phone-a-friend
Use when the user wants a second opinion on a plan or code change from the other CLI, Claude Code or Codex.
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-14
- Stars
- ★ 1.6k
- Forks
- 491
01 前言 Front Matter
- name
- phone-a-friend
- effort
- high
- description
- Use when the user wants a second opinion on a plan or code change from the other CLI, Claude Code or Codex.
- argument-hint
- [plan path | review scope]
- user-invocable
- true
- disable-model-invocation
- true
02 资源文件 1
- 参考文档agents/openai.yaml351 B
03 SKILL.md 正文
Phone a friend
Get a second opinion from the other model family before sharing the first reviewer's conclusions. Use one rebuttal round to resolve findings with evidence. Escalate material disagreements that remain; never run a third round.
Scope and partner
Plan mode reviews the given plan path, else the active task plan. Review mode reviews the requested scope, else the current branch changes. Ambiguous → ask.
If PHONE_A_FRIEND=1 is set, refuse: this session already is the partner.
Detect the host and call the other family, without permission prompts (the partner runs unattended):
- From Claude Code:
PHONE_A_FRIEND=1 codex exec -c model_reasoning_effort="xhigh" --sandbox read-only --cd <repo> --json --output-last-message <file> "<packet>", then the rebuttal with the same env guard and flags before the subcommand:PHONE_A_FRIEND=1 codex exec -c model_reasoning_effort="xhigh" --sandbox read-only --cd <repo> resume <session-id> "<rebuttal>"(session id from thethread.startedJSON event;resumeitself accepts no sandbox flag).execnever prompts; the read-only sandbox blocks writes instead of asking. Codex uses the model configured in~/.codex/config.toml; add-m <model>only when the user names one. - From Codex:
PHONE_A_FRIEND=1 claude -p --model opus --effort xhigh --dangerously-skip-permissions --output-format json "<packet>"(theopusalias selects the latest Opus; the result JSON carries thesession_id; plain text output does not), then the same flags with--resume <session-id>for the rebuttal.
The partner keeps its own MCP servers: the same read access to code, docs, and data as the host. If the partner CLI is missing or errors, report it and stop — a same-family review is not a second opinion.
Neutral packet
Send: the user's requirements verbatim; the artifact (plan text or diff); paths of referenced files and applicable conventions. Send the artifact and requirements without the host's conclusions, so the reviewer can make an independent assessment. Keep every working file (packet, verdicts, event logs) outside the reviewed repository, in a temp directory: a reviewer that finds another reviewer's verdict on disk is no longer independent.
State these rules to the partner:
- Material findings only; zero findings is a valid outcome; most important first.
- Every finding: location · falsifiable claim · evidence · impact · falsifier ("what would change my mind") · likelihood and confidence as two separate fields. No location or no evidence → not a finding. Reasoning is not evidence.
- Strictly read-only: no commit, merge, or push; no file writes; no writes to any database; no calls to endpoints that write or change state on any production service. Reads are fine everywhere — code, docs, services, logs, and databases alike.
- Missing context: say what is missing; do not guess.
- Plan mode only: a contested section gets a rewrite mandate (redo this, because…), not accept/reject.
Review mode: two independent reviewers
Also dispatch the clean-context development-skills:staff-reviewer (reuse its verdict if this session already produced one for the same diff). The reviews must not see each other. Merge by location and claim: cross-validated / external-only / host-only.
One rebuttal round
Answer each finding once, as the author: accept (update the plan or fix list) · counter with new evidence · uncertain. The partner closes each point on session resume: confirmed / withdrawn / uncertain. A changed verdict without new evidence remains unsupported; record that evidence gap.
Close and escalate
Severity follows the review categories, derived from impact and reversibility — never taken from a model's own label. A "critical" without location and evidence is a red flag, not a verdict. Accepted findings follow the normal loop's authority; withdrawn ones leave the report but stay in the record with the dissent verbatim.
Escalate to the user through the development loop's platform input behavior every question the protocol could not settle — no more, no fewer — each carrying both positions, the evidence each side holds, and the cost of each choice:
- still disputed or uncertain at material severity after the rebuttal;
- costly to reverse (migration, published contract, auth, data deletion) — even when the models agree;
- the disagreement is about what the work should achieve, not how.
Record
Report the full outcome in chat: findings table (location, claim, source tag, resolution), kept dissent, escalated decisions. Record the same in the active task chronicle. With no active chronicle, still report everything in chat and write it to a dedicated file in the project's docs/ (its chronicle convention if present, else docs/phone-a-friend__YYYY-MM-DD__<slug>.md).
04 同仓库更多 Skills
worktable-review-with-annotations
Create, inspect, reply to, update, or resolve situated Worktable annotations. Use when feedback belongs to a Doc section, block, text range, or HTML Doc, or when the user asks to address existing review feedback and coordinate a requested artifact edit.
worktable-find-and-synthesize-context
Find, assess, and synthesize context from Worktable Docs and Records. Use when the user asks what a workspace contains, wants workspace-backed research, needs current facts reconciled across artifacts, or requests a cited synthesis without changing workspace state.
worktable-create-or-update-docs
Create, edit, rename, or maintain narrative Worktable Docs. Use for plans, decisions, briefs, notes, research syntheses, and other durable prose when the user requests a new artifact or a safe update to an existing one.
worktable-create-or-manage-records
Organize repeated structured information as durable Worktable Records. Use for research datasets, inventories, directories, pipelines, trackers, and other collections whose items need identity, fields, filtering, links, validation, or individual updates.
worktable-create-or-manage-interactive-html
Create or improve durable visual and interactive artifacts in Worktable. Use for dashboards, trackers, calculators, mini apps, slide presentations, and other requests that benefit from a polished HTML Doc.
worktable-collaborate-threads
Read, answer, and hand off messages through durable Worktable threads with users or other authenticated participants. Use when the user asks to check addressed messages, reply to a participant, continue a thread, request another agent's help, wait briefly for a response, or watch for comments and messages sent to the agent.