Repository navigation
fix(codex): do not double-count a spawned sub-agent's replayed parent history - #1383
Merged
Merged
Conversation
roy-tong
added a commit
to roy-tong/codeburn
that referenced
this pull request
Sep 16, 2026
…per review Re-verified on a fresh spread sample of codeset-release-evals (21 sessions, 404 events): exactly one info:null rate-limit ping per session, zero events with info present but total_token_usage absent — matching the reviewer's 137k-event corpus. The missing-cumulative branch is a no-op on real data: - CODEX_CACHE_VERSION stays 15 (no cold reparse for identical numbers; v16 never shipped in a release) - drop the -missing-cumulative-v1 parse suffix - Quirks: null-info pings take the estimate path, not the dedup path; the earlier 54-of-1313 figure conflated them with partial-info events - dedup layer 3 now documents the forkedFromId/getagentseal#1383 sub-agent caveat
iamtoruk
pushed a commit
that referenced
this pull request
Sep 16, 2026
…ng (#1264) * fix(codex): keep token_count events in sessions that never write total_token_usage * fix(codex): preserve distinct requests when cumulative usage is missing * fix(codex): narrow missing-cumulative fix - byte-identical re-emissions still collapse (#257) Real-data answer to the review's question: Codex does re-emit byte-identical token_count events. Scanned 53 public rollout sessions (codeset-ai/ codeset-release-evals, 1313 token_count events): 603 (46%) are byte-identical repeats of their predecessor (identical total AND last usage), 0 cumulative monotonicity violations, 54 events carry no total_token_usage at all. The missing-cumulative path now collapses a record whose info payload is byte-identical to the previous event's (#257 semantics restored); only records with differing payloads count as distinct requests. prevInfoIdentity is persisted in the resume state. The #1088 invariant note is restored at the dedupKey in adapted form; docs/providers/codex.md documents the three-layer dedup and the accepted fork-cutoff trade-off. * fix(codex): revert cache/parse version bumps, correct null-info docs per review Re-verified on a fresh spread sample of codeset-release-evals (21 sessions, 404 events): exactly one info:null rate-limit ping per session, zero events with info present but total_token_usage absent — matching the reviewer's 137k-event corpus. The missing-cumulative branch is a no-op on real data: - CODEX_CACHE_VERSION stays 15 (no cold reparse for identical numbers; v16 never shipped in a release) - drop the -missing-cumulative-v1 parse suffix - Quirks: null-info pings take the estimate path, not the dedup path; the earlier 54-of-1313 figure conflated them with partial-info events - dedup layer 3 now documents the forkedFromId/#1383 sub-agent caveat --------- Co-authored-by: Roy tong <tongxiarui@elliwooddeMacBook-Air.local>
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.
Fixes #1380.
Codex writes a spawned sub-agent's rollout by replaying the parent's history first, keeping every
event_msg/token_countline with its original timestamp and dropping the parent'stoken_usage_recorditems (codex-rsagent/control/spawn.rs). The child names its parent undersource.subagent.thread_spawn.parent_thread_id, notforked_from_id, so the fork guard never engaged and every child re-counted its parent's prefix. That is the multiplier the report describes; ccusage documents and handles the same replay in its Codex adapter.The parser now reads the parent id from either field (nested object on the plain JSON path, flattened on the compact decoder for oversized meta lines) and the existing fork guard and content-keyed dedup take over. One new test: a spawned child replaying its parent with original timestamps counts parent plus its own work only. It fails on main and passes here; the 48 existing Codex tests still pass.
Verified on this machine's corpus, 1,378 Codex rollouts of which 381 are spawned children: lifetime Codex totals drop by exactly 405 calls and 403,712 cache tokens, and those numbers equal the replayed prefixes summed directly from the files. Every other total is unchanged.
Two notes for the reporter's table: codeburn's "input" column is uncached input, so an exact count that includes cached tokens under input will always read higher; and the per-turn growth in
turn_token_usageis per-response accumulation, not streaming, as the Codex source shows.