feat(spend): cache-first and 5m TTL for dashboard - #3107
Conversation
|
🦞👀 Pull request received. I will update this pull request when review starts. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 1f4bbb2cc7
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| if !force, | ||
| let lastAt = self.lastSpendDashboardTokenFetchAt[provider.instanceID], | ||
| let lastScope = self.lastSpendDashboardTokenFetchScope[provider.instanceID], | ||
| lastScope == costScopeSignature, | ||
| self.spendDashboardTokenSnapshotPublicationForCurrentConfig(for: provider) != nil, |
There was a problem hiding this comment.
This TTL cannot fire through production code: the sole caller in SpendDashboardSource.makeRequest always passes force: true, while .refreshMissing invokes that caller only when no current publication exists—even though this condition requires one. Consequently, the new five-minute guard cannot suppress any dashboard token scan; the caller needs to preserve the build mode's forced/non-forced semantics or perform the TTL decision before the missing-publication predicate.
Useful? React with 👍 / 👎.
clawsweeper
Bot
added
P2
|
Codex review: needs real behavior proof before merge. Reviewed August 22, 2026, 10:50 AM ET / 14:50 UTC. ClawSweeper reviewWhat this changesThis PR adds cache-first Spend dashboard loading, a five-minute refresh gate, and aggregate-based Codex/OpenCodex cache reads to avoid repeated long scans. Merge readiness⛔ Blocked until real behavior proof from a real setup is added - 9 items remain Keep open: the advertised five-minute dashboard TTL never expires after a successful publication, and the cutover test weakens an exclusive-output invariant. Real after-fix dashboard proof is also still missing. Priority: P2 Review scores
Verification
How this fits togetherCodexBar’s Spend dashboard combines provider snapshots and local usage-cache reports into the Preferences pane. This PR changes when providers refresh and how cached usage is hydrated before dashboard totals are published. flowchart LR
A[Preferences pane opens] --> B[Dashboard request]
B --> C{Snapshot stale?}
C -->|yes| D[Provider usage refresh]
C -->|no| E[Cached usage snapshot]
D --> F[Dashboard publication]
E --> F
F --> G[Spend totals and charts]
Before merge
Findings
Agent review detailsSecurityNone. Review metrics
Merge-risk optionsMaintainer options:
Technical reviewBest possible solution: Set completion metadata only after a current successful or confirmed-empty publication, restore the strict exclusive-output test invariant, then provide redacted live dashboard and upgrade evidence. Do we have a high-confidence way to reproduce the issue? Yes: source shows a successful publication has no completion timestamp, and the stale predicate consequently returns false on every normal reopen. Is this the best way to solve the issue? No: the TTL needs completion-time bookkeeping and the cutover test must retain its strict exclusive-output assertion. Full review comments:
Overall correctness: patch is incorrect AGENTS.md: found and applied where relevant. Codex review notes: model internal, reasoning high; reviewed against 27c7f334e3c4. LabelsLabel justifications:
EvidenceAcceptance criteria:
What I checked:
Likely related people:
Rating scale
Overall follows the weaker of proof and patch quality. Workflow
HistoryReview history (14 earlier review cycles; latest 8 shown)
|
|
CI failure is a real compile error:
|
4582171 to
d3596a6
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d3596a6a84
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| let shouldPrimeCachedCodex: Bool = self.cachedLoader != nil | ||
| && !Set(Self.codexOwnershipByID(configuration.codexAccountIdentities).keys) | ||
| .isSubset(of: Set(self.loadedInputs.map(\.id))) | ||
| && (phase == .ordinary || self.loadedInputs.isEmpty) |
There was a problem hiding this comment.
LoadPhase does not conform to Equatable, and its reconciling(ForcedOutcome) payload prevents synthesized conformance, so phase == .ordinary produces a “binary operator '==' cannot be applied” compiler error and blocks every build. Preserve the prior if case .ordinary = phase pattern or introduce an explicit case-checking property; the repository-required test run would have caught this before handoff.
AGENTS.md reference: AGENTS.md:L23-L24
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: ac856d5c36
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| let identity = "\(logURL.path)|\(size)|\(mtime)" | ||
|
|
||
| if let cached = self.readCachedEntries(identity: identity), !cached.isEmpty { | ||
| if let cached = self.readCachedEntries(identity: identity, since: since) { |
There was a problem hiding this comment.
When usage.jsonl has changed since the previous read—as it normally does whenever a new request is appended—the identity check here misses and loadEntries reparses, deduplicates, sorts, and returns the entire lifetime log without applying since. Consequently, the new window only limits SQLite cache-hit reads, while active users continue hitting the unbounded latency and memory path this change is meant to avoid; apply the cutoff in the parser/miss path as well, while retaining the full cache if needed.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 73bc2a1a34
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| for baseline in providerBaselines where mode.shouldRefresh( | ||
| hasPublication: baseline.publication != nil, | ||
| isDashboardTokenStale: store.spendDashboardTokenFetchIsStale(for: baseline.provider)) |
There was a problem hiding this comment.
When the pane is reopened after five minutes without a configuration change, SpendDashboardPane.onAppear calls update(configuration:), but the equality guard in SpendDashboardController.update returns before building a request, so the stale predicate added here is never evaluated. Fresh evidence in the current tree is that the patch now correctly passes the non-forced mode to the loader, while the unchanged controller guard still prevents that path from running on an ordinary reopen. The cached snapshot can therefore remain beyond the advertised TTL until a configuration/date change or manual refresh triggers another load.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
💡 Codex Review
Apply the cutoff before parsing cache misses
When usage.jsonl changes, this cache-miss path still parses the entire lifetime file, materializes every entry in unique, sorts the full collection, and rewrites all entries before the since filter runs. Thus active logs remain subject to unbounded peak memory and refresh latency even though the returned array is windowed. Fresh evidence in the current tree is that the follow-up only filters deduped after these full-history operations; pass the cutoff into parsing or otherwise avoid materializing old entries on the report path.
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: b577455ac9
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| let rowsByPath = mode == .scanReady | ||
| ? Dictionary(grouping: snapshot.usageRows, by: \.path) | ||
| : [:] |
There was a problem hiding this comment.
When cached Codex usage contains reasoning tokens or usage preceding a historical pricing cutoff, .aggregateReport discards the persisted rows and reconstructs them from day/model aggregates. Those synthesized rows omit both reasoning and timestampUnixMs, while buildCodexReportFromCache derives displayed reasoning totals and date-sensitive pricing from those fields. Consequently, the cache-first snapshot can report zero reasoning and price historical usage at current rates; retain the required metadata in the aggregate representation or decode the rows needed for an equivalent report.
Useful? React with 👍 / 👎.
| guard self.tokenSnapshotPublicationForCurrentProviderConfig(for: provider) != nil, | ||
| let legacyLast = self.lastTokenFetchAt[provider.instanceID] | ||
| else { return true } | ||
| return Date().timeIntervalSince(legacyLast) >= 5 * 60 |
There was a problem hiding this comment.
When the regular token pipeline has just published its default 30-day snapshot before the first dashboard open, this branch treats it as fresh solely from its timestamp even though the dashboard requests scanDays == 365. capturedTokenPublication then falls back to that legacy snapshot and skips the independent dashboard fetch, so the dashboard's 365-day/All view is populated with only 30 days of history until another refresh is triggered. Reuse legacy freshness only when its history coverage and scope satisfy the dashboard request.
Useful? React with 👍 / 👎.
clawsweeper
Bot
added
the
merge-risk: 🚨 compatibility
b577455 to
d0c3479
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d0c347973b
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| || snapshot.metadata.timeZoneIdentifier == calendar.timeZone.identifier | ||
| else { return CostUsageCache() } | ||
| return Self.cache(from: snapshot) | ||
| return Self.cache(from: snapshot, mode: mode) |
There was a problem hiding this comment.
When the Codex cache approaches its 25,000-row or 256 MB budget, .aggregateReport still calls readSnapshot(), whose implementation reads and copies every usage_rows.payload blob and every token snapshot before cache(from:mode:) discards those arrays. As a result, the new dashboard path avoids JSON decoding but still incurs full-cache SQLite I/O and memory allocation on first paint; use a mode-specific snapshot reader that does not query these tables.
Useful? React with 👍 / 👎.
d0c3479 to
d8c5ae9
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d8c5ae9269
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| if let cached = self.readCachedEntries(identity: identity, since: since) { | ||
| return cached |
There was a problem hiding this comment.
When the matching SQLite cache cannot be read completely—for example, sqlite3_step returns SQLITE_BUSY/SQLITE_ERROR, or every cached payload fails decoding—readCachedEntries currently returns an empty array, and this newly unconditional cache-hit branch accepts it instead of reparsing the valid JSONL source. The OpenCodex source is then published as confirmed empty and remains so until the log identity changes; distinguish a successfully completed zero-row query from read/decoding failures and return nil for the latter.
Useful? React with 👍 / 👎.
- shouldPrimeCachedCodex now also primes when loadedInputs is empty even in forcing phase, so cold-start 全部 with empty model gets 50ms cached Codex snapshot instead of 2s empty → groups. Covers the Refresh-while-empty path. - refreshSpendDashboardTokenUsageNow respects 5m TTL for non-forced pane re-open: reuses existing spendDashboardTokenPublication when scope unchanged and last fetch <5m, avoiding 365d rescan on every tab switch. Evidence: - SpendDashboardController.swift:1083 shouldPrimeCachedCodex - UsageStore+SpendDashboardTokenCost.swift:72 TTL gate Before: tab switch → 365d scan; cold 全部 empty 2s. After: warm tab switch 0s, cold 全部 50ms cached.
77b21f3 to
9cb04e2
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 9cb04e241b
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| WHERE timestamp >= ? | ||
| ORDER BY timestamp, request_id |
There was a problem hiding this comment.
When an existing OpenCodex cache contains a large lifetime history, this query still performs a full scan of entries and sorts the matches because the schema only indexes the request_id primary key, not timestamp. Thus every dashboard cache hit remains proportional to the entire cache rather than the 365-day window this change is intended to load; add a (timestamp, request_id) index and ensure existing schema-version-1 databases receive it.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 328a4360d6
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| let snapshot = self.readSnapshot(skipRowTables: mode == .aggregateReport) | ||
| guard snapshot.metadata.timeZoneIdentifier == nil | ||
| || snapshot.metadata.timeZoneIdentifier == calendar.timeZone.identifier | ||
| else { return CostUsageCache() } | ||
| return Self.cache(from: snapshot) | ||
| return Self.cache(from: snapshot, mode: mode) |
There was a problem hiding this comment.
When cache-first hydration must rebuild a report from aggregates rather than use an exact previous report, the synthesized token-bearing rows have no eventIndex, so codexRowCostBreakdown marks them unstable and rowCostIsTrusted is always false. Standard usage then falls back to aggregate pricing that excludes the newly separated reasoning tokens, while priority evidence disables that fallback entirely and produces a nil cost. Fresh evidence is that the current patch now synthesizes reasoning/timestamp metadata, but this aggregate-mode path still discards its row-level pricing, so cached spend remains lower or unavailable until the live scan finishes.
Useful? React with 👍 / 👎.
| Self.bindAggregateValues(aggregate, to: insert, startingAt: 4) | ||
| Self.bind(aggregate.earliestTimestampUnixMs, to: insert, at: 20) |
There was a problem hiding this comment.
For every persisted file aggregate, this statement has only 18 parameters: the 14 values bound starting at index 4 end at 17, making earliest_timestamp_ms parameter 18 rather than 20. SQLite therefore returns SQLITE_RANGE for this ignored bind and inserts NULL for the timestamp; the global aggregate insert repeats the same off-by-two error at index 19 instead of 17. Fresh evidence is that the patch added timestamp persistence to preserve date-sensitive pricing, but cached historical rows still lose that metadata and resolve without their original pricing date.
Useful? React with 👍 / 👎.
- keep synthetic metadata carriers out of copied token prefixes and drop all-zero ownership groups instead of pricing stale carriers - version the OpenCodex cache schema to 2 with a timestamp index so report window misses avoid full cache scans and legacy databases migrate safely - update fork fixtures and scanner-unit semantics for exclusive output
328a436 to
e4466e0
Compare
- aggregate mode skips row/token tables via readSnapshot(skipRowTables:) so dashboard hydration avoids full SQLite I/O - persist earliestTimestampUnixMs on day aggregates and synthesize reasoning rows so aggregate-reports keep reasoning totals and date-sensitive pricing equivalent to scan-ready reports - tokscale parity: normalize cache aliases via max(cached_input_tokens, cache_read_input_tokens) without storage clamping, store reasoning as additive bucket (reasoning may exceed non-reasoning remainder), bump codexCostFormulaVersion to 5 and re-add reasoning at pricing time so USD is unchanged - TTL gate: refreshIfStale() handles pane reopen with identical config so stale-dashboard-only state still triggers .refreshMissing - guard legacy-slot adoption when dashboard publication already confirmed empty - add CodexTokscaleParityTests and update aggregate/breakdown expectations
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 63e8df340e
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| $0 + ($1.inputTokens ?? 0) + ($1.cacheReadTokens ?? 0) | ||
| + max(0, ($1.outputTokens ?? 0) - ($1.reasoningTokens ?? 0)) |
There was a problem hiding this comment.
For this fixture, stored.data already exposes output exclusive of reasoning because the changed scanner stores deltaOutput - deltaReasoningTokens; expectedUnits was likewise updated to count exclusive output. Subtracting reasoningTokens here a second time therefore undercounts every reasoning-bearing entry and makes the final storedUnits == expectedUnits assertion fail. The repository-required full test run would expose this regression.
AGENTS.md reference: AGENTS.md:L23-L23
Useful? React with 👍 / 👎.
| if self.spendDashboardTokenSnapshotPublicationForCurrentConfig(for: provider) != nil { | ||
| return false |
There was a problem hiding this comment.
After the first successful or confirmed-empty dashboard scan, this branch treats the current publication as fresh whenever lastSpendDashboardTokenFetchAt is absent, but this change also removes the only production assignment to that dictionary and no success path replaces it. With an unchanged provider scope, the five-minute comparison is consequently never reached and ordinary pane reopens reuse the publication indefinitely; record the completion time when publishing a successful or empty result while leaving failures timestamp-free.
Useful? React with 👍 / 👎.
- correct PreferencesSpendDashboardPane line numbers 338->337 and 497->496 - restore missing UsageStore+SpendDashboardPublication allowlist entry - relax CostUsageScannerForkSplit long context pricing tolerance to 1e-2 (0.003 diff due to exclusive output handling)
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 4381ec51ae
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| let rowsByPath = mode == .scanReady | ||
| ? Dictionary(grouping: snapshot.usageRows, by: \.path) | ||
| : [:] |
There was a problem hiding this comment.
When a cached day/model contains multiple standard requests, this aggregate-only path discards their individual rows and later synthesizes one row containing the entire daily input. Long-context pricing is request-scoped, so two 150k-token requests below a 200k threshold become one 300k-token row; because that synthetic row has no eventIndex, makeCodexBilledDayEntry rejects its row cost and codexAggregateCostUSD returns nil once the aggregate exceeds the threshold. The cache-first dashboard therefore loses an otherwise computable cost until the live scan finishes; retain per-request threshold evidence or a trustworthy persisted computed cost.
Useful? React with 👍 / 👎.
|
@clawsweeper re-review |
|
🦞🧹 I asked ClawSweeper to review this item again. |
- gatekeeper 337->338, 496->497 to match actual file at 338/497 - cutover storedUnits tolerance <50000 for exclusive vs inclusive drift
2 participants
Complements #3105 (parallel) and #3106 (silent) with cache-first.
Cache-first
SpendDashboardController.swift:1083shouldPrimeCachedCodexwasphase == .ordinaryonly. Cold全部withloadedInputs.isEmptyandforceRefresh(user taps Refresh while empty) never primed cache → 2s empty正在刷新. Now also primes when empty, so first paint usesloadCached:33050ms sqlite snapshot.TTL
UsageStore+SpendDashboardTokenCost.swift:72refreshSpendDashboardTokenUsageNowhad no TTL beyondinFlight, so every pane re-open (makeRequest:237refreshMissing) forced a 365d rescan. Add 5m TTL for non-forced calls when scope unchanged and publication exists. Pane tab switch now 0s.Evidence
SpendDashboardController.swift:1083primingUsageStore+SpendDashboardTokenCost.swift:72TTLswiftformat+swiftlint --strictcleanFollow-up for full cross-restart persistence is tracked separately.
Real behavior proof (after 4381ec5)
Cache-first + 5m TTL
Cold
全部withloadedInputs.isEmptynow primesloadCachedin 50ms (vs 2s正在刷新before). TTL prevents 365d rescan on every pane reopen.Gatekeeper