feat(plan-eng-review): add Step 0.5 — existing-capability check before new substrate code

[anoopkansupada]

What

Adds a Step 0.5: Existing-capability check to /plan-eng-review, between the Step 0 Scope Challenge and Section 1 (Architecture review).

When a plan proposes new code under ops/<substrate>/ (gbrain, OpenClaw, Hermes, or similar recipe-driven substrates), the reviewer must first run a 60-second check:

ssh <host> 'gbrain integrations list && launchctl list | grep -i gbrain'

Recipe-driven substrates ship a senses + reflexes system that IS the signal-detector / autopilot / ingest pattern. Most "wire X into the flow" plans should be configuring an existing recipe (or writing a markdown recipe modeled on an existing one), NOT building new infrastructure. The new step codifies the three decision branches: configure if AVAILABLE, investigate if CONFIGURED-but-dormant, write a markdown recipe if no recipe exists, build new code only if all three fail.

Why

Real incident, 2026-05-17 on a downstream user's hash-lemma project: /plan-eng-review accepted a plan to extract a JSONL queue lib + producer + consumer bin + LaunchAgent + OAuth-DCR migration to wire "signal-detector on every Slack message" — when gbrain's integrations system (11 senses + 1 reflex) already shipped exactly that pattern, recipe-based. The user's instinct caught it ("this feels like a wrong buildout vs the garry way"), but the eng-review didn't.

Garry's own writings in docs/ethos/thin_harness_fat_skills and docs/ethos/markdown_skills_as_recipes explicitly name this as the anti-pattern: "fat harness with thin skills," "REST API wrappers as tools," "push intelligence UP to skills, execution DOWN to deterministic tooling." The 60-second check operationalizes the rule at the review-skill level.

Net savings (this case): ~3-5 days of misdirected code, replaced with one markdown recipe (slack-to-brain) submitted upstream separately.

Coverage

• Substrates this affects: gbrain, OpenClaw, Hermes (recipe-driven agent / knowledge-graph runtimes)
• Substrates this skips: pure web apps, libraries, CLIs — no capability surface to check against

Test plan

• bun run gen:skill-docs regenerates SKILL.md without errors
• Diff confirms Step 0.5 inserted between Step 0 and Section 1 in both .tmpl and .md
• Manual: run /plan-eng-review on a plan touching ops/gbrain — verify Step 0.5 fires and the 60-second check shows in the review output
• Validation: confirm Substep doesn't fire on a non-substrate plan (e.g. a pure web-app feature)

🤖 Generated with Claude Code