feat(agents): enforce aggregate agent budgets

[joshua-berri]

TLDR

Problem this solves:

• Agent budgets needs an integrated gateway workflow

How it solves it:

• Configures an aggregate budget on each agent
• Enforces budgets during admission and deferred accounting
• Preserves budget windows across resets and concurrent settlement

Depends on #43149; followed by #43230. Reconciles existing implementation and review fixes; no additional feature scope

Diff sizes are being refreshed after the reviewed repairs

User Flow

Before: the complete workflow below is unavailable at this PR's parent

1. Configure an agent budget through the existing agent API
2. Call the agent and inspect its accumulated spend
3. Exhaust its budget and observe denied requests
4. Reset the budget window and verify new requests and delayed charges

After: this PR provides the workflow through the existing gateway surfaces

1. Configure an agent budget through the existing agent API
2. Call the agent and inspect its accumulated spend
3. Exhaust its budget and observe denied requests
4. Reset the budget window and verify new requests and delayed charges

Relevant issues

Original complete implementation: #43149

Affected release

Linear ticket

Pre-Submission checklist

• Existing implementation and review fixes are preserved
• Meaningful tests remain with their owning behavior
• Devin verification at the reconciled head is complete
• Required CI and completed Codecov pass
• Greptile, Veria and Bugbot reviews have no actionable findings
• User has accepted the hosted UI workflow

Screenshots / Proof of Fix

Matched live gateway checks used a controlled A2A upstream. With a zero-budget path target and a different target in the JSON model field, the earlier edb6e0f gateway allowed three requests and contacted the upstream

At c49737d, all four conflicting-target cases returned 422 before contacting the upstream. Reversing the targets allowed the request and billed the path-selected agent. Eight focused target-resolution tests passed; the same tests exposed two behavioral failures on the parent

The current restack preserves this PR's own patch exactly. Previous-tip Codecov covered 161/161 reported changed executable lines, Greptile scored 5/5 and Veria reported no open concerns. Current-tip hosted results remain pending

Type

New Feature

Caveats

• Part of the four-PR identity, budgets, SCIM and delegated-team stack; do not release the complete feature without fix(agents): preserve delegated team authorization context #43459
• Current-tip verification and reviews remain pending; fresh Bugbot review is waived due to team quota
• Optional selector cleanup is excluded and saved as an unpublished follow-up

Final Attestation

• Current-tip CI, coverage and review gates are complete

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[joshua-berri]

@greptileai Please review 39ddd00: the split preserves identity boundaries, isolates feature dependencies, and adds regression coverage for confirmed authorization gaps.

[joshua-berri]

bugbot run

Please review 39ddd00, including authorization boundaries, schema compatibility, failure handling, and regression tests for this split PR.

[greptile-apps]

Confidence Score: 5/5

[Critical risk] Adds database schema and budget enforcement for agents.

The PR appears safe to merge based on this review; no new actionable finding or outstanding previous finding remains.

Summary

This PR adds aggregate agent budgets across configuration, admission, spend settlement, window resets, and the dashboard. The only change since the previous review reformats agent rate-limit descriptor construction without changing its behavior. The previously raised cold-counter database-read finding was explicitly withdrawn and is not an open concern.

_{Reviews (13) · Last reviewed commit: "fix(agents): bill the path target on exp..."}

[cursor]

Stale Bugbot comment from a previous run.

[joshua-berri]

@greptileai Please review a67e88a: trusted billing, duration validation, budget-skip handling, and reset repairs. Shared reset concurrency limitations remain explicitly documented.

[joshua-berri]

bugbot run

Please review a67e88a for verified billing, budget-skip behavior, validated durations, and reset repairs; documented concurrency caveats remain open.

[cursor]

Stale Bugbot comment from a previous run.

[joshua-berri]

Budget counters now follow separate windows. Two-worker Redis testing preserves concurrent spend and late settlement. @greptileai and @cursor review please recheck 1d67236

[joshua-berri]

@greptileai @veria-ai Please review 5b64f51. This reconciles the approved three-PR stack and prior fixes. Verify this complete scope against its immediate parent.

[joshua-berri]

bugbot run

Please review 5b64f51, including prior findings and cross-feature interactions. The reconciled stack remains draft pending verification and user acceptance.

[veria-ai]

PR overview

All previously flagged issues have been addressed. No open security concerns remain on this pull request.

Security review

No open security issues remain on this pull request.

Fixed/addressed: 4 · PR risk: 0/10

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 5b64f51. Configure here.}

[cursor]

Lifetime budgets inherit window spend

Medium Severity

_budget_write only clears spend when a new budget_reset_at is stamped. Disconnecting a budget or switching it to a lifetime cap (no duration) sets spend_window to None but leaves the current window balance in place. That leftover window spend is then treated as lifetime usage, so a later lifetime cap can deny immediately even though prior reset windows were already zeroed away.

 

^{Reviewed by Cursor Bugbot for commit 5b64f51. Configure here.}

[joshua-berri]

@greptileai @veria-ai Please review the current tip and confirm resolution of the focused findings. Actor/target token settlement now reuses recorded reservation scopes; delegated tests use module-boundary isolation. No unrelated features or selector cleanup were added. Fresh verification runs in parallel with CI

[joshua-berri]

@greptileai @veria-ai Current tip c49737d applies the exact Ruff formatting correction only; behavioral patches are unchanged. Focused delegated tests168pass, rate settlement2pass, and actual Entra before/after confirm grant-team model, route and budget denial. Please confirm current-tip review outcomes; remaining live controls and CI are pending

[joshua-berri]

@greptileai @veria-ai Please review 91bc86a: own patch unchanged after restack. Live target-budget denial and focused regression tests pass.

[joshua-berri]

@greptileai Please refresh the summary after your explicit withdrawal of the cold-counter database-read finding. Recovery behavior and the reviewed commit remain unchanged.