ariobarin · Oct 4, 2026
diff --git a/.github/workflows/check.yml b/.github/workflows/check.yml
@@ -7,7 +7,9 @@ on:
       - main
 
 jobs:
-  portable:
+  catalog:
+    # These names are required by the repository's branch protection.
+    name: portable (${{ matrix.os }})
     strategy:
       fail-fast: false
       matrix:
@@ -22,22 +24,8 @@ jobs:
         with:
           fetch-depth: 0
 
-      - name: Run portable tests
-        shell: pwsh
-        run: ./scripts/test-all.ps1
+      - name: Check source catalog
+        run: python scripts/check.py
 
       - name: Check diff formatting
-        shell: pwsh
         run: git diff --check HEAD^
-
-  windows-powershell:
-    runs-on: windows-2025
-
-    steps:
-      - uses: actions/checkout@v6
-        with:
-          fetch-depth: 0
-
-      - name: Run portable tests
-        shell: powershell
-        run: ./scripts/test-all.ps1
diff --git a/AGENTS.md b/AGENTS.md
@@ -1,8 +1,8 @@
 # Repository Guidance
 
-Compass is the lean source of truth for portable Codex and Claude Code
-configuration. The bundle contains reviewed Codex defaults and only explicitly
-selected skills or agents.
+Compass is Ario's collection of agent skills and working preferences. Its job
+is to improve how work gets done: clear decisions, independent execution,
+observable results, and fewer unnecessary human handoffs.
 
 ## Philosophy
 
@@ -12,7 +12,7 @@ judgment, execution paths, tool use, and self-checking to current models.
 Every Compass instruction must earn its place. Use decisive words that change
 behavior. Delete noise.
 
-Add prompts, skills, agents, or configuration only for a demonstrated
+Add prompts, skills, or agents only for a demonstrated
 current-model gap. Re-evaluate them after material model changes, and remove
 scaffolding that duplicates or distorts native behavior. Use absolute rules
 only for true invariants.
@@ -25,25 +25,25 @@ Read `philosophy.md` for simplicity and `source-grounding.md` for research.
 
 ## Boundaries
 
-- `codex/AGENTS.md` is the permanent global Codex instruction source. It may be
-  empty, but the source path stays in Compass.
-- `codex/config.toml` contains reviewed portable keys. Install overlays those
-  keys without replacing unrelated live configuration.
-- Portable skills belong under `codex/skills/<name>/` and are selected by
-  `agents.skills` in `manifests/portable-files.json`.
-- Portable Codex subagents belong under `codex/agents/<name>.toml` and are
-  selected by `codex.agents`.
-- Direct Claude files, skills, and agents belong under `claude/` and are
-  selected by the corresponding `claude` manifest collection.
+- `codex/AGENTS.md` holds global user preferences.
+- Skills live in `codex/skills/<name>/`. Keep the catalog in
+  `manifests/portable-files.json` aligned with those directories.
+- Compass MCP reads these paths and `agents.skills` directly. Preserve that
+  interface and single-line skill names and descriptions in frontmatter.
+- Optional Codex roles live in `codex/agents/`. Put a skill's required
+  specialist prompt inside that skill so it can supply the prompt itself.
 - Project-specific behavior belongs in the project that uses it.
+- Models, runtime settings, and installation choices belong to the environment
+  running the agent.
 - Auth, sessions, logs, caches, databases, browser state, generated plugin
   state, and machine-only values stay untracked.
 
-The manifest describes only the current desired bundle. Do not add migration,
-retirement, adoption, receipt, or compatibility machinery for prior layouts.
-Install may replace a selected target after backing it up. It must not change
-unlisted runtime state.
+Adapt upstream skills deliberately. Preserve attribution and license notices,
+record the inspected revision, and explain consequential departures in the
+skill's source notes. Keep provenance out of the normal execution path.
 
-Use a focused pull request as the review unit. Run `git diff --check`,
-`.\scripts\test-all.ps1`, and the Windows PowerShell 5.1 form of the same suite
-before committing.
+Use a focused pull request as the review unit. Run `git diff --check` and
+`python3 scripts/check.py` before committing. For behavioral changes, exercise
+the decision the skill should improve with a fresh agent and realistic inputs;
+include a nearby task where it should stay out of the way. Structural checks
+alone do not establish that a skill helps.
diff --git a/CODE_OF_CONDUCT.md b/CODE_OF_CONDUCT.md
@@ -1,7 +1,7 @@
 # Code of Conduct
 
 This project expects direct, respectful, technical collaboration. The goal is to
-make the portable setup easier to understand, review, and safely adapt.
+make the skills easier to understand, review, and adapt.
 
 Contributors should focus criticism on code, docs, claims, and evidence. Do not
 use personal attacks, harassment, threats, discrimination, or private

diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
@@ -1,19 +1,31 @@
 # Contributing
 
-Compass carries the current portable agent setup selected by
-`manifests/portable-files.json`. Keep the source small, explicit, and directly
-installable.
+Compass carries reviewed agent skills and working preferences. Keep each skill
+focused on a decision or capability that benefits from explicit guidance.
 
-Do not add project guidance, hosted applications, source archives, templates,
-runtime state, or compatibility machinery for previous layouts.
+Project guidance belongs in its project. Runtime settings belong in the agent's
+environment. Keep source archives, generated state, and installation machinery
+out of Compass.
+
+For a skill change, name the observed gap and inspect the current upstream
+artifact when one exists. Keep useful source wording, preserve license notices,
+and record the exact revision and material adaptations in `references/sources.md`.
+Keep provenance separate from instructions the agent needs during execution.
+
+Exercise consequential behavior changes with a fresh agent given a realistic
+task and the relevant skill. Include a nearby task that should not trigger it.
+For a new intervention, compare with a run that does not receive the skill.
+Report what changed and any limits; do not turn a few trials into a claim about
+all models. Keep disposable experiments out of the repository unless they
+protect a recurring risk.
 
 Before opening a pull request, run:
 
-```powershell
+```sh
 git diff --check
-.\scripts\test-all.ps1
-powershell.exe -NoProfile -ExecutionPolicy Bypass -File .\scripts\test-all.ps1
+python3 scripts/check.py
 ```
 
-Portable changes must preserve preview-first installation, backups before
-replacement, unlisted live state, path containment, and round-trip coverage.
+The source checker requires Python 3.11 or newer. It protects the skill catalog
+and local source references; it does not measure instruction quality. Keep PRs
+cohesive and explain the changed behavior and useful evidence.
diff --git a/README.md b/README.md
@@ -1,57 +1,57 @@
 # Compass
 
-Compass is reviewed source for portable Codex and Claude Code configuration,
-instructions, skills, and agents. The bundle contains portable Codex defaults
-and only explicitly selected skills or agents.
-
-It is an allowlist, not a runtime-home backup. Auth, sessions, logs, caches,
-databases, browser state, generated plugins, and machine-only values stay local.
-
-## Portable Sources
-
-`manifests/portable-files.json` selects the current bundle:
-
-- `codex.files` installs files from `codex/` into the Codex home.
-- `codex.config` overlays reviewed keys from `codex/config.toml` while keeping
-  unlisted live configuration intact.
-- `codex.agents` installs selected `codex/agents/<name>.toml` subagents without
-  copying repository documentation.
-- `agents.skills` installs `codex/skills/<name>/` into `$HOME/.agents/skills`.
-- `claude.files`, `claude.skills`, and `claude.agents` install direct Claude
-  definitions.
-
-`codex/AGENTS.md` carries reviewed global user preferences. `codex/config.toml`
-carries durable global defaults while machine paths, generated plugin state,
-project trust, and app-local settings stay in the live file. Skills and agents
-are added only when a reviewed change includes both the definition and its
-manifest entry.
-
-## Commands
-
-Preview the current bundle:
-
-```powershell
-.\scripts\install.ps1
-```
-
-Install it:
-
-```powershell
-.\scripts\install.ps1 -Apply
-```
-
-Verify installed targets:
-
-```powershell
-.\scripts\verify-live.ps1 -RequireInSync
+Compass is my collection of agent skills and working preferences. It captures
+the judgment I want agents to reuse: settle consequential decisions, work
+independently, prove the result, and make the next human visit useful.
+
+The collection draws on [Matt Pocock's skills](https://github.com/mattpocock/skills),
+[Lauren Tan's pstack](https://github.com/cursor/plugins/tree/main/pstack), and
+[Theo Browne's workflows](https://www.youtube.com/watch?v=D8PikZ1KhUo&t=2723),
+adapted to [my preference for simplicity](philosophy.md). Each adapted skill
+records its sources and the decisions behind the adaptation.
+
+## Choose The Skill For The Job
+
+| Need | Skills |
+| --- | --- |
+| Settle an important decision | [grilling](codex/skills/grilling/SKILL.md), [ground-in-sources](codex/skills/ground-in-sources/SKILL.md) |
+| Explore an uncertain idea | [micro-experiment](codex/skills/micro-experiment/SKILL.md), [design-prototype](codex/skills/design-prototype/SKILL.md) |
+| Break down and deliver substantial work | [to-tickets](codex/skills/to-tickets/SKILL.md), [implement-spec](codex/skills/implement-spec/SKILL.md) |
+| Diagnose and prove behavior | [diagnosing-bugs](codex/skills/diagnosing-bugs/SKILL.md), [test-for-risk](codex/skills/test-for-risk/SKILL.md), [blast-radius](codex/skills/blast-radius/SKILL.md) |
+| Make project verification repeatable | [create-verification-skill](codex/skills/create-verification-skill/SKILL.md) |
+| Carry work through checks and review | [babysit-pr](codex/skills/babysit-pr/SKILL.md), [monitor](codex/skills/monitor/SKILL.md) |
+| Improve the next session | [retro](codex/skills/retro/SKILL.md), [write-a-skill](codex/skills/write-a-skill/SKILL.md), [handoff](codex/skills/handoff/SKILL.md) |
+| Refine prose and comments | [unslop](codex/skills/unslop/SKILL.md), [no-comments](codex/skills/no-comments/SKILL.md) |
+| Change this collection | [compass](codex/skills/compass/SKILL.md) |
+
+These are tools to select, not stages every task must pass through. A small fix
+can go straight to implementation. A larger change might use grilling to
+settle its decisions, to-tickets to expose dependencies, implement-spec to
+deliver independent slices, and babysit-pr to follow the review.
+
+## Use What Helps
+
+Read a skill and copy its folder into the skill directory your agent uses.
+Keep its references and license notices with it. Global preferences are in
+[codex/AGENTS.md](codex/AGENTS.md); the optional Codex role is in
+[codex/agents](codex/agents/README.md). Review personal preferences before
+adopting them.
+
+Compass MCP consumes `codex/AGENTS.md`, `codex/skills/`, and the skill catalog in
+`manifests/portable-files.json`. Those paths remain its source interface; the
+manifest now only lists skills. Model selection and runtime setup belong to
+the environment running the agent.
+
+## Maintain The Collection
+
+Add instructions for an observed gap. Prefer a structural fix when the project
+can make the mistake impossible. Remove scaffolding that newer models no
+longer need. [Source Grounding](source-grounding.md) explains how evidence
+becomes guidance; [Contributing](CONTRIBUTING.md) covers review and validation.
+
+Run the read-only source checks with Python 3.11 or newer:
+
+```sh
+python3 scripts/check.py
+git diff --check
 ```
-
-Run the portable test suite:
-
-```powershell
-.\scripts\test-all.ps1
-```
-
-Install backs up a selected target before replacing it. It does not remove or
-inspect unlisted runtime state. Git contains the configuration history, so
-Compass carries no retirement or backwards compatibility database.
diff --git a/SECURITY.md b/SECURITY.md
@@ -1,13 +1,13 @@
 # Security Policy
 
-Compass contains reviewed source for a portable Codex setup. It should never
+Compass contains reviewed agent skills and working preferences. It should never
 contain auth tokens, session data, logs, browser state, generated plugin caches,
 SQLite state, or private machine-only runtime files.
 
 ## Reporting
 
 Please do not open a public issue for secrets, credentials, private machine
-state, or install behavior that could expose local data.
+state, or instructions that could expose local data.
 
 Use GitHub private vulnerability reporting if it is enabled. If it is not
 available, contact the maintainer privately through the channels listed on their
@@ -22,14 +22,13 @@ issue without exposing new secrets.
 In scope:
 
 - tracked files in this repository;
-- install and verification scripts;
-- portable config fragments and manifests;
-- guidance that could cause Codex to copy local-only state into git;
+- the skill catalog, source checks, and optional role definitions;
+- guidance that could cause an agent to expose local-only state;
 - guidance that could grant stronger authority than the docs explain.
 
 Out of scope:
 
 - private forks or local edits that are not proposed back to this repo;
-- generated files already excluded by the repo allowlist and denylist;
+- untracked runtime state;
 - issues in the Codex product or third-party services independent of this
   repository.
diff --git a/SUPPORT.md b/SUPPORT.md
@@ -1,16 +1,16 @@
 # Support
 
-Compass is a personal portable Codex setup shared as a reference. It is
-not a general Codex support channel and it is not meant to be installed blindly
-on every machine.
+Compass is a personal collection of agent skills shared as a reference. Support
+here concerns the collection's behavior and source, rather than general agent
+product questions.
 
-Open an issue when the question is about this repository's files, scripts,
-manifests, docs, or portability boundaries. Include the command you ran, the
-relevant file paths, and the behavior you expected.
+For a skill issue, include the skill, task, model and harness when known, the
+behavior you expected, and the relevant result or correction. A concrete
+session is more useful than a general claim that an instruction is too long
+or too short.
 
-Use `SECURITY.md` for secrets, credentials, private machine state, or unsafe
-install behavior.
+Use `SECURITY.md` for secrets, credentials, private state, or guidance that
+crosses an authority boundary.
 
-For questions about adapting the setup, describe the change you want and why it
-belongs in Compass instead of a local Codex home, target project repo, or plugin
-cache.
+For an adoption request, link the upstream artifact and describe the gap it
+would address. Project-specific behavior belongs in the project using it.
diff --git a/codex/AGENTS.md b/codex/AGENTS.md
@@ -18,6 +18,13 @@
 - Prefer an authenticated CLI over a connector when both can complete the task.
 - Prefer non-forked subagents with self-contained assignments. Fork context only
   when the task depends on prior history.
+- Use parallel agents to reduce my handoffs. Keep one owner responsible for
+  integration and the evidence behind the final result.
+- Complete authorized work before handing it back. Make the result easy to
+  inspect, with evidence for the current revision and a usable preview or PR
+  when the task calls for one.
+- Ask for missing decisions that materially affect the outcome or authority.
+  Do not ask me to reconfirm permission I have already given.
 - Keep plans, goals, and other working documents concise and complete enough to
   understand as a whole. Draft freely, then prune before execution.
 - Every line becomes maintenance. Delete, consolidate, or reuse before adding.

diff --git a/codex/agents/README.md b/codex/agents/README.md
@@ -1,6 +1,10 @@
-# Portable Subagents
+# Optional Codex Role
 
-Add each Codex subagent as `codex/agents/<name>.toml`, then add its name to
-`codex.agents` in `manifests/portable-files.json`.
+[simplicity-critic.toml](simplicity-critic.toml) defines a read-only review role
+for finding what can be removed without weakening the intended outcome. Copy
+it into your Codex agent definitions if useful. It inherits the model selected
+by the caller.
 
-Compass installs selected subagents into the Codex home.
+Specialists required by a skill live inside that skill. No Comments supplies
+[its own Comment Sicko prompt](../skills/no-comments/references/comment-sicko.md)
+to a fresh subagent without requiring a configured custom role.
diff --git a/codex/agents/simplicity-critic.toml b/codex/agents/simplicity-critic.toml
@@ -1,14 +1,10 @@
 # Sources:
-# GPT-5.6: preserve outcomes and constraints while favoring leaner prompts.
-# https://developers.openai.com/api/docs/guides/prompt-guidance-gpt-5p6
 # Agile Manifesto: maximize work not done.
 # https://agilemanifesto.org/principles
 # Musk engineering sequence: question requirements and delete before simplifying.
 # https://everydayastronaut.com/starbase-tour-and-interview-with-elon-musk/
 name = "simplicity-critic"
 description = "Finds what can be removed without weakening the intended outcome."
-model = "gpt-5.6-sol"
-model_reasoning_effort = "high"
 sandbox_mode = "read-only"
 developer_instructions = """
 Answer one question: what can be removed without weakening the intended outcome?

diff --git a/codex/config.toml b/codex/config.toml