Verification Toolchain Selector

Verification Toolchain Selector

Picks how a project's steps get verified — the test/QA harness — and freezes that decision before the implementation loop starts. The goal is to eliminate the failure mode where agents reach for an inefficient verification tool (slow input→signal, no realtime error console visible to the agent, GUI-only) when a far tighter loop exists.

Canonical example. Expo / React Native app. Efficient: Maestro flows + the simulator on this machine (scriptable, realtime logs the agent can read, short feedback loop). Inefficient: running it through the Xcode GUI and debugging there (slow input, the agent can't see the error console in realtime). The selector must pick the former and forbid the latter.

Selection criterion (the one metric that decides)

Rank candidate harnesses by shortest agent-visible feedback loop:

1. Realtime error/console visibility to the agent — the agent can read failures programmatically (stdout/logs/exit codes), not via a human-only GUI.
2. Latency input→signal — how fast a change produces a pass/fail signal.
3. Headless / scriptable — runs without a GUI; automatable in the loop and in CI.
4. CI compatibility — the same harness runs unattended.
5. Maintenance cost — setup/flakiness/upkeep.

A tool that fails (1) or (3) is disqualified regardless of other merits (this is why Xcode GUI debugging loses).

Mode is the user's choice

When this runs, ask the user: Default or Escalate. Do not auto-pick the mode.

Default mode — curated lookup

Map the stack to its established harness. Confirm with the user, then freeze. Starter table (extend as needed):

If the stack isn't in the table or the user is unsure → recommend Escalate.

Escalate mode — role-structured cross-vendor consensus (HARD GUARDRAILS)

Based on stochastic-multi-agent-consensus, but cross-vendor (different model families = different blind spots) and role-structured (not 30 blind votes — that's weak signal per token for a convergent question). Defaults: 5 agents per vendor (15 total), web search mandatory.

Mechanism (non-negotiable)

• Claude agents → spawned via the Task tool (general-purpose, sonnet).
• Codex agents → via omc ask codex "<prompt>" (CLI; skill nesting unsupported).
• Gemini agents → via agy -p "<prompt>" (the Gemini CLI; -p/--print = non-interactive single prompt). Do not use omc ask gemini or a bare gemini — Gemini is always invoked through agy.
• Codex output lands in .omc/artifacts/ask/; agy prints to stdout — capture that directly. Claude reads both and aggregates.
• If a vendor's CLI is absent → run that vendor with 0 agents and note the missing perspective (never silently treat a 2-vendor run as full consensus).

The pipeline (run in order)

1. Propose — each agent independently proposes a harness for the stack + a one-line rationale. (5×Claude via Task, 5×Codex via omc ask codex, 5×Gemini via agy -p.) Use framing variations (risk-averse, first-principles, CI-focused, DX-focused, contrarian) so same-vendor agents aren't identical.
2. Web-ground — a subset (2–3 per vendor) MUST web-search current best practice (tools change fast; cite year). Findings are shared back so proposals are grounded, not from stale memory.
3. Score — every proposal is scored on the 5 criteria above (each 1–5). This makes aggregation mechanical. Disqualify any option failing criterion 1 or 3.
4. Cross-critique — each vendor critiques the other vendors' top picks (catches vendor bias).
5. Consensus — aggregate: mode (the winner most agents + highest mean score converge on), divergence (genuine judgment calls → surface to user), outlier (creative single-agent ideas → note, don't auto-adopt).

Output

A ranked recommendation with the winning harness, its criteria scores, the dissent, and citations from the web-grounding step. Write the consensus report to .omc/artifacts/verification-toolchain/consensus.md.

HARD GUARDRAILS (apply to BOTH modes)

These are strict — violating any one means the selection is invalid:

1. Decide once, freeze. The chosen harness is written to memory-bank/tech-stack.md under a ## Verification Toolchain section, before the implementation loop starts. It does not change mid-loop.
2. Enforce via AGENTS.md. Add the always-rule: "Always verify via the toolchain defined in memory-bank/tech-stack.md. Never substitute a slower or GUI-only tool (e.g. Xcode GUI debugging) for the defined harness. If the defined tool genuinely cannot run, stop and ask the user — do not improvise a slower path."
3. Criterion is mandatory. The decision MUST be justified against the 5 criteria, with the feedback-loop length explicit. "It's familiar" is not a valid reason.
4. User confirms. Both modes end with the user approving the frozen choice — it's a long-lived decision.
5. Disqualify GUI-only / no-realtime-console tools unless the user explicitly overrides with a reason recorded in tech-stack.md.
6. Escalate ≠ silent partial. If a vendor is unavailable, the report must say so; never present a degraded run as full cross-vendor consensus.

Configuration

{
  "verificationToolchain": {
    "agentsPerVendor": 5,        // Escalate: Claude/Codex/Gemini each
    "webSearchRequired": true,   // Escalate: grounding step is mandatory
    "vendors": ["claude", "codex", "gemini"]  // DeepSeek added later
  }
}

Related

• stochastic-multi-agent-consensus — the base voting mechanism this extends cross-vendor.
• ccg — Codex+Gemini+Claude synthesis; the lighter cousin used for spec review (Phase 0.5).
• vibe-coding — calls this at the required Phase 0.6 (Default vs Escalate; always runs before the loop) and consumes the frozen tech-stack.md toolchain at the step green gate.
• expo-react-native-verification-gate-generator, verification-core-claim-evidence — reuse for mobile/native specifics rather than rederiving.