can1357/oh-my-pi v16.3.2

on GitHub

@oh-my-pi/pi-ai

Changed

• Removed automated injection of reasoning suppression prompts in OpenAI responses

@oh-my-pi/pi-catalog

Fixed

• Fixed ZenMux model discovery to run without a ZENMUX_API_KEY, so newly published ZenMux models (for example anthropic/claude-fable-5-free) auto-update into the runtime models.db cache instead of waiting on a regenerated models.json.
• Fixed ZenMux runtime discovery to query the /api/v1/models endpoint even when the resolved provider base URL points at the Anthropic-compatible route, so discovery no longer requests a non-existent /api/anthropic/models path.

Removed

• Removed reasoning suppression prompt logic for GPT-5 models

@oh-my-pi/pi-coding-agent

Breaking Changes

• Changed search tool paths parameter to a single semicolon-delimited path string parameter
• Changed the grep, glob, and ast_grep tools to take a single optional path argument instead of a paths array. path accepts one path or a semicolon-delimited list (src; tests); omitting it searches the workspace root (.). Multi-path search, delimited expansion, and internal-URL scopes are unchanged. (ast_edit continues to take paths.)

Added

• Added speech.enhanced setting to rewrite assistant output into natural spoken prose
• Added speech.enhanced setting: assistant output is rewritten into natural spoken prose by the tiny/smol model before synthesis — code blocks become one-clause descriptions, links speak their label or site name, numbers and symbols read naturally, lists become flowing sentences. Blocks are rewritten fence-aware and coalesced (bounded to two concurrent completions); any failed or timed-out rewrite falls back to the mechanical cleanup so speech never blocks on the model.

Changed

• Reduced extension startup cost, especially on Windows, by reading each extension source-graph module from disk once per load instead of twice (the graph scan now feeds the load-time rewrite hook) (#4196).
• Redesigned speech vocalization for low latency and clean spoken content. Assistant markdown now runs through a speakable-text pipeline before synthesis: code blocks and tables are silent, links speak their label, bare URLs speak their host, inline-code ticks/emphasis/heading/bullet markers are stripped, and long file paths collapse to their basename. Segmentation is now parent-side and emits at sentence boundaries immediately (the previous engine-side splitter held each sentence until the next one arrived), with clause-level cuts for long sentences and an idle flush when generation stalls mid-sentence. macOS gains a gapless streaming playback backend (ffmpeg AudioToolbox, sox fallback) instead of spawning afplay per sentence.

Fixed

• Fixed ALL-CAPS acronyms (e.g. CNPG, ETL, JWT) being lowered to title case in auto-generated session titles. reconcileTitleCasing (packages/coding-agent/src/tiny/text.ts) now maps ALL-CAPS source tokens into an acronyms table and restores them when the model produces a title-cased artifact (Cnpg), while still declining restoration on shouty input (FIX the BUG NOW, ALL ERROR HANDLING) via a consecutive-ALL-CAPS heuristic. Title prompts also instruct the model to preserve ALL-CAPS acronyms verbatim. (#4220)
• Fixed cold-start --model resolution for extension providers whose catalogs come only from fetchDynamicModels, so fresh cached runtime models are available before session startup falls back or hard-fails. (#4216)
• Fixed plugin and legacy extension discovery repeatedly re-reading plugin manifests and walking extension node_modules by caching results until plugin cache invalidation. (#4197)
• Fixed discoverExtensionPaths invoking every registered extension-module provider (claude, codex, gemini, opencode) on startup and discarding all non-native results. The extension-module capability is now loaded with providers: ["native"], skipping four foreign directory walks per session — noticeable on Windows where the walks are slowest (#4198).
• Fixed /move overlay running an fs.statSync per directory entry per keystroke; the directory listing cache now stores Dirent[] and classifies entries without a syscall, falling back to statSync only for symlink entries (#4199).
• Fixed default model switches being persisted without changing the active goal-mode session when the current context exceeded the target model window. (#4219)
• Fixed live tool preview spinners staying pinned to their first frame for eval and shell-style renderers. (#4170)
• Fixed isolated task merges failing when the parent working tree carried WIP for a file the isolated subagent also touched. commitPatchToBranchWorktree now tries plain apply and git apply --3way first (agent-only outcome when the WIP-side blob is tracked in HEAD), then falls back to seeding the temp worktree with the baseline WIP so the delta patch's HEAD+WIP context matches, and rewinds WIP-only files afterward so they don't leak into the branch commit. Covers untracked WIP files, staged-new WIP files, and overlaps --3way cannot resolve. (#4136)
• Fixed discoverAgents() skipping agents/ subdirectories inside OMP extension packages, so agents shipped by omp plugin install-ed npm plugins (e.g. loom) and --extension/extensions: settings roots now load the same way their sibling skills/, hooks/, tools/ directories already do. The new scan goes through listOmpExtensionRoots, so Claude marketplace installs continue to flow through the claude-plugins provider without being double-counted. (#3920)
• Fixed plan mode hanging without converging on ask/resolve after advisor cards, idle IRC messages, or follow-on turns. Plan-mode decision enforcement ran on only the non-synthetic prompt() return; continuation/wake paths settled via agent_end and bypassed it. Advisor cards and idle IRC are now recorded into context without waking an autonomous turn, and the ask/resolve decision is enforced at the universal agent_end terminal settle via a bounded-retry counter (provider-neutral required, both tools kept available) that reminds-then-forces a fixed number of times and then yields to the user — never looping, never silently ending plan mode un-converged. An irc send await:true to an idle plan-mode session now answers the sender through the existing ephemeral side-channel auto-reply instead of stranding it until its wait timeout, and a queued forced plan decision is dropped when its continuation is skipped or plan mode exits. (#3910)
• Reduced subagent streaming CPU cost: the recent-output window no longer re-splits the full (up to 8 KB) tail on every streamed text token. Fragments without a newline extend the current last line in place, and a full recompute runs only when line boundaries actually change.
• Reduced task-render CPU cost: the task result frame (repainted ~30×/sec via the spinner) previously did 7+ full passes over the result set (some/filter/reduce); a single pass now derives the status booleans, footer counts, and request total, and incremental review extraction reuses the yield data the caller already normalized instead of re-normalizing it.
• Reduced model-resolution cost: resolveModelRoleValue now builds the preference context (an O(n) model-order map over all available models) once and reuses it across every fallback pattern instead of rebuilding it per pattern, and matchModel hoists the case-folded pattern once instead of .toLowerCase()-ing it for every candidate across each filter pass.
• Reduced read-tool allocation: line counting counts newlines directly instead of allocating via split("\n"), and the hashline formatter no longer counts the same content twice.
• Fixed the assistant-message streaming fast path dropping the transient flag, which disabled the transient render path (code-highlight skip and streaming prefix caches) on every same-shape streaming tick. In-flight renders now correctly skip per-tick syntax highlighting; highlighting applies once at message finalization.
• Fixed hidden goal-mode todo context: phase names and task text are now sanitized before prompt injection (no raw newlines or control characters forging extra context lines), and the block is only rendered with tool-accurate guidance when the todo tool is active or discoverable instead of unconditionally instructing the agent to call an unavailable tool.
• Fixed custom tool loading treating process.exit() from a tool module's import or factory as a host process exit instead of a recoverable load failure. Custom tools now load under the shared extension exit guard, so an exiting tool is skipped with a load error while remaining tools still load (#1704).
• Fixed stuttering/latency in speech by running synthesis chunks through the player gaplessly
• Fixed race condition causing EPIPE errors and broken pipes during speech playback
• Fixed interrupted speech audio by ensuring segments queue and drain in order
• Fixed speech vocalization starting only after the entire reply was synthesized: ONNX inference blocks the TTS worker's event loop, so per-segment IPC audio chunks queued unflushed and arrived in one burst. Streaming sends now drain the IPC channel before the next segment's inference, cutting time-to-first-audio to ~1.5s regardless of reply length.
• Fixed an unhandled EPIPE: broken pipe, write rejection at the end of speech playback: the streaming player's stop() raced an un-awaited FileSink.end() against the backend SIGKILL, and mid-session writes never awaited the flush. Writes now await the flush (so a dead backend is detected and the chunk replays on the next candidate or the per-file path) and stop() swallows the expected teardown rejection.

@oh-my-pi/collab-web

Changed

• Updated the glob, grep, and ast_grep tool cards to read the new single path argument, falling back to the legacy paths array so historical transcripts still render their search scope.

@oh-my-pi/omp-stats

Added

• Added a Tools tab to the omp stats dashboard (/#/tools): per-tool call counts, error rates, result/argument payload sizes, per-model breakdown, and a stacked calls-over-time chart. Token and cost columns attribute each invoking turn's real provider usage evenly across that turn's tool calls. Existing databases re-parse sessions once on the next sync to backfill historical tool calls.

@oh-my-pi/pi-utils

Fixed

• Fixed parseJsonWithRepair failing tool calls whose streamed arguments contain an unquoted string value (e.g. {"paths": packages/foo/*, "i": "…"}). Final parsing now recovers such barewords in object/array value position as strings, terminating at , / } / ] / newline. Recovery deliberately refuses anything that could mask real structure or bad data — truncated values, tokens containing " / { / [ or a key-like : (URL :// and Windows :\ colons stay literal), and non-finite atoms (NaN, Infinity, undefined) — and streaming partial parses still roll back unfinished barewords instead of committing them.

What's Changed

• Fix todo HUD and goal context follow-ups by @jeffscottward in #3777
• perf: streaming-reveal/render throughput + core hot-path optimizations by @oldschoola in #3843
• fix(session): converge plan mode on ask/resolve across continuation paths by @metaphorics in #3911
• fix(task): scan OMP extension agents/ dirs in discoverAgents by @roboomp in #3922
• fix(coding-agent): stopped isolated task merges failing when working tree carries WIP for files the agent also modifies by @roboomp in #4140
• fix(tui): animate live tool spinners by @roboomp in #4172
• fix(robomp): run sandbox setup/teardown off the event loop safely by @metaphorics in #4184
• fix(coding-agent): scope discoverExtensionPaths to native extension-module provider by @roboomp in #4202
• fix(model-discovery): auto-update ZenMux models into models.db without a key by @metaphorics in #4204
• fix(coding-agent): cache plugin extension resolution by @roboomp in #4209
• fix(providers): hydrate runtime model cache before selection by @roboomp in #4217
• fix(session): keep model switches active after rate limits by @roboomp in #4221
• fix(coding-agent): guard custom tool process exits during load by @roboomp in #1706
• fix(tui): stop /move overlay from statting every entry per keystroke by @roboomp in #4200

Full Changelog: v16.3.1...v16.3.2