fix(ai): claude quota transform silently overwrites duplicate limits entries (last wins) — diverges from the #1750 keep-first convention

Summary

The Claude quota transform (packages/ai/src/quota/providers/claude.ts, toClaudeUsage) silently overwrites duplicate limits entries — last wins. addWindow assigns target[key] = ... unconditionally (claude.ts:152–164), and scoped model windows are stored by unconditional assignment models[modelName] = ... (claude.ts:185). That diverges from the keep-first convention the suite has now adopted three times for the identical shape: #1750 (opendesign coding-plan windows), #1769 (cline-pass limit entries, fixed in PR #1773), and #1762's collision handling for codex. Claude's transform was pinned in #1750 for the legacy-fallback behavior, but no test exercises duplicate limits entries.

Evidence (verified this run, main `32fc03b`)

Probe on the laptop (~/git/fleet/probe-2026-11-05-claude-quota/probe.ts, kept) importing toClaudeUsage:

CASE1 — limits [session 20% (reset 2026-12-01), session 80% (reset 2026-12-02), weekly_all 30%, weekly_all 70%]: published 5h = 80% with the second reset, 7d = 70% — both first entries silently dropped.
CASE2 — two weekly_scoped entries for the same model ("Claude Sonnet", 30% then 60%): published model window = 60% with the second reset; the 30% entry is gone.
CASE3 (control) — a duplicate session entry with percent: null does not overwrite (20% survives), because addWindow returns early when percent is null and there is no valueLabel. The overwrite therefore happens exactly when the duplicate carries a real figure — the case that matters.

Why it matters

Claude is a plan pool the fork rotates across. If Anthropic's usage payload ever carries repeated limit entries (per-model splits, plan migrations, a kind repeated during a rollout), the published usage silently becomes whichever entry sorts last, with nothing recording that a choice was made — the exact failure #1750/#1769 treated as a defect in the sibling providers. The transform is otherwise well-pinned after #1750; this is the one input shape left unpinned and convention-divergent.

Acceptance criteria

Duplicate limits entries keep the first entry per kind (and per model for weekly_scoped), per the test(ai): quota suite pins none of claude / meta-billing / opendesign's pure parsers — billing banner rejects numeric strings, claude transform drops legacy windows when limits is non-empty, coding-plan publishes a zero-limit window #1750 convention — or a different disposition is chosen and documented in the code + this issue's closing note. The legacy-fallback precedence from test(ai): quota suite pins none of claude / meta-billing / opendesign's pure parsers — billing banner rejects numeric strings, claude transform drops legacy windows when limits is non-empty, coding-plan publishes a zero-limit window #1750 (limits-produced windows win over five_hour/seven_day) is unchanged.
packages/ai/test/quota.test.ts pins: duplicate session, duplicate weekly_all, duplicate weekly_scoped for one model, two different models' scoped entries both surviving, and the null-percent duplicate control (first entry survives).
Existing quota suite stays green; gate state not regressed (the three quota-era gates already red on main belong to fix(ci): three gates red on main — PR #1722's quota subcommand landed 16 undispositioned sites, and turso bypasses the PR's own bounded reader #1724 — do not conflate).

🤖 Agent note · filed

Agent: new-issue-creator job (new-issue-creator-opencode)
Session / ID: cron new-issue-creator-opencode run 2026-11-05T21:45Z
Found via: filed this run
Task: Filed after probe verification on the laptop (~/git/fleet/probe-2026-11-05-claude-quota/probe.ts): toClaudeUsage with duplicate session entries (20 then 80) published 80, duplicate weekly_all (30 then 70) published 70, duplicate weekly_scoped for one model (30 then 60) published 60 — silent last-wins at claude.ts:158/185, diverging from the test(ai): quota suite pins none of claude / meta-billing / opendesign's pure parsers — billing banner rejects numeric strings, claude transform drops legacy windows when limits is non-empty, coding-plan publishes a zero-limit window #1750 keep-first convention (cf. fix(ai): cline-pass quota silently overwrites duplicate limit entries (last wins) — diverges from the #1750 keep-first convention, and the mapping is unpinned #1769 cline-pass). A null-percent duplicate does not overwrite (addWindow early return) — that control is in the issue. Triage packet enqueued to repo-health-opencode-pilot.
Expectation: effort/s — triage confirms the keep-first disposition, then the fix + pins land per the acceptance criteria.
Links: issue fix(ai): claude quota transform silently overwrites duplicate limits entries (last wins) — diverges from the #1750 keep-first convention #1781; evidence body in the goal workspace (hidden_files/bodies/opencode-claude-quota-duplicate-limits.md); convention source test(ai): quota suite pins none of claude / meta-billing / opendesign's pure parsers — billing banner rejects numeric strings, claude transform drops legacy windows when limits is non-empty, coding-plan publishes a zero-limit window #1750

🤖 Agent note · adopted as phase 2 of design #1794 (one window contract for the provider quota parsers; doc: files/designs/opencode-quota-window-contract.md, keep-the-four-repos-healthy goal workspace). This issue is the contract's claude application: rule 2 (duplicates keep the first — the #1750/#1769 convention — per kind and per model scope). Linked as a sub-issue of #1794; its queue lifecycle is untouched and it is not re-enqueued. When it lands, design phase 1's inventory row for claude duplicates flips from known-lossy to conforming.

Partial verification — staying open (2026-11-07 done-task-only audit, tracking: developing-today/code#7016).

Queue task(s) [10788] were marked done, and some work did land:

PR test(ai): quota window contract — parser inventory + shared probe matrix (#1795) #1796 — test(ai): quota window contract — parser inventory + shared probe matrix (test(ai): quota window contract — parser inventory + shared probe matrix for the 13 quota providers (design #1794, phase 1) #1795) (merged 2026-11-06T03:06:24Z): test(ai): quota window contract — parser inventory + shared probe matrix (#1795) #1796

No merged change closes this issue's full scope, so the done status is not treated as completion. Remainder re-queued as triage task 11507.

🤖 Agent note · picked up

Agent: repo-worker-opencode (OpenChamber scheduled lane)
Session / ID: ses_e5e468b38a6ffezBpdzAu38uDA · lane 3fe37e8a-c61b-419c-9e97-769ecd264d5d
Found via: queue repo-health-opencode-pilot, task 11854
Task: Revalidate issue fix(ai): claude quota transform silently overwrites duplicate limits entries (last wins) — diverges from the #1750 keep-first convention #1781's remainder against current main and convert any remaining work to a scoped small-fix packet. The issue is open; PR test(ai): quota window contract — parser inventory + shared probe matrix (#1795) #1796 is merged for phase-1 inventory (test(ai): quota window contract — parser inventory + shared probe matrix for the 13 quota providers (design #1794, phase 1) #1795), not this duplicate-limits fix.
Links: fix(ai): claude quota transform silently overwrites duplicate limits entries (last wins) — diverges from the #1750 keep-first convention #1781 · test(ai): quota window contract — parser inventory + shared probe matrix (#1795) #1796

🤖 Agent note · triaged

Agent: repo-worker-opencode (OpenChamber scheduled lane)
Session / ID: ses_e5e468b38a6ffezBpdzAu38uDA · lane 3fe37e8a-c61b-419c-9e97-769ecd264d5d
Task: Revalidated on current main 24200c3. The remainder is actionable: direct toClaudeUsage probes still show last-wins for duplicate session (20→80), weekly_all (30→70), and same-model weekly_scoped (30→60); distinct model scopes survive. PR test(ai): quota window contract — parser inventory + shared probe matrix (#1795) #1796 is phase-1 inventory/shared probes for test(ai): quota window contract — parser inventory + shared probe matrix for the 13 quota providers (design #1794, phase 1) #1795, not the fix(ai): claude quota transform silently overwrites duplicate limits entries (last wins) — diverges from the #1750 keep-first convention #1781 fix.
Disposition: Created scoped small_fix queue task 12828 (repo-health:v1:fix:developing-today-forks/opencode:issue:1781). It covers first-entry deduplication and the requested regression pins while preserving test(ai): quota suite pins none of claude / meta-billing / opendesign's pure parsers — billing banner rejects numeric strings, claude transform drops legacy windows when limits is non-empty, coding-plan publishes a zero-limit window #1750 legacy fallback precedence.
Verification: Direct probes reproduced the defect and confirmed that a real-percent entry followed by a null-percent duplicate retains the first value. The targeted command bun run --cwd packages/ai test quota.test.ts could not load the suite in this fresh worktree because package effect is missing; 0 tests executed. This is a dependency/setup limitation, not a code failure.
Links: fix(ai): claude quota transform silently overwrites duplicate limits entries (last wins) — diverges from the #1750 keep-first convention #1781 · test(ai): quota window contract — parser inventory + shared probe matrix (#1795) #1796

Merge pull request #2043 from developing-today-forks/queue/12828-1781…

Navigation Menu

Repository navigation

fix(ai): claude quota transform silently overwrites duplicate limits entries (last wins) — diverges from the #1750 keep-first convention #1781

Summary

Evidence (verified this run, main `32fc03b`)

Why it matters

Acceptance criteria

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

fix(ai): claude quota transform silently overwrites duplicate limits entries (last wins) — diverges from the #1750 keep-first convention #1781

Description

Summary

Evidence (verified this run, main 32fc03b)

Why it matters

Acceptance criteria

Activity

dezren39 commented on Nov 5, 2026

dezren39 commented on Nov 6, 2026

dezren39 commented on Nov 8, 2026

dezren39 commented on Nov 8, 2026

developing-today commented on Nov 8, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

Evidence (verified this run, main `32fc03b`)