
  --repo-path "$REPO" \
  --executor grok-cli \
  --authorization-profile local_mutating \
  --body "$BRIEF" \
  --no-wait

agpair task watch TASK-123 --json
agpair task wait TASK-123 --json

Each async $BRIEF must be clear enough to identify the goal, scope, allowed changes, and expected evidence. Use the structured shape for mutating work when those boundaries are known.

watch --json emits state changes and raw evidence paths. Do not stream full executor logs into the main Claude context unless the terminal receipt or raw path needs inspection.

wait --json reports outcome, agent_result, recovery_decision, and whether the controller wait lease expired. Treat controller_lease_expired and soft_no_progress as background-running outcomes when recovery_decision.action=wait_background; otherwise follow switch_executor, native_fallback, repair_executor, or retry_same_executor. Inspect task status --json rather than burning model turns in a polling loop.

Workflows

Use agpair workflow start only for high-value multi-part, parallel, adversarial, or long-running work. Workflow manifests are declarative; AGPair rejects arbitrary script fields and creates normal AGPair child tasks.

Workflow ready_for_review means AGPair has an evidence pack for Claude Code verification, not final user-facing success.

Authorization

Pick the narrowest dispatch-time authorization profile that can finish the task:

• local_readonly: inspect-only work.
• local_mutating: normal local edits and tests.
• local_test_heavy: long or heavy local validation.
• external_network: work that needs external network access.

AGPair does not pause a running executor for live approval. If an executor needs more authority, it must return blocked(approval_required).

Blocked Retry

When a task is blocked for approval, do not keep polling. Retry with structured block context:

agpair task retry TASK-123 \
  --from-block \
  --authorization-profile local_mutating

--from-block carries the original brief, blocked reason, terminal receipt, journal tail, git status, diff/commits, and the new authorization profile into a fresh attempt.

If an external attempt fails or is low quality, retry naturally, switch to another external executor, or use Claude Code native subagents as fallback/review.

Review Gate

Treat ready_for_review, evidence_ready, and committed as review gates, not automatic completion.

Before reporting success:

• inspect agpair task status TASK-123 --json;
• inspect changed files, git status, and relevant diff/commit evidence;
• read receipt and raw log paths when the claim is surprising or high-risk;
• run the narrowest meaningful local verification.

Use recovery_decision as the controller-facing next step, agent_result as the evidence quality state, and artifact_result as the evidence map. Follow recovery_decision.action: use_result for reports, review_then_apply for isolated implementation diffs, wait_background for live background tasks, switch_executor for the next external executor, native_fallback for native subagents/direct controller work, repair_executor for auth/binary health, and inspect_evidence when artifacts need manual inspection. Read artifact_result when a result is partial, malformed, or surprising: report/stdout_salvage can still be incorporated, blocked diff must not be applied, and nothing_useful means retry/switch unless a global hard blocker requires inspection. protocol_result and adoption_result remain compatibility/debug surfaces; do not make low-risk protocol warnings override useful evidence.

For isolated implementation or test-fix tasks, review and apply the executor diff explicitly:

agpair task diff TASK-123
agpair task apply TASK-123 --check
agpair task apply TASK-123

task apply leaves changes in the controller worktree for normal Claude Code review and verification. It does not auto-accept the AGPair task.

After verification, close the loop:

agpair task accept TASK-123 --adoptable-result yes --controller-rework none

If the protocol failed but report/stdout evidence is still useful, record explicit salvage. This updates artifact_result and agent_result so the controller can use the report without treating the executor as perfectly protocol-compliant:

agpair task adopt TASK-123 --from-report --adoptable-result partial --controller-rework minor

Claude Code remains accountable for final quality even when AGPair executors did the edits.

Claude Code Integration

Install or print the managed settings snippet:

agpair claude config
agpair claude config --install --scope project --repo-path "$REPO" --sync-skill

Managed hooks:

• UserPromptSubmit: injects external-first routing context.
• SubagentStart: advisory fallback-scope context.
• SubagentStop, TaskCreated, TaskCompleted: observability-only.
• SessionStart and PreCompact: lightweight status/compaction guardrails.

Stop is optional. Pass --include-stop-hook only when you explicitly want a post-answer hard guardrail for actionable AGPair terminal states such as ready_for_review and approval_required.

Hooks fail open when AGPair state is unavailable. They preserve unrelated Claude Code settings and remove only AGPair-managed entries on uninstall.

Claude Code worker auth mode is auto: OAuth/subscription first, then the current Claude provider selected in CC Switch. Probe timeout is not the same as auth failure; check agpair doctor --fresh last_failure_type for executor_probe_timeout or executor_hook_interference before changing credentials.