Repository navigation
[pull] main from can1357:main - #110
Merged
Merged
Conversation
- Queried tmux client_termtype once at startup and reused the existing terminal profile mapping. - Covered the SSH pane environment with an executable tmux shim and documented required tmux features. Fixes #12524
An event-loop stall drains the pty backlog in one stdin read (or several reads in one tick before the 10ms classification timer can fire, since the loop was exactly what was blocked). `isRawMultilineBurst` then sees an ESC-free batch with two breaks plus trailing content and coalesces it onto the paste channel, so every typed Enter becomes a literal newline and nothing submits. The heuristic exists only for terminals that do not bracket pastes, so where bracketed paste is confirmed it is pure downside. Probe DEC mode 2004 via the existing DECRQM plumbing and disable StdinBuffer's raw-paste classification once the terminal confirms support; terminals that ignore DECRQM keep the heuristic. Fixes #12540
The steering watch loop wraps checkSteering() with a defensive .then(() => true, () => false), but the trailing await checkSteering() at the end of runTool was unguarded. A rejecting host hasSteeringMessages/hasIrcInterrupts callback made that runTool task reject after the tool already ran, so the start.then(() => runTool()) ordering chain rejected and every subsequent chained (exclusive/serial) record was never invoked. The tail sweep then emitted phantom "Skipped due to pending steering message" results for tools that were never executed and never steered. Mirror the watch-loop guard: await checkSteering().catch(() => undefined). The probe is best-effort and its failure is already surfaced by the dedicated watch path. Fixes #12544
agentLoop and agentLoopContinue ran the aside-commit loop and the agent_start push outside the settlement try, so a throwing host ASIDE_MESSAGE_COMMIT hook escaped as an unhandled rejection and left the returned EventStream unsettled, hanging callers awaiting stream.result(). runLoopBody also called discardAsides bare in its finally, letting a throwing discard hook replace any in-flight loop error. - Moved the settlement-critical work inside each IIFE try so every throw routes through stream.fail. - Wrapped discardAsides in the finally so a throwing discard hook is logged, never load-bearing. Fixes #12545
…panicking A sloppy edit whose ⟪…⟫ markers resolve to overlapping or same-start selection spans made rewrite_selection_spans splice a multibyte (CJK) replacement into the mutated content[start..end] buffer and re-index it on a mid-char edge, panicking the worker instead of returning a match error. - Added selection_spans_usable: a candidate's spans must be ordered, non-overlapping, distinct-start sub-ranges of [start, end]. - Search::visit now drops candidates that fail the invariant, so the miss surfaces as an EditError like the #12268 contract requires. - Regression test over CJK content that reproduced the panic pre-fix. Fixes #12529
- Preserved credential subtype metadata when broker refreshes write back bare OAuth grant responses. - Covered consecutive broker-backed MCP token rotations through the HTTP refresh endpoint. - Documented the user-visible fix in the coding-agent changelog. Fixes #12563
Made member email decisive when broker usage reports also carry email, so shared Team workspace identifiers cannot cross-wire quotas. Keyed usage overlays by email before shared account IDs and covered report routing and overlay merging with a regression. Fixes #12560
- Added the Automation entitlement required for first-time Xcode MCP access. - Documented and regression-tested the signed release contract. Fixes #12572
On case-insensitive filesystems Windows reports a project's cwd with
drifting casing across launches (C:\...\Documents vs ...\documents).
`encodeProjectPath` collapsed both into one on-disk memory root, but the
scope keys derived from cwd — `global:${cwd}` job keys, `threads.cwd`, and
the `t.cwd = ?` output filter — compared with SQLite BINARY collation, so
one project split across two scopes sharing a single directory. Phase 2 for
one casing saw zero outputs, wiped the shared MEMORY.md/memory_summary.md/
skills, and advanced its watermark, permanently stranding the other casing's
extracted outputs.
Add `normalizeScopeCwd`, applied at `globalJobKey`, `upsertThreads`,
`listStage1OutputsForGlobal`, and `getMemoryRoot`, to case-fold cwd on
win32 so one physical directory maps to one scope key.
Fixes #12596
The Anthropic prompt-cache keep-alive refresh captures a completed turn's
payload and replays it with `max_tokens: 0`. When that turn forced a tool
(e.g. a forced yield pinning `tool_choice: {type:"tool", name:"yield"}`), the
zero-output replay inherited the forced selector, and Anthropic rejects
`tool_choice` type `tool`/`any` with `max_tokens: 0` (400
`tool_choice ... cannot be used when max_tokens is 0`). The keep-alive request
then failed and the prefix was not kept warm.
Strip a forced (`tool`/`any`) `tool_choice` when assembling the zero-output
refresh params; a keep-alive produces no tokens, so the selector is meaningless
there. Added a regression test asserting the refresh body omits the forced
choice while the originating turn keeps it.
Fixes #12597
The Codex SSE path cleared the pre-response watchdog the instant fetch resolved headers, so a non-2xx error body was consumed with no live deadline: parseCodexError's response.text() (403) and fetchWithRetry's retry-status response.clone().text() (503) could hang the turn until the caller/global backstop when a server sent error headers then stalled the body. - Keep the pre-response guard armed for a non-2xx (or thrown) fetch and disarm only for a successful 2xx body, which the iterator idle watchdog governs. - Disarm a retried non-2xx attempt's guard before arming the next one. - Clear the guard after CodexApiError.fromResponse consumes the error body, and on the send() failure path. - Add a regression test covering the 403 (fromResponse) and 503 (fetchWithRetry inspection) error-body paths. Fixes #12664
…mocks InteractiveMode.init() schedules the tiny-title prewarm probe behind a setImmediate. When that callback outlives its test case, it runs with getSessionName() already restored and the old mode's tiny role still configured, so it calls tinyTitleClient.prewarm() while a later case's spy is installed — 'does not prewarm an unconfigured default row' then fails with one unexpected call on loaded CI runners. Flush one immediate tick in afterEach, while the case's own spies are still installed, so no deferred prewarm crosses a test boundary.
Esc during Hindsight auto-recall left the admitted submission stuck for the full recall timeout. abort() bumped #promptGeneration so the cooperative isCurrent() checks turned false, but the blocking recall fetch could not observe that until it resolved. - Thread an AbortSignal from prompt setup through buildSystemPromptForAgentStart -> MemoryBackend.beforeAgentStartPrompt -> HindsightSessionState.beforeAgentStartPrompt -> recallForContext -> HindsightApi.recall, which already forwards options.signal via withTimeoutSignal. - Add #promptSetupAbortController to AgentSession; the direct prompt path installs it and abort() aborts it, cancelling the recall fetch so the turn unwinds immediately. - Mirror the signal parameter on the mnemopi backend/state, bailing cooperatively after the local recall when aborted. Fixes #12668
…and a bg run A background bash job settled only when the native Shell.run promise resolved, which waited for brush stdout to fully drain into JS. Under high-volume output the reader parks on the bounded bridge (send_async) whenever the JS chunk pump falls behind; if a single call_async wedges, the bridge stays full, the pipe reader parks, the child blocks in write(2) and never exits, so run_string/Shell.run/executeBash never return. A timeout:0 job has no deadline, so the job stays running and hub wait loops forever. Give pump_chunks a per-forward stall deadline (FORWARD_STALL_TIMEOUT). A slow-but-progressing consumer resets it per chunk, so output still drains losslessly; only a genuinely wedged forward is abandoned, which drops rx, disconnects the bridge, unblocks the reader and child, and lets the run settle instead of stranding. Fixes #12657
…against rejection (@roboomp)
…ans instead of panicking (@roboomp)
…l-batched enter (@roboomp)
Replaces the hand-maintained bind list from #12856 with a prototype walk so newly added methods cannot be left unbound.
…lty from google transports (@roboomp)
…before restoring mocks (@andrebrait)
…ns from dashboard opt-in gate (@roboomp)
- Delete redundant test file for umans provider which has been superseded by integrated model discovery tests.
- Implemented dynamic Claude Code version tracking, allowing the runtime to adopt newer versions requested by Anthropic's API via `claude_code_version_too_old` rejections. - Introduced automatic retry logic in `streamAnthropicOnce` that re-initializes client fingerprints when a version mismatch is detected. - Updated Claude Code version management to fetch and pin current versions via an improved version-spoofing check script. - Promoted default models across multiple providers in `catalog` from Opus 4.8 to Opus 5.5.
- xkb.rs tests include_str! testdata/fr.xkb and us-fr.xkb (added in 95776f0); cargo reads them from the source tree but the bazel sandbox only sees declared compile_data, so //crates/pi-natives:pi_natives_test failed to build on linux CI.
- Removed tests pinning zenmux's defaultModel to a specific id; the default is catalog data, not a contract.
- map_or_else and const fn per clippy in xkb.rs; allowed non_send_fields_in_send_ty on the documented unsafe Send impl for Libei, matching the macOS input pattern. - The linux-only wayland module (95776f0) had never run under the bazel clippy-strict aspect.
…llback test - The startup-fallback regression test hardcoded claude-opus-4-8; 2282226 moved the anthropic default-model to claude-opus-5-5, so the literal broke while the behavior under test stayed correct. Read DEFAULT_MODEL_PER_PROVIDER.anthropic instead.
…redemption - Implemented Claude usage reset credit discovery and redemption infrastructure, supporting Cedar and Juniper programs. - Added automated Claude reset redemption planning, including strategies for recovery from usage blocks and salvaging expiring credits. - Generalized reset UI and CLI commands to support multi-provider (Anthropic/Codex) account identification via organization and credential IDs. - Enhanced usage dashboards and status line components to display banked credits, eligibility status, and expiry details. - Sanitized reset-related output strings for account labels and error diagnostics to ensure CLI and TUI consistency. - Updated slash commands to require provider-prefixed target arguments (e.g., anthropic/) for explicit reset control.
- 3d4d323 generalized /usage reset from Codex-only to any provider; the RPC metadata test still pinned the Codex wording and usage string.
Anthropic's plain `/api/oauth/usage` body lists every promo program as a key whose value stays `null` until the matching probe query asks for an evaluation. `parseClaudeResetCreditsFromUsagePayload` keyed off `Object.hasOwn` and treated such a `null` as an authoritative "no resets", so `fetchClaudeUsage` short-circuited on a synthetic zero inventory and never issued `?cedar_ember=1&skip_spend=1` / `?at_wall=1&skip_spend=1` — the 18.2.9 discovery path could not run on a real account, and accounts holding a redeemable Cedar grant saw no reset row. Only an object block now counts as an answer; anything else returns `null` so the caller falls through to `listClaudeResetCredits()`, which is the authoritative source for the empty/ineligible case (including the server's `reason`). Fixes #12883
…ull-program-probe'
…ance optimizations - Introduced `intent` parameter to `judge_batch` for descriptive job progress labeling in the UI. - Added real-time judgment batch progress HUDs to the TUI to monitor bulk eval tasks. - Optimized `judge_batch` `drain()` to coalesce near-simultaneous item settlements into single batch deliveries. - Implemented event-loop-safe status event coalescing to prevent UI stalls during high-frequency batch progress updates. - Added TTL-based candidate list caching for `resolveJudge` to reduce redundant catalog filtering in bulk operations.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
See Commits and Changes for more details.
Created by
pull[bot] (v2.0.0-alpha.4)
Can you help keep this open source service alive? 💖 Please sponsor : )