Agent
Mid-cost interpretation — trace how code works across files, reproduce and analyze test failures, research dependency changelogs and upgrade impact, run code-graph and architecture queries. Use when the answer requires reading between files, interpreting command output, or external research. Also the escalation target…
Agent
Read-only codebase exploration — file discovery, code search, structure questions. Overrides the built-in Explore agent to pin automatic exploration to haiku for cost control.
Agent
Coordinator-owned bounded git/GitHub operations — stage, commit, push, open/reconcile a PR, watch CI, and merge only on an exact PR/CI/review head match. Never edits repository files and never force-pushes. Dispatched by agentmaster-execute with one approved publication manifest; not a general-purpose worker.
Agent
Executes exactly one task group from an approved implementation plan — edits only the files the group owns, runs each task's verification, reports results. Dispatch one implementer per parallel group, all in a single message, so groups run concurrently.
Agent
Fresh-context adversarial review of a draft implementation plan. Dispatched by agentmaster-plan with the goal, constraints, draft skeleton, and evidence ledger. Assumes the plan is wrong and hunts for the flaw, verifying claims against the repository read-only.
Agent
Cheap mechanical retrieval — locate files and symbols, list dependencies and versions, run a single command or test and capture output, extract specific facts from specific files. Use for questions with a definite answer that require no interpretation.
Agent
Executes an approved agentmaster-plan plan end to end. Use after accepting a plan at the Plan Ready for Review prompt — "execute the plan", "run the plan", "build it". Dispatches one implementer subagent per conflict-free parallel group, gates on per-task verification, then performs the full frontier code review…
Agent
Cost-tiered planning coordinator. Use to produce an implementation plan, spec, or root-cause investigation before code changes — multi-file features, bug hunts, refactors, dependency problems. Reasons and decides only; delegates all repository and dependency evidence to scout and code-analyst subagents, stress-tests…
Agent
Transcript-driven analyze-fix-verify loop for the agentmaster skill suite itself. Use after a pipeline run leaves artifacts behind — .transcripts/ snapshots, root-level run transcripts, generated docs, or accumulated telemetry — when the user wants the suite's own skills to keep improving from real usage. Grades…
Agent
Cost-tiered adversarial code review. Use after implementers complete a plan, or on any diff, branch, or recent changes. Assumes the code is bad — poorly implemented, doesn't scale, violates YAGNI, SOLID, or DRY, hard to test, badly structured, insecure — and makes it prove otherwise with evidence gathered by scout and…
Skill Claude CodeCodex
Execute an agentmaster plan. Use when the user asks to execute, implement, or run a plan produced by agentmaster-plan, or invokes /agentmaster-execute. Honors the plan's execution mode (sequential default, justified parallel with optional pilot), gates every task on verification, re-verifies independently, runs the…
Skill Claude CodeCodex
Cost-tiered planning pipeline. Use when the user asks to plan a feature, refactor, migration, or fix with agentmaster, or invokes /agentmaster-plan. A frontier orchestrator dispatches cheap scout/code-analyst workers for evidence, detects the project toolchain, drafts a parallel-safe plan with an evidence ledger, and…
Skill Claude CodeCodex
Transcript-driven analyze-fix-verify loop for the agentmaster skill suite. Use when the user wants agentmaster's own skills improved from real usage artifacts (.transcripts/ snapshots, root-level run transcripts, generated docs, telemetry), or invokes /agentmaster-retro.
Skill Claude CodeCodex
Adversarial code review with the agentmaster evidence pipeline. Use when the user asks for an agentmaster review of a diff, branch, or change set, or invokes /agentmaster-review. Assumes the code is guilty until proven working across correctness, structure, testability, flexibility, and security axes, with severity…