feat(plan-check): pre-implementation plan review agent + skill - #3
Merged
Conversation
… skill plan-reviewer agent (fresh-eyes, findings-only) plus the plan-review skill: 8 dimensions distilled from real vault-cortex planning failures — premise audit, alternatives + do-nothing baseline, guard/control arithmetic (the OAuth client-cap denial-vector case), concurrent-writer analysis, post- incident overcompensation check, verification blast radius. Upstream counterpart to ship-check: it reviews premises before code exists; ship-check reviews mechanisms after. Registered in marketplace.json, root README table, and AGENTS.md structure. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
![umm-actually[bot]](https://avatars.githubusercontent.com/in/4233716?s=60&v=4){kind=small}
|
umm-actually re-reviewed at 1 new finding(s) posted (4 tracked finding(s) across all runs). Context notes
umm-actually · deepseek/deepseek-v4-flash-0731 |
…lass Step 6 said dimensions 4 and 6a are both conditional on the change classification, but 6a's own trigger is whether the plan responds to a recent incident — orthogonal to the five classes. A post-incident plan classified 'feature' would have skipped the overcompensation check. Reworded step 6 and the agent orientation to bind 6a to the incident property. (umm-actually finding on PR #3.) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The plan-review skill instructs the agent to settle checkable-now
premises itself ("a read or grep settles it"), but the tool allowlist
omitted Grep and Glob — the same gap fixed for the ship-check phase
agents in #4.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
SKILL.md frontmatter does not recognize a tools: key (that field belongs to agent definitions), so the declaration was inert. allowed-tools is the recognized skill field — it pre-approves the listed tool for inline invocation without restricting anything. Applied across plan-review and the five ship-check skills carrying the same key. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
AGENTS.md trigger structure requires action + condition + boundary (or an explicit always-applies). Three flags had neither: D1 no-one-named (now shares the mechanical-task carve-out), D2 sourceless constraint (trace it yourself when one search settles it), and D2 observation-promoted-to- requirement (explicit reasoned preservation is fine; the flag is for silent conversion). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Dimensions use Flag/Block verbs while the output contract counts blockers / must-answer / recommendations, with only two dimensions mapping between them explicitly. Added the mapping rule at the top of the Dimensions section: Block is always a blocker; Flag resolves via the guess test unless the dimension assigns severity itself. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… set The premise audit reads task notes, boards, backlinks, and linked research — the four-tool subset forced Bash workarounds for graph, listing, and task queries. Adds the 20 read-only vault-cortex tools under both server prefixes; write tools stay excluded (findings only). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
New
plan-checkplugin — the upstream counterpart to ship-check. One agent (plan-reviewer, fresh-eyes, findings-only) plus one skill (plan-review) that critiques implementation plans before any code exists.Why
Premise flaws survive post-implementation review by construction: the code faithfully implements whatever the plan assumed, and mechanism review pronounces it well-built. Founding case: the OAuth client-registration cap passed all four ship-check phases; only manual review computed that one address under the rate limit fills the cap in ~75 minutes and locks the owner out.
Design
governance-roadmap-pr), produced 2 must-answer + 3 recommendation findings with quoted premises, and returned execution feedback — all five feedback items (verification-state pinning, docs two-sources-of-truth variant, checkable-now stop condition, must-answer/recommendation boundary, mandatory verification-basis line) are folded into this commit.Registered in marketplace.json, root README table, and AGENTS.md structure tree. Version pinned at 1.0.3 for lockstep; next release bumps all fields together.
🤖 Generated with Claude Code