Skip to content

fix(agent): make budget exhaustion visibly incomplete - #49309

Closed
spiky02plateau wants to merge 1 commit into
NousResearch:mainfrom
spiky02plateau:fix/budget-exhaustion-finalizer
Closed

fix(agent): make budget exhaustion visibly incomplete#49309
spiky02plateau wants to merge 1 commit into
NousResearch:mainfrom
spiky02plateau:fix/budget-exhaustion-finalizer

Conversation

@spiky02plateau

@spiky02plateau spiky02plateau commented Jun 19, 2026

Copy link
Copy Markdown
Contributor

Summary

Budget-exhausted turns now surface as visibly incomplete instead of letting a fluent model summary pass as a verified final answer. Both max_iterations and shared IterationBudget exhaustion return completed=false with structured budget metadata, and the final visible response is wrapped in runtime-owned wording that says the task is not verified complete and can be resumed with continue.

This keeps downstream behavior honest without introducing startup-context optimization or a new resume system: gateway resume markers stay uncleared, cron treats the run as failed, and final-send suppression no longer hides the deterministic incomplete-state warning after streamed/interim content.

Design notes

Concern Decision
Completion truth budget_exhausted=true, failure_reason="budget_exhausted", budget_used, and budget_max are emitted on every finalizer result.
User-facing response Model-authored summary text is demoted under “Progress summary from the model”; runtime wording remains the completion authority.
Gateway delivery Budget-exhausted responses bypass normal “already streamed” suppression so the warning reaches chat even after interim content.
Scope discipline The budget fix is independent of startup context audit code/config and does not add provider/model or auto-resume behavior.

Tests

  • python -m pytest tests/agent/test_turn_finalizer_budget_exhaustion.py tests/run_agent/test_turn_completion_explainer.py tests/gateway/test_restart_resume_pending.py tests/cron/test_scheduler.py -q -o 'addopts='
  • python -m py_compile agent/turn_finalizer.py agent/chat_completion_helpers.py gateway/run.py run_agent.py tests/agent/test_turn_finalizer_budget_exhaustion.py tests/run_agent/test_turn_completion_explainer.py tests/gateway/test_restart_resume_pending.py tests/cron/test_scheduler.py

Post-Deploy Monitoring & Validation

  • Validation window: first 24 hours after deployment; owner: Hermes ops.
  • Healthy signals: exhausted turns include budget_exhausted=true, completed=false, and user-visible text starting with Iteration budget exhausted; cron runs with exhausted budgets record failure rather than ok.
  • Log checks: search gateway/agent logs for iteration_budget_exhausted, max_iterations_reached, budget_exhausted, and Suppressing normal final send around exhausted turns.
  • Failure signals: budget-exhausted turns marked completed=true, cron jobs marked successful after exhausted turns, or chat users receiving only interim/tool-progress text without the deterministic warning.
  • Rollback trigger: any increase in normal successful turns receiving budget warnings, or plugin/gateway delivery errors caused by the new result metadata.

Compound Engineering
gpt--5.5

@alt-glitch alt-glitch added type/bug Something isn't working comp/agent Core agent runtime: loop, agent_init, prompt builder, context-compression, responses endpoint P1 High — major feature broken, no workaround labels Jun 19, 2026
@teknium1

teknium1 commented Jun 21, 2026

Copy link
Copy Markdown
Collaborator

Thanks @spiky02plateau. The core data-loss in #8049 is now fixed on main via #50009 (merge commit 9f67ba1): the unguarded cleanup chain in finalize_turn no longer lets a trajectory/resource/SQLite failure escape run_conversation and discard the partial response.

This PR goes considerably further than the reported bug — ~670 lines reworking completion semantics (budget_exhausted metadata, runtime-owned "incomplete" wording, gateway final-send suppression bypass, cron PARTIAL handling). That's a meaningful design direction, but it's a separate feature from the silent-exit fix #8049 asked for, and it doesn't itself guard the cleanup chain. We'd rather not couple the budget-completion-truth redesign to this P1 bug.

Closing against #8049 since the data-loss is resolved. If you want to pursue the "visibly incomplete budget" behavior, please open a focused PR scoped to just that and we'll review it on its own merits — thanks for the thorough write-up.

@teknium1 teknium1 closed this Jun 21, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

comp/agent Core agent runtime: loop, agent_init, prompt builder, context-compression, responses endpoint P1 High — major feature broken, no workaround type/bug Something isn't working

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

3 participants

@spiky02plateau @teknium1 @alt-glitch