Repository navigation
[pull] main from can1357:main - #138
Merged
Merged
Conversation
When Cursor's stream closed without its turnEnded frame (for example an HTTP/2 reset that settles as a clean stream end), the turn failed with "Cursor stream ended before turnEnded". The message matched none of the premature-close patterns, so a turn whose tool calls had all run was treated as replay-unsafe and stopped instead of continuing. The Cursor message now joins the OpenAI/Codex premature-close class, which is retried with completed tool results preserved. Refs #6772
…t grammar, forward tool args in background job progress
The in-process `git worktree add` tests in crates/pi-shell/src/git.rs either failed or stopped testing the builtin, depending on the developer's git setup: 1. Global git config reached the fixture. `fixture()` pinned only `user.*`. A global `tag.gpgSign = true` made the helper's lightweight `git tag v1` fail, so worktree_add_reproduces_git_target_forms panicked at git.rs:738 with `fatal: no tag message?`. A global `commit.gpgSign` without a usable key fails the fixture's first commit. A global `core.hooksPath` hides the fixture's `post-checkout` hook from git and from the builtin, whose gix config reads include the global file. The fixture now sets `commit.gpgSign false`, `tag.gpgSign false` and `core.hooksPath` repo-locally. The hooks path is absolute because the builtin's hook lookup joins a relative value onto the linked worktree's root. 2. Ambient `GIT_CONFIG*` variables made the builtin tests vacuous. The one-shot shell exports the test process's environment (shell.rs:827), and the builtin hands a command to the git binary while any `GIT_CONFIG*` variable is exported (git.rs:93-95, 586-593). Under, for example, a `GIT_CONFIG_COUNT` include, every target form ran through git and the tests passed without the builtin running. `run_with` now unsets those variables inside the shell before the command. Setting the process environment instead would race the other tests' threads. 3. Global worktree and branch settings made the builtin tests vacuous too. The builtin hands a form to git when `worktree.useRelativePaths` (git.rs:301), `worktree.guessRemote` (git.rs:370) or `branch.autoSetupMerge=always|inherit` (git.rs:598) asks for behavior pi-vcs does not reproduce, and git's own output then satisfies the assertions. The fixture pins all three to git's defaults locally. 4. The test helper `git()` and the builtin's gix reads config from different sources, so isolating the helper's config could produce a false failure where the builtin behaves correctly. `GIT_CONFIG_GLOBAL`, `GIT_CONFIG_SYSTEM` and `GIT_CONFIG_NOSYSTEM` redirect the plain `git` binary the helper spawns, but gix's `git_prefix: Permission::Deny` blocks all three (pi-vcs open.rs:154-162; gix-0.85.0 config/cache/init.rs:259-280), so gix keeps reading the real global and system files underneath. `GIT_CONFIG_PARAMETERS` reaches the helper (it inherits the test process's environment directly) but gix never reads it at all. `GIT_CONFIG_COUNT`/`KEY_n`/`VALUE_n` are the one source gix does read from the process environment (gix-config from_env.rs:39-62), so `git()` must keep seeing those. `git()` now removes `GIT_CONFIG_GLOBAL`/`SYSTEM`/`NOSYSTEM`/`PARAMETERS` from its own child's environment, leaving `GIT_CONFIG_COUNT` untouched. Proof, on a box whose global config sets tag.gpgsign and commit.gpgsign and whose environment exports GIT_CONFIG_COUNT: - Before: worktree_add_reproduces_git_target_forms fails at git.rs:738, `git ["tag", "v1"]: fatal: no tag message?`. After: git::tests passes 5/5 with the ambient environment, with GIT_CONFIG_COUNT=1 GIT_CONFIG_KEY_0=x.y GIT_CONFIG_VALUE_0=z, and with every GIT_CONFIG* unset. In all three, a logging git wrapper on PATH shows only the two forms the builtin leaves to git (`--no-checkout ../bare-wt` and `../again main`) reaching the binary. - A mutant that misspells the builtin's `Preparing worktree` line passes git::tests without change 2 under GIT_CONFIG_COUNT=1 x.y=z, with all six `worktree add` calls reaching git. With change 2 it fails both worktree tests, under that environment and with GIT_CONFIG* unset. - Under a HOME whose .gitconfig sets key-less signing and a global core.hooksPath: the unmodified fixture fails at its first commit; dropping only the local core.hooksPath fails the `hook stdout reaches the transcript` assertion; the fixture with its pins passes 5/5. - Under a HOME whose .gitconfig sets only worktree.useRelativePaths=true, the fixture without change 3 passed with all six `worktree add` calls reaching git; guessRemote=true sent `../wt` there, and autoSetupMerge=always sent `../wt` and `-B old ../old`. With change 3, each of those configs, and one combining them with key-less signing and a global core.hooksPath (also under GIT_CONFIG_COUNT=1 x.y=z), passes 5/5 with only the two unsupported forms reaching git. - A `GIT_CONFIG_GLOBAL=/dev/null` override (a common isolation idiom) with `core.abbrev=12` in the real `~/.gitconfig`, and separately `GIT_CONFIG_PARAMETERS="'core.abbrev=12'"` with no such file: before change 4 both failed `worktree_add_creates_git_compatible_worktree` and `worktree_add_reproduces_git_target_forms` at the `HEAD is now at` and `was at` assertions, because the helper's `rev-parse --short` and the builtin's abbreviation disagreed on length. After change 4 both pass 5/5, with the same two forms reaching git. - cargo nextest run -p pi-shell: 900 passed, 1 skipped. cargo clippy -p pi-shell --tests --no-deps -- -D warnings and cargo fmt -p pi-shell -- --check are clean. Omp-Session: 01a0dba6-1c3e-7537-80b0-6494b21465f1
…y permission
Bring the branch up to date with main, which now carries the opt-in subagent live preview from #3821. The branch is reduced to what that change does not cover: tool calls tracked by call id, path arguments shortened by argument key, the last completed call kept with its success or error mark, edit target paths, and the current tool's arguments forwarded through background task job progress.
…etups Save every model role assignment plus `defaultThinkingLevel` under a name and switch between them with `/presets save|switch|delete|list`, pick one with a bare `/presets`, or press `s` in the `/models` Roles view. Switching writes roles the way the model hub does (per `modelRoleStorage`, clearing roles the preset omits, replacing `--model` runtime overrides), then switches the live model and applies the preset's thinking level. The preset's default model is validated before anything is written; roles another layer still decides are reported rather than hidden, and a failed live switch is reported as a failure. Ported from the original `/profiles` PR onto the current settings registry; renamed to presets because `omp --profile` now means an isolated user base.
…wn intent The card and the native description preferred lastIntent, so a call that carried no intent was labelled with an earlier call's and hid its own path argument. They now use the per-call currentToolIntent (recent calls use their own intent), matching the pinned HUD.
…progress Detached task progress copied the current tool's arguments, key and start time but not its intent, so a background subagent's row lost the call's own label and could show an earlier one.
…layered deletes - applyModelPreset pre-validates the preset's default through a role lookup overlaying the preset's own roles, so `default: "@slow"` sees the state the write produces: a preset that fixes `slow` switches, and one that breaks it is refused with settings untouched. - deleteModelPreset reports `project` when another layer still defines the name after the global entry is removed, instead of a false `deleted`. - An `:inherit` default suffix no longer clears the session thinking level; setModel's reapply stands.
…ting - getModelPreset resolves each preset whole from the highest layer owning its name (new Settings.getOwnedModelPreset); same-name entries across layers no longer deep-merge. A global save shadowed by a higher layer says so instead of looking applied. - A preset with no default resolves its automatic fallback before writing; with no authed model it is refused with settings untouched. - Global-mode switch skips roles neither the global layer nor a runtime override owns, instead of dirtying the save queue with no-op deletes. - Default-role mutations share one tail with the model hub (acquireModelRoleMutation), so /presets switch serializes with hub assign/unassign.
…AMA_CPP_BASE_URL The runtime discovery suite cleared the Ollama variables but not LLAMA_CPP_BASE_URL, so a value exported in the shell or loaded from ~/.omp/agent/.env redirected the probes and failed 11 tests. Clear and restore it per test like the others.
…switching role setups (@abilliontokens) feat(coding-agent): add model presets for saving and switching role setups
…s from ambient LLAMA_CPP_BASE_URL (@andrebrait) test(coding-agent): isolate llama.cpp discovery tests from ambient LLAMA_CPP_BASE_URL
…eloper's git config (@sjawhar) test(shell): isolate git worktree tests from the developer's git config
…enai dialect (@alphastorm) docs(models): clarify thinking off on the generic openai dialect
…efore turnEnded (@eggpeat) fix(coding-agent): continued Cursor turns that end before turnEnded
…most macOS app focused (@will-bogusz) fix(natives): stop marking every window of the frontmost macOS app focused
…sent literally (@alphastorm) docs(models): warn that an unset apiKey variable is sent literally
…characters `mask_non_prose` replaces every hidden byte with a space, so a masked multi-byte character fans out into several spaces and the masked text has more chars than the original. Pairing the original's `char_indices()` with the masked text's `chars()` by index then misclassifies every character after the first masked multi-byte span: a CJK prompt that quotes code in backticks or fences loses most of its prose score, so `is_learnable` refuses to learn it. Compare each character over its own byte range in the masked text instead.
….homedir spy Bun fixes os.homedir() at process start, so setting HOME in beforeAll left plugin and user skill discovery reading the developer's real home and leaked 110 installed skills into 15 tests. Spy on os.homedir for the file, re-armed before each test because inner tests restore their own homedir spies.
The file-wide os.homedir spy pins the home to the isolated temp dir, so the WSL test can no longer pass on Windows by accident via USERPROFILE. The /mnt/<drive> mapping it exercises cannot resolve on a Windows host.
…y typed With the argument completion popup open, Enter on e.g. `/mcp list` re-accepted the identical `list ` completion and returned without submitting, so the first Enter appeared to do nothing. Partial arguments (`/mcp li`) still complete without submitting.
…the lifecycle reviver (@iliaal) fix(task): release parked subagent sessions held by the lifecycle reviver
…al home via an os.homedir spy (@andrebrait) test(coding-agent): isolate skills tests from the real home via an os.homedir spy
…asked multi-byte characters (@jchanghong023) fix(pi-predict): keep prose scoring aligned across masked multi-byte characters
…p a line around one survivor
…ment Per review: fully typed /mcp test (usage <name>) now inserts 'test ' so the user continues to the server name, instead of submitting a bare command the handler rejects. Optional-only usages ([...]) still submit.
…nnamed containers (@will-bogusz) fix(natives): stop ax() dropping everything inside unnamed containers
…t is already fully typed (@H4vC)
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 : )