architrave-review
Skill2 starsChanged 7 days agoMITView on GitHub
Explicitly run a read-only Architrave rubric review of a frozen proposal or implementation and return PASS, REVISE, or FAIL with evidence.
In this repository14 other files it gives its agents.
- AGENTS.mdAGENTS.md
- architrave-agent-authoringSkill
- architrave-ctoSkill
- architrave-mcp-integrationSkill
- architrave-skill-authoringSkill
- architraveSkill
- architrave-tournamentSkill
- architrave-workSkill
- skill-creatorSkill
- subagent-driven-developmentSkill
- systematic-debuggingSkill
- verification-before-completionSkill
- webapp-testingSkill
- writing-skillsSkill
The file
We split this skill into the rules it gives an agent
3 rules. About 0 are common, 3 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 12
Keep private packets in host session files.
- Distinctive: line 23
Prefer the host-native reviewer role where available, using a visible sidebar session for independent proposal/implementation review.
- Distinctive: line 64
Run, receipt or report solely to record a review.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
For a session/run performance review, use docs/session-audit.md: collect
explicitly selected real host sessions (optionally one turn) with
python "<kit-root>\harness\architrave_cli.py" session-audit collect, assess
original goals, quality, tokens and approach using the packet's review prompt,
then save the source-bound advisory analysis with session-audit record for
the existing scoped board. Resolve kit root from this loaded skill (../..).
Keep private packets in host session files. Internal invocation uses the same
path at a task boundary; no extra supervisor, automatic every-turn audit,
instruction changes, permission changes or new completion gate.
The review remains read-only. If the owner explicitly approves applying a
chosen finding, the existing lead applies critical-path recovery from
knowledge/execution-policy.md, using the current goal, exact hold and next
observable action. Reuse the owning session and host messaging, not another
coordinator. Verify subsequent execution separately from sending the correction;
an audit finding or approval never expands existing operation permissions.
Prefer the host-native reviewer role where available, using a visible sidebar
session for independent proposal/implementation review. The existing Judge is
the rubric source/fallback, not an additional duplicate critic. Final development
review uses the lead's two independent sidebar sessions after all code is written;
bounded interim review follows the applicable risk policy. Do not silently
replace requested visible review with a hidden native subagent.
Pass the frozen review subject, original owner messages and attached reference assets,
acceptance criteria, gates/rubric.md, and deterministic evidence.
Require one PASS/REVISE/FAIL verdict with findings
by severity and uncovered requested criteria. Use the rubric's compact output;
do not repeat the contract or score irrelevant dimensions.
Compare reference identity, intended rendered width and component coverage
separately from prose; unavailable sources remain unverified, not satisfied by
an agent-narrowed SPEC. Do not start a recursive program board or redeclare
the parent's scope.
Recheck accepted fixes on the integrated candidate, not an earlier worker patch.
For domain/service changes or application audits, use the existing Service
Architect's evidence and knowledge/backend.md; do not launch another architect.
Compare authoritative rules and current corrections with real mutation paths
and tests, including permitted exceptions, contextual vocabulary, translation
and relevant concurrency. Code and its generated tests can share a wrong rule.
Challenge every alleged violation with a credible valid alternative and
discriminating evidence. Preserve actual ADR status; missing evidence is unknown,
not noncompliance. Keep audit dispositions separate from the overall verdict,
include source/code/scenario/confidence/remedy/falsifying test, and never repair
target code or compute a compliance percentage during read-only review.
Refresh only affected findings and producer/consumer paths when later integration
or dependency changes can invalidate the reviewed behavior. Reuse unchanged
evidence; do not restart a full review after every task or harmless merge.
A worker's fixed claim, passing checks, file preservation or PR publication is
not a reviewer PASS. State the actual reviewed source and remaining findings;
an unavailable reviewer is an explicit review hold, not a coding approval stage.
adversarial-judge (architrave_judge in Codex) stays the rubric source and
the fallback when no native reviewer exists. The Codex role is advisory and
inherits parent skills, MCP, and permission policy.
For an adopted canonical Run, record its result with gate-record --type semantic --reviewer
host-native|architrave-judge --family <model family> --effort <requested>:<mapping>.
For ordinary direct work, return the review in conversation; do not create a
Run, receipt or report solely to record a review. This does not waive required
source-bound admission for an existing Run or authorized external operation.
For R3/R4 request effort: high, mapped only to what the host exposes (auto
tier or reasoning_effort), else none; never name a model. For bounded interim
review one independent review suffices; final development still requires the
lead's two contexts. Only when review.crossFamily is true (building Architrave
itself) does R3/R4 need two different families: get the second through a model
override on the same native reviewer when the host allows it, otherwise through
adversarial-judge. Do not invent a model pin or bypass host selection to obtain
different families. Unavailable cross-family evidence remains a scoped review
hold; the Run runtime rejects duplicate same-family R3/R4 gate admission
(DUPLICATE_REVIEW_FAMILY). Two final contexts are still required by the lead;
two contexts alone do not prove cross-family qualification.
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
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.

