haiku-split

Constraint — read this first

This skill is the one place where subagents are launched, and only after the user confirms that specific run (§6). The confirmation covers that one run, never a later run or another skill; a resume after a pause (§8) is a new run and is confirmed again. Every Haiku segment follows the context-barrier rules encoded in workflow.js.tmpl and the hard-rules block. The skill itself never reads project source code and never edits the target skill.

Usage: /haiku-split <skill-name | path> [--limit <tokens>] [--task-size <tokens>] [--max-segments <n>] [--args "<args for the target skill>"]

Defaults: --limit 100000, --task-size 20000, --max-segments 10. Haiku model ID is exactly claude-haiku-4-5-20251001. It is set explicitly on every agent() call: an omitted model inherits the controller's model and breaks the Haiku bound.

1 — Resolve

1. <skill-name> resolves to ~/.claude/skills/<skill-name>/SKILL.md. If the argument contains a path separator or ends in SKILL.md, use it as that path.
2. If nothing resolves: Glob ~/.claude/skills/*/SKILL.md, list the names, and stop.
3. If the target is haiku-split itself, refuse and stop.
4. Read the target SKILL.md and every _shared or reference file it names. Read no project source.
5. Name the single scope AREA: <skill-name> plus the --args value, e.g. plan-feature:ADO-123.
6. Run folder: <cwd>/.claude/haiku-split/<skill-name>/<YYYYMMDD-HHMM>/. Create it.

2 — Classify every step

Walk the target skill step by step and classify each step. Label each step exactly one of:

• haiku — mechanical, bounded and verifiable: reading named files, grep/glob, running a command and recording its output, writing a templated file, transcription, applying a fully specified edit.
• main — judgment, scoring, design, security or risk calls, anything gated by the target's model tier. Stays with the main session.
• pause — a hard pause, an approval, or a question to the user.

If a haiku step cannot be given a mechanical done-check, relabel it main. If no step is haiku, report that, show the classification, and stop. Do not generate a workflow.

3 — Split into tasks

Cut haiku steps into tasks of at most --task-size tokens (estimate: reading a file about 1 token per 4 characters, plus output, plus about 3k overhead). Rules:

• A task inlines the target skill's wording. A segment never invokes the target skill and has no model gate.
• Every task has: Source step, Class, Depends on, Inputs, Instructions, Outputs, Done-check, Est tokens, Side effects (the templates/TASKS.md block).
• Depends on lists every earlier task whose Outputs this task reads.
• main and pause steps become M<n> / P<n> blocks in their original position.
• Collect every question a segment might ask (confirmations, parameters) and answer them now by asking the user once (AskUserQuestion). These become the pre-answered parameters.

4 — Write the run folder

Write into the run folder:

1. TASKS.md from ~/.claude/skills/haiku-split/templates/TASKS.md: header, hard rules (from templates/hard-rules.md, deleting any line that cannot apply, e.g. the browser line), tasks.
2. handoff-0.md from templates/handoff.md: Done = none, Remains = all task ids in order, Current state = "not started", Gotchas from your reading of the target, Side effects = none.
3. notes.md: # Findings — AREA=<area>, empty body.
4. progress.log: empty file.

5. workflow.js from templates/workflow.js.tmpl, substituting: {{LIMIT}}, {{MAX_SEGMENTS}} as bare numbers; {{AREA}}, {{RUN_DIR}}, {{PARAMS}}, {{HARD_RULES}} as JSON string literals including quotes (escape newlines as \n). Use forward slashes in {{RUN_DIR}}. Render with Node so strings are escaped correctly (node -e has no script slot, so arguments start at process.argv.slice(1)):

   node -e 'const fs=require("fs");const [t,out]=process.argv.slice(1);
   const m={LIMIT:"100000",MAX_SEGMENTS:"10",AREA:JSON.stringify("<area>"),RUN_DIR:JSON.stringify("<run folder>"),
   PARAMS:JSON.stringify("<params>"),HARD_RULES:JSON.stringify("<rules text>")};
   let s=fs.readFileSync(t,"utf8");for(const k in m)s=s.split("{{"+k+"}}").join(m[k]);fs.writeFileSync(out,s)' \
   ~/.claude/skills/haiku-split/templates/workflow.js.tmpl <run folder>/workflow.js

Then check the script parses: wrap it as (async () => { ... })() with export const meta changed to const meta, and run node --check on it. Fix and retry until it passes.

5 — Show the plan

Report in a short table: counts of haiku / main / pause tasks, estimated segments (ceil(total est tokens / --limit)), AREA, LIMIT, MAX_SEGMENTS, the run folder, and the first main/pause block that will stop the run. Do not paste the files.

6 — Confirm the launch (hard pause)

Ask with AskUserQuestion: "Launch N Haiku segments for <skill> now? This starts subagents on claude-haiku-4-5-20251001, capped at <LIMIT> tokens each." Options: Launch / Don't launch. On anything but Launch, stop and leave the files in place. The yes covers this run only.

7 — Launch

Call Workflow with scriptPath = the run folder's workflow.js (no name). Then wait for the completion notification. Read progress.log and notes.md only after it arrives.

8 — Pauses and resume

The workflow returns blocked when the next task is a P/M block, depends on a FAILED task, or the counter was not visible. For each case:

• P/M block: do it yourself (or ask the user), then append AREA=<area> | main | DONE-MAIN <id> - <result> to progress.log.
• FAILED dependency: report the failure and ask the user how to proceed.
• Counter not visible: report that the barrier cannot fire and stop; do not relaunch.

To continue, repeat §6 (confirm again), then call Workflow with the same scriptPath and args: { startSegment: <last segment number + 1> }.

9 — Final review

When the workflow returns done: read notes.md, progress.log and the produced files. Check each FAILED line and each Outputs path from TASKS.md exists. Report using the global response format. Anything you cannot verify, say so.

Rules

• Never launch without the §6 confirmation.
• Never run a segment on a model other than claude-haiku-4-5-20251001.
• No automatic escalation to a larger model. A failed or ambiguous task goes to the main session.
• Segments run sequentially, never in parallel.
• Never modify the target skill or any file under ~/.claude/skills other than this skill's own.