Delete the turn-budget mechanism and deprecate task() (CL-7004)

[TheGreatAxios]

No max turns anywhere. Net 147 insertions / 1436 deletions (−1289) across 71 files.

Peer coding agents have no turn cap — one resolves max-steps to Infinity unless an operator opts in, the other has no turn-count constant in its loop. Both end a run on the model's own finish signal, an operator interrupt, or a hard error. Ours ended runs on a turn count that truncated real work: a measured healthy lane used 59 assistant turns, 58 carrying tool calls, with a clean final report — which explore (35), critique (45) and every 40 would have killed; intern (20) three times over.

Removed

• Director packages — nudge.maxTurns from all 17
• Resolution chain — resolveSubAgentMaxTurns, resolveDefaultSubAgentMaxTurns, clampSubAgentMaxTurns, validateTaskMaxTurns, Settings.subagentMaxTurns and its schema/load/save plumbing
• Stops — subAgentTurnLimitExceeded, the turn-budget stop reason and report text, TURN_BUDGET_STOP_AFTER_DISPATCHES, shouldStopTurnBudgetRedispatch, the turn-budget branch of appendSubAgentParentHints
• Wrap-up — thrashForceReport, evaluateThrashStop, report-forced, REPORT_FORCED_WRAP_UP_NUDGE
• Tool args — maxTurns on task() and spawn_agent, plus the profileMaxTurns threading added in Thread the director package turn budget through spawn_agent (CL-7001) #622
• Config — AgentProfile.maxTurns and its arktype schema entry
• Diagnostics — InterventionRecord.state.maxTurns and the forensics "early stop" column that read it
• Eval harness — the global soft turn-budget rubric: evaluateSoftBudget, overBudget, budgetError, OVER_BUDGET, --max-turns, and maxTurns from all four tier cases

Why the eval rubric went too

It never stopped an agent — it failed a case after the fact for exceeding 15/25/30/40 turns. Since the healthy reference run took 59, the instrument we planned to measure this release with would have scored a correct, complete run as a failure at every tier. --agent-timeout-ms is retained as the per-eval bound, and turnsUsed is still captured and reported — just no longer part of pass/fail. Verdict is now agentExitCode === 0 && verify.exitCode === 0 && requireBehaviorCheck.ok, which still fails closed on a broken run.

On report-forced

Deleted rather than kept as a plain turn-count nudge. Without maxTurns there is no denominator for "near budget", so a nudge keyed to an arbitrary turn number would be a new limit wearing guidance clothing.

task() deprecation

Marked deprecated in both the model-visible tool description and code, pointing at spawn_agent/wait_agents. Not removed — much still routes through it. This is what CL-6942 specified: available during migration, then removed.

Prompt text

The bullet teaching the model to pass maxTurns and re-dispatch with a higher one is replaced with guidance on the real judgement: if a worker comes back without finishing, change the brief rather than repeating it — narrow the scope, name the files, or state the done-when more sharply. Prompts guide output quality; code does enforcement.

Also swept

Stale comments in intervention-log.ts, session-store.ts, fleet-report.ts, report.ts and tool-execution-watchdog.ts asserting deleted mechanisms. explore's prompt claim that "the default budget is real" is gone. 15 package tests asserting specific budget numbers deleted rather than weakened.

Gate

bun run check green: 5278 pass, 0 fail — including the vendor reactor snapshot test that fails on main.

Follow-up filed: CL-7006 — stopReasonFromReport still parses stop reasons out of report prose, a second source of truth #610 was meant to end.

[linear-code]

CL-7004