[Kemalau]

Summary

• Fixes Extend agenda token accounting to the auto_research execution path #42. auto_research workers now enter agenda_scope before they spend, so execution-path LLM calls are attributed to the agenda that authorized them.
• A scoped legacy call_llm checks the agenda cap before the provider request, writes agenda_token_ledger, and bumps token_spent. NULL, 0, and negative budgets stay accounting-only; crossing a positive cap flips the agenda to paused_budget, and raising the budget unpauses on the next check.
• Grant-backed role calls keep the existing reservation clock: token_spent still moves when the parent grant settles, so usage is not counted twice. Each settled sub-reservation now appends a ledger row immediately, and in-flight grant usage counts toward the cap before that settlement.

Test plan

• python -m unittest tests.test_agenda_execution_budget
• python -m unittest tests.test_ingestion_extraction_retry.ScopedAttemptRegressionTests tests.test_ingestion_extraction_retry.RoleCheckpointRegressionTests tests.test_llm_client
• Confirm a live execution grant writes agenda_token_ledger rows as each role call settles, and that token_spent still moves only when the parent grant settles

Made with Cursor

[cla-assistant]

All committers have signed the CLA.

[Kemalau]

I have read the CLA.md and agree to the terms.