fix(global-discover): bucket codex sessions by originator

[0xDevNinja]

Summary

Buckets codex sessions by payload.originator. scanCodex normalizes it into { desktop, exec, claude_code, other } and surfaces the breakdown at tools.codex.originators (global) and codex_originators (per-repo), plus the summary view. Existing codex totals are untouched — the change is purely additive, so no consumer breaks.

Unknown, missing, null, and non-string originators land in other rather than being dropped, so the buckets always sum to the codex total. Whitespace is trimmed before matching.

Refs #1315 — this covers Problem 1 only. See scope note below.

Why

/retro global narrated "codex was the primary execution tool, 414 sessions across 7 repos" when codex actually drove dev for one repo's middle phase. The other ~309 codex_exec entries were CC firing codex as a cross-model review subagent. A single bucket can't tell those apart.

Scope

Originally this PR carried two patches from #1315. It has since been rebased onto current main and narrowed to Problem 1:

• Problem 1 (here): originator bucketing. Still missing on main — rg "originator" returns nothing in bin/gstack-global-discover.ts.
• Problem 2 (dropped): the extractCwdFromJsonl cwd-truncation fix landed independently via v1.41.1.0 fix wave: 7 HIGH bugs from external audit + regression tests (PR #1169 follow-up) #1592 / v1.41.1.0 (64KB / 30-line / drop-partial). Main's trade-offs there are deliberate and have regression tests; no reason to re-litigate them from this branch. The CLAUDE_PROJECTS_DIR test knob went with it.
• Problem 3 (not addressed): the /retro global "sessions = file count, not interactive dev" annotation. Belongs in the retro skill template, and probably deserves its own issue.

Because Problems 2 and 3 aren't both resolved here, this uses Refs rather than Fixes so merging doesn't auto-close #1315.

Tests

test/global-discover.test.ts — 35 pass, 0 fail. Covers each bucket, the whitespace-trim case, missing / null / non-string originator → other, and a per-repo sum invariant (desktop + exec + claude_code + other == sessions.codex).

Thanks

@Akagilnc for the review that caught the staleness, and for the .trim() and null-originator suggestions — both are in.

[Akagilnc]

Heads-up that this PR is now stale against main. The Problem 2 fix landed independently via #1592 / v1.41.1.0 on 2026-05-20: main's extractCwdFromJsonl uses a 64KB cap, scans up to 30 lines, and drops the trailing partial segment before parsing. The current PR is marked CONFLICTING / DIRTY.

The two implementations make different trade-offs rather than one being a strict superset:

• This PR's 128KB read covers a hypothetical >64KB first-line case that main's 64KB read doesn't.
• Main's 30-line scan handles cases where cwd lands past line 15 (which this PR caps at).
• Main's trailing-partial-drop is cleaner at the buffer boundary, but drops a file whose final line has no trailing newline — a corner this PR's implementation does happen to handle.

That said, main is the one that's merged, has regression tests, and chose the 64KB / 30-line / drop-partial mix deliberately. Keeping this PR's Problem 2 patch would mean re-arguing those defaults from scratch.

The unique remaining value here is Problem 1 — bucketing Codex sessions by payload.originator. That part is genuinely missing on main (rg "originator" returns nothing in current bin/gstack-global-discover.ts) and the design looks good: case-insensitive normalization, unknown values land in other rather than being dropped, additive (existing totals unchanged), 7 new tests with sum invariants.

Suggestion: rebase onto main, drop the extractCwdFromJsonl / CLAUDE_PROJECTS_DIR / CC-bigline-test changes, keep only the originator bucketing. Two small additions worth picking up while rebasing:

• normalizeCodexOriginator should .trim() before lowercasing — currently "Codex Desktop " (with a stray space) lands in other.
• Add explicit tests for missing / null / non-string originator values mapping to other.

Happy to open a follow-up PR with just the originator half if that helps — no rush from my side, just flagging since the PR has been quiet for ~16 days.

[0xDevNinja]

Rebased onto current main and narrowed the scope to just the codex-originator bucketing, per @Akagilnc's review.

Dropped (now redundant): the extractCwdFromJsonl 8KB→128KB change and the CLAUDE_PROJECTS_DIR test knob. That cwd-truncation fix landed independently via #1592 / v1.41.1.0 (64KB / 30-line / drop-partial), and main's deliberate trade-offs there are the ones with regression tests. No reason to re-argue them from this branch.

Kept (unique, missing on main — rg "originator" returns nothing): the four-bucket classifier so /retro global can separate real interactive dev (Codex Desktop) from subagent / scripted runs (codex_exec) and CC-driven calls (Claude Code).

Picked up both follow-ups you flagged:

• normalizeCodexOriginator now .trim()s before lowercasing, so "Codex Desktop " maps to desktop instead of other.
• Added explicit tests for missing / null / non-string originator → other, plus a whitespace-trim test.

Buckets are additive (existing totals unchanged) and always sum to the codex total. 35 tests pass.

[Akagilnc]

Thanks for the rebase and for picking up both follow-ups — the .trim() and the missing/null/non-string tests are exactly what I had in mind, and narrowing to the originator bucketing makes this much easier to reason about.

Three small housekeeping suggestions, all non-blocking and entirely up to you and the maintainers:

1. The PR description is still the pre-rebase version. It describes both patches (including the extractCwdFromJsonl 8KB→128KB change and the CLAUDE_PROJECTS_DIR knob) and reports 26 pass, 0 fail, whereas the rebased branch drops patch 2 and you mention 35 passing. Might be worth refreshing so reviewers reading the description see the current scope.

2. Fixes #1315 in the description vs Refs #1315 in the commit. The commit message has it right — Fixes is a closing keyword, so merging as-is would auto-close gstack-global-discover: session counts conflate originator types and undercount CC by ~5x #1315. Since this PR now only covers Problem 1, that would close the issue with Problem 3 (the /retro global "sessions = tool invocations" annotation, which you flagged as out of scope) still unaddressed. Changing the description to Refs #1315 would keep the two in sync.

3. Problem 2's fix never got linked back to the issue. It landed in v1.41.1.0 fix wave: 7 HIGH bugs from external audit + regression tests (PR #1169 follow-up) #1592 / v1.41.1.0 on 2026-05-20, but that PR doesn't reference gstack-global-discover: session counts conflate originator types and undercount CC by ~5x #1315, so the issue currently has no record that half of it is already resolved. Worth a note on gstack-global-discover: session counts conflate originator types and undercount CC by ~5x #1315 at some point, and Problem 3 probably deserves its own issue.

Either way, the bucketing work here looks good to me. Thanks for sticking with it.

[0xDevNinja]

good catches, both real. updated the description: scope is originator-only now, 35 tests, and swapped Fixes -> Refs so merging doesn't close #1315 with problem 3 still open. the commit already said Refs, i just forgot the body when i rebased.

agreed on the third one too. i'll drop a note on #1315 pointing at #1592 for problem 2 and open a separate issue for the /retro global annotation, that's a retro-template change and doesn't belong here.

[Akagilnc]

Thanks for following through on this. Really appreciate the cleanup and for linking everything back together — much easier to follow now.

[time-attack]

@16francej Merge this version because it correctly buckets Codex sessions by originator and passed the Mac checks.