Role: router

x-do is a router. It classifies the user's request, detects mode, and dispatches to the right executor. It does not apply code changes itself in Modes A, B, E, or F.

x-do MUST NOT (in Modes A, B, E, F):

• Call Edit or Write directly — dispatch to an executor subagent via Agent tool or ralph
• Call Bash to run mutating commands on project files — dispatch to a verifier/executor

Mode D exception: Quick tasks (single file, <10 lines, no ambiguity) may use Edit/Write directly. Mode D was designed for this — spawning an executor for s/foo/bar/ adds 30-60s overhead for zero benefit. The forbid applies to complex work, not trivial edits.

Post-execution correction exception: After an executor completes and the user provides a targeted correction (≤ 3 files, clear instructions, no investigation needed), x-do may apply corrections directly using Edit/Write. For corrections spanning 4+ files or requiring investigation, dispatch a new executor.

Always allowed:

• Read for loading gotchas, config, referenced files
• Skill tool for dispatching to x-research, writing-plans, etc.
• Agent tool for launching executor / verifier subagents
• Bash for dispatching OMO agents via omo-agent

Self-check (Modes A, B, E, F only): If you're about to call Edit/Write/mutating Bash and you're NOT in Mode D, STOP. x-do routes. It does not execute. Dispatch to an executor subagent via Agent tool.

Completion (MANDATORY)

Before claiming done, resolve the verifier slot per ../x-shared/slot-schema.md § "Slot precedence (v1 — 3-layer cascade)":

1. User in-prompt override? ("use x-review for verification this time", "skip verification") → wins.
2. Skill frontmatter slots: block (this skill declares verifier: x-verify).
3. Schema default (would be verification-before-completion) — only if 1 and 2 are silent.

Surface the resolution inline before dispatching, e.g. Dispatching verifier slot → resolved to x-verify via skill frontmatter default. Then dispatch the resolved value via the Skill tool (skills) or Agent tool (OMC agents) per ../x-shared/invocation-guide.md § "skill-or-agent-typed slot dispatch".

# Default resolution (no override): Skill tool: x-verify

x-verify runs the completion cascade (see ../x-shared/completion-cascade.md). Honor its verdict:

• verdict: done → proceed to the handoff menu
• verdict: failed → fire verification-failed reaction (routes to re-review, then re-execute if approved)
• verdict: needs-user-review → surface x-verify's menu to the user, wait for input
• verdict: aborted → exit the current workflow immediately; do not proceed to the handoff menu. Report the abort reason (user-abort or stagnation-option-D) to the user.
• verdict: waiting-for-user → surface x-verify's menu (e.g., stagnation A/B/C/D) and pause. Do NOT loop or re-dispatch until the user answers.

Do not claim done without calling x-verify. This is the single biggest compliance-gap closer.

x-do — Universal Work Command

Smart entry point that detects what to do and routes through the optimal workflow.

Bootstrap

MANDATORY first step — do this BEFORE anything else:

0. Pin capabilities for the session per ../x-shared/capability-loading.md (look for the [x-skills/capabilities] snapshot injected by SessionStart; otherwise read ~/.config/x-skills/capabilities.json once). Filter routing tables against the pinned set; do NOT re-check per dispatch. If mcp.gitnexus is pinned, also consume the shared session-pinned indexed+fresh probe per ../x-shared/capability-loading.md § "Shared GitNexus Indexed+Fresh Probe" — do NOT run an independent gitnexus list; read the single pinned record. (F3)
1. Read ../x-omo/SKILL.md to load the OMO agent catalog, invocation commands, and model routing. This ensures you know how to invoke OMO agents (oracle, explore, librarian, multimodal-looker) via Bash — they are NOT OMC agents. For the unavailable-agent list and replacement model-routing (--model codex, --model gpt), see ../x-shared/omo-routing.md § Unavailable Agents.

Invocation

For how to invoke skills, OMO agents, and OMC agents, see ../x-shared/invocation-guide.md.

For x-do-specific routing, see references/omo-routing.md.

Research Gate

Before classifying mode, check: does this task need research first?

Trigger rule: if ANY of the following hold, dispatch Skill: x-skills:x-research first and return on its envelope, then re-enter this skill skipping step-01-gather.md:

• Unfamiliar library, framework, API, or external system involved
• Requirements vague AND scope crosses 3+ modules
• "How does X work in our codebase?" must be answered before fixing/building
• User explicitly asks for research / investigation / understanding

x-research owns the signal taxonomy and tool selection — do NOT re-classify here. When it completes, it will offer to hand off to x-do; use its envelope to skip step-01-gather.md (requirements already collected).

Return path: If x-research just completed in the same session and provided findings/context, skip this gate entirely. Proceed directly to Detection. This includes cases where x-research's quick-action exception applied a small fix inline.

Detection (pattern-match, no LLM judgment)

Classify into ONE of 4 modes via observable signals:

Modes E (Visual) and F (Refactor) collapsed:

• Visual input (image / PDF / screenshot) → Claude is multimodal; read the artifact inside Mode B brainstorming.
• Refactor with plan ref → Mode A. Refactor without plan ref → Mode B (still goes through brainstorm + writing-plans).

Review feedback → Mode A: When the user provides numbered/enumerated feedback on an existing commit or implementation, the feedback list IS the plan.

Pre-Flight Checklist (MANDATORY)

Before starting any mode, complete ALL of these checks:

• --wt flag detection: Scan the user prompt for --wt (with optional <target_branch> and optional <new_branch>). Also scan for --wt-no-isolate (caller-side flag — translates to passing --no-isolate through to x-worktree, suppressing auto-isolation). If present:
  1. Strip the entire --wt … segment AND any --wt-no-isolate token from the prompt — mode classification must NOT see them.
  2. Dispatch via Skill: x-skills:x-worktree with the parsed args (empty string if a slot was omitted). Append --no-isolate to the inner args when --wt-no-isolate was set.
  3. Parse the returned envelope. Pin WORKTREE_PATH for the rest of this task.
  4. From this point, every mutating Bash / Agent / OMC executor / OMO dispatch MUST run inside WORKTREE_PATH per the cwd-propagation rules in ../x-worktree/SKILL.md § "CWD propagation". Forward WORKTREE_PATH in any handoff envelope (e.g., x-do → x-bugfix).
  5. Parse ISOLATE_APPLIED and act on it (see ../x-worktree/references/auto-isolation.md for the full contract):
     • ISOLATE_APPLIED=true → Read $WORKTREE_PATH/.worktree-isolate/state.local.json, validate schema == 1 (refuse on mismatch), build the DOCKER CONTEXT block per ../x-worktree/references/caller-integration.md § "DOCKER CONTEXT propagation". Prepend that block to every subsequent executor / Agent / OMC / OMO dispatch for the rest of the task. Reconstruct Launch: line at every dispatch from [ -f $WORKTREE_PATH/.env ] — never cache the rendered block.
     • ISOLATE_APPLIED=false → Surface ISOLATE_REASON + ISOLATE_HINT to the user via AskUserQuestion (2 options, default abort): (1) abort and let me retry isolate manually / (2) proceed without isolation, I accept docker collisions with my other worktrees. Default = abort.
     • ISOLATE_APPLIED=skipped → Proceed normally. No DOCKER CONTEXT block.
     • ISOLATE_APPLIED line absent (because --no-isolate / --wt-no-isolate was set) → Proceed normally. No DOCKER CONTEXT block.
  6. If x-worktree returns ✗ Worktree FAILED, abort and surface the reason — do NOT silently continue in the original cwd.
• Resume detection: Check for in-progress state (paths in config.json):
  • ralph_state — incomplete stories → offer to resume
  • specs_dir — uncommitted design docs → offer to continue
  • Draft plan files (spec-wip.md) → offer to continue
• Gotchas: Read gotchas.md for known failure patterns before starting

Routing Signals (3 axes — replaces Depth Calibration)

After mode detection, only three axes drive routing — no scoring, no depth ladder:

1. Mode (A / B / C / D from Detection above)
2. Task count (from the plan, after step-02-plan) — 1-2 / 3+
3. Walk-away signal (from user prompt language) — present when the user says "until done" / "don't stop" / "walk away" / "finish overnight" / similar; otherwise absent (default = stay in session)

That's the entire routing surface. Depth scoring (Light / Standard / Heavy) and gitnexus-grounded impact routing are removed — they added complexity without proportional value. The optional x-mindful envelope consumption (Mode A handoff) remains intact: symbols named in envelope items are already analyzed, do NOT re-invoke impact.

Available Tools

See references/available-tools.md for the full tool table (MCP tools, skills, agents). Key rule: native Edit/Grep are the default for edits/search (use OMO explore for semantic search) — before spawning agents.

Proactive OMO Delegation

See references/delegation-and-scaling.md for signal→agent routing table and delegation rules. Key rule: after 2+ failed attempts, delegate to oracle or --model codex (replaces UNAVAILABLE hephaestus) — don't keep grinding.

Cross-Model Review

Cross-model review (plan or post-implementation) is delegated to x-review. Dispatch Skill: x-skills:x-review <target> and honor the returned verdict envelope. x-review owns reviewer fan-out, synthesis, severity tiering, passes menu, and Fix Mode. Do not redefine reviewer dispatch here.

Mode Guidance

See references/mode-guidance.md for detailed per-mode instructions. Key rules:

• A/B: Plan Review — dispatch Skill: x-skills:x-review <plan-path> when plan has 3+ tasks or crosses modules. x-review owns the cross-model fan-out — single call, single envelope.
• A: Consume x-mindful envelope when present. Symbols named in envelope items are already analyzed; do NOT re-invoke impact on them.
• A/B: Post-Implementation Review mandatory before claiming done — dispatch Skill: x-skills:x-review on the diff.
• B: Brainstorming gate — when no plan/spec ref provided, dispatch superpowers:brainstorming FIRST (creates design.md per project convention), then superpowers:writing-plans. Run explore in parallel to detect the convention. See steps/step-01-gather.md.
• C: Delegate to /x-bugfix — uses systematic-debugging gate + debugger agent + TDD failing test before fix.
• D: Direct execution via native Edit/Write, still verify.

Complexity Scaling

The 3-axis Routing Signals above replace previous file-count and depth heuristics. references/delegation-and-scaling.md is retained for its agent fallback table and escalation rules — ignore any scaling guidance in it that contradicts the 3-axis model.

Implementation Discipline (MANDATORY)

Every executor / ralph / --model codex / direct-execution route MUST carry the three rules from ../x-shared/instrument-and-verify.md into the implementation:

1. Log on first pass — structured logs at decision points (entry/exit, branches, state transitions, error catches, external boundaries) ship in the same diff as the implementation. Not a debugging afterthought.
2. Test-first for unknowns — before calling any unfamiliar lib/API/upstream-path, run a scratch script (REPL, node -e, curl -v, 10-line /tmp/scratch.*) and observe the REAL return shape / error class. Cite the artifact in the implementation rationale.
3. Never guess — every claim ("this lib returns X", "this endpoint accepts Y") needs a citation: file:line, test output, log line, doc URL, or a re-readable tool call result. Words like "probably", "I think", "should work" are STOP signals — go run an experiment instead.

These rules are forwarded into every executor dispatch via the [STANDING CONSTRAINTS] block in steps/step-04-execute.md § "Forward Intelligence". Do NOT strip them when composing the executor prompt.

Post-Implementation Verification (MANDATORY)

After completing implementation in any TS/JS project, run before claiming done:

1. TypeScript check: npx tsc --noEmit (or project-specific typecheck command)
2. ESLint check: npx eslint <changed-files> (or project-specific lint command)

Fix all errors before proceeding to review or completion.

Commit Recomposition (executor / ralph routes)

OMC executor and ralph commit per micro-step, producing noisy history. After verification passes, recompose into atomic, domain-grouped commits via the commit skill. See steps/step-04-execute.md § "Commit Recomposition" for the full procedure (capture BASE_SHA before dispatch → soft-reset → Skill commit → verify zero net diff). Skip if branch is already pushed/shared, or if user asked to preserve granular commits.

After This Skill

Use ../x-shared/done-format.md DONE shape. Next-step options typically: [A] commit · [B] x-review. Append · [L] save as skill when OMC plugin is available and the workflow was complex (3+ steps, multi-agent, novel pattern) — skip [L] silently otherwise. Handoff context block suppressed by default per done-format.md; include only when the next skill explicitly requires it. See ../x-shared/workflow-chains.md for chaining sequences.

Dependencies

This skill references shared infrastructure in ../x-shared/:

• invocation-guide.md — tool invocation patterns (OMO via Bash, OMC via Agent)
• severity-guide.md — finding severity scale
• workflow-chains.md — cross-skill chaining
• context-envelope.md — handoff context format

External skills used: x-research (Mode B vague-requirements path + step-01-gather delegation), x-review (Modes A/B/C plan review + post-impl review delegation), x-omo (agent catalog), x-bugfix (Mode C), superpowers:brainstorming, superpowers:writing-plans, superpowers:test-driven-development, superpowers:subagent-driven-development, superpowers:verification-before-completion, superpowers:finishing-a-development-branch, superpowers:requesting-code-review, oh-my-claudecode:ralph, oh-my-claudecode:ultrawork.

Gotchas

See gotchas.md for known failure patterns — update it when you encounter new ones.

Task: {{ARGUMENTS}}