# Ce Conventions

> Shared conventions and references for the ce-* skill family. Shared by Engineering skills as a sibling directory; not invoked directly.

- Skill: `sitaggart/ce-conventions` (Agent Skill, multi-file: 3 files)
- Install (CLI): `npx skillmds@latest add sitaggart/ce-conventions`
- Raw SKILL.md: https://api.skillmd.com/api/skills/sitaggart/ce-conventions/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: SiTaggart (https://skillmd.com/u/sitaggart)
- Updated: 2026-09-10
- Page: https://skillmd.com/skills/sitaggart/ce-conventions

---


# ce-* Shared Conventions

Engineering skills share this file and `references/` through
`../ce-conventions/`. Git and Knowledge are separately installed plugins.

## Companion skills

Resolve companion skills by name from the host's installed skill catalog.
Never build a relative filesystem path into another plugin: hosts can cache
plugins in separate versioned directories. Before offering or invoking a
companion, check that it is available. If absent, continue the authorized step
with native tools and the current skill's guidance. If its missing instructions
are essential, state the gap instead of claiming the step ran. Installing a
companion does not authorize commits, pushes, messages, or other side effects.

## Sub-agent dispatch

Skills delegate persona work by spawning a sub-agent via the platform's
subagent primitive (`Agent` in Claude Code, `spawn_agent` in Codex, `task` in
OpenCode, `subagent` in Pi via the `pi-subagents` extension) with a prompt that
tells it to read a persona file and apply it. Rules every dispatch site
follows:

- **Resolve persona paths to absolute paths in the parent.** A spawned
  sub-agent starts in the user's project, not the skill directory, so a bare
  `references/<file>.md` in its prompt resolves to nothing and the persona
  silently never loads. The parent knows where its own SKILL.md lives; expand
  from there before writing the prompt.
- **Standalone-skill personas:** have the sub-agent load the skill by name;
  on platforms where sub-agents cannot load skills, point at the absolute
  path of `skills/<name>/SKILL.md` instead.
- **Run research and review personas read-only.** Use the platform's enforced
  mechanism where one exists (a read-only agent type in Claude Code, a
  read-only sandbox in Codex); a review or research worker inheriting write
  access is a dispatch bug, not a convenience.
- **Permissions:** do not override the sub-agent's permission mode — let the
  user's configured permission settings apply.
- **Spawn workers fresh.** Dispatch personas into clean sub-agent contexts,
  not forks carrying the parent's conversation history — independent workers
  must stay independent.
- **No sub-agents available:** run the persona inline and keep the report
  short.
- **Persona files are dispatch payloads, exempt from parent-reads-references
  rules.** On harnesses whose skill contract says the parent reads a skill's
  references itself (Codex), that rule does not apply to persona files: the
  parent may read a persona but must not need to — it passes the absolute
  path to the sub-agent. The parent reads a persona itself only when running
  it inline.

## Blocking questions

Use the platform's blocking question tool: `AskUserQuestion` in Claude Code
(load it via `ToolSearch` first if its schema isn't loaded), `request_user_input`
in Codex, `ask_user` in Gemini and Pi (Pi requires the `pi-ask-user` extension).
Fall back to numbered options in chat only when no blocking tool exists or the
call errors — not because a schema load is required. Never silently skip a
required question, and act on the selection rather than just announcing it.

## Next-step menus

After finishing, recommend the natural next skill for what the work proved and
offer to run it. Always include a "Done for now" option. Route by judgment:
never offer shipping routes while behavior is unproven, and never offer
commit/checkout routes for targets reviewed remotely. Skip the menu entirely
when invoked as a sub-step from another skill — return results to the caller.

## Output-mode resolution (document-producing skills)

`OUTPUT_FORMAT` resolves: `output:` CLI token > the skill's own active config
key in `.compound-engineering/config.local.yaml` (`plan_output`,
`brainstorm_output`; commented lines are never active) > `md`. Unknown values
fall through with a one-line note naming the format actually resolved.
Pipeline and sub-step invocations always force `md`. Only literal flag tokens
(`output:`, `mode:`, `delegate:`) are consumed from arguments — `feat:`-style
prefixes pass through as description text. Regression evals for this contract
live in `../ce-plan/evals/`.

## Artifact paths

Use repo-relative paths inside written artifacts. Confirm a created artifact
to the user with its absolute path so the reference is clickable.

## Slack context (opt-in)

Run the installed `ce-slack-research` skill from Knowledge only when the user
asks for Slack context. That skill resolves its own persona path. If absent,
use available Slack tools for the authorized search and report access gaps.
Fold the returned digest into the current phase's material; do not paste it raw.

## Shared references

- `references/markdown-rendering.md` — markdown presentation for ce-plan and
  ce-brainstorm documents.
- `references/html-rendering.md` — HTML presentation for the same documents.