parallel-delegation

Parallel Delegation

This is the trigger-loaded parallel-lane law. Worktree lifecycle and Git rules remain in agenticloop/AGENTIC_LOOP.md.

Parallel Opportunity Scan

Trigger. Every authorized multi-task work unit (phase, group, milestone, epic, task set, or other bounded multi-task unit) receives one current scan after decomposition, once per meaningful readiness snapshot. Source plans, roadmaps, architecture documents, issue groupings, and maintainer proposals are inputs only; none is a scan result or parallel authorization.

Before selecting an execution order, the orchestrator records the scan in the canonical ## Concurrency Plan, an authorized task-record surface, or a single-writer coordination output. Do not duplicate it into every task or create a shared mutable findings ledger. The maintainer supplies per-task code/collision and joinability classifications through ## Parallel Safety; the orchestrator verifies required task, artifact, host, liveness, and join inputs, then records and routes the batch decision. The orchestrator does not originate or override a semantic/code classification.

Decide inventory completeness before ready count. A ready count over an inventory never proven complete is not an answer:

1. incomplete inventory or decomposition -> incomplete; never an eligibility answer, however few tasks are visible. Repair the evidence and rescan.
2. complete inventory, zero ready tasks -> no_eligible_work.
3. complete inventory, one ready task or no valid candidate pair -> not_currently_eligible, recorded as not currently eligible - <n> ready task(s) plus a rescan trigger.
4. only a complete, fresh, fully accounted inventory reaches parallel_candidates; then perform the assessment below.

Rescan when inventory membership or enumeration coverage, task carrier digests, base or dependency evidence, ready membership, task scope or ownership, shared design assumptions, base artifact, host capability, backend collision state, or a cross-lane finding changes.