st-generate-tasks

st-generate-tasks

Drive the end-to-end decomposition of an existing Strikethroo plan into atomic Markdown task files. The skill is assistant-agnostic and self-contained: every script it invokes lives under this skill's scripts/ directory and is referenced by relative path.

Inputs

The user supplies the numeric plan ID conversationally. Treat it as the only authoritative source of intent. Do not invent answers to clarifying questions — prompt the user instead.

Operating Procedure

1. Locate the strikethroo root

Run scripts/find-strikethroo-root.cjs from the user's working directory. The script walks up looking for .ai/strikethroo/.init-metadata.json and prints the absolute path of the resolved root on success.

If the script exits non-zero, the working directory is not inside an initialized strikethroo workspace. Stop and ask the user to run the project initializer (e.g. npx strikethroo init) before continuing. Do not attempt to generate tasks outside of a valid root.

For every subsequent step, treat the path printed by this script as <root>.

2. Resolve the plan

Run scripts/validate-plan-blueprint.cjs <plan-id> planFile to obtain the absolute path of the plan file. The same script also accepts these field names (single-field output mode) and exposes them on demand:

• planDir — absolute path of the plan directory
• taskCount — number of existing task files in that plan's tasks/
• blueprintExists — yes or no
• taskManagerRoot — absolute path of <root>
• planId — the resolved numeric plan ID

If the script exits non-zero, stop and ask the user to confirm the plan ID. Do not guess a different ID.

3. Load project context

Read these files, in order:

• <root>/config/STRIKETHROO.md — directory conventions for plans, tasks, and the archive layout.
• The plan body at the path returned by step 2 — this is the contract for what tasks must exist.
• <root>/config/templates/TASK_TEMPLATE.md — every task file you emit must conform to this template's frontmatter schema and section structure.
• <root>/config/shared/anti-rationalization.md — apply in step 4.

4. Analyze and decompose the plan

Read the entire plan. Identify all concrete deliverables explicitly stated. Decompose each deliverable into atomic tasks only when genuinely needed.

Task minimization (mandatory):

• Create only the minimum number of tasks necessary. Target a 20–30% reduction from comprehensive lists by questioning the necessity of each candidate.
• Direct Implementation Only: a task corresponds to an explicit requirement, not a "nice-to-have".
• DRY Task Principle: each task has a unique, non-overlapping purpose.
• Question Everything: for each task, ask "Is this absolutely necessary to meet the plan objectives?"
• Avoid Gold-plating: resist comprehensive features the plan does not require.

Antipatterns to avoid:

• Separating "error handling" from the main implementation when it can be inline.
• Splitting trivially small operations into multiple tasks (e.g. "validate input" + "process input" as separate units).
• Adding tasks for "future extensibility" or "best practices" the plan does not mention.
• Comprehensive test suites for trivial functionality.

Apply <root>/config/shared/anti-rationalization.md to this rationalization table:

You catch yourself thinking…	The binding rule
"One extra task won't hurt."	It violates the 20–30% minimization target. Every task traces to an explicitly stated deliverable or it does not exist.
"This edge case deserves its own task."	Fold it into the task that owns the behavior. Do not split trivially small operations into separate units.
"I'll add a test suite to be safe."	Comprehensive tests for trivial functionality are gold-plating. Follow the test philosophy — meaningful tests only.
"Future extensibility justifies this task."	YAGNI. The plan does not mention it, so it is not a task.

5. Apply granularity and skill rules

Each task must be:

• Single-purpose — one clear deliverable.
• Atomic — cannot be meaningfully split further.
• Skill-specific — executable by an agent with 1–2 technical skills.
• Verifiable — has explicit acceptance criteria that include at least one concrete, runnable verification step (a command plus its expected output, or another observable signal). Never settle for a vague "works correctly".

Skill assignment (kebab-case, automatically inferred from the task's technical requirements):

• 1 skill — single-domain task (e.g. ["css"], ["vitest"]).
• 2 skills — complementary domains (e.g. ["api-endpoints", "database"], ["react-components", "vitest"]).
• 3+ skills indicates the task should be broken down further.

6. Test philosophy: "write a few tests, mostly integration"

When generating test tasks, keep this constraint:

Definition. Meaningful tests verify custom business logic, critical paths, and edge cases specific to this application. Test your code, not the framework or library.

When TO write tests:

• Custom business logic and algorithms.
• Critical user workflows and data transformations.
• Edge cases and error conditions for core functionality.
• Integration points between components.
• Complex validation logic or calculations.

When NOT to write tests:

• Third-party library functionality.
• Framework features.
• Simple CRUD operations without custom logic.
• Trivial getters/setters or static configuration.
• Obvious functionality that would break immediately if incorrect.

Test task creation rules:

• Combine related test scenarios into a single task (e.g. "Test user authentication flow" not separate tasks for login, logout, validation).
• Favor integration and critical-path coverage over per-method unit tests.
• Avoid one test task per CRUD operation.
• Question whether simple functions need a dedicated test task.

If any test task is generated, restate this section verbatim or near-verbatim in that task's "Implementation Notes" so the executing agent applies it.

7. Dependency analysis

For each task, identify:

• Hard dependencies: tasks that MUST complete before this one can start.
• Soft dependencies: tasks that SHOULD complete for optimal execution.

A task B depends on A if B requires A's output or artifacts, modifies code created by A, or tests functionality implemented by A. Validate that the final dependency graph is acyclic.

8. Complexity analysis

For every candidate task, assign a complexity_score (integer 1–10) before writing any file. Base the score on these four dimensions:

Score	Skill breadth	Acceptance-criteria clarity	Integration surface	Decomposition depth
1–3	One well-known skill	Criteria are concrete and observable	None or a single file/module	No further split possible
4–5	One primary skill plus a familiar adjacent skill	Criteria are clear with few edge cases	One component or API boundary	Already atomic
6–7	Two distinct skills, or one skill with ambiguous requirements	Criteria need clarification or have multiple edge cases	Multiple components or contracts	Could still be split
8–10	Three or more skills, or cross-cutting design decisions	Criteria are vague, unknown, or depend on unresolved choices	Wide integration surface or external systems	Must be decomposed further

Pre-emission sanity rules — apply these before any task is written:

• 3+ skills assigned → split the task into smaller tasks, each with 1–2 skills.
• Vague acceptance criteria → sharpen them until they include at least one concrete, runnable verification step.
• Trivially small adjacent tasks → merge them into a single task.
• Score ≥ 8 → decompose further; do not emit as-is.
• Score 6–7 → either sharpen criteria or split; do not emit without an explicit reason.

Required frontmatter:

• Every emitted task MUST include complexity_score (integer 1–10).
• Optionally include complexity_notes when the score needs justification, such as "Ambiguous API contract" or "Decomposed from a higher-score parent".

Loop-back rule:

After applying split, sharpen, or merge, re-run dependency analysis and re-score the adjusted tasks. Repeat this loop no more than three times. If complexity is still unresolved after three passes, stop and surface the blocker to the user.

9. Allocate task IDs

Run scripts/get-next-task-id.cjs <plan-id> to obtain the first available task ID. Allocate subsequent IDs by incrementing in-process; do not invoke the script repeatedly. Use the unpadded integer in the task frontmatter id field and the zero-padded form ({padded-id}--{slug}) for the filename.

The slug derives from a short task title: lowercase, alphanumeric and hyphens only, collapsed, trimmed.

10. Emit the task files

Write each task to:

<root>/plans/<plan-dir-name>/tasks/{padded-id}--{slug}.md

Each file must conform to <root>/config/templates/TASK_TEMPLATE.md, including required frontmatter fields:

• id (integer)
• group (string)
• dependencies (array of task IDs, possibly empty)
• status — pending for new tasks
• created (YYYY-MM-DD)
• skills (array of 1–2 kebab-case skills)

Required additional frontmatter:

• complexity_score (integer 1–10, required on every emitted task)

Optional frontmatter:

• complexity_notes (string) — include when the score needs justification, such as "Decomposed from a cross-cutting parent task" or "Ambiguous API contract".
• execution_profile (string) — optional durable routing profile metadata. Omit it during initial task emission; the routing helper writes it only after validating the complete task-to-profile mapping.

The body sections (Objective, Skills Required, Acceptance Criteria, Technical Requirements, Input Dependencies, Output Artifacts, Implementation Notes) must be filled with task-specific content. Place detailed implementation guidance inside a <details> block under "Implementation Notes" — write it so a non-thinking LLM could execute the task from that section alone.

11. Validation checklist

Before declaring task generation complete, verify:

• Each task has 1–2 appropriate technical skills assigned and inferred from its objectives.
• Dependencies form an acyclic graph; no orphan or circular references.
• Task IDs are unique, sequential, and start from the value returned by get-next-task-id.cjs.
• Groups are consistent and meaningful.
• Every task's Acceptance Criteria includes at least one concrete, runnable verification step (command + expected output / observable signal), not a vague "works correctly".
• Every explicitly stated deliverable in the plan is covered.
• No redundant or overlapping tasks.
• Minimization applied (20–30% reduction target).
• Test tasks focus on business logic, not framework functionality.
• No gold-plating: only plan requirements are addressed.
• After writing the task files, run scripts/validate-plan-blueprint.cjs <plan-id> complexityScoresValid. Stop unless it prints yes; if it prints no, run scripts/validate-plan-blueprint.cjs <plan-id> invalidComplexityTasks to see which files are missing, non-integer, or out-of-range, fix them, and re-run. Every generated task must carry an integer complexity_score from 1 to 10.
• Add complexity_notes only when a score needs explanation (typically atomic tasks scoring greater than 4).

12. Route task execution

Read <root>/config/hooks/TASK_EXECUTION_ROUTING.md and follow its instructions together with this procedure:

1. Run scripts/route-task-execution.cjs profiles <plan-id> and interpret its JSON result. On no-config or disabled, routing is off: skip the remaining routing steps and continue. On invalid-config, stop and surface the errors to the user — do not generate the blueprint.
2. Classify every task in the plan's tasks/ directory against the configured profile descriptions. For tasks generated in this run, use the task content already in your context — objective, acceptance criteria, technical requirements, skills, and complexity_score; do not reread the emitted task files to reconstruct information you already hold. If the plan carried task files from an earlier generation run, read those files (and only those) to classify them — the mapping must cover every task in the plan. Assign each task ID exactly one configured profile name. Never invent a profile name, model, or harness.
3. Write the complete task-ID-to-profile mapping as a JSON object to a temporary file, for example {"1": "routine", "2": "demanding"}.
4. Run scripts/route-task-execution.cjs apply <plan-id> <mapping-file>. The helper validates the mapping (every task exactly once, only configured profiles), writes one execution_profile frontmatter field per task, and verifies the written files. Target selection and resolver execution happen later at task dispatch, never during generation.
5. On routed, delete the temporary mapping file and continue. On any failure result (invalid-assignments, invalid-tasks, routing-failure, infrastructure-failure), stop and surface the JSON errors to the user. Never proceed to blueprint generation with partially routed tasks.

Profile names are durable routing labels. Persist them only through the helper's execution_profile field; never hand-write a concrete execution target into task frontmatter or task bodies.

13. Run the POST_TASK_GENERATION_ALL hook

Read <root>/config/hooks/POST_TASK_GENERATION_ALL.md and follow its instructions. Run it only after routing succeeded or reported routing off. This typically requires:

• Appending an Execution Blueprint section to the plan document, including a Mermaid dependency diagram and explicit phase groupings (Phase 1 contains zero-dependency tasks; each subsequent phase contains tasks whose dependencies all live in earlier phases). Use <root>/config/templates/BLUEPRINT_TEMPLATE.md for structure.

14. Emit the structured summary

Conclude with exactly this block as the final output:

---
Task Generation Summary:
- Plan ID: [numeric-id]
- Tasks: [count]
- Status: Ready for execution

The summary is consumed by downstream automation; keep the format exact.

Failure Modes

• No strikethroo root found. Stop, instruct the user to initialize the project. Do not write any files.
• Plan ID does not resolve. Stop and surface the script's stderr to the user. Do not guess a different ID and do not write any files.
• User declines to clarify a blocking ambiguity. Mark the affected tasks with status: "needs-clarification" and document the open question in the task's "Implementation Notes". Do not invent answers.
• A helper script fails unexpectedly. Surface stderr to the user and stop — do not fall back to manual ID allocation or path discovery.
• Execution routing fails. Surface the routing helper's JSON errors and stop before blueprint generation. Do not guess profile assignments, do not hand-write execution_profile or execution frontmatter, and do not continue with partially routed tasks.