cadence-first

Cadence-First Protocol

Assigns one of three tiers per block:

• Tier 1 Full — flow-first → library-first → plan-first → execute → ship-first
• Tier 2 Partial — library-first → plan-first → execute → ship-first
• Tier 3 Plan-focused — plan-first → execute → ship-first

Rules live in CADENCE_HEURISTICS.md (colocated). This SKILL is the executor; the manifest is the source of truth, read conditionally.

Does NOT execute the block itself. Decides cadence, then hands off.

---

Step 0 — Check task context

Are {log_dir} and {slug} known in the current session?

If yes → continue. If no → list latest task folders, ask which is current, set {slug} and {log_dir}. Cannot skip — without {log_dir} the artifact cannot be saved.

---

Step 1 — Parse invocation mode

1.1 Detect standalone vs batch

Signal	Mode
ARGUMENTS has task_id=X block_nums=[...]	Standalone
Invoked from generator context (parent inherits state)	Batch
No args, no context	Standalone (ask user)

1.2 Extract parameters

Standalone: parse task_id, block_nums from ARGUMENTS. Batch: inherit from parent. If missing → ask.

1.3 Anti-recursion guard

If task_id == {slug} (framework session's own task) → warn and default to abort. Prevents writing into the framework session's own task.

1.4 Verify task exists

Query the routing DB for id={task_id}. Empty → abort with "task not found".

---

Step 2 — Read sources (minimal by default)

2.1 Read target task_blocks

sqlite3 {routing_db} \
  "SELECT block_num, title, type, atom_type, status, recommended_cadence
   FROM task_blocks
   WHERE task_id='{target_task}' AND block_num IN ({block_nums_csv})
   ORDER BY block_num"

Defensive column check: PRAGMA table_info(task_blocks) | grep recommended_cadence. Column exists → write_to_db=true. Missing → markdown only.

2.2 Read target task.md

Extract task type/scope, ## Blocks table if present, HIGH-risk tags.