Guardrails — Uber Report

A deliberately lightweight adaptation of the house uber-report pattern. This repo is small and single-track: don't pad — a short report on a small repo is correct.

The Reality Gate (fill from EVIDENCE, first, always)

Until all three pass, the verdict is hard-capped at "walking skeleton not proven" — no doc counts, no milestone counts in the summary.

Procedure

1. Prior context: glob .claude/tasks/*.findings.json; carry forward open slugs; mark satisfied ones completed (confirm by inspection).
2. WIP survey (brief): git -C <repo> status --short, log --oneline -15, unmerged branches. Single-session repo — this is usually one line, but say it.
3. Reality Gate (above).
4. Two-axis status:
   • Harness vs docs/plans/03-roadmap.md milestones (M1–M7, exit-criteria-based — Built ≠ Verified; check the domain-knowledge Status section, then verify claims).
   • Skill maturity, per skill: missing / drafted / validated-against-example / battle-tested-on-real-plan.
5. Slim dispatch (read-only, parallel; skip any with nothing to assess): guardrails-harness-developer — code health of src/** (TODO/dead code/oversized, seam violations, test gaps); guardrails-test-author — suite health, passing-but-blind hunts; guardrails-architect — plan-doc drift vs as-built; guardrails-devils-advocate — risk register against the seeded risks in 03-roadmap.md (every seeded row appears, N/A'd if inapplicable). Each proposes ≤ 5 findings, cited path:line, falsifiable.

6. Report → .claude/tasks/YYYY-MM-DD-uber-report-vN.md (never overwrite a prior version):

   # Guardrails — Status & Readiness
   **Date / HEAD / WIP one-liner**
   ## Reality Gate            (the table, with evidence)
   ## Executive summary       (3–5 sentences; no counts-as-progress)
   ## Milestones M1–M7        (exit-criterion verdicts, evidence cited)
   ## Skill maturity
   ## Findings                (merged, deduped, ranked; ≤ ~10)
   ## Risk register           (seeded rows from 03-roadmap.md + run-specific)
   ## Proposed next actions   (slug-keyed)

7. Findings index → same basename .findings.json: { "sourceOfTruth": "plan-docs", "findings": [ { slug, kind, priority, effort, route, status, dependsOn, notes } ] } per finding; carry-forward joins on slug. route ∈ direct | execute-tasks | guardrails | decide (see Routing below).

Routing

Each finding carries a route the orchestrator proposes and the human confirms — most findings are small and never need a plan:

• direct — trivial, self-evidently correct (rename a test, fix a typo). Just do it.
• execute-tasks — standard work an agent implements under normal review.
• guardrails — plan-worthy (multi-step, touches a contract/invariant, needs authored tests). Graduates to this repo's own /plan-breakdown → guardrails-review → guardrails run.
• decide — needs a human call or a research pass before it's actionable.

Honesty rules

• Built ≠ verified-running: milestone verdicts come from exit criteria demonstrated, not code existing.
• Findings are falsifiable (path:line or a named contract) — "looks fine" is not a finding.
• The Reality Gate's token-spending check is either RUN or marked UNVERIFIED. Never inferred.