Repository navigation
[pull] main from can1357:main - #133
Merged
Merged
Conversation
Parallel calls whose tool_call hooks return the same additionalContext (e.g. one guidance line per search call) produced a developer message repeating that text once per call. joinAdditionalContext now keeps the first occurrence of each distinct value.
… one call Several tool_call handlers on the same call returning identical additionalContext were joined before the batch-level dedup saw them, so the repeat survived as one composite value. Aggregate through joinAdditionalContext as well, and align the hook docs.
Batch dedup compares each call's joined context, not individual handler values; narrow the contract wording to match and add the missing coding-agent changelog entry.
Scrollback retirement renders the published reply once per batch instead of once per row, the transcript stops rescanning session history every frame, finalized Markdown releases its streaming caches, provider decoders drop per-character and per-delta scans, and tool previews/outputs avoid whole-payload rework per frame or chunk.
Addresses review: TailBuffer marked verbatim truncateTailBytes results clean, so a budget-sized chunk with a lone surrogate lost the next character on trim. Also typed ServerSentEvent.raw readonly (observers copy it for onSseEvent), replaced a double cast with Reflect.get, and dropped test assertions pinning sanitizeText's U+FFFD side effect.
Benchmarked at 1.04x on a 1.7M-char ConPTY replay frame and no change on typical frames; not worth the extra chunkForConPTY parameter.
…tribute changelogs
…nce per batch (@andrebrait) fix(agent): deliver identical passive tool context once per batch
- Renamed the Codex image model identifier from `openai-codex/gpt-image-1` to `openai-codex/gpt-image-2`, updating resolution of saved selectors. - Added a provider-alias entry in `_collapse.kdl` mapping the old model id to the new one, ensuring backward-compatible selector resolution. - Updated the compatibility rules JSON file to include the new model identifier mapping.
- Introduced `git` builtin in `pi-shell` that intercepts `git worktree add` for faster, in-process worktree creation. - Leveraged `pi-vcs` to perform copy-on-write clones (APFS/reflink/ReFS), preserving ignored build artifacts like `target/` and `node_modules/` in new worktrees. - Implemented transparent fallback to the system `git` binary for unsupported options or repository configurations. - Enabled the builtin via the `PI_SMART_GIT` environment variable. - Maintained compatibility with existing `post-checkout` hooks and standard `git worktree list` output reporting. - Strictly limited in-process handling to `add` operations; all other git subcommands or global overrides remain handled by the original git binary.
- Batch task calls now start subagents speculatively as their task items stream, reducing execution latency. - Introduced a speculative launch coordinator that manages lifecycle adoption and aborts for out-of-sync subagents. - Added host authorization checks to ensure speculative launches only occur under safe, auto-allow conditions. - Optimized stream speculation and tool call handling by ensuring message transforms preserve tool call structures. - Updated task tool to support speculative launch integration and concurrent job adoption during dispatch. - Added terminal-native UI feedback for task execution status within the Tern Surface Protocol.
…ng behavior - Changed hardened git CLI args: reads now receive --no-optional-locks, writes now pin fsmonitor and untrackedCache off. - Increased git status refresh interval to 10 seconds in the status line component. - Added generation counter to git status fetching to prevent stale counts after HEAD moves. - Reset git status cache and incremented generation on branch invalidation to ensure fresh data. - Used generation token to decide whether fetched git status timestamps should be applied.
…ry catalog rebuilds - Added ModelRegistry.refreshIfStale method that synchronizes the model catalog only when background refresh finishes or models.yml changes. - Updated ModelPickerComponent to call refreshIfStale and synchronize UI only when the catalog changed. - Fixed alt+p and `/switch` model picker latency by avoiding full catalog rebuilds. - Added tests confirming refreshIfStale returns true after model configuration changes and false otherwise.
…ool choice Command Code's /responses route rejects an object tool_choice and returns empty image_generation_call results, so its gpt-* rows now opt out of hosted image and mark named tool choice unsupported. `hosted-image #false` now clears a baked hostedImage flag. Adds deepseek-v4.1-flash-fast from command-code@1.67.0 and the newly served claude-sonnet-5-5 with Anthropic's first-party metadata.
…d tool choice OpenAI-style web search now sends tool_choice "required" for models whose compat marks named tool choice unsupported, so Command Code GPT models can run hosted web search.
…d related capabilities - Added compatibility rules for Anthropic Sonnet family (>=5.5 <6) with forced tool choice disabled and thinking prefix binding enabled. - Added mid-conversation system, tool change, turn-scoped system, per-message effort, and thinking binding capabilities for Anthropic Sonnet models. - Added support for Claude Sonnet 5.5 model with reasoning, text and image inputs, and defined cost and token limits.
- Annotated the it.each table so expected infers string[] instead of readonly tuples, fixing TS2769 in check:types.
- Applied cargo fmt to crates/pi-vcs so the bazel rustfmt check in Validate Rust workspace passes.
…ider-refresh # Conflicts: # packages/catalog/src/compat/rules.json # packages/catalog/src/models.json
- Login rejects only a 401 from /alpha/whoami, matching command-code@1.67.0 validateCommandApiKey. A new opt-in `trust-forbidden=#true` on `validate` lets an optional probe trust a 403; only Command Code sets it. - Discovery drops listed rows whose id matches a seed, so typesafe/jev stays one System One judge row. - `image-model #false` clears the class gpt-image-2 grant from the GPT rows. - The Responses route and GPT rules list the 10 ids exactly, so gpt-oss-style ids stay on chat completions. - Changelog covers the CLI price corrections.
- omit absent resetsAt/limit/usedFraction instead of emitting undefined, which the usage wire schema rejects - record the billing org on each limit scope; credits are queried per org - drop the unused ?limits=1 on /alpha/whoami - fix stale gpt-* routing and optional-probe comments
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 : )