菜单
Check whether the plan(s)/PRD(s)/issue-slice(s) under discussion still match the decisions made in this session — a sonnet subagent reads the current session log, later decisions win, and reports drift read-only. Use for "/verify-plan", "does my plan still match what we decided?".
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | verify-plan |
| description | Check whether the plan(s)/PRD(s)/issue-slice(s) under discussion still match the decisions made in this session — a sonnet subagent reads the current session log, later decisions win, and reports drift read-only. Use for "/verify-plan", "does my plan still match what we decided?". |
| argument-hint | [optional: which plan/PRD/issue to check; empty = infer from the conversation] |
| model | inherit |
| allowed-tools | Read, Grep, Glob, Bash, Agent |
Check whether the written artifacts under discussion still match the decisions made in this session. A fresh sonnet subagent reads the session transcript and each target, then reports contradictions and omissions. Read-only — nothing is edited.
Run this in a Bash block. The key is sha1 of the canonical --git-common-dir — identical
from the primary checkout and every linked worktree, so the stash survives EnterWorktree
(the hook keys the same way):
gcd="$(git rev-parse --git-common-dir 2>/dev/null)"
if [ -n "$gcd" ]; then root="$(cd "$gcd" && pwd -P)"; else root="$(pwd)"; fi
key="$(printf %s "$root" | sha1sum | cut -c1-16)"
stash="${TMPDIR:-/tmp}/verify-plan-session-$key.path"
log="$([ -s "$stash" ] && cat "$stash" | tr -d '\n')"
# Stale-path fallback: entering a worktree moves the transcript to another
# ~/.claude/projects/<slug>/ dir — relocate it by session id (the filename).
if [ -n "$log" ] && [ ! -s "$log" ]; then
log="$(ls -t "$HOME"/.claude/projects/*/"$(basename "$log")" 2>/dev/null | head -1)"
fi
printf '%s\n' "$log"
If $stash is missing or empty, or $log resolves to no readable file: hard error,
stop. Tell the user:
The session pointer isn't set yet — this usually means the hook hasn't fired yet. Run
/reload-plugins, submit one prompt (anything), then re-run/verify-plan.
Do nothing else until the user addresses this.
Using the $log path resolved in Step 1, measure its size:
bytes="$(wc -c < "$log")"
marker="${TMPDIR:-/tmp}/verify-plan-oversize-$key"
bytes ≤ 600000 (≈ 150k tokens): remove any stale marker (rm -f "$marker"),
proceed to Step 3.bytes > 600000 and $marker absent: write the marker (touch "$marker"),
then hard-stop. Report the measured size (e.g. "session log is ~410k bytes, over the
~150k-token cap") and tell the user that re-running /verify-plan once more will
override the guard and proceed.bytes > 600000 and $marker present: remove the marker (rm -f "$marker"),
then proceed to Step 3 — this is the override run.From the conversation, gather the concrete refs of the plan/PRD/issue-slices in play:
PLAN.md, a handoff-plan file under ~/.claude/plans/, a PRD
markdown file in the repo)prd-labeled tracking issue and/or its slices, read
live with gh issue view <n>)$ARGUMENTS, if provided, narrows or overrides what to check. Pass refs, not pasted
content — the subagent reads each source live so the review is never against a stale
paraphrase. If you can't identify any target from the conversation and no $ARGUMENTS
were given, ask the user which artifact to check rather than guess.
Call the Agent tool with subagent_type: "general-purpose" and model: "sonnet".
(There is no effort parameter on Agent — don't try to pass one.)
Keep the prompt short. A terse prompt is both far faster (a verbose, heavily structured prompt makes the subagent read the log in tiny chunks and crawl — ~10× slower) and sharper. Use roughly this, filling in the refs from Step 3 and the path from Step 1:
Read the target(s)
<refs assembled in Step 3 — file paths;gh issue viewfor GitHub issues>and the session log<resolved transcript path>(JSONL of the conversation that produced them).Does each target match the decisions made in the session? Later decisions override earlier ones. Report any contradictions or omissions, read-only. Lead with
VERDICT: ALIGNEDorVERDICT: MISMATCHES (n).
Relay the subagent's output verbatim to the user. Do not summarize, filter, or editorialize.
Run one task through the plan→build→review pipeline — a Step-0.5 classify-task call routes the planner/implementer/reviewer models to the task's complexity tier, then the plan is written (planner subagent or inline), the implementer builds it in an isolated worktree, the my-review agent reviews the diff, and findings route by severity (low→issues, medium→triaged fix-list, high→collective replan, critical→own cycle). For single tasks not worth slicing into an issue graph. Use for "/pipeline <issue#|task>", "pipeline this".
Deep, security-weighted review of your changes or a PR — routes the reviewer model to the change's complexity (opus by default; fable for a complex diff, on ask), then spawns the my-review agent (xhigh reasoning) on that model. Read-only, report-only. Use for "/my-review", "/my-review <PR#>".
Classify one task or issue into a complexity tier — trivial, standard, or complex — and emit that tier plus a rationale; consumers resolve the tier's {model, effort} roster via the plugin's resolve-tier.sh. Grounds the call by fanning out 1–3 Explore subagents over the touched codebase areas, then asks you to confirm or override. Use for "/classify-task <issue#|brief>", "classify this task".
Run N rounds of the autonomous issue-solving loop — pick the ready set (blockers closed, skip hitl), fan out parallel implementers in isolated git worktrees, hand the completed branches to a merger that merges in dependency order and resolves conflicts under the done-check, close finished issues, then a reviewer files blocking follow-ups. Use for "/orchestrate", "run the loop", "build the ready issues".
Capture the just-approved plan after exiting plan mode and hand it to a fresh session — write the plan to a file plus the resume pointer the workflow plugin reads, then tell me to /clear and send `go` to execute it. Use for "/handoff-plan", "hand off this plan".
Capture a rich handoff before /clear — write a markdown handoff doc (work done, in-flight state, next steps, key files, gotchas) plus the resume pointer the workflow plugin reads, then tell me to /clear and send `go`. Use for "/handoff", "hand this off", "save state and clear".