DESIGN_STRATEGIST.md §4. Replace the fixed-round hypothesis-directed
loop with a belief-driven strategist loop that runs the strategist
agent once per round and dispatches the leads it proposes.
New helpers on Orchestrator:
_budget_exceeded() hard budget caps (tool_calls,
wall_clock_minutes), complementing
strategist self-throttling.
_execute_strategist_lead(lead) dispatch one lead serially; the
next strategist round sees the
cumulative effect of this lead's
graph mutations.
_phase3_strategist_loop() main loop. Open round, run strategist,
exit on declare_complete or empty
proposals, otherwise dispatch each
lead, judge new phenomena, close round,
apply yield/budget checks.
_phase3_legacy_loop() fallback when strategist.enabled is
false. Identical to the
pre-DESIGN_STRATEGIST behaviour.
The run() entry point branches on strategist_cfg.enabled (default
true) and always follows up with _retry_failed_leads() + Gap
Analysis + mark_remaining_inconclusive() regardless of variant.
Orchestrator.__init__ also wires graph.budgets and
graph.run_start_monotonic from config so the budget_status tool
sees real numbers.
Integration tests use a mock strategist + mock workers to verify
declare_complete, propose_lead -> worker dispatch, zero-yield-streak
hard stop, and budget-cap-stops-the-loop.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
a103c17bdb