| name | itsol-execution-policy |
| description | Resolve and preserve cost-aware model, reasoning, delegation, parallelism, review, completion, and stop boundaries for ITSOL engineering work. Use after itsol-workflow-mode at task intake, before delegation, for long-running agent work, and whenever the user sets agent, budget, model, reasoning, review-round, or stop-point limits. |
ITSOL Execution Policy
A commit-only/repository-inspection follow-up or .itsol.md initialization is bounded administration. Reuse prior policy when applicable; do not create a new execution policy, agent budget, delegated review cycle, application QA, or completion gate solely for that operation.
Resolve this policy after itsol-workflow-mode. Keep resource limits independent from decision authority and planning gates.
Record and preserve:
execution_policy:
preset: economy | standard | deep | custom
policy_sources:
base: explicit-user-task-instruction | repo-default | agent-default
constraints: []
model_profile: economy | balanced | frontier
model_control: enforced | advisory
reasoning_profile: low | medium | high
reasoning_control: enforced | advisory
max_subagents: unlimited | 0..64
max_parallel: 0..10
max_review_rounds: 0..2
stop_after: <resolved named stage>
budget_escalation: forbidden | ask
done_when:
- <observable criterion with evidence>
Use standard when no explicit or repository policy exists. Preset model/reasoning intent is advisory by default: a harness-specific configured profile+role mapping may override the preset reasoning level. Set reasoning_control: enforced only for an explicit user or repository hard ceiling; never infer it from the preset name. standard and deep default to max_subagents: unlimited; do not invent a numeric agent-type ceiling. max_subagents counts distinct identities/types, not executions, so one type may handle several independent work items. Keep max_parallel: 3 as the execution-instance scheduling bound. A numeric identity ceiling is valid only when the user, an explicitly selected restrictive preset, or repository policy requests it. Apply platform constraints and explicit repository restrictions by tightening fields; never expand a resolved enforced ceiling automatically. Native model capability is always authoritative. Report advisory model or reasoning control honestly.
Do not set maxTurns or use a turn count as completion. Accept completed only after validating every done_when criterion and required evidence. Preserve partial, blocked, and failed results.
Only the main agent delegates. Count distinct child identities separately from execution instances, reuse stable work_item_id values for follow-ups, bound concurrency, keep one writer per file or semantic contract, and prohibit nested delegation. A required independent review that does not fit an explicit numeric policy ends incomplete; do not weaken it. An unlimited identity budget should select all required specialists and schedule them in bounded parallel batches without asking for budget expansion.
Read:
Propagate the full execution state through task context, plans, compaction, task packets, continuations, reviews, and final handoff.
