orchestration-rules

TheOneKit Orchestration Rules

Classify every user request and route to the matching T1K command. The authoritative role→agent and keyword→command routing lives in the t1k-routing-*.json fragments — read those to resolve.

Decision Tree

Classify the request, then route: feature/implement → /t1k:cook · planning/architecture → /t1k:plan · bug/error/compile → /t1k:fix · run tests → /t1k:test · investigate/debug → /t1k:debug · code review → /t1k:review · documentation → /t1k:docs · git ops → /t1k:git (cm|cp|pr|merge) · skill/agent mgmt → /t1k:issue, /t1k:sync-back · triage issues/PRs → /t1k:triage · brainstorm/ideation → /t1k:brainstorm · technical question → /t1k:ask · explore codebase → /t1k:scout · session review → /t1k:watzup · registry validation → /t1k:doctor · usage guide → /t1k:help · module mgmt → /t1k:modules · parallel multi-agent → /t1k:team · stuck/blocked → /t1k:problem-solve · structured reasoning → /t1k:think.

Priority order: T1K commands (registry-routed workflows) → skills (auto-activated by keyword context) → standard tools (Read/Write/Edit/Bash, trivial tasks only).

Task-Type → Agent Routing

Assumes the Agent/Task tool is available this session. When features.agentFanOut: false is set in t1k-config-core.json, or a session-level instruction has disabled the tool, this whole table has no reachable target — treat its routing as advisory, run the work same-context, and say so in the output rather than silently degrading. Default true (current documented posture unchanged); the flag makes an explicit, inspectable opt-out possible for a consumer that genuinely cannot fan out.

Delegating to a specialist agent is the default for substantive work, not an option — inline stays reserved for the four exceptions in delegate-to-preserve-context.md (design judgment, sub-agent-result synthesis, sequential full-output chains, patches under ~20 lines). Pick the narrowest specialist; general-purpose is the LAST RESORT and is wrong for kit-specific tasks (no kit context). Core-universal rows — never general-purpose for these:

Task pattern	Canonical agent
Skill creation / SKILL.md updates / agent definition edits	`t1k-skills-manager`
*`rules/.md` edits** — create, trim, dedupe, context-budget work	`t1k-rules-manager`
Doctor checks, kit install-scope findings, SessionStart scope remediation	`t1k-doctor-manager`
Module/kit install-update-audit state, both global and project scope	`t1k-modules-manager`
Cross-kit work (release-action scripts, CLI, kit fragments, sync-back, issue filing)	`t1k-kit-developer`
Markdown/frontmatter authoring outside skills — docs, READMEs, session memory files + `MEMORY.md` index	`t1k-docs-manager`
Mechanical cross-source reconciliation — “do these two systems’ numbers agree?”: query/sum/diff/tabulate against a stated tolerance	`t1k-data-verifier` (cheap tier — prefer over `t1k-researcher` whenever the comparison method is already specified)
Aggregate git/gh/log activity metrics into a table or scoreboard — counts, PRs merged, coverage, build size, velocity over a stated window (reports figures, does not interpret)	`t1k-metrics-reporter`
Retrieve stated facts from a NAMED source and report them verbatim — vendor report for given dates, documented quota/pricing tier (quotes, does not evaluate)	`t1k-fact-fetcher`
Poll CI runs / PR checks / deploys until terminal and report status per check (watches, does not diagnose, retry, or merge)	`t1k-ci-watcher`
Label every item in a supplied list against a rubric the brief states explicitly — CONFORMANT/DRIFT, REUSE/EXTEND, dead/live (applies, does not devise)	`t1k-list-classifier`
Ad-hoc fact-gathering against a live system — DB query, HTTP/API probe, log or telemetry inspection	`t1k-researcher`
Prior-art / “do we already have this?” sweep — corpus-first, grep-second, reuse verdict before scoping or building	`t1k-corpus-scout`
Phase coordination, task tracking, finalization, consumer-decision handoff	`t1k-project-manager`
Read-only structural audit — dead-code / orphan-asset sweep, reachability or dependency mapping, duplicate-implementation diffing	`t1k-code-reviewer`
Mechanical edits — rename, format, boilerplate, edit-per-plan	`t1k-fullstack-developer`
Run a test suite + report pass/fail	`t1k-tester`
Commit / push / branch ops, PR-fleet sweeps, merge-sequencing	`t1k-git-manager`
Handoff save/resume/list	`t1k-handoff-manager`

Tier = the agent’s pinned model:, never a per-spawn choice (skills/t1k-agent-creator/references/architecture-rules.md § 0.3; history: docs/orchestration-rules.md).

t1k-rule-creator / t1k-skill-creator are SKILLS, not spawn targets. Route rule work to t1k-rules-manager and skill work to t1k-skills-manager; each drives the matching skill itself.

t1k-doctor-manager vs t1k-modules-manager: doctor diagnoses install-SCOPE problems (a kit in the wrong place, or in two places at once) and remediates by removing a copy; modules-manager acts on install STATE (versions, add/remove, presets, t1k modules update) in a scope that is otherwise correctly placed.

Diagnostic disambiguation: a reported error/failure to root-cause → t1k-debugger; numbers to compare with the method already specified → t1k-data-verifier (cheap; “do they agree?” not “why don’t they?”); one source, count/aggregate → t1k-metrics-reporter; named source, quote verbatim → t1k-fact-fetcher; state over time (CI/deploys) → t1k-ci-watcher; given rubric applied to a list → t1k-list-classifier; open-ended fact-gathering with no failure reported → t1k-researcher; “has this already been built?” before scoping/implementing → t1k-corpus-scout (corpus-first, then grep — not the same question as t1k-researcher’s live-system lookup); codebase/file search → Explore; judging what that search found (dead vs live, orphan vs referenced, duplicate vs canonical) → t1k-code-reviewer. Explaining WHY / recommending → none of the cheap four; that is t1k-debugger (root cause) or t1k-code-reviewer (judgment).

Filing an issue / opening a PR via gh: use the /t1k:issue skill (/t1k:sync-back for fragments), never a raw general-purpose gh sub-agent — full routing rows: docs/orchestration-rules.md.

Engine/designer kits supply their own routing rows (via t1k-routing-*.json fragments + kit-local rules). Rule of thumb: a task naming a kit-specific concept (DOTS, ECS, Unity, scenes, shaders, wiki, game design) goes to that kit’s canonical agent; general-purpose only when genuinely engine-agnostic.

Mandatory Skill Usage (NEVER bypass)

/t1k:sync-back (sync .claude/ changes to kit repos) and /t1k:issue (report skill/agent bugs) — background sub-agent only · /t1k:triage — process issues/PRs; never manually browse issues · /t1k:git — ALL git operations (commit, push, branch, PR) route through t1k-git-manager, never raw git in the main session, except the patch-under-~20-lines case in delegate-to-preserve-context.md · /t1k:handoff — background t1k-handoff-manager only, never inline, never ask first (see skill’s “Delegation”).