os-sdlc: reviewer is a named agent with hook-assembled context packet

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VgG2g9ooqbnBiZYGAonGYG
This commit is contained in:
jared 2026-07-16 16:18:28 -04:00
parent 4cc524b3a5
commit 5c18ba696f
1 changed files with 12 additions and 3 deletions

View File

@ -25,9 +25,18 @@ Manually triggered by `/implement <ticket>`. Stages, in order:
5. **Green-assert gate** (deterministic hook) — suite + lint must pass. Failures are
auto-fed back into the same programmer's next turn; a max-iterations cap escalates to the
human rather than looping forever.
6. **Reviewer agent** — two ordered checks: first tests align with the spec/ticket, then code
aligns with both tests and spec. Order matters — misaligned tests invalidate a "code passes
tests" result.
6. **Reviewer agent** — a *named agent* (not a skill), so its tool surface is minimal and
its context never leaks into the orchestrator's (see
[agent-design-principles](agent-design-principles.md)). It reads only a hook-assembled
context packet — test status from the green-assert gate (passing, unless max-iterations
escalated), a mechanically generated diff, and the spec/ticket — gathered into one
document by a code-actor hook before the agent is invoked; the reviewer runs nothing
itself. Two ordered checks: first tests align with the spec/ticket, then code aligns
with both tests and spec. Order matters — misaligned tests invalidate a "code passes
tests" result. ADR-0037's `/os-sdlc:review` (standards-conformance + spec-fidelity +
Fowler refactor smells) is satisfied by this same agent; if standalone review of an
arbitrary diff is wanted outside a pipeline run, a thin skill wrapper dispatches the
agent with a hand-assembled packet — one reviewer, two entry points.
7. **Cleanup & report** (agent/code mix) — trims scratch artifacts, produces a summary for the
human. What gets written, where, and for how long is defined in
[pipeline-observability](pipeline-observability.md).