8000 [pull] main from can1357:main by pull[bot] · Pull Request #131 · ericrihm/oh-my-pi · GitHub
[go: up one dir, main page]

Skip to content

[pull] main from can1357:main - #131

Merged
pull[bot] merged 97 commits into
ericrihm:mainfrom
can1357:main
Sep 28, 2026
Merged

pull[bot] merged 97 commits into
ericrihm:mainfrom
can1357:main

Conversation

@pull
@pull pull Bot commented Sep 28, 2026 •
Copy link
Copy Markdown

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 : )

jchanghong and others added 30 commits September 26, 2026 12:46
…ity 429

Google's Cloud Code Assist backend (daily-cloudcode-pa.googleapis.com)
deterministically returns 429 RESOURCE_EXHAUSTED when systemInstruction
contains the exact case-insensitive string:

  RFC 2119: MUST, REQUIRED, SHOULD, RECOMMENDED, MAY, OPTIONAL.

which is the opening line of all three bundled prompt templates. Any
minimal rewording clears it (verified against the live endpoint, 3x each).
Reword to 'RFC 2119 keywords: ...' - semantics unchanged.
Used the shared frontmatter parser for invocation bodies so CRLF skills render without YAML metadata.

Covered user-invoked and autoload messages with a CRLF fixture.

Fixes #13590
Kept the writing handle open through sync_all and closed it before the atomic rename, avoiding Windows access-denied errors.

Fixes #13589
Disabled frontmatter repair when extracting skill prompt bodies so body HTML comments survive while CRLF normalization still applies.

Fixes #13590
- Added `-c` flag to the `cp` builtin, mapping to existing `--reflink=auto` functionality.
- Enabled copy-on-write overwriting on macOS by swapping the destination file with a clone, preserving permissions and hard-link behavior.
- Fixed `block-clone` isolation on Windows to correctly handle files exceeding 4 GiB, sparse files, and non-cluster-aligned sizes.
- Synchronized destination integrity-stream and sparse settings during Windows block-cloning to match source configuration.
8000
- Added `pi_iso::cow` to abstract native file-cloning primitives (Linux `FICLONE`, macOS `clonefile`, Windows `FSCTL_DUPLICATE_EXTENTS_TO_FILE`).
- Refactored `cp` builtin to use the shared `cow` module, improving robustness and simplifying platform-specific logic.
- Enabled native cloning on Windows (ReFS/Dev Drive), allowing `cp --reflink` to succeed where previously it failed.
- Moved macOS cloning logic from `cp` into the unified `cow` module to support reliable in-place overwriting.
Translated grouped Grep output and metadata into Cursor wire results.

Preserved raw reads, negative offsets, full edit materialization, shell codes, and deleted file sizes.

Covered native bridge behavior with regression tests.

Fixes #13600
Capped the edit-owned whole-file read at the read tool snapshot limit and returned an actionable read error beyond it.

Sized the read buffer from the open handle so a growing file cannot exhaust memory.

Fixes #13600
composerChips() tested chip presence with text.includes(label); the
compact label "<icon> #1" is a substring of "<icon> #10", so keeping
attachment #10 also surfaced chip #1 in the attachment band.

- Added referencedAttachments(), which parses the buffer once with
  COMPOSER_TOKEN_REGEX and collects whole #<n> indices per kind.
- composerChips() now keys image, video, and paste chips on that set.

Supersedes #11029, which targeted the pre-move coding-agent path.

Fixes #13605
…handles (@roboomp)

fix(predict): synced ngram snapshots using writable handles
…Tab and cursor keys on an empty narrowed list (@jchanghong023)

fix(tui): stop the @ completion popup from trapping Tab and cursor keys on an empty narrowed list
…ceiling (@jchanghong023)

fix(ai): clamp thinking budgets to a model's output ceiling
… to flush (@jchanghong023)

fix(pi-shell): give cancelled Windows pipelines room to flush
…released

Tests for EOF frontmatter and no-frontmatter bodies adapted from #13601 by @sonnivasquez.
referencedAttachments() scanned with COMPOSER_TOKEN_REGEX, which only
knows the built-in preset glyphs, so a theme overriding chip.* symbols
hid every chip whose token used the override.

- Scanned with a regex built from preset glyphs plus the active theme's
  chip glyphs (activeChipIcon, shared with chipLabel).
- Extracted glyphSource() for the word-glyph lookbehind used by skill
  chips, attachment chips, and the scanner.

Fixes #13605
…ompts (@roboomp)

fix(skills): stripped CRLF frontmatter from skill prompts
A theme switch changed the active glyph used by referencedAttachments,
while draft atoms and pending paste labels still held the prior glyph.
A subsequent draft edit hid those cards even though their tokens remained.

- Included recorded paste labels and image/video atom labels in the
  scanner before matching current and preset glyphs.
- Matched registered labels as complete numbers to avoid prefix collisions.
- Covered theme switches after image and ten text attachments.

Fixes #13605
can1357 and others added 27 commits September 28, 2026 12:57
- Removed unused `last_used_at` field from model usage tracking queries to reduce database payload size.
…mode settings

- Updated #handleInlineMouse to read the TUI mouse flag from the controller's context settings instead of the global settings object.
- Adjusted test harnesses to set cfgTuiMouse on mode.settings rather than the shared settings import.
- Removed unused Settings export from the test import list after switching to mode-specific settings.
…ing live execution events

- Added tool_execution_end events that fire as each tool call settles, enabling live UI updates.
- Changed emission of tool result messages to follow the original call order rather than completion order.
- Regenerated via bun run gen:bazel-lock; fixes the lock freshness CI check.
Moved the Darwin arm64 release and cache warmer to the Xcode 27 image so the addon compiles the Swift bridge. Kept x64 on macOS 15 and rejected arm64 addons missing the FoundationModels link.

Fixes #13610
…istic

- Marker-mode fixture first launch now never answers initialize, so the test can use a 1s client timeout instead of 100ms.
- A 100ms timeout also bounded every retry's bun cold start, which exceeded it on CI and exhausted the 5s budget via the reconnect ladder.
pi-vcs tests compare gix output against real git. A global diff.external
(difftastic) replaced git diff's patch output, and a global core.hooksPath
redirected hook lookup away from the fixture repo, so test:rs failed on any
machine with those settings. The runner now points git at empty global/system
config, and the mutate fixture pins core.hooksPath locally.
8BD0
- Hoisted repair-generated assistant notes across preceding output runs.
- Synthesized required reasoning items for each native replayed assistant turn.
- Added regressions for both captured DeepSeek request shapes.

Fixes #13083
- Dropped opaque encrypted_content (reasoning and compaction items) from the remote compaction size estimate, matching excludeEncryptedReasoning, so the #13509 fit gate no longer refuses Codex histories the server accepts.
- Added a regression test with blob-heavy native history.

Fixes #13611
- Replaced root-cause narration and diagnostic token estimate with the user-visible symptom that was fixed.

Fixes #13611
…om swallowing

- The `@` autocomplete popup no longer displays the "Searching..." placeholder while a fresh file search is pending; it remains hidden until results arrive.
- `Editor.isShowingAutocomplete()` now returns false when the popup is hidden, reflecting the hidden state.
- Pressing Escape cancels autocomplete without swallowing the key when the popup is hidden.
- Added internal helper `#visibleAutocompleteList()` to expose the open list only when it has candidates.
- Rendering now skips the empty list, preventing placeholder rows from appearing.
- Tests added to verify hidden popup behavior and Escape handling.
…git config (@ParadaCarleton)

test(vcs): isolate Rust test oracles from developer git config
…ss comment-only lines (@jchanghong023)

fix(pi-ast): keep block-range detection working across comment-only lines
Once DECRQM confirmed bracketed paste (#12811), StdinBuffer stopped
classifying unmarked multiline bursts entirely, so IME/dictation commits
(Ghostty since ghostty#13817 delivers them as typed input) were split into
one submit per line (#13344). The terminal now installs the TUI's
LoopWatchdog stall probe instead: a burst on a responsive loop lands as one
paste, while keystrokes a stall batched into one read still submit (#12540).

Also reverted the BracketedPasteHandler changes that treated every
unbracketed CR/LF chunk as a paste (Enter could never submit) and broke
uppercase Ctrl-letter decoding (cp - 96).
…op stalls (@goransh-walia)

Fix: TUI: multiline IME/dictation input is split into separate single-line
@pull pull Bot locked and limited conversation to collaborators Sep 28, 2026
@pull pull Bot added the ⤵️ pull label Sep 28, 2026
@pull
pull Bot merged commit cabfa74 into ericrihm:main Sep 28, 2026
7 checks passed
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

6 participants

0