8000 [pull] main from can1357:main by pull[bot] · Pull Request #111 · ericrihm/oh-my-pi · GitHub
[go: up one dir, main page]

Skip to content

[pull] main from can1357:main - #111

Merged
pull[bot] merged 27 commits into
ericrihm:mainfrom
can1357:main
Sep 23, 2026
Merged

pull[bot] merged 27 commits into
ericrihm:mainfrom
can1357:main

Conversation

@pull
@pull pull Bot commented Sep 23, 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 : )

can1357 and others added 27 commits September 22, 2026 22:25
…ayback

- Introduced `/record` slash command to capture terminal sessions into `.ompcast` (JSONL) files.
- Added `omp play` CLI to replay screen recordings in the terminal with configurable speed and idle-time limits.
- Added `● REC` status indicator to the TUI to reflect active recording state.
- Refactored stream output into a shared `StreamPaintEncoder` to support both live streaming and file-based recording.
- Enabled multiple simultaneous TUI paint observers, allowing recording and streaming to operate concurrently.
… per-kind metrics

- Integrated comparison table into the live benchmark status output, updating results in real-time as runs finish.
- Added dedicated columns for TTFT p50, p95, and specific throughput metrics based on the selected benchmark profile.
- Updated `bench` profile default from `mix` to `chat` and clarified CLI flag documentation.
- Added `prefill` throughput metric reporting for prefill-focused benchmark runs.
- Automated leaderboard ranking by sorting models based on the lead challenge kind's median performance.
- Refactored worker queue management to support incremental report updates in the UI.
…mp.sh

- Implemented `omp clip` command that validates flags, uploads a recording, and prints the clip URL.
- Extended recording header with optional description and owner fields and added server upload support.
- Updated CLI help, command registration, status messages, and documentation to reflect the new clip feature.
- Added `signal` to FileLockOptions; aborting stops the retry wait without entering the critical section or releasing the current owner.
- Turned find.enabled into an auto/on/off enum; auto enables find only when the judge role resolves first to a native System One model.
- Migrated boolean settings to on/off and routed tool creation and sibling prompt hints through isFindEnabled.
- Summed every judgment attempt's cost, retries and failures included, into batch status, progress events, the job summary, and the HUD row.
- sessionJudge accepts an onUsage observer alongside the session ledger.
- Replaced the file pass plus hunk pass with one score judgment per hunk, untracked file, or binary, contrasted against the full changed-path list.
- Removed the per-file question prompt template.
- Mix-profile tests now request the mix profile explicitly; the cache test expects the chat default.
- Shadow-plan fixtures await the tool coroutine; call-site ids shift accordingly.
…bals

- Top-level declarations in async-wrapped cells become assignments on the worker global, with function declarations installed at wrapper entry, so persisted functions see later cells' updates instead of stale wrapper locals.
- A filename containing CR/LF/U+2028/U+2029 could end the sourceURL comment and inject executable source; those separators are now percent-encoded.
- Standalone `%load <path>` cells are read on the host (quoted paths, local:// URLs) and run with their filename: Python sets __file__, adds the script dir to sys.path, and reports tracebacks against the script; JS resolves relative imports from the script.
- Runner `%load` now executes in the retained namespace with top-level await and no source echo; `%pip` pauses the cell watchdog while pip runs; missing-module errors hint at distribution names.
- JS `%bun add` installs into a per-project OMP-managed package environment (lifecycle scripts disabled) used as a module-resolution fallback; `%environment project|managed` switches the target; `eval.autoProvision` gates creating the managed environment.
- A process-local eval state registry feeds a hidden eval-state-context message describing live kernels and loaded scripts into rebuilt and resumed agent context.
- Percent commands render verbatim in the eval tool preview.
- Integrated a multi-stage verification loop for `ai-stage` to filter out false positives by comparing high-scoring units against a "none of these" criterion.
- Increased the initial staging threshold to 0.6 and added boundary truncation for improved token efficiency during judgment.
- Renamed `StreamCredential` to `StencilCredential` and moved to `src/stencil/` to reflect broader usage across platform services.
- Deleted tests for cached discovery, context window restoration, and special provider model loading.
…parse cache

Type-less script dependencies were re-parsed with Babel on every startup to check for unshadowed require/exports, bypassing the persistent parse cache. Store the result alongside the cached analysis (schema v3) so warm loads never reparse.
Interactive mode now warms the native highlighter after first paint, and the warmup also compiles the regexes of commonly highlighted languages (ts, tsx, js, bash, python, rust, markdown) on the worker pool, so the first code block no longer stalls rendering.
- b5d692e pruned packages/wire/test/constants.test.ts, the package's only test; bun test exits 1 on "No tests found!", failing the Test TS workspace fast job.
- Removed wire from fastWorkspacePackages and dropped its test script, matching the other test-less packages (stats, browser-relay).
- Linux CI runners find a project artifact in the workspace after the managed install (2/2 runs since ff74eac), while macOS, Linux containers, and the full native bucket at CI concurrency all pass.
- Replaced the per-artifact access loop with one assertion over the workspace listing and the managed manifest's dependencies, so a failure shows which entries leaked and where bun installed.
…ource

rewriteLegacyExtensionSource re-analyzed the import-rewritten source to resolve require() targets. On POSIX that text embeds the per-load ?mtime= tag, so every load produced a fresh parse-cache key and reparsed with Babel. Resolve requires against the original source and apply both replacement sets in one pass.
- The test required the project dir to stay free of node_modules, which fails on hosts whose Bun global cache is unwritable (the kata runner image sets BUN_INSTALL=/usr/local): Bun's runtime auto-install then falls back to <cwd>/node_modules/.cache.
- Eval has resolved bare imports with Bun.resolveSync from the project cwd since before ff74eac, so this is long-standing eval behavior, not a regression; the test premise is host-dependent.
perf: faster extension loading and off-thread highlighter warmup
…-default

fix(ai): send per-message effort when an Anthropic session starts on the API default
@pull pull Bot locked and limited conversation to collaborators Sep 23, 2026
@pull pull Bot added the ⤵️ pull label Sep 23, 2026
@pull
pull Bot merged commit da58b16 into ericrihm:main Sep 23, 2026
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.

2 participants

0