Repository navigation
feat: add Codex skill discovery parity #386
Description
Overview
Make assistant and workspace-local skills discoverable by Codex without duplicating their canonical instructions or breaking Claude discovery. Extend the existing Brain discovery generator to inventory registered repositories, produce stable collision-free Codex adapters, neutralize workspace-local canonical skill locations, and drift-check parity across both harnesses.
Plan
- Extend the shared Brain discovery generator instead of adding per-repository installer scripts.
- Generate repository-namespaced, hyphenated Codex adapters for every public assistant skill while keeping flat assistant Markdown canonical.
- Move workspace-local skill bodies to neutral
skills/locations and generate both Claude and Codex discovery adapters. - Preserve all existing Claude skill names and links; exclude helper/internal Markdown from public discovery.
- Add inventory parity, name validity, collision, idempotence, drift, and broken-link coverage.
- Update skill-bootstrap instructions so newly added skills acquire both harness adapters.
- Implement as one task in two waves: generator/contract first, then repository rollout.
Detailed implementation plan
Affected Repositories
- PyAutoBrain (primary)
- autofit_assistant
- autogalaxy_assistant
- autolens_assistant
- autocti_assistant
- autofit_workspace
- autogalaxy_workspace
- autolens_workspace
Work Classification
Workspace
Branch Survey
| Repository | Current Branch | Dirty? | Conflict |
|---|---|---|---|
| PyAutoBrain | main | clean | codex-hook-parity |
| autofit_assistant | main | user-owned untracked scripts/compose_model_gaussians_exponentials.py |
codex-hook-parity |
| autogalaxy_assistant | main | clean | codex-hook-parity |
| autolens_assistant | main | user-owned untracked scripts/cluster_model_composition.py |
codex-hook-parity |
| autocti_assistant | main | clean | codex-hook-parity |
| autofit_workspace | main | clean | none |
| autogalaxy_workspace | main | clean | none |
| autolens_workspace | main | clean | mass-field-workspace-sweep (active dirty task worktree) |
Suggested branch: feature/codex-skill-discovery
Planned state: serialized behind codex-hook-parity and mass-field-workspace-sweep; do not create worktrees until both claims clear.
Implementation Steps
- Extend
PyAutoBrain/bin/install.shproject-discovery machinery to identify skill-hosting repositories through the registered workspace/body map and inventory their supported canonical skill shapes. - Define deterministic repository-namespaced Codex names using lowercase hyphenated components, validate the Codex name grammar/length, and fail on duplicate generated names.
- Treat assistant
skills/*.mdfiles with valid public frontmatter as canonical. Generate thin.codex/skills/<stable-name>/SKILL.mdadapters that carry valid Codex frontmatter and route to that one canonical body. ExcludeREADME.md, underscore-prefixed helpers, and non-skill Markdown. - Move the four existing workspace-local canonical skills from
.claude/skills/*/SKILL.mdinto neutralskills/*/SKILL.mdlocations. Regenerate the original Claude discovery paths and add namespaced Codex adapters, preserving user-facing Claude names. - Update the four assistant
AGENTS.md,skills/README.md, and_bootstrap_skill.mdinstructions so adding a public skill updates the canonical inventory and the shared discovery generator rather than only adding a Claude symlink. - Add Brain unit/fixture tests for assistant/workspace inventory parity, stable names, collision detection, helper exclusion, idempotent generation, drift detection, and resolving canonical links. Add repository-level checks where existing CI has a suitable hook.
- Generate and review all committed adapters across the seven consumer repositories, verify existing
.claude/skillspaths still resolve, and run focused repository tests plus the independent review gate. - Ship one PR per affected repository through
ship_workspace; Heart and merge remain separate gates.
Key Files
PyAutoBrain/bin/install.sh— shared cross-harness discovery generator and drift checkerPyAutoBrain/tests/test_skill_install.py— generator, collision, parity, and drift coverage*_assistant/skills/*.md— canonical flat assistant skill bodies*_assistant/.codex/skills/*/SKILL.md— generated Codex discovery adapters*_assistant/skills/_bootstrap_skill.md— new-skill registration instructions*_workspace/skills/*/SKILL.md— neutral workspace-local canonical bodies*_workspace/.claude/skills/and.codex/skills/— generated harness adapters
Original Prompt
Click to expand starting prompt
Make assistant and workspace skills discoverable in Codex
Type: maintenance
Target: assistants
Repos:
- @PyAutoBrain
- @autofit_assistant
- @autogalaxy_assistant
- @autolens_assistant
- @autocti_assistant
- @autofit_workspace
- @autogalaxy_workspace
- @autolens_workspace
Difficulty: large
Autonomy: supervised
Priority: high
Status: draft
Filed: 2026-09-17
Parent: draft/maintenance/organs/agent_harness_agnostic_setup.md
Blocked-by: codex_hook_parity
Request
Can you check that all Agent setup (skills, md files, etc) is agnostic to whether its claude or codex, been using claude for a while but gonna codex for the forseeable so worth an agnostic sweep
Scope
Keep the existing flat assistant skill Markdown as canonical for this phase and
generate Codex-discoverable <hyphenated-name>/SKILL.md adapters for every
public assistant skill and the workspace-local skills. Drive discovery from
the repository/body map rather than another hardcoded installer list. Update
bootstrap instructions so new skills acquire both Claude and Codex adapters.
Acceptance
- Public skill inventories have tested Claude/Codex parity.
- Every Codex adapter has a stable valid name, description and single
canonical body; helper/internal Markdown is not exposed as a public skill. - Name normalization is collision-checked across installed assistants.
- Existing Claude skill names and links continue to work.
Activity
Jammy2211 commentedon Sep 19, 2026
codex-skill-discovery
Summary
Added registry-driven Claude/Codex discovery for shared assistant and workspace skill bodies; all 105 Codex adapters were discovered by codex-cli 0.155.1. Existing Claude links and canonical bodies were preserved. Clone boundaries exclude parent-specific generated wrappers from newborn copy sets.
Validation and limits
Focused tests and independent review CLEAN. Discovery installer: 22 tests; clone profiles: 37; Mind ledger: 26; template/privacy: 111; hook adapters: 8; shared guard fixtures: 68; Brain Memory faculty: 8; Memory board: 62. Tenant firewall, scoped discovery drift, Memory structure/wikilinks, and lensing clone boundary passed. Actual Codex smoke covers discovery, not end-to-end scientific execution or every harness. Eleven pre-existing inference-assistant clone-boundary gaps remain outside this task. No template publish or release was performed.