Repository navigation
[pull] main from can1357:main - #115
Merged
Merged
Conversation
- Increased probe count to ~639 k and adjusted base token count to 53 622 in data files and validation constant. - Added numerous new fixture entries covering whitespace, punctuation, and rare characters, and updated documentation to reflect the revised counts.
…t Jev token limits - Imported `countTokens` and `Encoding` from pi-natives and added `JUDGED_CONTENT_MAX_TOKENS` constant. - Implemented `jevPrefix` function to truncate output content to a token budget without splitting surrogate pairs. - Replaced character-based slicing of judged content with token-aware prefix, reducing max content fro 8000 m 60,000 characters to 32,000 Jev tokens. - Exported `JUDGED_CONTENT_MAX_TOKENS` and updated tests to verify that output is trimmed to the token limit while preserving full English output.
…turn recursion guard - emitContext rejects when its signal aborted instead of returning untransformed context - onTextDelta runs outside the recursion guard so work it starts is not permanently blocked - add regressions for hook recursion rejection and onTextDelta-started consultations - move changelog entry from released 18.1.20 to Unreleased; refresh side-turn docs
…w-ups test(compat): address capability policy review follow-ups
…-out Expose ephemeral turns to extensions with a tool opt-out
…ble interrupts - Introduced `softInterrupts` flag to differentiate immediate and wait interrupt modes. - Modified steering abort logic to respect `softInterrupts`, aborting only when appropriate. - Adjusted `executeToolCalls` watch conditions to include soft in 10BC8 terrupts and interruptible records. - Added conditional `wait` tool hint in eval prompts when the tool is available. - Removed automatic inclusion of `wait` tool from subagents, limiting it to top-level agents.
- Introduced `TOOL_INTERRUPT_ABORT_REASON` to distinguish tool signals caused by queued steering, peer messages, or background completion from user-triggered run aborts. - Fixed job duration reporting to display accurate run duration for completed, failed, or cancelled jobs instead of age-based duration. - Corrected foreground-to-background lifecycle: foreground-backed jobs are now hidden from listings until backgrounded, preventing foreground commands from surfacing as abandoned background jobs. - Ensured background job IDs are non-recyclable across a session to prevent ID collisions in `proc://`. - Enhanced `wait` to detect and return completed job results that are settled but not yet delivered to the transcript, preventing missed or delayed completion reports.
- Captured the regular-file identity backing builtin stdout across Unix and Windows. - Excluded that identity from rg search and file-list traversal while keeping pipes and terminals unchanged. - Added a bounded shell regression that exposes self-search without runaway output. Fixes #13032
The compiled-binary module resolver reports realpath-resolved filenames, while a registered runtime root is whatever the caller configured. installRuntimeModuleResolver compared the two with raw string math, so a runtime cache reached through a symlink/junction (or a case-differing Windows path) failed both containment checks: the stock hit on a package's TS source entry was never corrected to its manifest entry. In the fastembed runtime cache that meant a compiled binary loaded @huggingface/hub's root index.ts instead of dist/index.js, and that TS graph's transitive 'import "@huggingface/xetchunk-wasm"' resolves through Bun's ESM resolver, which never consults the patched Module._resolveFilename — the embed worker died with ResolveMessage: Cannot find package. Both sides are now canonicalized (realpath, Windows case folding) through a memoized helper before comparison. Fixes #13034
Side turns (runEphemeralTurn, handoff) mint a fresh session id per call, so a burst of them pushed the live conversation's control baseline out of the 16-entry LRU and forced the next request to re-declare tools, system, and top-level effort. Eviction now drops keys no request came back to before any conversation that was continued at least once. Control transitions are also immutable once sent: recordAnthropicControlTransition no longer merges into a transition already on the wire, and materializeAnthropicControlTransitions no longer folds a later control into the system message an earlier transition produced. Fixes #13028
- Replaced the continued-flag eviction with two plain LRUs: a new baseline enters pendingControlStates and is promoted to controlStates when a later request resolves it, so one-shot side turns only churn the pending tier. - Abandoned continued baselines (compaction, system-prompt change) now age out again instead of pinning the cache and letting a single side request evict a fresh conversation's baseline. - Added a regression test for a post-compaction baseline surviving an idle recap once abandoned baselines fill the cache. Fixes #13028
openai-models-list discovery read only input modalities, so a /v1/models row advertising an embeddings or image output was registered with the provider's chat API and no kind: the chat picker offered a model its endpoint cannot serve, and the embedding role never saw it. Rows whose sole advertised output is embeddings or an image now build with kind embedding/image on openai-embeddings/openai-images against the same OpenAI-compatible root, with no tool support and no output-token budget. Rows that also emit text stay chat, and audio/video outputs stay chat because the modality alone cannot name an endpoint. Fixes #13021
…asterizer noise - Added PIXEL_CHANNEL_TOLERANCE constant to treat ±2 per-channel differences as unchanged. - Introduced pixelChanged helper that compares two PNG pixels with the tolerance. - Modified pngPixelChangeRatio to use pixelChanged, ignoring rasterizer noise in the change ratio. - Updated createPngDiff to rely on pixelChanged for detection and adjust diff color coding. - Documentation for pixelChangeRatio fields now clarifies that rasterizer noise is ignored.
… auth - Migrated Anthropic compaction to `compact-2026-09-04` API with signature-based block validation. - Added support for Anthropic User Profiles API with persistent session attribution. - Implemented credential-aware auth resolution to track and propagate `credentialId` across request retries and streams. - Added cross-credential thinking block redaction to prevent leakage across different API accounts. - Enhanced agent session maintenance to handle native compaction snapshot branches and rewinds. - Expanded compaction support to include `claude-mythos-preview`.
- Replaced empty-content `write` cancellation with explicit `/kill` path for `proc://` jobs and services. - Added mandatory `content` requirement for all `write` operations except `proc://<id>/kill` to prevent accidental file truncation. - Updated `proc://` protocol to disambiguate service stdin input from lifecycle commands. - Improved `write` tool error messaging for invalid or ambiguous protocol paths. - Updated system prompts and documentation to reflect the new `proc://` usage contract.
- Introduced `SentToolDefinitions` to maintain persistent tool states across requests, preventing premature tool removal. - Added `inactiveTools` tracking in `Context` to support robust tool rebasing and prevent history-based tool definition loss. - Updated Anthropic message conversion to preserve assistant control messages during interruptions, improving response stability. - Mandated `getApiKeyWithCredentialForProvider` in `ApiKeyResolverRegistry` to ensure consistent authentication resolution. - Strictly scoped agent process kill requests to prevent accidental interference with non-owned processes. - Integrated schema-based validation for Anthropic User Profile API responses to ensure data integrity.
- Added `messageIndex` field to AnthropicRequestControls to record the response's own index. - Extended control record type with `live` flag and updated collection logic. - Modified tool control planning to skip non-live records, preventing duplicated controls from rewritten history. - Adjusted effort control planning similarly to ignore non-live records. - Set `messageIndex` in built request controls to current message count. - Updated tests to include `messageIndex` in expected requestControls objects.
`stat.rs`'s `#[cfg(all(test, windows))] mod win_tests` called `tempdir()` six times without defining it — the `(TempDir, PathBuf)` helper lives in the sibling unix-only `tests` module as `canonical_tempdir`, so `cargo test -p pi-builtins` and `cargo clippy --all-targets` failed with six E0425 on Windows before any test ran. - Added a local `tempdir()` to `win_tests` that keeps the path as `tempfile` reports it; `fs::canonicalize` would hand the Win32 stat and volume backends a `\\?\` verbatim cwd. - Imported `output_handle` into `host.rs`'s `#[cfg(test)] mod testing`, which `Host::set_test_stdout` calls. That break is platform-independent and kept the crate's test target from compiling anywhere; `check:rs` misses it because its clippy pass omits `--all-targets`. Fixes #13048
- Added deprecated `getApiKey` method delegating to `AuthStorage.keys.get` for backward compatibility. - Added deprecated `reload` method delegating to `AuthStorage.credentials.reload` for backward compatibility. - Updated `scripts/rewrite-changelog.ts` to use `storage.keys.get` and include explicit `sourceLabel` in error messages. - Updated `scripts/session-stats/audit.ts` to use `storage.credentials.reload` and `storage.keys.get`.
- Allowed missing_const_for_fn on apple_fm_cancel: on non-macOS the stub cancel is const, so clippy wanted the #[napi] export const, which napi cannot generate. - Allowed unnecessary_wraps on the non-macOS generate stub, whose Result signature mirrors the fallible macOS bridge. - Broke the bazel "Validate Rust workspace" job on Linux for v18.3.0.
- Replaced the maxTokens inequality between the resolved kimi-k3 reference and the ClinePass fixture: the catalog regen raised bundled fireworks kimi-k3 maxTokens to 1048576, matching ClinePass, so the test failed without any code regression. - Now asserts the KDL isolation predicate and that the bare-id resolver never returns a ClinePass entry, independent of upstream limits.
- Retargeted the xAI and Gemini web-search rotation tests to getApiKeyWithCredentialForProvider, which createApiKeyResolver calls since 7d5e994; the stale getApiKeyForProvider spy let every retry reuse the stored key, so all attempts saw 401. - Mocks return ResolvedApiKey to match the registry signature.
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 : )