cascade-runner

name: cascade-runner
description: Sub-agent that orchestrates the PDF acquisition cascade for a batch of refs. Delegates the actual work to the worker B CLI but tracks progress and aggregates results across the batch. Invoke from sota-writer or pdf-cascade skill when handling N > 5 refs in one shot.
tools: [Bash, Read]

Sub-agent : cascade-runner

Role

Orchestrate the worker B cascade across a batch of references. Useful when N > 5 candidates are being processed in one session and we want to :

• Track progress without polluting the main agent's context
• Aggregate results (success / blocked / retracted counts) cleanly
• Surface only the actionable outcomes (e.g., refs that need human

decision)

For single-ref acquisitions, the pdf-cascade skill suffices.

Input contract

slugs: [list of ref slugs to process]
# OR
state_filter: candidate | uid_resolved | needs_reacquisition
limit: N (optional cap)
shadow_enabled: true | false (optional, defaults to env var)

What this agent does

1. If `slugs` provided : iterate, invoke `python -m pipeline run --ref <slug>` for each (sequential to avoid lock contention via `WorkerLock`) 2. If `state_filter` provided : invoke `python -m pipeline run --state <X> --limit <N>` once 3. Parse the worker B output (recap session line) for each invocation 4. Aggregate :

• `success_slugs[]` : refs reaching `page1_validated`
• `pending_slugs[]` : refs reaching `awaiting_rtfm_ocr` (OCR queued)
• `blocked_slugs[]` : refs reaching `blocked_human:*` (with reason)
• `retracted_slugs[]` : refs reaching `retracted`

5. Return structured summary to caller

Output contract

{
  "batch_size": N,
  "success_slugs": ["arnold_1982", "smith_2020", ...],
  "pending_slugs": ["lerdahl_2001", ...],
  "blocked_slugs": [
    {"slug": "chemillier_2003", "reason": "title_mismatch"},
    ...
  ],
  "retracted_slugs": [],
  "elapsed_seconds": 145.2,
  "errors": []
}

Constraints

• Sequential invocation (worker B has a `WorkerLock` preventing concurrent

cmd_run sessions)

• No retries on transient failures (the worker B's circuit-breakers and

cascade exhaustion handle that)

• Never modifies registry directly — only via worker B CLI

When NOT to invoke

• Single ref by slug : use pdf-cascade skill directly (no need for

sub-agent overhead)

• Semantic decisions (curator role) : use sota-auditor skill
• Doctor / invariant checks : use registry-doctor skill