fix(codex): attribute resumed token counts to later model

[IvGolovach]

Summary

Fix Codex resumed-session attribution so early token_count events emitted before turn_context inherit the later model instead of staying under unknown.

Fixes #490.

Why

Codex 0.124 resumed sessions can replay event_msg / token_count rows before the first turn_context row that carries payload.model. The previous parser emitted those early rows immediately with modelId = unknown, which also meant pricing stayed at 0 even after the real model appeared later in the same session file.

Diff scope

• crates/tokscale-core/src/sessions/codex.rs: buffer model-less Codex token_count messages until the parser observes a model, then flush them with that model; preserve unknown only when no model ever appears.
• crates/tokscale-core/src/lib.rs: avoid writing Codex source-cache entries when a file still has unresolved model-less token events.
• crates/tokscale-core/src/message_cache.rs: bump the source-message cache schema from 8 to 9 so stale cached unknown Codex rows are invalidated.

This is a narrow runtime parser/cache change. No frontend, release, workflow, migration, or CLI command behavior changed.

Branch integrity

• Base branch: main
• origin/main: 97809e755b58b07e21c0eaceb9005f0ba073de0c
• Ahead / behind from git rev-list --left-right --count origin/main...HEAD: 0 behind / 1 ahead
• Merge base: 97809e755b58b07e21c0eaceb9005f0ba073de0c
• git merge-base --is-ancestor origin/main HEAD: PASS
• Fast-forward safety: branch contains the current validated base.

Commit integrity

• Introduced commit: c3ca2ce9cfc113bc465dff8965f49da6cb83d865 fix(codex): attribute resumed token counts to later model
• One logical change per commit: yes.
• Ledger-Id trailers: Not applicable — scripts/ledger/ does not exist in this repository.
• Duplicate Ledger-Id check: Not applicable — scripts/ledger/ does not exist in this repository.

Ledger proof

Not applicable — scripts/ledger/ does not exist in this repository.

Diff hygiene

git diff --name-status origin/main...HEAD:

M	crates/tokscale-core/src/lib.rs
M	crates/tokscale-core/src/message_cache.rs
M	crates/tokscale-core/src/sessions/codex.rs

• git diff --check origin/main...HEAD: PASS, no output.
• No .env files, local environment files, secrets, tokens, credentials, cache files, build artifacts, or unrelated generated files are included.
• Required metadata note: CACHE_SCHEMA_VERSION was bumped because the interpretation of cached Codex source messages changed; invalidating old cache entries prevents previously cached unknown rows from surviving the parser fix.

Validation tier and test proof

Validation tier: Tier 2 — narrow runtime parser/cache change.

Reason: the change is localized to Codex session parsing and source-message cache eligibility. It does not alter migrations, auth, permissions, payment/billing logic, global configuration, CI, release tooling, production infrastructure, or external service contracts.

Run:

• scripts/ledger/check: Not applicable — scripts/ledger/ does not exist in this repository.
• git fetch --no-tags origin main:refs/remotes/origin/main: PASS.
• cargo test -p tokscale-core token_count: PASS, 10 passed.
• cargo test -p tokscale-core codex_cache: PASS, 2 passed.
• cargo test -p tokscale-core: PASS, 599 passed, 1 ignored; codebuff integration tests 10 passed; hermes integration tests 3 passed; doc-tests 0 passed.
• cargo clippy -p tokscale-core --all-features -- -D warnings: PASS.
• cargo fmt --all --check: PASS.
• git diff --check: PASS.
• git diff --cached --check: PASS.
• Cyrillic scan for touched source files: PASS, no matches.

Not run:

• Full workspace test suite — not required for this Tier 2 parser/cache change because the touched runtime surface is covered by tokscale-core tests and clippy.
• Native build matrix — left to GitHub CI; no native build configuration changed.

External gates:

• Pending — required GitHub checks can run only after this PR is opened.

Verification-pack proof

Not applicable — no infra, governance, replay, invariant, or verification-pack files changed.

Migration notes

Not applicable — no DB migration changed.

CI context confirmation

• CI context names unchanged.
• Not applicable — no CI/workflow context changes.
• Required contexts: Pending — PR has not been opened yet, so GitHub has not created check runs for this head branch.
• Live branch-protection status: unavailable from the local check; GitHub API returned 404 for the branch-protection required-status-checks endpoint.

Runtime safety

Reviewed runtime files/modules:

• crates/tokscale-core/src/sessions/codex.rs
• crates/tokscale-core/src/lib.rs
• crates/tokscale-core/src/message_cache.rs

Safety reasoning:

• No new blocking locks are introduced; parsing still uses the existing sequential reader flow.
• No new unbounded queues are introduced beyond a per-file in-memory buffer for model-less token events before the first model appears; the buffer is drained as soon as a model is observed or at EOF.
• No parser invariants are removed; token delta logic, zero-token filtering, timestamp fallback tracking, workspace attribution, and provider/agent propagation are preserved.
• Cache writes are disabled only for unresolved model-less Codex files, preventing temporary unknown attribution from being persisted.

No invariant regression introduced.

Documentation integrity

Not applicable — no docs, runbooks, public commands, release procedure, deployment procedure, or user-facing documented behavior changed.

Rollback plan

• Rollback command after merge: git revert <post_merge_commit_sha>.
• Replace <post_merge_commit_sha> with the final squash or merge commit SHA after merge.
• DB downgrade required: no.
• Data repair required: no.
• Operational caveats: rollback would restore cache schema version 8; users may keep regenerated cache entries from the fixed parser until the cache naturally refreshes or is removed.

Known residual risks

No known residual risks.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
tokscale	Ignored	Preview	May 3, 2026 8:29am

[cubic-dev-ai]

1 issue found across 3 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="crates/tokscale-core/src/sessions/codex.rs">

<violation number="1" location="crates/tokscale-core/src/sessions/codex.rs:238">
P2: Pending model-less token_count rows can remain buffered when a headless JSON line sets `current_model` but returns `None`, so earlier rows are later emitted as `unknown` instead of the discovered model.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[junhoyeo]

Thanks!