.agents/skills/orchestrator/SKILL.md
Codex wrapper for the canonical WF orchestrator. Use when the user wants project initialization, stage status, gate checks, rollback, or decision logging around `PROJECT_STATE.json`.
npx skillsauth add linzhe001/Harness-Research orchestratorInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
Read these first:
../../../.agents/references/workflow-guide.md../../../.agents/references/language-policy.md./references/project-state-schema.json./references/stage-gates.md../../../CLAUDE.md../../../PROJECT_STATE.jsonUse this skill when the user wants project-level workflow control rather than direct implementation work.
Interpret natural-language requests as one of these canonical intents:
initstatusnextrollbackdecisionPROJECT_STATE.json is the only stage-flow source of truth.iteration_log.json is read-only from this skill; $iterate owns experiment writes.project_map.json is read-only here; $build-plan and $code-debug own structure updates.init$init-project in init mode to generate the minimal CLAUDE.md.PROJECT_STATE.json using the canonical schema.statusPROJECT_STATE.json.iteration_log.json for latest and best iteration status..auto_iterate/state.json exists, include loop state (current round, goal, progress) in the report (read-only — orchestrator never writes to that file).nextdocs/Baseline_Report.md and populated baseline metrics.$validate-run.$validate-run passes, orchestrator should auto-trigger a $auto-iterate-goal check so that an iteration goal is set before WF8 begins.CONTINUE.NEXT_ROUND → stay in WF8 (do not advance stage); ordinary improvement roundDEBUG → stay in WF8 (do not advance stage); fix technical issuesCONTINUE → can advance to WF9PIVOT → rollback to WF2ABORT → terminate projectrollbackcurrent_stage back without deleting artifacts, and append a rollback event to history.decisionAskUserQuestion.$init-project, $validate-run, and $iterate instead of Claude slash commands.../../../.agents/references/language-policy.md for reply language; keep workflow IDs, schema keys, and explicitly English-only fields unchanged.Use the local references above as the workflow contract. Do not depend on .claude at runtime.
development
WF7.5 training pipeline validation. Before entering WF8 iteration, first use Codex to review code for baseline equivalence, then run a 100-step smoke test to verify end-to-end pipeline functionality.
business
WF1 Inspiration survey and gap analysis. Takes the user's research idea, performs literature search, gap analysis, competitor analysis, and feasibility scoring, then outputs Feasibility_Report.md. Use when the user has a new CV research idea that needs a feasibility assessment.
tools
WF10 Submission/Release Tool. Multi-scene training, result packaging, filename validation, dry-run submission checks. Used after ablation experiments are complete and before competition submission.
development
WF2 Architecture refinement and MVP design. Reads the feasibility report, analyzes the base codebase architecture, designs plug-and-play new modules, defines the MVP, provides A/B/C alternative plans, and outputs Technical_Spec.md. Use when a research idea needs to be translated into a concrete technical architecture design.