Problem
The wiki is consulted at plan time only. skills/wiki-plan/SKILL.md and
skills/wiki-query/SKILL.md contain no occurrence of "review", and INDEX.md
opens with:
Route by matching your current task to a "route here when" line
Every review surface in dev-loop therefore reviews a diff with none of the
accumulated failure knowledge routed to it:
| Surface | Reviewer | What it actually checks |
|---|
loop-implement step 6 | the session that wrote the code | bugs/edges/leaks + only the pages the plan named |
loop-implement step 6.5 | independent (pinned model) | test quality only |
orchestrate Phase 4 (SKILL.md:479-482) | coordinator | one sentence: "Review each worktree diff" |
Proposal
Add a review-time routing pass, symmetric to wiki-plan's decision→page map but
with a different input:
wiki-plan routes from the intent ("what we set out to build").- Review must route from what the code actually does — files touched, APIs and
flags introduced, and the concurrency / IO / schema / auth / process-boundary
signals visible in the diff.
Then load those pages and check the diff against their failure-mode and edge-case
rows.
The delta is the point. Pages the reviewer routes to that the plan never named
are the unplanned risk surface. If review simply re-reads the plan's decision→page
map, that delta is invisible by construction — which is the current behaviour.
Why this specific gap
From "Reviewing code is a skill"
(read 2026-08-13): the three defects a frontier LLM reviewer missed all required
context outside the diff — a remembered prior incident, the tool version on the
machine that actually runs the code, and cross-object failure ordering. In
dev-loop, the wiki is precisely the container for that context, and review is the
one phase that never opens it.
Scope
Evidence status
Gap is verified by inspection (grep -n "review" skills/wiki-plan/SKILL.md skills/wiki-query/SKILL.md → no matches). The proposed remedy is a design
proposal, not a measured improvement.
Source
Problem
The wiki is consulted at plan time only.
skills/wiki-plan/SKILL.mdandskills/wiki-query/SKILL.mdcontain no occurrence of "review", andINDEX.mdopens with:
Every review surface in dev-loop therefore reviews a diff with none of the
accumulated failure knowledge routed to it:
loop-implementstep 6loop-implementstep 6.5orchestratePhase 4 (SKILL.md:479-482)Proposal
Add a review-time routing pass, symmetric to
wiki-plan's decision→page map butwith a different input:
wiki-planroutes from the intent ("what we set out to build").flags introduced, and the concurrency / IO / schema / auth / process-boundary
signals visible in the diff.
Then load those pages and check the diff against their failure-mode and edge-case
rows.
The delta is the point. Pages the reviewer routes to that the plan never named
are the unplanned risk surface. If review simply re-reads the plan's decision→page
map, that delta is invisible by construction — which is the current behaviour.
Why this specific gap
From "Reviewing code is a skill"
(read 2026-08-13): the three defects a frontier LLM reviewer missed all required
context outside the diff — a remembered prior incident, the tool version on the
machine that actually runs the code, and cross-object failure ordering. In
dev-loop, the wiki is precisely the container for that context, and review is the
one phase that never opens it.
Scope
AGENTS.mdrouting protocol — add the diff-as-input route.INDEX.mdpreamble — "current task" is not the only entry point.Evidence status
Gap is verified by inspection (
grep -n "review" skills/wiki-plan/SKILL.md skills/wiki-query/SKILL.md→ no matches). The proposed remedy is a designproposal, not a measured improvement.
Source