fix(server): stop querying Claude context usage after turns - #8610
Conversation
|
Important Review skippedAuto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the ⚙️ Run configurationConfiguration used: Repository UI Review profile: CHILL Plan: Pro Plus Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
Comment |
Thread transfer impact✅ Thread transfer remains within every enforced ceiling.
Baseline: Scenario and decoded snapshot size10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.
Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed. |
ApprovabilityVerdict: Would Approve Macroscope's review found this PR approvable — This is a localized Claude adapter bug fix that removes unintended context-usage queries and adds targeted usage/compaction coverage without changing schemas, defaults, deployment, or sensitive paths. A separate unresolved Medium finding flags lost automatic-compaction metadata, leaving a correctness risk recorded for follow-up. Not approved because:
Adjust the Minimum Blocking Severity for this repo — including turning it Off — in Settings. You can add or adjust custom eligibility rules. Learn more. |
Main's #8610 (stop querying Claude context usage after turns) only touches the v1 ClaudeAdapter layers this branch replaces with orchestration-v2; the v2 Claude adapter never issued the post-turn getContextUsage control query, so the fix has no v2 counterpart to port. Keep the deletions. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Merges 58 upstream commits (`badae6a5c` → `6a9d9f988`, upstream v0.0.34 → v0.0.36) through the `fork-upstream-merge` skill. `merge-stats.mjs` reconciles exactly: **431 files landed** against **431 in the upstream range**, no gap to explain. Fork delta is 611 files. ## Conflicts Three, plus one git resolved silently and wrongly. - **`Sidebar.tsx`** — upstream replaced the project picker's `Menu`/`MenuRadioGroup` with a searchable `Combobox` (pingdotgg#5931). Took upstream's structure and re-applied only the `FEATURES.projectManagement` gate on "New project", now carrying the `// Fork:` marker it had been missing. Upstream has since grown the mobile touch-target span itself, so the Mobile Touch Delta has nothing left to re-apply here — one convergence, unprompted. - **`ProviderSettingsPanel.tsx`** — upstream split provider settings into list and editor (pingdotgg#8380, pingdotgg#8472), moving `ProviderLastChecked` and the refresh button out of `headerAction` into the list footer. Took upstream's and re-applied the gate as `!readOnly && FEATURES.serverAdministration`. - **`pnpm-lock.yaml`** — took upstream's and re-ran `vp i`, which restored the fork's `@t3tools/moatless-api` workspace edge. - **`packages/contracts/src/orchestration.test.ts`** — auto-merged, no marker, broken. Both sides appended `OrchestrationMessage` to the same import list and the same `decodeOrchestrationMessage` const at different offsets, so git took both. Surfaced as a parse error in lint, typecheck and test at once. This is the case `merge-stats.mjs`'s conflict-candidate list exists to catch. `HostedBrowserFrame.tsx` (fork-only) also needed the `renderingActive` prop pingdotgg#8567 made required. Upstream suspends a parked webview unless background audio, PiP or a recording still needs it painted; a frame has none of those to read and is the app's only copy of the preview page, so it passes `true` and keeps today's behavior. ## Sweep Five keyword hits, all false positives: - `apps/web/src/connection/clientMetadata.ts` + test — reports the client's OS/browser/device on connect (pingdotgg#8481). Auth-adjacent, but it rides `ClientPresentation` on the relay and remote-bearer bootstraps; the fork's primary environment sends none of it. - `packages/client-runtime/src/relay/errorPresentation.ts` + test, `connection/errors.test.ts` — explain DPoP failures (pingdotgg#8351). Relay only, and T3 Connect is decided out. ## Feature classification **Usable as-is** - Searchable project-filter combobox in the sidebar (pingdotgg#5931). - Long task drawers stay usable on small screens (pingdotgg#8313) — directly relevant, the fork's phone story is `apps/web` in mobile Safari/Chrome. - Toggleable confirmation before unpinning a thread (pingdotgg#7313); toggle thread pin from the keyboard. - Back button in project settings (pingdotgg#8168); the configured stash shortcut is shown (pingdotgg#8437). - No more red-x summaries for ordinary tool failures (pingdotgg#8395); PRs open in the browser when loading fails (pingdotgg#8507). - Project default models are honored in new threads (pingdotgg#6011). - Provider settings split into list and editor (pingdotgg#8380, pingdotgg#8472) — landed, though `/settings/providers` is itself gated behind `serverAdministration`. **Unsupported in Moatless / needs implementation** - **Non-image file attachments** (pingdotgg#8235) — a turn may now carry any file up to 50MB, advertised as `capabilities.fileAttachments.maxUploadBytes` and sized by `PROVIDER_SEND_TURN_MAX_FILE_BYTES`. Moatless advertises neither this nor `attachmentUploads`, so the composer's attach affordance stays off. Costs nothing today (upstream's own web composer offers images only; `ChatAttachment` widened just far enough to typecheck), but it is the second capability key to report when uploads land. Extends the existing **Attachment uploads** entry in `docs/fork/gaps.md`. - **Connected-client platform analytics** (pingdotgg#8481) and **DPoP failure explanations** (pingdotgg#8351) — relay and T3 Connect surfaces, already decided out of the fork. No new WS methods entered the contract in this range, so no new `UnsupportedMethodError` union entries. `unsupported-methods.mjs` reports `ADD 0`, `DROP 1` (`scripts.run`), `KEEP 2`. The `scripts.run` DROP is **not** actioned — it is the documented exception in the gaps register: that union entry answers for `apps/server`, which still stubs the method (verified surviving in `ws.ts`), not for Moatless. **Backend behavior to consider reproducing in Moatless** - **Replay all un-applied events during projection bootstrap** (pingdotgg#7538) — a one-line pipeline fix upstream; Moatless runs its own projection and the same class of bug applies. - **A draft can retry its first send after a failed bootstrap** (pingdotgg#8226) — Moatless owns turn start, and a draft wedged by a failed bootstrap is the same dead end here. - **Recover stale Codex approval callbacks** (pingdotgg#5195) — Moatless owns the agent runtime in the sandbox. - **Stop querying Claude context usage after turns** (pingdotgg#8610) — a cheap correctness/cost fix in the same place Moatless polls. - **Server-side accept and size-limit non-image uploads** (pingdotgg#8235), the other half of the gap above. - Auto-settling churn (pingdotgg#8321 made it opt-in, pingdotgg#8596 reverted) nets to no change — noted so the next merge does not re-derive it. ## Verification `verify.mjs`: `tripwires`, `fmt:check`, `lint`, `typecheck` pass. `test` reports `@t3tools/web` **flaky, passed alone** — in the full run `MessagesTimeline.test.tsx` skipped all 34 tests on a 30s module-import timeout under `ChatMarkdown.tsx`; alone the package is 297 files / 3117 tests green. Load, not the merge. `inventory-check.mjs` clean. Model: Claude Opus 5, via Claude Code in a Moatless sandbox. --- Moatless task: https://moatless.soaplabstest.com/tasks/a0c041bb-1426-4591-9296-6a4b0cfa2eff
Brings in 192 upstream commits through v0.0.36 (053affb). Most of the 52 conflict hunks were additive: the fork's thread `group` landing on the same lines as upstream's new `linkedPullRequest`. Both sides kept. Real tie-breaks, all resolved for upstream: - Projector bootstrap paging dropped for upstream pingdotgg#7538, which drains the whole stream in one read. BOOTSTRAP_PAGE_SIZE went with it. - Claude `getContextUsage` dropped for upstream pingdotgg#8610. The fork's plan usage meter sat next to it and stays; only the context-usage pull is gone. - The Windows logs-directory assertion in server.test.ts dropped for upstream's path.basename form, which handles both separators anyway. The fork's thread group migration moves from 042 to 044, above upstream's new 042 and 043. The migration runner skips any id at or below the highest one already recorded, so a fork migration parked on an id upstream later reuses masks upstream's silently. Existing databases that recorded 42 as ProjectionThreadsGroupKey need that row dropped once so upstream's 042 and 043 can run; all three migrations are idempotent. Upstream also stopped passing the raw activity list into the composer, so the plan usage meter now takes a derived snapshot prop like the context window does, computed in ChatView.
Resolves an import conflict in ClaudeAdapter.ts: pingdotgg#8610 dropped the detailed context-usage query, so SDKControlGetContextUsageResponse no longer has a consumer and is gone; SDKRateLimitInfo stays, since the usage-limit surfacing this branch adds still uses it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…gg#8610 Turn completion now prefers the last assistant usage over the cumulative result total, so the parent's used count here is its real 12,000 rather than the clamped window. The subagent total this guards is unchanged.
Upstream pingdotgg#8610 (`stop querying Claude context usage after turns`) brought six `Option.isNone`/`isSome` call sites into ClaudeAdapter, but the merge dropped the accompanying `effect/Option` import. That is a ReferenceError at runtime in the context-usage and stop-acknowledge paths, not just a type error — the package had nine TS2552 errors. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Range f6f2be3..053affb, v0.0.36. 411 upstream files, 632 fork files, 81 touched by both, 19 conflicted. Competing implementations resolved toward the fork's superset: - Claude context usage (pingdotgg#8610). Upstream deleted `queryCurrentContextUsage` and `normalizeClaudeContextUsageApiSnapshot` outright. `getContextUsage` is the ONLY source of the compaction facts (`autocompactSource`, `autoCompactThreshold`, `isAutoCompactEnabled`) the Vitals gauge's compaction note and marker render from, and they travel on the wire contract, so the call is kept. Upstream's `latestAssistantUsage` tracking and its post-compaction guard are adopted BELOW it as the next-best fallback: `contextUsageSnapshot ?? latestAssistantSnapshot ?? ...`. Upstream removed `getContextUsage` from `ClaudeQueryRuntime`, the `SDKControlGetContextUsageResponse` import and the `Option` import outside every conflict marker; all three restored. - Provider settings editor (pingdotgg#8472). Upstream's `useEffect` + `lastPublishedEnvironmentRef` re-seed is the superset of the fork's render-phase `seededRef` (8fe3190, one of four silent-data-loss fixes). The merge had kept BOTH mechanisms; collapsed to upstream's, with the fork's half-typed-row null guard on the publish path. `environmentKey` deleted with the code it served. - Sidebar project combobox (pingdotgg#5931). Upstream replaced the menu-based project filter, whose `MenuRadioItem` carried the fork's project-actions button. Rebuilt on upstream's `ComboboxItem` with the ellipsis button grafted back (invariant 6). The dead `Menu*` import is gone. - Composer attachments (pingdotgg#8236). Upstream's `addComposerAttachments` pipeline adopted; the fork's drop split (image -> attach, other -> absolute path with upload fallback) kept and repointed at the new handler. - Mid-turn steer. Upstream's `steeringTurnState` branch rejected again (invariant 5); its side of that hunk also opened an `if` whose brace lives below the hunk. - Thread deletion drain (pingdotgg#8226). Upstream replaced `drain` with `drainThrough(sequence)`; the fork's test retargeted. Its new drain test observed `ProviderService.stopSession`, which this reactor never calls (it dispatches `thread.session.stop`), so it now counts engine dispatches and provides the two services the fork's reactor needs. Three defects the resolutions introduced, each caught by a check that is not the suite: - `shouldShowBranchMismatchBanner` was resurrected into two import lists by resolving them toward upstream. Fork commit cfdf255 replaced that function; the import had no definition. Found by the RESURRECTED sweep. - `scriptPath` ended up declared twice in CodexCollabRuntime.integration.test.ts: `both` reintroduced upstream's fixture-dir copy that the fork deliberately moved to a temp dir. Found by the RESURRECTED sweep. - `setProjectScopeMenuOpen` survived in the fork's `handleProjectActions` after upstream replaced that state with a reducer, outside every marker. Found by typecheck. Upstream regressions of fork fixes, kept fixed: - Raw NUL bytes returned to ChatComposer.tsx. A raw NUL renders as nothing in grep and aborts BSD sed mid-file; the fork had converted all six to `\0` escapes and carries a guard test for it. The guard failed, and the offending line reads as space-separated in any normal diff. Converted back to escapes. - `defaultTheme` / `defaultThemeSetAt` (pingdotgg#8569) are absent from `ServerSettingsPatch`; the fork's patch-parity guard caught it. They are owned by `t3 theme set`, which rewrites settings.json directly and which clients only read, so they are recorded as deliberately unpatchable rather than mirrored. - The fork's mobile held-messages strip used `dark:` variants that upstream's new Uniwind theme rule (pingdotgg#7327) forbids; moved to the `text-foreground-muted` semantic token rather than added to the rule's allowlist. Invariants: all re-probed against the merged tree. 1 (49 migrations, ids unique and monotonic, max 50, 34 burned), 2 (the fork-only @effect/platform-node patch survived the Expo 57 rewrite of patchedDependencies, is pinned, resolved and applied on disk; effect stays at beta.103 so no re-pin), 3, 4, 4b (4 `isSendBlocked=`), 5, 5b, 5c, 6, 7, 8, 9, 10, 11 all hold. Sweeps: resurrected 3, dropped 21, fork-loss 45, both-kept 0. Every entry named: the 3 resurrected and the fork-loss in ProviderInstanceCard / ChatComposer / Sidebar are the decisions above; the OpenCode fork-loss is 11 comment lines whose guard upstream extracted into `deleteContextIfCurrent` (the rationale is restored on the helper); pnpm-lock churn is the install. Test declarations 18,691 -> 19,081; one file lost one, upstream's own pingdotgg#8484. Gate: pnpm run verify green, exit 0 read from the captured log line. 14 blocks, 10,744 passed, 20 skipped, 0 failures. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Four new invariants and one correction, each from something this merge got wrong once before a check caught it. 12. The Claude adapter still calls getContextUsage. Upstream pingdotgg#8610 deleted it; it is the only source of the compaction facts the Vitals gauge renders, and upstream's replacement is adopted underneath rather than instead. Names the three symbols upstream removed outside every conflict marker, and why two of its tests are retargeted. 13. The provider-settings re-seed is upstream's on purpose. The merge kept both it and the fork's older one; restoring the fork's is adding a second, not restoring a fix. 14. Upstream keeps reintroducing raw NUL bytes into ChatComposer.tsx. The line reads as ordinary spaces in every diff and grep; only the guard test and cat -v see it. 15. defaultTheme / defaultThemeSetAt are deliberately unpatchable, so the patch-parity guard needs to know rather than be relaxed. Also: invariant 6's two project buttons now live in a ComboboxItem, and the Surface counts are redated instead of quoting stale figures.
Claude queried detailed context usage after every completed turn. If native token counting failed, that query could trigger extra model requests.
This change stops the automatic detailed query. Completion now uses the final main assistant usage when available, keeps cumulative result usage as processed tokens, and falls back to the existing result, iteration, and stream snapshots.
Focused Claude adapter tests verify usage, cumulative totals, turn completion, and zero detailed context queries.
Tradeoff: the context meter omits its automatic compaction note because normal SDK messages do not expose the effective policy. Claude Code's automatic compaction behavior is unchanged.
Refs #7338
Built by GPT-5.6 Sol in the Codex harness.