Director Pattern

v1.1.0 · October 27, 2026. The project Director stays in conversation, keeps its unfinished commitments, and accepts each outcome. Each complete outcome has one accountable owner, normally a Request Manager; specialists execute; a shared recorder preserves state. Tracked is default; lean is optional.

"Director Pattern" names the package. Assignment roles stay worker and director: the project Director and a Request Manager bounded to one outcome both carry Role: director, and Request Manager and Recorder are maintained responsibilities, not new role values. A specialty such as engineering, testing, or release is a task inside an outcome, not a reason for another manager. The Director can optionally dispatch substantial independent analysis that returns a proposal. Before staffing, an installation records enduring context, the canonical role index, capability and authority policy, Recorder routing policy, and the identity roster locator. Runtime controls are resolved when used; model/team/host observations stay in assignment or receipt evidence, outside PROJECT and startup instructions. It adds no runtime or live-agent guarantee. Updates preserve project policy and historical evidence without treating past environment mappings as current configuration.

New here? Start with the website for the overview and copyable requests to install, check, or update it through your project's existing agent. Website source and hosting notes live in site/.

For an agent: read the website's guide to continue to the canonical GitHub installation or update protocol.

Here to install? Start with pattern/README.md. That subtree stands alone. Install only its manifest-listed payload into the target project's .director/; the rest of this repository is not part of the installation. A temporary clone may transport files, but copying the checkout is not adoption.

Here to develop the pattern? Read development/RATIONALE.md for the reasons, tradeoffs, and open decisions, then follow this repository's authoring guidance. Maintainer material is not required for installation. Local adoption is separate: when configured, the optional local startup hook activates an installation. A source checkout alone does not install the pattern or confer director authority.

Candidate authorship, target installation, fresh-session verification, and public publication are separate steps. The source repository is bladnman/director_pattern; the agent installation entrypoint is pattern/README.md.

For an existing installation, pin published GitHub main to an exact reviewed commit and use that revision's update protocol, even when installed guidance predates incremental updates. Preserve project-owned rules and adaptations. Routine updates use focused checks and applicable prior evidence; full startup audits and fresh-agent trials run only for affected changes. While active, only the checked project Director arranges a tiny metadata check at most once per rolling 24 hours and defaults to a short notification. Say “always update director pattern” to enable project-scoped automatic updates through the same safe update protocol. No scheduler or new software is installed.

Optional global skill

This optional lifecycle pointer requires explicit user invocation. Ordinary task relevance must not select or run it. On request, copy the complete selected host skill below to its per-user director-pattern/ skill directory, preserving an existing copy before replacement. Do not install a project-local skill by default.

Host	Canonical source and required native control	Per-user destination	Explicit invocation
Codex	skill/director-pattern/, including agents/openai.yaml with policy.allow_implicit_invocation: false	$CODEX_HOME/skills/director-pattern/ when set, otherwise ~/.codex/skills/director-pattern/	$director-pattern
Claude Code	skill/claude-code/director-pattern/, with disable-model-invocation: true and user-invocable: true in SKILL.md	$CLAUDE_CONFIG_DIR/skills/director-pattern/ when set, otherwise ~/.claude/skills/director-pattern/	/director-pattern

A clear user request to use the Director Pattern skill also meets its invocation gate. Use the host command above or direct the agent to the selected skill file for a reliable loading route; a bare name may not load a skill hidden from model context. Native controls are required: description wording alone does not disable automatic selection. Preserve these controls during every global-skill update; do not copy only Codex's SKILL.md and omit its policy file. For another host, establish its supported explicit-only policy before installing this optional skill.

The pointer routes the requested operation to the current canonical guide; installing or invoking it does not install the pattern in any project. Adopted project startup hooks remain a separate instruction route. Already running chats may retain previously loaded instructions; these file changes do not establish a fresh-session behavior check.

On a request to remove the global skill, remove only the selected host's managed files: SKILL.md and, for Codex, agents/openai.yaml. Remove managed agents/ and director-pattern/ directories only if empty. Preserve unrelated files, other skills, and every project installation.