st-full-workflow

st-full-workflow

Critical Rule

Execute all three steps sequentially without waiting for user input between them.

Inputs

The user supplies the work order conversationally.

Operating Procedure

Step 1: Plan Creation

1. Locate the strikethroo root

Run scripts/find-strikethroo-root.cjs from the user's working directory. If it exits non-zero, stop and ask the user to run npx strikethroo init.

Treat the path it prints as <root> for every subsequent step.

Delegated execution workers skip this step and do not emit update notices.

Run scripts/check-for-updates.cjs "<root>" as a separate command using this skill's script path. Keep the user's working directory. Read the one JSON line on stdout. When notice is present, retain its exact text for the final user-facing response. Never pause for permission and never run an update.

2. Load project context

Read <root>/config/STRIKETHROO.md for this project's directory conventions. Read <root>/config/hooks/PRE_PLAN.md and execute its instructions before proceeding. Read <root>/config/templates/PLAN_TEMPLATE.md; the plan must conform to it.

Also read <root>/config/shared/anti-rationalization.md.

3. Analyze the work order

Identify:

• Objective.
• Scope and explicit boundaries.
• Success criteria.
• Dependencies, prerequisites, blockers.
• Technical requirements and constraints.

4. Clarification loop

If any critical context is missing, ask the user targeted questions. Keep looping until you have no further questions. Explicitly confirm whether backwards compatibility is required. Never invent answers; never paper over a missing answer.

If the user declines to clarify a blocking question, stop and report the plan as needing clarification. Do not produce a partial plan.

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

You catch yourself thinking…	The binding rule
"I can reasonably assume the answer."	An assumption is not an answer. Ask the question; never invent answers.
"Asking again is annoying."	A question the user can decline is recoverable; a silent wrong assumption is not. Ask.
"The user implied it, so it's settled."	An implication is not a confirmation. Surface it as a question and get an explicit answer.

5. Allocate the next plan ID

Run scripts/get-next-plan-id.cjs for the next available plan ID. Compute its zero-padded form for the directory name ({padded-id}--{slug}) and use the unpadded integer in the plan frontmatter.

6. Emit the plan

Write the plan to <root>/plans/{padded-id}--{slug}/plan-{padded-id}--{slug}.md, conforming to <root>/config/templates/PLAN_TEMPLATE.md in both frontmatter and sections. Include no time estimates, task lists, or code samples; those belong to the task-generation step.

Derive <slug> from the plan summary: lowercase, alphanumeric and hyphens only, collapsed, trimmed.

7. Run post-plan hook

Execute <root>/config/hooks/POST_PLAN.md.

8. Emit the Step 1 structured summary

Conclude Step 1 with exactly this block (a retained update notice, when present, follows after the workflow's final summary):

---

Plan Summary:
- Plan ID: [numeric-id]
- Plan File: [absolute-path-to-plan-file]

Parse the Plan ID value from this output and pass it to Step 2.

Step 2: Task Generation

Using the plan ID from Step 1:

1. Resolve the plan

Run scripts/validate-plan-blueprint.cjs <plan-id> planFile for the plan file's absolute path. Another field name prints that field instead.

If the script exits non-zero, surface its stderr to the user and stop the workflow. Do not guess a different ID.

2. Load project context

Read these files, in order:

• <root>/config/STRIKETHROO.md — project and directory conventions.
• The plan body at the path returned by step 1 — the contract for what tasks must exist.
• <root>/config/templates/TASK_TEMPLATE.md — the schema every task file must match.
• <root>/config/shared/anti-rationalization.md — apply in step 3.

3. 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.
• Every task corresponds to an explicit requirement, never a nice-to-have, a "best practice", or future extensibility the plan does not mention.
• Each task has a unique, non-overlapping purpose.
• Keep error handling inside the task that owns the behavior it guards.

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.

4. 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".

Infer skills in kebab-case from the task's technical requirements. Use one skill for a single-domain task and two for complementary domains, such as ["api-endpoints", "database"]. Three or more means the task must be broken down further.

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

Write tests for:

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

Do not write tests for:

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

Combine related test scenarios into one task ("Test user authentication flow", not separate tasks for login, logout, and validation). Favor integration and critical-path coverage over per-method unit tests. Never create one test task per CRUD operation.

Copy these rules into the "Implementation Notes" of every test task you generate.

6. Dependency analysis

Task B depends on task A when B requires A's output or artifacts, modifies code created by A, or tests functionality implemented by A. Record it as a hard dependency when B cannot start before A completes and as a soft dependency when B merely runs better after A. Validate that the final dependency graph is acyclic.

7. Complexity analysis

For every candidate task, assign a complexity_score (integer 1–10) before writing any file. Read references/complexity-rubric.md before scoring; it defines the four dimensions each band is judged on. Then apply these rules:

• Score ≥ 8 → decompose further; do not emit as-is.
• Score 6–7 → sharpen or split; do not emit without an explicit reason.
• Vague acceptance criteria → sharpen them into a concrete, runnable verification step.
• Trivially small adjacent tasks → merge them.

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

8. Allocate task IDs

Run scripts/get-next-task-id.cjs <plan-id> once for the first available task ID, then allocate the rest by incrementing in-process. Use the unpadded integer in the task frontmatter id field and the zero-padded form in the filename. The slug derives from a short task title: lowercase, alphanumeric and hyphens only, collapsed, trimmed.

9. 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. Add complexity_notes only when the score needs justification. Never write execution_profile; the routing helper writes it.

Fill every body section with task-specific content. Place detailed implementation guidance inside a <details> block under "Implementation Notes", written so a non-thinking LLM could execute the task from that section alone.

10. Validation checklist

Before declaring task generation complete, verify:

• Every explicitly stated deliverable in the plan is covered by a task.
• No two tasks overlap in purpose.
• Dependencies form an acyclic graph, with no orphan or circular references.
• Groups are consistent across the plan.
• 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.

11. 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>. 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. Assign every task in the plan's tasks/ directory exactly one configured profile name. Classify the tasks generated in this run from the content already in your context. If the plan carried task files from an earlier generation run, read those files to classify them.
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>. Target selection happens 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 continue to blueprint generation with partially routed tasks.

Never hand-write a concrete execution target into task frontmatter or task bodies.

12. Run the POST_TASK_GENERATION_ALL hook

Read <root>/config/hooks/POST_TASK_GENERATION_ALL.md and follow its instructions, using <root>/config/templates/BLUEPRINT_TEMPLATE.md for the Execution Blueprint structure. Run the hook only after routing succeeded or reported routing off.

Step 3: Blueprint Execution

Using the plan ID from Step 1:

1. Resolve the plan

Run scripts/validate-plan-blueprint.cjs <plan-id> planFile for the plan file's absolute path. Another field name prints that field instead.

If the script exits non-zero, surface its stderr to the user and stop the workflow. Do not guess a different ID.

Run scripts/validate-plan-blueprint.cjs <plan-id> planDir and treat the printed path as <plan-dir>.

2. Validate tasks and blueprint existence

Run scripts/validate-plan-blueprint.cjs <plan-id> taskCount and scripts/validate-plan-blueprint.cjs <plan-id> blueprintExists.

3. Auto-generate tasks and blueprint if missing

If taskCount is 0 or blueprintExists is no:

Notify the user: "Tasks or execution blueprint not found. Generating tasks automatically..."

• Execute the full task generation procedure from Step 2 for this plan ID.
• Re-run the planFile, planDir, taskCount, and blueprintExists queries to refresh the resolved paths and counts.

If the plan still has no tasks or no blueprint, stop and report failure.

4. Optionally create a feature branch

Run scripts/create-feature-branch.cjs <plan-id> once before phase execution. A skip is not a failure. Continue on the current branch, and never create the branch by hand. Uncommitted or untracked changes are permitted only inside the repository-root .ai/strikethroo subtree. An error result halts execution; report it.

Then run scripts/capture-base-commit.cjs <plan-id> once to record the commit the review gate diffs against. A skipped result continues execution and means the review gate will skip. Only an error result halts.

5. Load project context and execution blueprint

Read these files, in order:

• <root>/config/STRIKETHROO.md
• The plan document at the path from step 1, including its Execution Blueprint section, which defines the phase groupings and task dispatch order.
• <root>/config/shared/verification-gate.md and <root>/config/shared/anti-rationalization.md, both applied in the phase loop below.

6. Execute phases in order

Use an internal task or todo tracker to monitor progress. For each phase defined in the Execution Blueprint:

6a. Phase pre-execution

Run scripts/check-phase-readiness.cjs <plan-id> <phase-number>. If the script exits non-zero, halt the phase and report the blocking issues before continuing.

Read <root>/config/hooks/PRE_PHASE.md and execute its instructions before starting the phase.

6b. Task dispatch

Identify all tasks scheduled for this phase whose dependencies are fully satisfied. Read <root>/config/hooks/PRE_TASK_ASSIGNMENT.md and follow its instructions for agent selection before dispatching tasks.

Resolve every selected task's execution route first. Invoke one resolver per selected task simultaneously in a single parallel tool operation:

scripts/dispatch-task-execution.cjs resolve <task-file> <current-harness> <workspace> <plan-id> <task-id>

Resolvers never launch external processes. After interpreting all route results, issue every external execution and every sub-agent together in one parallel tool operation. External execution uses:

scripts/dispatch-task-execution.cjs execute <handoff> <task-file> <current-harness> <workspace> <plan-id> <task-id>

<current-harness> is the exact supported harness identifier running this skill; <workspace> is the project working directory.

Interpret the one-line JSON result and act on its kind exactly once:

kind	Required action
native-default	Dispatch natively with no execution-setting prose.
native-override	Dispatch natively, explicitly requiring the exact returned model. Require the returned reasoningEffort only when that property is present.
external-override	Run the execute command with the returned handoff, then read its result against this same table.
fallback	Nothing launched. Record the returned reason and detail visibly, then dispatch natively with no execution-setting prose.
launched-success	The external process exited zero. Do not dispatch natively; review status and evidence as you would for a native agent.
launched-failure	A failed task. Set its status to failed and run <root>/config/hooks/POST_ERROR_DETECTION.md. Never retry it natively.
infrastructure-failure	Handle exactly as the preceding row.

Handoff rules:

• Pass the exact opaque handoff string the resolver returned for that task. Never reconstruct one.
• Never reuse a handoff for another task, and never rerun resolution after launches begin.

Deploy all remaining native sub-agents simultaneously. Each sub-agent must read and execute <root>/config/hooks/PRE_TASK_EXECUTION.md before any implementation work, then execute the task and update its status. Do not run scripts/check-for-updates.cjs; delegated workers do not consume update notices.

6c. Phase completion verification

Ensure every task in the phase has status completed and collect its outputs. Do not accept a subagent's report of success as proof. Apply the evidence gate in <root>/config/shared/verification-gate.md before marking the phase complete.

6d. Phase post-execution

Read <root>/config/hooks/POST_PHASE.md and execute its instructions. Do not proceed to the next phase until this hook succeeds.

Update the phase status to completed in the plan's Execution Blueprint section.

Repeat for the next phase until all phases are complete.

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

You catch yourself thinking…	The binding rule
"The subagent reported success, so the task is done."	A report is a claim, not evidence. Apply the verification gate before marking the phase complete.
"The tests probably pass."	"Probably" is a red flag. Run the proving command, read its output and exit code, then state the result.
"I'll verify later, after the next phase."	A phase is not complete until POST_PHASE.md succeeds against verified evidence. Verify now; do not advance on an unverified phase.

7. Post-execution validation

Read <root>/config/hooks/POST_EXECUTION.md and execute its instructions. If validation fails, halt execution. The plan remains in plans/ for debugging.

Before declaring execution complete, apply the evidence gate in <root>/config/shared/verification-gate.md to the plan's Success Criteria and Self Validation steps.

Run the code review gate

After POST_EXECUTION.md reports green, follow the st-code-review skill and run its bundled mechanism:

code-review.cjs <plan-id> <current-harness>

Resolve code-review.cjs from the st-code-review skill's sibling scripts directory and pass the exact supported harness identifier running this skill. If the st-code-review skill is not installed, record that outcome in the execution summary and continue to summary and archival.

Handle the one JSON line it prints on stdout, in this order:

1. Copy it verbatim into the execution summary's review outcome. Do not reformat it or omit fields.
2. Follow its top-level action. If it is halt, stop, report the top-level detail, and emit the terminal Execution Summary with the review result. If it is continue, proceed to the execution summary and archival.
3. Only when verdict.kind is review-recorded, read <plan-dir>/review/review.xml and <plan-dir>/review/findings.json, then decide which findings to act on. severity and confidence are advisory labels, not instructions.

Never report an uncertified review as clean.

Hard rules:

• The review runs once. Do not re-run the gate to check a fix.
• Detection uses the reviewer route. Dispatch each fix on the implementer route; the reviewer does not fix its findings.
• After any fix, re-run POST_EXECUTION.md in full before declaring execution complete. The earlier green result no longer applies.

8. Append execution summary

Append an execution summary section to the plan document, filling every field of <root>/config/templates/EXECUTION_SUMMARY_TEMPLATE.md. Include a Code Review entry using the result’s codeReview field, or Failed with the reason no result was produced. Under Noteworthy Events, always record the review gate's JSON line verbatim, then which findings you acted on versus ignored and why.

9. Archive the plan

Move the completed plan directory from <root>/plans/<plan-folder> to <root>/archive/<plan-folder>, preserving the entire folder structure. If the move fails, log the error but do not fail the overall execution.

Failure Modes

• Plan directory already exists for the allocated ID in Step 1. Re-run the next-plan-id script and retry once. If the conflict persists, stop and report.
• Execution errors. If a task fails, read <root>/config/hooks/POST_ERROR_DETECTION.md, document the error in Noteworthy Events, halt the phase, and request user direction before continuing.

Execution Summary

Conclude with exactly this block (a retained update notice, when present, follows separately):

---
Execution Summary:
- Plan ID: [numeric-id]
- Status: [Archived or Halted]
- Location: [absolute path to current plan directory]
- Code Review: [codeReview from the review result, or Failed with the reason no result was produced]
---

Emit this block even when the review gate halts execution. Use the actual status and location; never claim archival after a halt. Copy codeReview without dropping the harness, model information, or rejection reasons. Findings acted on do not turn Address into Pass because the review is not re-run. If execution stops before review, report Failed with "Review was not run" and the blocking reason.

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

When the retained update notice is present, append that exact sentence after the structured summary block, or after your final response when this skill emits no summary block. Nothing may follow the notice.