fix: full budget-checkpoint coverage across review + adversarial (Codex #2) - #9
Merged
Merged
Conversation
#2) The cost-controller only recorded executor token spend; self-review (3 persona calls) and the adversarial gate were invisible to both the budget cap and the reported total — System A could spend ~3x what the checkpoint saw, and the kill switch couldn't fire mid-task. System B (papa) recorded an aggregate only AFTER the thinker fan-out, with no pre-flight check. Changes: - CodexCaller contract gains optional token fields; createCodexCaller parses OpenAI's usage block, createOpenAICompatibleCodexCaller estimates chars/4 when the server omits usage. The adversarial gate's spend was previously untracked AT THE SOURCE (the interface returned only { content }). - SelfReviewResult + AdversarialReviewResult carry tokenUsage. selfReview aggregates the three persona calls; adversarialReview surfaces the codex call's tokens. - loop.ts records every stage against the checkpoint: forceCheck before the self-review fan-out, recordAndCheck after self-review and after adversarial, kill→budget-exceeded on any over-budget result. The reported totalTokenUsage now sums executor + review + adversarial on ALL outcome paths (success and failure). - papa.ts forceChecks before the thinker fan-out so a request already at the ceiling never launches N parallel calls. The codex model id is free-form (not a ModelTier); cost-estimator returns $0 for ids absent from the pricing table (PR #7), so tokens are tracked even though the codex wire-cost line is $0. Adding codex pricing is a follow-on. Also adds docs/design/2026-06-17-criticals-sandbox-and-budget.md — the design for this fix (#2) AND critical #1 (sandbox hardening, Level 1 chosen). #1 ships as a separate follow-on PR. Tests: 380 → 384. New: cloud + local codex token surfacing, loop total-token accounting includes review+adversarial. No skips. Refs CODEX_REVIEW.md (#2).
This was referenced Jun 17, 2026
Merged
telivity-otaip
added a commit
that referenced
this pull request
Jun 17, 2026
…) (#11) #7 — DomainAnswerStore keyed answers by question TEXT only, so two different files asking the same question (e.g. "What is the grace period?") collapsed to one answer; answering it in file A silently unblocked file B with A's answer. hashQuestion now takes (filePath, text) and hashes the composite. Same wording in a different file is a distinct question again. #9 — Captured transcripts wrote full prompts/responses to disk unredacted; they can carry pulled-in source and secrets. New redactSecrets() masks known token shapes (sk-ant-, sk-/sk-proj-, ghp_/gho_/…, github_pat_, AKIA/ASIA, AIza, xox*, Bearer values, and SECRET_NAME=value assignments). instrumentLLMCaller/CodexCaller redact by default (opt out with { redact: false }); the events file is created mode 0600. Tests: 387 → 400. New: hashQuestion file-scoping + path-normalization; redactSecrets per token shape + non-secret passthrough; instrumented callers redact by default and honor redact:false. No skips. Refs Codex review #7, #9 (private KB).
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes critical finding #2 from the Codex review. Also lands the design doc for both criticals (#1 + #2); #1 (sandbox hardening, Level 1) follows as a separate PR.
The bug
The cost-controller created one checkpoint per task but recorded only the executor's tokens. Everything downstream was invisible to both the budget cap and the reported total:
CodexCaller.call()returned only{ content }, no token fields at all.papa) — recorded an aggregate only after the thinker fan-out, with no pre-flight check.Net: System A could spend ~3× what the checkpoint saw; the kill switch couldn't fire between the executor and the PR.
The fix
CodexCallergains a token surface (optional fields → existing mocks still compile).createCodexCallerparses OpenAI'susage;createOpenAICompatibleCodexCallerestimates chars/4 when the local server omits usage.SelfReviewResult+AdversarialReviewResultcarrytokenUsage.selfReviewaggregates the three persona calls;adversarialReviewsurfaces the codex call's tokens.loop.tsaccounts every stage:forceCheck()before the self-review fan-out;recordAndCheck()after self-review and after adversarial; any over-budget result →budget-exceeded+ break. The reportedtotalTokenUsagenow sums executor + review + adversarial on all outcome paths (success and failure) — previously even the report undercounted.papa.tsforceChecks before the thinker fan-out so a request already at the ceiling never launches N parallel calls.On codex pricing
The codex model id is free-form (
gpt-4o,codex-mini), not aModelTier.calculateCallCostreturns$0for ids absent from the pricing table (landed in PR #7), so adversarial tokens are tracked even though the codex wire-cost line reads$0. Adding a real codex pricing row is a clean follow-on.Tests
createCodexCallersurfacesusagetokens (+ zero-when-absent); localcreateOpenAICompatibleCodexCallertoken surfacing + chars/4 estimate; loop-level test asserting the reported total exceeds the executor-only tokens (proving review+adversarial are accounted).tokenUsagefield.Design doc
docs/design/2026-06-17-criticals-sandbox-and-budget.md— covers this fix and critical #1 (execution sandbox). #1's chosen depth is Level 1 (process hardening:--ignore-scripts, env allowlist so API keys/GH token never enter the pnpm/git child env, credential separation for PR creation). Ships next as its own PR.🤖 Generated with Claude Code