CLI Collaboration

CLI Collaboration

Use a written handoff as the source of truth when CLI agents, or one agent across sessions, work in the same project. The goal is continuity without destroying state left by the user or another agent.

Core Rule

Read AGENT_HANDOFF.md before interpreting git status. Never infer ownership from a dirty worktree alone.

If collaboration is paused via .cli-collaboration-off or Status: paused in the handoff, still preserve existing work and ask before touching shared files.

Language Preference

Run this once, as the final step of the activation cycle, so the whole session (and every later activation, by any agent) speaks the user's chosen language. The preference lives in one global file: ${XDG_CONFIG_HOME:-$HOME/.config}/cli-collaboration/config.json.

1. Run scripts/lang.sh get from the installed skill directory.
2. If it prints a value (exit 0): continue the session in that language. Do not ask again.
3. If it exits non-zero (not configured): ask the user, in English: Choose your language: — then run scripts/lang.sh set "<user's answer>" --by <YOUR_AGENT_NAME> and continue in that language from then on.

The script never prompts on its own — in a CLI-LLM session there is no TTY, so you (the LLM) mediate the prompt in chat. To change the language later, the user runs scripts/lang.sh set "<new>" --force. Non-interactive runs (CI) use scripts/lang.sh ensure-noninteractive, which reads CLI_COLLAB_LANG or falls back to en.

Scope note: this honors the language at each skill activation, not in every unrelated conversation. Projecting the preference into each agent's native memory (CLAUDE.md/GEMINI.md/AGENTS.md) is a deferred Phase 2.

Start Gate

Before editing, state:

Handoff read: <path, last-updated timestamp>
Current task: <one line>
Files I will touch: <explicit file list>
Expected red test: <test name, or no test with reason>
Reserved zones confirmed: <user-reserved/frozen areas>
Stop condition: <task-complete | context-budget | blocker>

If you cannot fill a field, read more or ask. Do not improvise.

Bootstrap Without Handoff

If no handoff exists, create or propose AGENT_HANDOFF.md before normal work. Include the user's request, current path, git status --short --branch if available, declared ownership, and the next concrete action. Mark it as bootstrap state, not authoritative history.

If AGENT_HANDOFF.md is missing but a known workflow seed file such as brainstorming.md exists in the same folder, read that seed file before creating a generic handoff. Follow its bootstrap instructions to create the initial AGENT_HANDOFF.md.

For alternate workflows such as multi-LLM brainstorming, load references/alternate-workflows.md only when the user explicitly asks for that workflow, the handoff declares it, or a known seed file declares it.

Ownership

Use this exact line shape:

- <path-or-glob>: <agent-name> — <reason>

Patterns are bash case patterns: * matches any sequence of characters including /, so scripts/* matches both scripts/foo.sh and scripts/sub/foo.sh. Use explicit path segments when you need to scope to a single directory level. ** is not a recognized token.

Sections:

• ### agent-owned: files assigned to Codex, Claude, Gemini, Grok, or another named agent. Globs are allowed. Parenthetical notes like (pending Phase 2) are advisory.
• ### user-reserved: user-owned files. Stop before editing.
• ### frozen: published or protected files. Stop before editing.

Run scripts/check-ownership.sh --agent <name> <files...> when the file list intersects handoff ownership or before risky edits. Exit 0 means no conflict detected, 1 means ownership conflict, 2 means malformed handoff or usage error.

Conflict Handling

When a file is owned by someone else, user-reserved, frozen, or unexpectedly changed:

1. Stop before editing.
2. Name the file, owner, and risk.
3. Offer reassignment, deferment, or a smaller scope.
4. Wait for the user or owning agent to decide.

Do not resolve ownership conflicts by overwriting, reverting, stashing, or cleaning.

Work Discipline

• Keep edits inside declared files.
• If a new file becomes necessary, announce it before editing.
• Use tests for implementation work when possible. A failing test is a good handoff artifact.
• Prefer a small complete task plus precise handoff over a broad half-finished refactor.
• The single-agent case still uses the handoff: it is project memory between sessions.
• Only one parent session updates AGENT_HANDOFF.md; parallel subagents report back instead of editing it concurrently.

User Supersession

The user's latest instruction wins. If it changes scope mid-turn, write a brief supersession note in the final handoff and adjust the declared file list before editing new files.

Destructive Actions

Never run or recommend these without explicit user approval:

• git reset --hard
• git clean
• git stash
• git restore
• git checkout --
• lateral overwrite of files you did not author

Adapters

Load adapter notes only when relevant:

• references/codex-adapter.md: Codex install paths, AGENTS.md, and script usage.
• references/claude-adapter.md: Claude Code hooks and command guidance.
• references/gemini-adapter.md: Gemini activation and GEMINI.md guidance.
• references/grok-adapter.md: Grok Build discovery, tools, Bluefin paths, and agent identity.

End Of Shift

Update AGENT_HANDOFF.md last with:

Agent: <Codex | Claude Code | Gemini CLI | Grok | other>
Date/time: <ISO 8601 with timezone>
Task: <what you took on>
Status: <done | in-progress | blocked>
Files changed: <explicit list>
Tests red: <names + reason>
Tests green: <names>
Open concerns: <concrete risks>
Next agent starts from: <exact next step>
Do not touch: <reserved files/areas>

Keep recent history useful: preserve the last three detailed handoff blocks and summarize older entries if the file gets large.