feat: centralize AI monthly budget guard by Pigbibi · Pull Request #73 · QuantStrategyLab/AIAuditBridge

Pigbibi · 2026-07-11T16:33:49Z

Summary

add ai_budget_decision.v1 monthly provider/repo/task budget preflight with fail-closed freshness checks
add Codex primary/secondary 30/20/10% thresholds, 10% reserve, and atomic reservation ledger
make quota exhaustion deferred_budget; disable automatic paid-provider fallback by default
route gateway analyze/review/execute gates through budget preflight and document deterministic controls

Validation

python3 -m pytest tests -q (634 passed, 1 skipped, 16 subtests)
python3 -m ruff check .
actionlint
git diff --check

This PR is intentionally separate from Contract Oscillation Guard PR #72.

github-actions · 2026-07-11T16:36:45Z

🤖 Codex PR Review

🚫 Merge blocked: 3 serious issue(s) found in high-risk files

⚖️ Codex Review Arbitration

🚫 block: At least two current blocking findings remain valid against the PR diff. In service/openai_admin_usage.py, the unchanged logic still applies project_ids and optional api_key_ids only to usage_params, while cost_params for /organization/costs remains organization-wide; the snapshot returned to QuotaManager.check(..., budget_guard=True) still carries that value under organization_costs.total_cost, and AIBudgetGuard._usage() explicitly uses that nested total. That contradicts the new documented contract in docs/ai_autonomy_architecture.md that API-key budgets are scoped by provider/key-project scope, so unrelated org spend can still trigger false budget defers. In service/anthropic_admin_usage.py, the unchanged if not api_key_ids and not workspace_ids: gate still skips the cost report whenever filtered Anthropic scopes are configured, returning costs=None. QuotaManager.check(..., budget_guard=True) passes that snapshot into AIBudgetGuard.preflight(), and _usage() only accepts top-level cost fields or nested organization_costs/costs/usage fields, not the returned messages usage payload, so filtered Anthropic requests still fail closed as usage_cost_missing. The current service/ai_budget_guard.py finding does not hold because preflight() now carries fresh spend in aggregate_observed_usage and reserve() rechecks the hard limit with that value even when aggregate_scope == scope, but that does not clear the other two valid blockers. No current suggestion conflicts with the prior blocking findings; the prior OpenAI/Anthropic findings were about billing-timezone month boundaries, not these scope/cost-path issues.

🚫 Blocking Issues

These issues must be fixed before this PR can be merged:

1. 🟠 [HIGH] Logic in `service/ai_budget_guard.py`

preflight() writes observed_usage=0.0 into API budget decisions, but reserve() later rechecks the hard limit against decision["observed_usage"]. When a provider snapshot already shows non-zero spend and the budget entry has no provider_project_limit_usd, two concurrent reservations can both succeed against the same snapshot and push total monthly usage past the hard limit because the second reservation ignores the existing used_usd. (line 400)

Suggestion: Carry the numeric observed spend for the reservation scope in the decision object and use that value inside reserve() when recomputing observed + settled + reserved + requested <= hard_limit.

2. 🟠 [HIGH] Logic in `service/openai_admin_usage.py`

The new guard now enforces OpenAI monthly budgets from organization_costs.total_cost, but cost_params does not include the same project_ids or api_key_ids filters that are applied to the usage query. In a multi-project org, unrelated project spend will be counted against every scoped budget decision, causing false deferred_budget blocks for repositories that are still within their own project budget. (line 148)

Suggestion: Fetch or derive cost at the same scope as the usage snapshot (same project/API-key filters), or refuse to enable budget enforcement when only organization-wide OpenAI cost is available.

3. 🟠 [HIGH] Logic in `service/anthropic_admin_usage.py`

When CODEX_AUDIT_SERVICE_ANTHROPIC_ADMIN_API_KEY_IDS or ...WORKSPACE_IDS is configured, this function skips the cost report and returns costs=None. QuotaManager.check(..., budget_guard=True) passes that snapshot into AIBudgetGuard, whose _usage() then returns usage_cost_missing, so every Anthropic request is fail-closed even though a filtered usage snapshot exists. (line 168)

Suggestion: Provide a scoped USD cost value for filtered Anthropic snapshots, or explicitly disable Anthropic budget enforcement for filtered snapshots until a billable cost source is available instead of returning a snapshot that always blocks execution.

Review by Codex PR Review bot • PR

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: cbce3c594f

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-07-11T16:42:41Z

+            snapshot = (
+                self._anthropic_account_snapshot()
+                if provider == "anthropic"
+                else self._openai_account_snapshot()
+            )


Use month-to-date usage for monthly budget checks

When budget_guard is enabled, this rolling account snapshot is used as the observed usage for a monthly hard-limit decision, but the OpenAI and Anthropic readers default to a 7-day window. In a monthly billing period, spend from earlier than that window is ignored, so a repo can pass the guard after already consuming the monthly budget earlier in the month. Request or validate a snapshot that matches the decision billing period before allowing execution.

Useful? React with 👍 / 👎.

Sign in to view

+            review_quota = get_quota_manager().check(
+                review_repo,
+                reviewer_model,
+                req.prompt,
+                task_class="review",
+                budget_guard=_budget_gate_enabled(),
+            )


chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c14d0b2a95

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Sign in to view

+            get_quota_manager().record(
+                review_repo,
+                reviewer_result.model,
+                req.prompt,
+                reviewer_result.output if reviewer_result.success else "",
+            )


chatgpt-codex-connector · 2026-07-11T18:39:35Z

+            f"{provider}:{provider_scope}:{repo}",
+            f"{provider}:{provider_scope}",
+            provider,
+            "default",


Resolve nested budgets before default fallback

When monthly_budgets contains both a global default entry and a provider-specific nested entry such as openai -> project -> repo -> task, this candidate returns the default before the nested traversal below runs. That ignores the more specific repo/task budget and can enforce the wrong monthly hard limit for configured repos; the default fallback should come after the nested provider lookup.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-07-11T18:39:35Z

+                        hard_limit = _number(decision.get("hard_limit"), float("inf"))
+                        if hard_limit != float("inf") and (
+                            observed + stored_settled + stored_reserved + requested > hard_limit + 1e-12
+                            or observed + aggregate_settled + aggregate_reserved + requested > hard_limit + 1e-12


Don’t cap all repos at one repo’s API budget

For API-provider decisions, aggregate_scope is shared across repos for the same provider/project/task, but this reservation check compares that aggregate spend to the current repo’s hard_limit. With two repos each configured for a $10 monthly review budget, one $6 reservation leaves the second repo’s $6 reservation failing here even though both repos are under their own limits. Use a separate provider-project cap for aggregate enforcement, or keep repo budgets isolated from the aggregate reservation check.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-07-11T18:39:35Z

+
+            budget_guard_instance = get_ai_budget_guard()
+            snapshot = (
+                self._anthropic_account_snapshot()


Fetch costs for scoped Anthropic budgets

When CODEX_AUDIT_SERVICE_ANTHROPIC_ADMIN_WORKSPACE_IDS or CODEX_AUDIT_SERVICE_ANTHROPIC_ADMIN_API_KEY_IDS is configured, read_anthropic_admin_usage() returns filtered message usage but leaves costs as None, so passing that snapshot into the budget guard makes _usage() report usage_cost_missing and blocks every Anthropic request under budget_guard for scoped deployments. Fetch a cost snapshot that matches the Anthropic scope, or avoid using the filtered usage-only snapshot for hard-limit decisions.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-07-11T18:39:35Z

-          CODEX_PR_REVIEW_API_FALLBACK_DEFAULT: ${{ vars.CODEX_PR_REVIEW_API_FALLBACK_ENABLED || 'true' }}
-          CODEX_PR_REVIEW_DIRECT_API_PRIMARY_DEFAULT: ${{ vars.CODEX_PR_REVIEW_DIRECT_API_PRIMARY_ENABLED || 'true' }}
+          CODEX_PR_REVIEW_API_FALLBACK_DEFAULT: ${{ vars.CODEX_PR_REVIEW_API_FALLBACK_ENABLED || 'false' }}
+          CODEX_PR_REVIEW_DIRECT_API_PRIMARY_DEFAULT: ${{ vars.CODEX_PR_REVIEW_DIRECT_API_PRIMARY_ENABLED || 'false' }}


Preserve direct API primary for self-review

For AIAuditBridge’s own pull_request_target runs, CODEX_PR_REVIEW_REUSABLE_CALL is false, so the shell resolver uses this default rather than the workflow_call input. If the repo has no CODEX_PR_REVIEW_DIRECT_API_PRIMARY_ENABLED variable and no CODEX_AUDIT_SERVICE_URL—the legacy direct-API setup with only OpenAI/Anthropic secrets—this exports CODEX_PR_REVIEW_DIRECT_API_PRIMARY_ENABLED=false, causing run_codex_review_with_fallback() to fail with no backend instead of using the available API keys. Keep this default true for self-review while leaving reusable callers opt-in.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-07-11T18:39:35Z

+        if reservation_id:
+            payload["_budget_reservation_id"] = reservation_id
+        try:
+            job = _submit_job(claims, payload)


Release async reservations when jobs dedupe

For /v1/ai/execute/jobs, this call happens after a Codex budget reservation is put in payload, but _submit_job() can return an existing active job from the dedupe branch without creating a new job that owns that reservation. In that duplicate-request scenario no worker ever settles or releases the newly reserved ID, so the shared budget ledger keeps phantom Codex headroom consumed until manual cleanup/reset.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-07-11T18:39:35Z

+            if any(
+                (parsed := _reset_timestamp(value)) is not None and parsed <= self._clock()
+                for value in reset_at


Keep Codex reservations until both windows reset

Codex calls consume both primary and secondary rate-limit windows, so when only the shorter window’s resets_at has passed while the longer window is still active, this any() branch clears all Codex reservations and settled usage anyway. That lets new work reuse secondary-window headroom that was already consumed, weakening the protected reserve for accounts with staggered reset windows; clear the ledger only for windows that have all reset or track reservations per window.

Useful? React with 👍 / 👎.

Sign in to view

        "[quota_or_capacity_failure]" in message
        or "usage limit" in message
        or "daily budget exceeded" in message
+        or "deferred_budget" in message


chatgpt-codex-connector · 2026-07-11T18:39:35Z

+            snapshot = (
+                self._anthropic_account_snapshot()
+                if provider == "anthropic"
+                else self._openai_account_snapshot()


Use scoped OpenAI cost snapshots for project budgets

When CODEX_AUDIT_SERVICE_OPENAI_PROJECT_SCOPE is set, this path enforces a project-scoped budget but still feeds _openai_account_snapshot() into the guard; inspecting read_openai_admin_usage() shows its cost request populates organization_costs without project/api-key filters. A project can therefore be blocked by unrelated organization spend even though its own monthly budget has capacity, so the snapshot used for this scoped decision must include matching scoped costs.

Useful? React with 👍 / 👎.

Co-Authored-By: Codex <noreply@openai.com>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 12b24bc616

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Sign in to view

            self._save_records_locked()
+            self._record_failures.discard(repo or "unknown")


Sign in to view

+            if codex_quota is not None:
+                estimated_total += float(codex_quota.get("cost_estimate_usd") or 0.0)


chatgpt-codex-connector · 2026-07-11T22:07:53Z

        reasoning_effort = _resolve_codex_reasoning_effort(payload, str(payload.get("task") or TASK_EXECUTE))
+        # Persist a conservative possible-dispatch marker before invoking the
+        # external process; crash recovery must not refund ambiguous spend.
+        job["dispatch_started"] = True


Mark Codex dispatch only after the process starts

If the Codex command is unavailable or _codex_command() raises before subprocess.run() is reached, this pre-call marker makes the exception path settle the budget reservation as spent even though no Codex request consumed rate-limit headroom. In a misconfigured service host, each failed job therefore burns the protected Codex budget until reset instead of releasing the reservation; only set this after the process has actually been launched, or have the adapter report pre-dispatch failures distinctly.

Useful? React with 👍 / 👎.

Sign in to view

+                matched_result = next(
+                    (result for result in remaining_results if result.model == reviewer_model),
+                    None,


Co-Authored-By: Codex <noreply@openai.com>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: bba62d060b

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-07-11T22:57:12Z

+            get_ai_budget_guard().settle(
+                reservation_id,
+                float(qr.get("cost_estimate_usd") or 0.0),
+            )


Handle failed reservation settlement before returning

When budget_guard is enabled and the analyze provider call has already dispatched, settle() can still return False (for example, a transient SQLite/OSError in the shared ledger). This path ignores that return value and proceeds to return the model result, but no persisted job state retains the reservation id for retry, so the ledger can keep the amount in reserved and make later requests see phantom monthly budget consumption until reset. Check the result and persist/retry cleanup or fail closed, as the async job helper does.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-07-11T22:57:12Z

+                    get_quota_manager().check(
+                        review_repo,
+                        reviewer_model,
+                        req.prompt,
+                        estimated_output_tokens=DEFAULT_MAX_TOKENS,
+                    ).get("cost_estimate_usd") or 0.0


Resolve reviewer models for the combined daily precheck

When /v1/ai/review uses a model override or alias that resolve_model() maps to a priced model (for example, an unknown string falling back to Anthropic), this estimate calls check() without budget_guard, so the cost is computed from the raw string and can fall back to the cheaper default estimator. The later per-reviewer checks use the resolved model but do not reserve daily budget, so two reviewers can both run when the remaining daily budget is between this underestimated total and the real total, exceeding the API-key daily cap. Use the same resolved-model cost path for this combined precheck.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-07-11T22:57:12Z

+            if reset_event is not None and reset_event <= self._clock() and reset_event > self._codex_reset_events.get(reset_key, 0.0):
+                self._codex_reset_events[reset_key] = reset_event


Clear Codex reservations after missed reset windows

This reset cleanup only runs when the snapshot's current resets_at is already in the past. If no request arrives during the short period where a cached pre-reset snapshot is still served, the next fresh Codex snapshot normally points at the next future reset, so settled reservations from the previous window are never cleared and later low-volume accounts can remain deferred despite restored rate-limit headroom. Track reset advancement or reservation creation times instead of requiring the current snapshot reset time to be past.

Useful? React with 👍 / 👎.

Co-Authored-By: Codex <noreply@openai.com>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 3aa075a3c1

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Sign in to view

+        if result.success or bool(getattr(result, "dispatch_started", False)):
+            try:
+                quota.record(source_repo, resolved_model, req.prompt, result.output if result.success else "")


chatgpt-codex-connector · 2026-07-11T23:32:53Z

+                if payload.get("_budget_cleanup_pending") or payload.get("_budget_reservation_id"):
+                    continue


Mark expired uncertain jobs terminal before preserving them

When an expired queued/running async job still has a reservation that finalizes to pending_uncertain, this continue preserves the job file without changing its status. _active_job_count() still counts queued/running files, so enough expired ambiguous jobs can permanently consume all active-job slots and make new submissions fail with the max-active error. Mark the expired job failed/terminal before keeping it for budget reconciliation.

Useful? React with 👍 / 👎.

Co-Authored-By: Codex <noreply@openai.com>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 8f9f9da091

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Sign in to view

+            if not review_result.success and not bool(getattr(review_result, "dispatch_started", False)):
+                continue


chatgpt-codex-connector · 2026-07-11T23:45:33Z

+            for reservation_id, estimated_cost, reviewer_index in review_reservations:
+                matched_result = llm_results[reviewer_index] if reviewer_index < len(llm_results) else None
+                if matched_result is not None and bool(getattr(matched_result, "dispatch_started", False)):
+                    get_ai_budget_guard().settle(reservation_id, estimated_cost)


Check review reservation settlement before returning

When /v1/ai/review has already dispatched an API reviewer, settle() can still return False (for example, a transient SQLite/OSError in the shared budget ledger), but this path ignores the failure and then drops review_reservations. Because this synchronous review response has no persisted job state to retry cleanup, the reservation can remain in reserved and reduce monthly headroom for later requests until reset; handle a failed settlement by retrying/persisting cleanup or failing closed.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-07-11T23:45:33Z

+        if reservation_id and bool(getattr(result, "dispatch_started", False)):
+            from service.ai_budget_guard import get_ai_budget_guard
+
+            get_ai_budget_guard().settle(reservation_id, 0.10)


Handle sync Codex settlement failures

For the synchronous /v1/ai/execute compatibility path, a successful Codex dispatch can still hit a failed settle() call here (for example, a transient SQLite/OSError in the shared budget ledger), but the return value is ignored and there is no persisted job record to retry cleanup. In that case the reservation can remain in reserved after the response, reducing Codex headroom for later requests until reset; fail closed or persist/retry cleanup when settlement fails.

Useful? React with 👍 / 👎.

Co-Authored-By: Codex <noreply@openai.com>

chatgpt-codex-connector Bot reviewed Jul 11, 2026

View reviewed changes

Pigbibi and others added 22 commits July 12, 2026 02:52

feat: add centralized AI budget guard

c9fc649

Co-Authored-By: Codex <noreply@openai.com>

fix: enforce budget reservations through gateway

0dedcf8

Co-Authored-By: Codex <noreply@openai.com>

fix: close reservation lifecycle gaps

a32152e

Co-Authored-By: Codex <noreply@openai.com>

fix: reconcile settled budget usage safely

6122b6a

Co-Authored-By: Codex <noreply@openai.com>

fix: share provider reservations across repos

ab5dd82

Co-Authored-By: Codex <noreply@openai.com>

fix: enforce codex reservation headroom

f46acf8

Co-Authored-By: Codex <noreply@openai.com>

fix: reconcile provider spend and codex failures

30afc38

Co-Authored-By: Codex <noreply@openai.com>

fix: reconcile provider usage deltas

f57ff3b

Co-Authored-By: Codex <noreply@openai.com>

fix: clean up budget reservations on all exits

69eba6c

Co-Authored-By: Codex <noreply@openai.com>

fix: make budget reservations durable and conservative

3e6d231

Co-Authored-By: Codex <noreply@openai.com>

fix: persist budget reservations across workers

4c63c89

Co-Authored-By: Codex <noreply@openai.com>

fix: scope ledger rows by repo and billing period

bb42d6e

Co-Authored-By: Codex <noreply@openai.com>

fix: attribute review budget to authenticated repo

f7d3885

Co-Authored-By: Codex <noreply@openai.com>

fix: aggregate provider reservations and gate codex explicitly

8e58416

Co-Authored-By: Codex <noreply@openai.com>

fix: reconcile aggregate reservation balances

bd165a2

Co-Authored-By: Codex <noreply@openai.com>

fix: expire codex reservations across reset windows

2d8090e

Co-Authored-By: Codex <noreply@openai.com>

fix: fail closed on ledger errors and preserve API primary

354845a

Co-Authored-By: Codex <noreply@openai.com>

fix: fail closed on malformed usage snapshots

83955a8

Co-Authored-By: Codex <noreply@openai.com>

fix: reconcile monthly API quota accounting

c417901

Co-Authored-By: Codex <noreply@openai.com>

fix: separate repository and aggregate budget limits

de84b53

Co-Authored-By: Codex <noreply@openai.com>

fix: close review budget rollover leaks

499656f

Co-Authored-By: Codex <noreply@openai.com>

fix: close budget period and dispatch lifecycle gaps

b024478

Co-Authored-By: Codex <noreply@openai.com>

Pigbibi force-pushed the codex/ai-budget-guard branch from 43b9e31 to b024478 Compare July 11, 2026 18:52

Pigbibi and others added 4 commits July 12, 2026 02:52

fix: reconcile aggregate settled baselines

6448fc3

Co-Authored-By: Codex <noreply@openai.com>

fix: share provider reservation scope and preserve errors

8b6b2c3

Co-Authored-By: Codex <noreply@openai.com>

fix: release pre-dispatch job reservations

37ac73a

Co-Authored-By: Codex <noreply@openai.com>

fix: release deduped reservations and preserve baselines

06aa96a

Co-Authored-By: Codex <noreply@openai.com>

Pigbibi and others added 20 commits July 12, 2026 03:22

fix: preserve safety actions and reconcile batch budgets

64c426f

Co-Authored-By: Codex <noreply@openai.com>

fix: fail closed on quota persistence errors

cb46de6

Co-Authored-By: Codex <noreply@openai.com>

fix: settle async post-dispatch failures safely

855798e

Co-Authored-By: Codex <noreply@openai.com>

fix: settle ambiguous provider failures

954ce97

Co-Authored-By: Codex <noreply@openai.com>

fix: release non-billable AI reservations

cbc1cad

Co-Authored-By: Codex <noreply@openai.com>

fix: preserve budget reservations across failure states

5f02c2c

Co-Authored-By: Codex <noreply@openai.com>

fix: validate review budget ownership

60f1629

Co-Authored-By: Codex <noreply@openai.com>

fix: reconcile shared scope once per snapshot

fc6d78a

Co-Authored-By: Codex <noreply@openai.com>

fix: classify provider dispatch before settlement

0058005

Co-Authored-By: Codex <noreply@openai.com>

fix: cover guarded completion output budgets

fa4473b

Co-Authored-By: Codex <noreply@openai.com>

fix: make budget ledger transitions idempotent

f6c2aea

Co-Authored-By: Codex <noreply@openai.com>

fix: scope usage snapshots and persist failure state

0d539b8

Co-Authored-By: Codex <noreply@openai.com>

fix: harden budget reset and dispatch accounting

cd2548e

Co-Authored-By: Codex <noreply@openai.com>

fix: settle provider responses and charge verifier execution

4429137

Co-Authored-By: Codex <noreply@openai.com>

fix: persist fail-closed cleanup state

6e80772

Co-Authored-By: Codex <noreply@openai.com>

fix: fail closed without shared budget ledger

d90f5ad

Co-Authored-By: Codex <noreply@openai.com>

fix: close budget guard failure recovery gaps

5769baf

Co-Authored-By: Codex <noreply@openai.com>

fix: account provider usage for repo budgets

083e1a8

Co-Authored-By: Codex <noreply@openai.com>

fix: settle expiring jobs and persist quota failures

12b24bc

Co-Authored-By: Codex <noreply@openai.com>

fix: retain expired jobs pending budget cleanup

6cf3a79

Co-Authored-By: Codex <noreply@openai.com>

chatgpt-codex-connector Bot reviewed Jul 11, 2026

View reviewed changes

fix: retain ambiguous budget reservations

bba62d0

Co-Authored-By: Codex <noreply@openai.com>

chatgpt-codex-connector Bot reviewed Jul 11, 2026

View reviewed changes

Pigbibi and others added 2 commits July 12, 2026 07:04

fix: retain ambiguous Codex dispatches

3e1e060

Co-Authored-By: Codex <noreply@openai.com>

fix: classify provider dispatch outcomes

3aa075a

Co-Authored-By: Codex <noreply@openai.com>

chatgpt-codex-connector Bot reviewed Jul 11, 2026

View reviewed changes

fix: release known local provider failures

8f9f9da

Co-Authored-By: Codex <noreply@openai.com>

chatgpt-codex-connector Bot reviewed Jul 11, 2026

View reviewed changes

fix: align budget usage windows

7dc36df

Co-Authored-By: Codex <noreply@openai.com>

Pigbibi mentioned this pull request Jul 12, 2026

fix: classify adapter dispatch outcomes #75

Open

		self._save_records_locked()
		self._record_failures.discard(repo or "unknown")

		if codex_quota is not None:
		estimated_total += float(codex_quota.get("cost_estimate_usd") or 0.0)

		if reset_event is not None and reset_event <= self._clock() and reset_event > self._codex_reset_events.get(reset_key, 0.0):
		self._codex_reset_events[reset_key] = reset_event

		if payload.get("_budget_cleanup_pending") or payload.get("_budget_reservation_id"):
		continue

		if not review_result.success and not bool(getattr(review_result, "dispatch_started", False)):
		continue

Uh oh!

Conversation

Pigbibi commented Jul 11, 2026

Summary

Validation

Uh oh!

github-actions Bot commented Jul 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🤖 Codex PR Review

⚖️ Codex Review Arbitration

🚫 Blocking Issues

1. 🟠 [HIGH] Logic in service/ai_budget_guard.py

2. 🟠 [HIGH] Logic in service/openai_admin_usage.py

3. 🟠 [HIGH] Logic in service/anthropic_admin_usage.py

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

chatgpt-codex-connector Bot Jul 11, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

This comment was marked as off-topic.

Uh oh!

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

This comment was marked as off-topic.

Uh oh!

chatgpt-codex-connector Bot Jul 11, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Jul 11, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Jul 11, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Jul 11, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Jul 11, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Jul 11, 2026

Choose a reason for hiding this comment

Uh oh!

This comment was marked as off-topic.

Uh oh!

chatgpt-codex-connector Bot Jul 11, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

chatgpt-codex-connector Bot Jul 11, 2026

Choose a reason for hiding this comment

Uh oh!

This comment was marked as off-topic.

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Jul 11, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Jul 11, 2026

github-actions Bot commented Jul 11, 2026 •

edited

Loading

1. 🟠 [HIGH] Logic in `service/ai_budget_guard.py`

2. 🟠 [HIGH] Logic in `service/openai_admin_usage.py`

3. 🟠 [HIGH] Logic in `service/anthropic_admin_usage.py`