# Conducty Codex > Always-on Conducty orchestration for Codex. Use by default for coding, debugging, refactoring, review, CI, deployment, production-change, multi-step planning, substantial research, agent coordination, tracer-first execution, vault memory, prompt logging, checkpoints, ship gates, or improvement loops. Prefer the conducty-codex MCP tools when available. Use when this capability is needed. - Skill: `tomevault-io/conducty-codex` (Agent Skill, multi-file: 2 files) - Install (CLI): `npx skillmds@latest add tomevault-io/conducty-codex` - Raw SKILL.md: https://api.skillmd.com/api/skills/tomevault-io/conducty-codex/raw - Safety review: pending (external: skill-scanner PASS, skillspector PASS) - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: tomevault-io (https://skillmd.com/u/tomevault-io) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/tomevault-io/conducty-codex --- # Conducty Codex Use this skill to run Conducty's loop inside Codex: Shape -> Plan -> Trace -> Execute -> Verify -> Improve -> Code Review -> Ship. ## Default Moves 1. Resolve the vault with `resolve_vault`; if it does not exist, call `bootstrap_vault`. 2. Shape the request: goal, appetite, no-go zones, acceptance criteria, verification. 3. Check prompt quality with `check_prompt_smells` before writing or dispatching prompts. 4. Create a plan with `create_plan`. Every plan needs a tracer and a verification command. 5. Execute the tracer first. If it fails, revise the plan before broader execution. 6. Log each prompt result with `log_prompt_outcome`. 7. Record group health with `record_checkpoint`. 8. Record learning with `record_improvement` when the plan ends or a failure pattern appears. 9. Create a pre-merge ship report with `create_ship_report` when verification evidence and residual risks are known. 10. Audit the vault with `audit_vault_graph` when notes, links, or closure signals need a health check. ## Codex Boundaries - Keep critical-path implementation local unless the user explicitly asks for subagents, delegation, or parallel agent work. - Use parallel agents only for independent work with disjoint write scopes. - Treat time budgets as circuit breakers: when work exceeds appetite, stop and reshape. - Preserve exact commands, file paths, acceptance criteria, no-go zones, and error strings. - Run the smallest meaningful verification before claiming completion. ## Prompt Shape Use compact prompts: ```text P1 - Add date helper Context: src/date.ts, tests/date.test.ts Accept: ISO input formats unchanged; invalid input returns null. No-go: no timezone library swap; no API rename. Verify: pnpm test tests/date.test.ts Budget: 25m Review: verify-only Tracer: yes ``` ## MCP Tools - `resolve_vault`: show active vault path and whether it exists. - `bootstrap_vault`: create Conducty vault indexes, accumulators, and note folders. - `get_cycle`: return the operating loop and phase routing. - `check_prompt_smells`: find missing acceptance, context, verification, no-go zones, and mixed concerns. - `create_plan`: write a timestamped plan note and update `Plans Index`. - `list_recent_notes`: list recent plans, improvements, code reviews, ship reports, or accumulator notes. - `log_prompt_outcome`: prepend a terse entry to `Prompt Log`. - `record_checkpoint`: append group health metrics to the plan note. - `record_improvement`: write an improvement kata note and update `Improvements Index`. - `create_ship_report`: write a green/yellow/red pre-merge verdict with verification evidence and residual risks. - `audit_vault_graph`: report broken wikilinks, duplicate basenames, orphan user notes, plans without ship reports, and plans without checkpoints. ## Finish Criteria Before final response, confirm: - plan or vault state was written when needed - tracer/verification result is known - prompt outcomes or residual risk are recorded - next Conducty move is obvious --- > Source: [robertbarclayy/conducty](https://github.com/robertbarclayy/conducty) — distributed by [TomeVault](https://tomevault.io).