agentleFS
augments-labsaugments-labs/sdlc-skills/skills/design/reviewing-plans/SKILL.md

reviewing-plans

Skill1 starsChanged 3 months agoMITView on GitHub

Gets an independent review of one exact plan version before the plan is presented for approval. Use when a written plan's route is high-risk and it is about to be shown for approval, or when the user says review this plan, is the plan sound, or check the plan before I approve it. Skip a plan whose files are not written yet, and skip editing the plan.

7 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

10 rules. 0 are common, 3 are distinctive, and 7 are very specific to this project.

  • 0

    Common

    Most of the 5 edited versions we read keep these.

  • 3

    Distinctive

    This author's own choices. Worth a look.

  • 7

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    If the caller is writing-plans presenting a plan whose route is not high-risk, ask the offer question below.

    line 25
  2. Very specific:

    Ask it with the version plan-version.sh prints:

    line 28
  3. Very specific:

    Ask this and end the turn.

    line 39
  4. Distinctive:

    If the answer is option 1, go to Step 1.

    line 40
  5. Very specific:

    If a predecessor version already has a review, bind that report too.

    line 60
  6. Very specific:

    If the action is unavailable, refused, or comes back empty, write not dispatched and ask the not-dispatched question from …

    line 75
  7. Distinctive:

    Read the report at the current attempt's location, opening the file when only its location came back and writing it there first when it …

    line 94
  8. Very specific:

    If a blocker is accepted, return to the caller's pending step.

    line 123
  9. Distinctive:

    If no blocker is open, return to the caller, or to the user who asked.

    line 127
  10. Very specific:

    Never re-enter writing-plans or start an executor from here.

    line 131

“Common” is measured: kept by at least 6 in 10 of the 5 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Reviewing Plans

This skill never edits the plan, approves it, or starts an executor.

When to use

  • A plan directory is written and about to be presented for approval, and its classification route is high-risk: writing-plans requires this review.
  • The user asks for an independent look at a plan, or accepts the review that writing-plans offers for a plan that is not high-risk.
  • An executor proposes a high-risk successor mid-run (Step 1.6).
  • Skip a plan with no printed version yet: its index or task files are not written.
  • Skip editing the plan. Corrections belong to the plan's author, as a successor version.
  • Skip a frozen code change; requesting-code-review reviews that.

Step 0: The offer, for a plan that is not high-risk

  1. If the caller is writing-plans presenting a plan whose route is not high-risk, ask the offer question below. If the user asked for the review directly, or the route is high-risk, skip Step 0 and go to Step 1.
  2. Ask it with the version plan-version.sh prints:

    Plan {{version}} is not high-risk.
    
    1. Review it independently before presenting it
    2. Present it without a review
    
    Recommendation: {{option}} — {{one sentence}}.

    Ask this and end the turn; clarifying-intent owns what closes it.

  3. If the answer is option 1, go to Step 1. If it is option 2, return to the caller, which presents the plan without a review.

Step 1: Bind the version, the inputs, and the reviewer

  1. Print the plan's version with the plan-version.sh script that writing-plans ships, run on the plan directory. If the script fails, name its error and stop; a version read from the index is not a version.
  2. Bind every approved input the index's Bound inputs line names, at the exact identity it names. An input missing at that identity is a finding for the author; never review against a newer one instead.
  3. Bind one reviewer role ID that did not write this plan. Pick its tier with the Model selection section of dispatching-parallel-agents.
  4. Bind the review boundary: read-only access to the plan directory, its inputs, and the repository at the Bound inputs code revision, the worker, provider, storage, and egress authority the user or project already gave, a deadline, and who cancels. If there is no such authority, the review is not dispatched (Step 2.3).
  5. Bind the report location in the plan's evidence directory: .sdlc-skills/evidence/{{plan directory name}}/00-index.review-{{version}}-{{attempt}}.md, attempt 1 first. It is the only path that attempt's reviewer may write.
  6. If a predecessor version already has a review, bind that report too; the new review reads the successor's delta (its Successor delta line, the tasks it adds or changes) and every section that depends on them, nothing else, and carries the earlier recorded decisions forward. An executor proposing a successor mid-run comes here the same way: the delta is reviewed, the approved remainder is not reread.

Step 2: Dispatch the review and keep its receipt

  1. Open assets/plan-reviewer.md after Step 1 is bound. Fill every Inputs slot with the paths and identities from Step 1, never a paraphrase.
  2. Send the filled prompt once through the harness's real dispatch action, as dispatching-parallel-agents Step 2 says. Record the tool-issued ID as the dispatch receipt, with the plan version it binds. A name or a prompt is not a dispatch.
  3. If the action is unavailable, refused, or comes back empty, write not dispatched and ask the not-dispatched question from dispatching-parallel-agents Step 2.2. A high-risk route requires an independent reviewer: omit its self-review option and say why. With no answer, the review stays pending.
  4. Poll that receipt until the deadline. Never say a review is running without a receipt.
  5. On a failure, a timeout, or a cancellation, the review stays pending. This skill may retry once, through a new receipt linked to the failed one and a report location with the next attempt number, after correcting the cause that attempt named; a late result from the failed attempt, or a file at its path, is rejected. On a second failure, return the pending review and both receipts to the caller.
  6. NEEDS_CONTEXT is not a failure and does not use that retry; Step 3.1 handles it. Its rerun happens once, at the next attempt number; a second NEEDS_CONTEXT returns to the caller as blocked, with its question.

Step 3: Receive the report, decide each finding, return

  1. Read the report at the current attempt's location, opening the file when only its location came back and writing it there first when it came back inline. If its Plan version differs from the bound version, or the report is missing or unreadable, the review stays pending. If the reviewer returned NEEDS_CONTEXT with a question, the review stays pending; return that exact question to the plan's author, at the caller's pending step (writing-plans Step 3.11, or the executor step that proposed the successor). A high-risk plan is not presented until the question is answered and the review is run again on the answered version.
  2. Print the plan's version again. If it changed during the review, the review is void; report the edit to the caller and keep the review pending.
  3. Give every finding one recorded decision (the report calls these its dispositions):
    • accepted: the author corrects it in a successor version.
    • rejected: record the path and symbol or the command result that refutes it; an argument alone makes it a decision.
    • decision: the index's Approval rule owner decides it, at presentation.
    • A finding that cites no task ID, section, or path:line: record it as not a finding: uncited.
  4. The report is the review record. Write into its header the dispatch receipt, the report path, every recorded decision, who decided each, and the date. Write no row in the plan's decision ledger: the plan's presentation and its later approval row cite the report path in Bound evidence.
  5. Return to the caller's pending step with the report path and its recorded decisions:
    • If a blocker is accepted, return to the caller's pending step; the correction is a proposed successor, and a high-risk successor comes back here before it is presented. That re-review is the last: a blocker it still raises is recorded as a decision.
    • If no blocker is open, return to the caller, or to the user who asked.
    • On a high-risk route, the caller presents nothing until this review has returned and every blocker has a recorded decision; a pending or void review holds the presentation.
  6. Never re-enter writing-plans or start an executor from here.

Gotchas

  • A review that reads only the plan and its inputs checks the plan against itself: nine such reviews passed a plan whose tasks could not be built as written on the base they were bound to. The Buildability lens opens the code, and a reviewer given no code revision returns NEEDS_CONTEXT.
  • Self-review is not this review. The author's own checks in writing-plans Step 3 read the plan with the assumptions that wrote it; counting them as the independent review presents a high-risk plan no second reader saw.
  • A reviewer that edits the plan voids the review: the version it read no longer exists, and its findings describe text nobody approved. Step 3.2 catches it by printing the version again.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.