8000 [pull] main from can1357:main by pull[bot] · Pull Request #125 · ericrihm/oh-my-pi · GitHub
[go: up one dir, main page]

Skip to content

[pull] main from can1357:main - #125

Merged
pull[bot] merged 62 commits into
ericrihm:mainfrom
can1357:main
Sep 27, 2026
Merged

pull[bot] merged 62 commits into
ericrihm:mainfrom
can1357:main

Conversation

@pull
@pull pull Bot commented Sep 27, 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 : )

andrebrait and others added 30 commits September 26, 2026 02:33
Upstream's new `omp skill` command owns the `skills` alias, so the listing
lives as its `list` action instead of a second top-level command. Reports
the skills a session in the given directory resolves, with `--json` adding
discovery warnings; internal metadata (`_source`, `containRoot`) stays out
of the result shape. omp-web will consume this instead of replicating
discovery rules it cannot keep in sync.
Plain `omp skill list` wrote discovery warnings and the empty-listing
notice to stdout alongside the name<TAB>description rows, so piping the
output through `cut -f1` or similar picked up non-skill lines. Rows stay
on stdout; notices and warnings now go to stderr. JSON output is unchanged.
`handleSkillList` is async but validated its target with
`fs.existsSync`/`fs.statSync`, blocking the event loop. It now awaits
`fs.promises.stat`, maps ENOENT/ENOTDIR to the same "Not a directory"
usage error as before, and rethrows any other filesystem error.
`omp skill list` honors `--json`, but the flag description still named
only search, info, and token, so `omp skill --help` hid the list mode.
The `omp skill list` entry had landed under a new `### Added` heading in
the released 18.2.3 section. Released sections are immutable, so the
entry now sits under Unreleased and 18.2.3 matches upstream again.
Skill discovery in skill-list.test.ts walked the real `os.homedir()` and
`getAgentDir()` (native user skills, managed auto-learn skills, and
Settings.init()'s config.yml/settings.json lookup), so a developer's or
CI runner's real agent config leaked into every assertion.

Isolate both seams per test via the established autolearn-managed-skills.test.ts
pattern (spyOn(os, "homedir") + setAgentDir into a fresh temp dir, restored in
afterEach), and tighten the customDirectories assertion from arrayContaining to
an exact list now that no ambient skill can leak in.
Matched the legacy desktop adapter types to its existing takeover behavior so the branch type gate can run.
Fenced automatic redemption by account across OMP processes and reused a recent successful reset for concurrent blocked turns.
Added a regression with separate auth stores and a final-credit stale-429 retry.

Fixes #13430
Keyed the cross-process reset lock and marker on the ChatGPT account identity instead of the local credential row id, so credential stores that reuse ids cannot adopt each other's reset.
Added a regression with two stores sharing credential id 1 for different accounts.

Fixes #13430
…luna

fix(web-search): prioritize GPT-6 Luna over GPT-5.6 Luna
- CollabHost reports a fatal relay close (send backlog, non-retryable close) through a new onEnded option.
- CollabController re-applies collab.autoStart for the current session, hosting the next generation once the ended room withdraws.
- Documented the replacement behavior in docs/collab.md.

Fixes #13445
…lity robustness

- Refined background input policy to avoid side-effect focus changes; clarified takeover as an explicit escalation.
- Enhanced AT-SPI hit-testing logic to prevent ambiguous window ownership in Linux.
- Aligned macOS input restoration to safely preserve user-selected focus when takeover is invoked.
- Standardized computer-use documentation and prompts on background-first input semantics.
- Normalized computer tool coordinate documentation across Windows (physical) and macOS (logical) platforms.
- Added failure-resilience in Win32 foreground input to verify ancestor windows and restore state accurately.
- CDP console text join no longer throws on JSON args like { toString: "x" }, which rejected the pending chain and broke every later tab.console()/tab.errors() read.
- cmux page-side console hook no longer throws from the page's own console.log for the same objects.
- Both fall back to JSON, then Object.prototype.toString.

Fixes #13319
- Object.prototype.toString throws on a revoked Proxy, so the last fallback in the cmux page hook and its host-side mirror now returns [unserializable] instead of throwing from the page's console call.
- Extended the cmux hook regression test with a revoked-proxy log.

Fixes #13319
… progress HUD

- Introduced `DownloadActivity` registry and HUD to surface asynchronous model downloads, tool installs, and side-runtime provisioning.
- Added support for SmolLM2-135M word-completion weights and associated background fetch scheduling.
- Unified download progress tracking across local models, STT, and TTS workers, removing redundant TUI overlays.
- Updated `omp tiny-models download` CLI to include word-completion models.
…vices and refine scope

- Updated `--all` flag description to indicate inclusion of exited global services.
- Modified help hint to mention exited global services when `--all` is not used.
- Reworked `collectReports` to always show the current project and live global services, hiding exited globals unless `--all` or `--global` is specified.
- Implemented early return for explicit `--dir` or `--global` targets to return only that scope's report.
- Adjusted TUI `PsTopComponent` scope label to display "(target)" for explicit targets and "(current + global)" for the default view.
…sing

- Implemented live steering mechanism to allow providers to accept and stream steering messages into ongoing responses.
- Updated agent loop to perform immediate soft-interrupts when live steering is accepted by a tool-executing response.
- Added logic to requeue unrecorded live-steered messages upon abort, ensuring inputs remain pending if not yet persisted in the transcript.
- Centralized steering queue state classification to share source-attribution logic between queue peeking and live delivery.
- Refactored input controller to trigger stream aborts based on interruptible input status including both queued and live-steered messages.
- Replaced arrayValuedLabels with yieldSectionShapes so assembly knows which declared sections are scalars.
- Repeated incremental yields of a schema-declared scalar section now replace the earlier value instead of assembling an array the schema rejects.
- Undeclared labels keep accumulating into a list; array-declared labels are unchanged.

Fixes #13448
- Relaunch after a self-ended room is immediate once, then waits 1 s, 2 s, 4 s, ... up to 60 s while rooms keep ending within five minutes of each other.
- /collab, session switches, /collab stop, and shutdown are not held behind the delayed relaunch.
- Documented the backoff in docs/collab.md.

Fixes #13445
… message take

- Added invocation of onLiveSteeringTaken after retrieving steering messages.
- Renamed AgentLoopConfig callback property from onLiveSteeringAccepted to onLiveSteeringTaken.
- Updated internal adoptLiveSteering method to accept "taken" messages instead of "accepted".
- Removed onAccepted hook from LiveSteeringChannel, simplifying live-steering flow.
- Adjusted live-steering tests to resolve when a steer is claimed rather than accepted.
- Updated documentation comments to reflect live steering taken semantics.
- Collected property shapes from allOf/oneOf/anyOf branches so JTD discriminator and composed caller schemas also keep the latest scalar section.
- Left labels declared array in one branch and scalar in another unshaped.

Fixes #13448
- Ignored relay close notifications while a host drains its explicit goodbye.

- Covered a terminal close racing the goodbye flush so auto-start stays off after stop.

Fixes #13445
- Kept override provenance only for surviving scalar sections while retaining override flags on accumulating sections.
- Covered strict-mode finalization after an overridden scalar is corrected following a terminal yield.
Fixes #13448
…gents

- Replaced formatted key hint import with static description for `/play` command, showing "space pauses, q quits".
- Changed ReadToolGroupComponent to display "2 more lines" summary instead of the previous Ctrl+O hint.
- Granted the `wait` tool to subagents when explicitly requested, allowing nested agents to pause execution.
…uto method

- Redefined `auto` completion mode to use Apple's native dictionary on macOS and N-gram on other platforms, deprecating automatic SmolLM weight downloads for this method.
- Implemented cold-start ingestion of foreign prompt histories (Claude Code and Codex) for new predictive engines to bootstrap user vocabulary.
- Removed typed-past exclusion logic from the N-gram engine, ensuring ghost text remains stable while users type through existing suggestions.
- Restricted SmolLM prefetching to explicit engine requests, preventing background resource usage for users not utilizing the SmolLM model.
Brit and others added 27 commits September 27, 2026 01:34
…k freshness check

- win32-arm64 addon via the Bazel msvc toolchain (aarch64 xwin splat,
  clang shim for ring, opus NEON defines) in native_addons_cross; binary
  cross-compiled on linux; release_smoke_win32_arm64 runs it on windows-11-arm.
  Replaces the ~13-20 min cargo leg.
- darwin release legs no longer save the bazel archive (bazel-cache-warm
  saves the same key for every release push); darwin-arm64 moves to macos-15
  with cross-scope restore fallback; bun store saves moved after the build.
- bazel_lock job fails when MODULE.bazel.lock is stale and uploads a refresh.
- OMP_TEST_TIMEOUT=120 for the slower kata runners; longest test bucket first.
- libei: two unit-returning statements from 09dee77 tripped
  semicolon_if_nothing_returned (Validate Rust / clippy-strict).
- rustfmt pi-predict (70cca81) and pi-natives spelling.rs (50c78ea),
  which fail the bazel rustfmt aspect on main.
PR runs install main's (or the npm release's) addons, which predate the stamp
slot until this lands on main:
- bazel-natives --source installs them unstamped instead of aborting; workspace
  loads skip version validation.
- embed-native accepts an exact legacy __piNativesV* export for the version,
  which the loader already accepts for that release; a longer sentinel
  (18_1_10 for 18.1.1) and an empty stamp slot are still rejected.
- ci-release-publish: a package publishes only after the workspace packages
  it depends on published; a failed prerequisite discards its dependents
  (reported failed) while unrelated packages still publish. Any prepare
  failure publishes nothing. Waiting on prerequisites holds no concurrency
  slot, so the declared order need not be topological; cycles are rejected.
- release.ts: the CI-green preflight only trusts main push/dispatch runs; a
  pull_request or branch run for the same SHA (PR CI skips Rust validation
  and native builds) no longer vouches for a main commit.
- natives: replace the lib.rs source-text test with an observable contract:
  the host load probe requires the addon to report the stamped version via
  __piNativesBuildVersion(), exercised on a real stamped copy of the addon.
- Replaced `theme.bg` with `theme.bgFill` in ToolExecutionComponent to preserve background color across ANSI resets.
- Ensured the pending-tool card background is reapplied after the truncate-to-width reset before the ellipsis.
- Added a test in the coding-agent suite that checks the background tint persists on the truncation ellipsis.
- The fix applies to both partial execution and benign skip states, covering all pending tool render paths.
- No functional changes beyond rendering; tool execution semantics remain unchanged.
…ting

- Simplified N-gram engine gating by removing legacy single-letter/two-letter threshold logic in favor of a constant 0.45 confidence floor.
- Deprecated macOS-specific `auto` behavior, now defaulting to the cross-platform N-gram engine on all operating systems.
- Removed `show_threshold_k1` and `show_threshold_k2_after_k1` from engine configuration, unifying completion confidence evaluation.
hostTargetName threw 'No pi_natives addon target for host win32-arm64' now
that win32-arm64 is a real addon target (review on #13450).
The runbook still described win32-arm64 as a Cargo/N-API build on
windows-11-arm and counted seven crate triples (review on #13450).
ci: release-time uplifts and green main (no native rebuild on bump, win32-arm64 cross-build, fresh bazel lock)
fix(coding-agent): scope focused /btw history and keep it rendering after leaving the view
- Pre-3.8 Qwen templates only honour the enable_thinking toggle, so the advertised low/medium/high/max ladder produced one identical request body; the ladder is now a single high rung beside off.
- Pinned the pre-3.8 ladder and its clamping in model-thinking tests.

Fixes #13454
…le skill listings (@andrebrait)

feat(cli): add omp skills command for machine-readable skill listings
- rg --crlf overwrote the matcher's CRLF terminator with LF for non-multiline searches, so grep-searcher rejected every search with a mismatched line terminator config error.

- Added a regression test asserting $ anchors before \r under --crlf.

Fixes #13457
…ering

- Enabled providers to report interim autocomplete suggestions via `onPartial` callback for improved responsiveness during high-latency file discovery.
- Introduced client-side filtering for `@` file completion lists, allowing the TUI to narrow stale lists in-place while awaiting asynchronous search results.
- Configured the TUI to show a 'Searching...' placeholder when `@` completions are being refreshed if no current matches exist.
- Fixed submission behavior to respect active autocomplete state, ensuring Enter correctly commits current input if the list is still pending.
Reparented completed sibling calls and their results onto the rewind branch, keeping provider history valid after the checkpoint cut.

Added a mock-provider regression for a batched rewind and task result.

Fixes #13458
- Integrated cross-engine prediction blending for SmolLM and N-gram providers to improve ghost text quality.
- Migrated SmolLM weights to GGUF (Q8_0) format for standardized loading and reduced storage requirements.
- Reduced N-gram show threshold from 0.45 to 0.15, prioritizing prediction recall.
- Implemented GGUF weights loader in `pi-predict` to support llama.cpp-compatible quantization while maintaining parity with bf16 research performance.
- Optimized prediction daemon to concurrently serve and blend SmolLM and N-gram suggestions.
- Updated SmolLM weight prefetching logic to manage obsolete `safetensors` cleanup and verification of pinned GGUF digests.
@pull pull Bot locked and limited conversation to collaborators Sep 27, 2026
@pull pull Bot added the ⤵️ pull label Sep 27, 2026
@pull
pull Bot merged commit 7f5e841 into ericrihm:main Sep 27, 2026
6 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.

5 participants

0