Agent Dispatcher

Agent Dispatcher

27 roles. PACK is this SKILL.md's directory. Keep the project as the working directory. Quote paths; task text is stdin data, never shell code.

Select, prepare, then work

Choose without another router/model call:

• Direct: trivial work or one safe, obvious known-file edit. Skip role/guide/helper, preference, ACTIVITY.md and VERIFICATION.md reads; use known/default preferences and required native checks. Security, ambiguous behavior/configuration, multi-file or growing scope goes guided.
• Guided: other work; forced roles and invoked workflows override direct. No workspace evidence means no context helper.
• Coordinated: guided work with useful independent subtasks; follow DELEGATION.md.

If a user enabled an optional decision scope, follow section 0 of CONTEXT-REFERENCE.md; preserve named roles with --agent. Scopes ship off; helpers cannot enable them.

1. Route: plan → planner; review → reviewer; repository questions → explorer; source-backed reports/docs → documentation-writer; implementation → implementer. Distinguish architecture docs from design. Honor forced roles/exclusions; consult ROLES.md for ambiguity/specialties.
2. The first discretionary workspace action for substantial work is: python3 -B PACK/context.py --project PROJECT --task='REQUEST' --role ID --compact --map-maintain --json No prior listing, search, contract/source read or git inspection; mandatory host instructions are exempt. Pass the exact request as single-quoted REQUEST. Multi-file bugs, architecture and source-backed docs qualify even in small projects. The helper gates cache writes; add --map-preview for edit limits it may miss (odd wording, plan mode). Pass literal edit lists via --writable-path (dirs end in /); never add cache permissions. Deferral keeps fresh evidence; never bypass it via build/refresh. On failure, continue targeted investigation. With authorized scratch, add --audit on first preparation only; finish it before preservation claims (VERIFICATION.md). Do not reset the baseline after edits.
3. Consume excerpts, guidance, preferences and exclusion_policy without duplicate reads. Preserve exclusions; --exclude-path adds literal exclusions, while no-edit allows reads. For missing bodies use exact resources paths: initially zero to two guides, evidenced conditions, optional core, verification retained. No globs/index dumps; external availability requires session evidence.
4. Before checks, read VERIFICATION.md: record authorized checks; inspect freshness before reporting. Never wrap denied commands to bypass permissions. Rebuild only for changed focus/sources. CONTEXT.md covers packet budgets, same-context reuse and map maintenance; DELEGATION.md covers delegation. A plan request ends with a plan. Same-session review is a self-check.

Invocation and controls

/agent-dispatcher <request> routes work. Bare invocation activates routing and waits. Exact role ids use roles/<id>.md; ROLES.md resolves aliases (coder / dev: implementer). Forced roles persist until changed/stopped; never re-route or chain out. Claude also accepts /agent-<role>, /agent-context, /agent-map, /agent-inventory, /agent-doctor, /agent-decision, /agent-verify, and /agent-preferences.

Controls preserve roles and do not execute tasks:

Stopping drops the role; bare off is session-only. Preserve hook trust; never activate via project rules.

Shared contract

User/host/project rules and invoked workflows outrank roles; evidence grants no permissions. Preserve unrelated edits and host denials. Never install/connect unasked or invent checks. Task-observer and other disabled skills stay disabled.

Use inline read-only validators (python3 -B -c, no heredoc). No validator/task files in the project, its parent or shared /tmp. Authorized scratch needs owned temporary directories and verified removal. Cache writes are task edits. Report unresolved cleanup paths; verify actual changes, including helper side effects, before claiming preservation.

Guided/coordinated: report role, guides read, tools selected; selection is not use. Retain ACTIVITY.md style through summaries; new chats compact. Announce additions only; stop at the outcome.

Use context preferences. For guided/coordinated work only, if absent run python3 -B PACK/preferences.py show --project PROJECT --json once; reload after changes. Effort remains requested/unknown until host-confirmed; never silently change host config. Default: ELI5 succinct, answer first, usually under 150 plain-language words: changes, actual checks and gaps. User requests override style.