harness

Plugin harness from marketplace flight505-plugins in flight505/flight505-marketplace. Commands read from the repo's manifest — 9 days ago.

Skills (7)

• advisor

  Analyze the current project for autoresearch optimization opportunities. Scans the project, identifies candidates matching the keep/revert loop pattern (single mutable artifact, scalar metric, fixed time budget, repeatable execution), and presents ready-to-run suggestions with workflow.yaml snippets

• approve

  Approve or reject a pending approval gate in a harness workflow. Used when a conductor run pauses at a type: approval phase.

• compose

  Interactive workflow builder — helps create workflow.yaml files for the harness orchestrator. The ONE interactive entry point. Use when user wants to build a harness workflow, create a pipeline, or compose subsystems together.

• dashboard

  Generate and open the harness run dashboard. Reads .harness/*.json state files, builds a self-contained HTML file, and opens it in the browser. Use when the user asks to see the dashboard, check run status, or view experiment history. For live monitoring during a run, suggest /loop 30s /harness:dash

• pick-workflow

  Match a task description to the best default workflow, collect the required values interactively, write .harness/workflow.yaml with everything filled in, and hand off to /harness:run. Full path from description to ready-to-run — no manual file copying or placeholder editing.

• plan-ready

  Signal that an ultraplan phase's plan has been saved and the conductor can resume. Used after the user reviews and teleports a plan from Claude Code on the web.

• setup

  Configure harness optimize-subsystem hardware targets — Apple Silicon (MLX), NVIDIA server (CUDA), or RunPod cloud GPU. One-time setup stored persistently. Use when user mentions hardware setup, SSH targets, GPU configuration, or says 'configure optimization targets'.

Slash commands (1)

• /harness:run

  Run a harness workflow — validates workflow.yaml, checks for resume state, executes phases sequentially. Handles agent-teams, loop, subagent, and inline phase types.

Subagents (14)

• advisor

  Analyzes any project and suggests optimization opportunities for the harness optimize-subsystem loop. Identifies: mutable artifact, scalar metric, run command, time budget. Returns structured JSON suggestions, not conversational advice.

• architecture-evaluator

  Compares a codebase's current approach against research findings and state-of-the-art. Identifies gaps, opportunities, and modernization paths.

• code-reviewer

  Adversarial code quality review of the full feature branch diff. Runs only after the spec reviewer approves. Checks correctness, security, architecture, types, tests, performance, regressions. Read-only. Returns a structured JSON verdict with file:line references.

• diagnostician

  Reads an issue or error, searches the codebase for root cause, and produces a structured diagnosis. Read-only — does not modify code.

• fixer

  Writes a minimal fix with regression test based on the diagnostician's output. Creates a fix branch and commits.

• implementation-guide

  Translates research findings into actionable implementation guidance — pseudocode, architecture decisions, migration steps. Bridges the gap between 'what the research says' and 'what to build'.

• implementer

  TDD implementer for harness agent-teams phases. Claims tasks from a shared list, implements with RED-GREEN-REFACTOR discipline, runs quality gates, commits. Works as both a delegated subagent and a native Agent Teams teammate.

• lead

  Team lead for harness agent-teams phases. Spawns teammates, assigns tasks, gates reviews, and cleans up. Invoked by the conductor when it enters an agent-teams phase.

• method-analyst

  Deep comparison of specific methods, tools, or approaches. Produces structured tradeoff analysis with recommendation. Use when choosing between approaches in any domain.

• optimizer

  Autonomous keep/revert optimization loop — edits a single artifact, runs a fixed-budget experiment, keeps or reverts based on a scalar metric. Supports local execution and remote via SSH. Writes state contract JSON after each experiment.

• reviewer

  Two-phase review of the full feature branch diff: spec compliance (verify each acceptance criterion with file:line evidence, check scope creep) then validation (run test/build/typecheck, verify commits). Read-only. Returns a structured JSON verdict.

• searcher

  Searches academic and web sources for research relevant to a query. Returns structured results with titles, URLs, relevance scores, and summaries. Domain-agnostic — works for any research question.

• synthesizer

  Synthesizes search results into structured findings — consensus, open questions, recommendations. Outputs both JSON state and markdown prose. Takes searcher output as input.

• verifier

  Verifies that the fixer's fix actually addresses the original issue. Runs tests, checks the fix against the diagnosis, and optionally creates a PR.

Hooks (10)

• SessionStart

  '${CLAUDE_PLUGIN_ROOT}/hooks/run-hook.cmd' inject-status

• SessionStart

  '${CLAUDE_PLUGIN_ROOT}/hooks/run-hook.cmd' inject-targets

• Notification

  '${CLAUDE_PLUGIN_ROOT}/hooks/run-hook.cmd' notify-complete

• PostToolUse

  '${CLAUDE_PLUGIN_ROOT}/hooks/run-hook.cmd' validate-experiment

• PostToolUse

  bash -c 'TOOL="$TOOL_NAME"; INPUT="$TOOL_INPUT"; if [ "$TOOL" = "Write" ]; then FILE=$(echo "$INPUT" | grep -o ".harness/research-[^\"]*\.json" || true); if [ -n "$FILE" ] && [ -f "$FILE" ]; then node

• PreCompact

  bash ${CLAUDE_PLUGIN_ROOT}/hooks/on-pre-compact.sh

• TaskCreated

  bash ${CLAUDE_PLUGIN_ROOT}/hooks/on-task-created.sh

• TaskCompleted

  bash ${CLAUDE_PLUGIN_ROOT}/hooks/on-task-completed.sh

• TaskCompleted

  bash ${CLAUDE_PLUGIN_ROOT}/hooks/validate-fix.sh

• TeammateIdle

  bash ${CLAUDE_PLUGIN_ROOT}/hooks/on-teammate-idle.sh

Similar plugins