multi-agent-consensus

Generate a report from the last consensus pipeline run (reads state.json + evidence manifest).

Show the three agent role definitions (Lead, Alpha, Bravo) and what each is calibrated to catch.

Run the full 3-agent consensus validation pipeline with unanimous gates.

Show the merged consensus configuration (.claude/consensus.local.md overlaid on defaults).

Run a single consensus gate check for one phase (no pipeline state advancement).

Use this agent when running a consensus gate check for the Alpha role — the code and logic specialist that reads implementation line-by-line for incorrect accumulation patterns, off-by-one errors, state machine bugs, race conditions, and API contract violations. Alpha is the detail-oriented auditor calibrated to catch the ChatViewModel += vs = bug and its siblings. Invoked exclusively by the consensus-pipeline skill in parallel with Lead and Bravo. <example> Context: The user ran `/consensus-run` and the pipeline reached the audit gate. user: "/consensus-run --target . --phases audit" assistant: "I'll invoke the Alpha agent via the Task tool for independent line-by-line logic validation of the audit phase, in parallel with Lead and Bravo." <commentary> Alpha reads every changed line. Its calibration is the += vs = class of bugs — correct-looking in isolation, broken at the interaction boundary. </commentary> </example> <example> Context: The user wants Alpha on a single-phase spot check. user: "/consensus-validate --target . --phase audit" assistant: "Invoking Alpha alongside Lead and Bravo via parallel Task calls for a one-shot audit gate check." <commentary> Same invocation pattern as /consensus-run but no fix cycle. </commentary> </example>

Use this agent when running a consensus gate check for the Bravo role — the systems and functional specialist that exercises the running system, verifies behavior under real conditions, catches regressions, and confirms fixes work in practice. Bravo is the only role with Bash because its mandate is to RUN the system (build, execute, curl, inspect output). Invoked exclusively by the consensus-pipeline skill in parallel with Lead and Alpha. <example> Context: The user ran `/consensus-run` and the pipeline reached the audit gate. user: "/consensus-run --target . --phases audit" assistant: "I'll invoke the Bravo agent via the Task tool for independent functional validation — it will run the target system and observe actual behavior, in parallel with Lead and Alpha." <commentary> Bravo does not just read code — it runs it. Use it when a FAIL should be a systems- level observation ("Four.Four." renders instead of "Four.") rather than a code-review finding. </commentary> </example> <example> Context: The user wants Bravo on a single-phase functional check. user: "/consensus-validate --target . --phase verify" assistant: "Invoking Bravo alongside Lead and Alpha via parallel Task calls — Bravo will execute the verify-phase runbook against the target." <commentary> Only Bravo has Bash. Security note: do not run Bravo against untrusted repos — its test commands could execute hostile code. </commentary> </example>

Use this agent when running a consensus gate check for the Lead role — the architecture and consistency specialist that validates cross-component coherence, pattern compliance, and regression detection at phase boundaries. Lead is invoked exclusively by the consensus-pipeline skill at gate checkpoints, in parallel with Alpha and Bravo, against a specific phase + target path. <example> Context: The user ran `/consensus-run` and the pipeline reached the audit gate. user: "/consensus-run --target . --phases audit" assistant: "I'll invoke the Lead agent via the Task tool for independent architectural validation of the audit phase, in parallel with Alpha and Bravo." <commentary> Lead is invoked exclusively at gate checkpoints inside the consensus-pipeline skill. Do not invoke Lead outside a gate context — it has no mandate for generic code review. </commentary> </example> <example> Context: The user wants to sanity-check a single phase without the full pipeline. user: "/consensus-validate --target . --phase audit" assistant: "Invoking Lead, Alpha, and Bravo in parallel via Task calls for a single audit-phase gate check." <commentary> Same invocation pattern as /consensus-run but one-shot, no fix cycle. </commentary> </example>

This skill should be invoked when the user asks to "run a consensus pipeline", "validate a phase with consensus", "run consensus gates", "check with three agents", "unanimous validation", "audit with Lead Alpha Bravo", or invokes /consensus-run or /consensus-validate. Orchestrates the three-agent unanimous-gate phase pipeline: for each configured phase, invoke Lead, Alpha, and Bravo in parallel, collect independent PASS or FAIL votes, enforce unanimity, trigger fix cycles on any FAIL with re-validation by all three agents, and persist evidence per phase and role. Works for CLI, web, API, and iOS targets — the triad validates the target system independent of its platform.

1 hook across 1 event