architecture-checkup — is this repo canonical, and will it actually run?

A structured audit of an agent codebase against the standards that make agent systems maintainable and safe. It does not rewrite anything — it reads, verifies against a fixed rubric, and reports findings ranked by severity, each tied to concrete evidence (file:line). The detailed rubric per axis lives in the sibling references/ files; this body is the procedure and the report shape.

The seven axes (each has a reference rubric)

| Axis | What it verifies | Rubric | |---|---|---| | 1. Manifests & SDD | AGENTS.md (root, agents.md standard), per-module agent manifests, SKILL.md (Anthropic spec), spec-before-code / intent-in-repo. | references/manifests-and-sdd.md | | 2. TDD & test strategy | Test pyramid, golden/eval lanes for LLM surfaces, E2E on every wire contract, one-PR-one-slice. | references/testing.md | | 3. Change-propagation | Every coupled artifact moves together; stale facts (models, ports, dims, hosts) across docs. | delegate to check-drift + the doc-drift hunt below | | 4. Architecture canon | Service boundaries, thin router, MCP domain-vs-capability, tools=MCP / reasoning=agent / instructions=skill / rules=data. | references/architecture-canon.md | | 5. Security doctrine | Untrusted-input-as-data, outbound confirm-gate, MCP trust boundary, OWASP LLM + MCP Top-10. | references/architecture-canon.md | | 6. Runtime & hardware fit | Model sizing vs RAM/GPU ceiling, embedding-model↔dim↔column consistency, resource caps, model-swap safety. | references/runtime-fit.md | | 7. Secrets & config hygiene | Secrets actually untracked (ask git, not .gitignore), template↔code parity, scanning/push-protection on public repos, deliberate visibility + licence. Pairs with axis 5. | references/secrets-and-config.md |

Procedure

Copy this checklist into your working notes and tick each axis as you finish it:

Checkup:
- [ ] 0. Scope: whole repo, or a diff? Which target hardware?
- [ ] 1. Manifests & SDD    (read references/manifests-and-sdd.md)
- [ ] 2. TDD & test strategy (read references/testing.md)
- [ ] 3. Change-propagation  (run check-drift + doc-drift hunt)
- [ ] 4. Architecture canon  (read references/architecture-canon.md)
- [ ] 5. Security doctrine   (same file, §Security)
- [ ] 6. Runtime & hardware  (read references/runtime-fit.md)
- [ ] 7. Secrets & config    (read references/secrets-and-config.md)
- [ ] 8. Compile report

1. Scope. Establish what is under audit (a whole repo vs a single change) and

the target deployment (host RAM/GPU, local vs cloud models). Runtime-fit findings are meaningless without a stated hardware target — get it first.

1. Per axis, read its rubric file, then verify against the real repo. Open

the actual manifests, test dirs, config/env, migrations — do not judge from the docs' claims; judge from what the code and config do. Every finding must cite a real file:line, never a hunch.

1. Change-propagation (axis 3). Run the check-drift skill if present. Then

do the doc-drift hunt it cannot: grep the whole repo for facts that live in more than one place — model tags, ports, embedding dimensions, host sizes, RAM ceilings — and flag any value that disagrees with the source-of-truth file. Note whether the repo's change-map.yaml even has a coupling for each drifted fact; a missing coupling is itself a finding (the automat has a hole).

1. Runtime & hardware fit (axis 6). This is where "looks fine on paper" meets

"won't boot". Apply references/runtime-fit.md literally — especially the embedding-dimension chain (model output dim ↔ config dim ↔ DB vector(N) column) and the model-memory budget vs the host ceiling.

1. Compile the report (shape below). Rank by severity, not by axis order.

Output — the report

For each finding: severity · one-line claim · evidence (file:line) · why it matters · fix. Then one prioritized recommendation table. Severity legend:

• 🔴 Blocker — will fail at build/boot/runtime, or a security hole. (e.g. a

vector(384) column fed 768-dim embeddings → every recall insert rejected.)

• 🟠 Drift — an artifact that did not move with its coupled change; misleads

future work but does not crash today.

• 🟡 Risk — load-bearing assumption with no guard (e.g. a model swap that

needs a clean unload but has no test).

• 🟢 Hygiene — naming, an over-claimed "compatible", a stale comment.

End with: verdict (one line) + the prioritized fix table (# · severity · action). If an axis is clean, say so in one line — do not invent findings to fill it.

Triggering

SHOULD fire: "сделай архитектурный чекап проекта"; "проверь, каноничная ли архитектура и заработает ли на этом железе"; "audit the repo against agents.md / SDD / the skill spec before the milestone".

SHOULD NOT fire: "fix this drift" (that is check-drift + edits); "scaffold a new module" (new-module); "write a SKILL.md" (new-skill). This skill finds, it does not fix — hand its findings to those skills.

Extending

A new standard to check = a new references/<axis>.md + one row in the axes table above. Keep the rubric in the reference file, not in this body — same discipline as check-drift keeping couplings in change-map.yaml.