Run Agentic Dse

Execute rounds

1. Read the current directive, architecture decisions, population, lineage, and contributions. Validate their benchmark, policy, and numerical-contract consistency. Check required tools and input files without launching expensive work just to inspect status.
2. Main selects parents from immutable, evidence-backed candidates and defines one bounded assignment per role. With zero parents, use three explicitly labeled seed assignments; with one parent, Innovator uses a single-parent seed variant. Only call an operation crossover when it actually has two parents.
3. For each actual DSE round, create fresh Explorer, Exploiter, and Innovator workers using prompts/explorer.md, prompts/exploiter.md, and prompts/innovator.md. Include the shared references/worker-workflow.md, absolute input/output paths, identities, numerical contract, policy, allowed clocks, parents, goals, and total attempt budget. Batch roles if concurrency is limited; do not omit a role or start nested workers.
4. Workers run only policy-permitted stages in their own workspaces. Default to three total attempts per assignment: initial attempt plus two retries, shared across all clocks, edits, and stages. Use a different explicit task budget when supplied.
5. Wait for all three workers to finish. Main checks reports against the candidate input digest and stage receipts, functional pass evidence, clock and metric units, and the selected policy. An exit code or synthesis estimate alone is not formal acceptance.
6. Before reusing any worker workspace, archive its result, source/header/test data, configuration, input manifest, stage receipts, and reports under a new round/candidate path. Use src/artifacts.py or an equivalent checked snapshot. Never overwrite an existing archive. Pareto and parent references must point to immutable archives, not mutable workspace files.
7. Main alone updates state, lineage, contributions, and learned knowledge. Admit only fully validated formal successes to the formal Pareto population. Keep cosim-only results in a separate provisional population; do not promote them or mix their metrics into formal HV.
8. If a fixed, benchmark-specific HV configuration is available, run src/hypervolume.py. Record the actual algorithm, configuration ID, population, reference point, result status, and value. Keep missing metrics as null; never substitute a platform limit or a guessed clock.
9. Apply the stopping mode below. For the next authorized round, archive the previous round and retire its workers before creating fresh identities. No per-round reapproval is needed within the existing scope and budget.
10. Report actual rounds, accepted/rejected/provisional candidates, best validated metrics, preserved artifacts, remaining failures, and the termination reason. Task completion and engineering success are separate: a requested cosim-only or fixed-budget run can finish without producing a formal feasible design.

Stopping and completion

• Fixed rounds: run dse <benchmark> N means exactly N rounds unless a user stop, exhausted execution limit, unavailable essential capability, or authority/input blocker prevents continuation. Without N, run one round. A low HV improvement alone does not truncate a fixed-round request.
• Until converged: requires a finite maximum-round budget. Use an existing project budget; if none exists, obtain that consequential limit before starting an open-ended run. The default criterion is three consecutive relative improvements below 2%, computed over comparable, valid, nonempty formal populations with the same HV configuration ID.
• Compute relative improvement against the preceding positive HV. Missing, empty, invalid, zero-baseline, unavailable, or differently configured values do not count as low improvement. A decrease requires consistency/regression diagnosis, not a convergence claim.
• When convergence cannot be measured, continue safe work within the authorized budget using available engineering evidence; report that convergence is undetermined.
• Distinguish fixed_rounds_complete, converged, budget_exhausted, blocked, and user_stopped. Separately record whether there is a formal_feasible_point, provisional_only, or no_feasible_point. Do not call a plateau, failed run, or incomplete report set successful closure.

Status and recovery

Read only the evidence that exists. The HV helper returns structured unavailable/error states and does not install dependencies or write state. Its absence or missing optional numerical libraries must not hide the benchmark, completed rounds, available points, failures, or missing reports.

Use host-provided tools and respect their sandbox/approval decisions. If an essential command is denied, complete unaffected in-scope work and report the precise blocker. Preserve the best candidate and archive before any requested reset; never reset or modify an unrelated project or installed skill.