@oh-my-pi/pi-coding-agent 17.0.1 → 17.0.3

package/CHANGELOG.md

@@ -2,11 +2,143 @@
 
 ## [Unreleased]
 
+## [17.0.3] - 2026-07-17
+
+### Changed
+
+- `omp usage` and the in-session `/usage` view now show the Anthropic organization next to the account for org-scoped credentials (with `--redact` masking applied per part in the CLI, falling back to the org id when no display name is available), attribute "no usage data" rows per organization, and match the "in use by this session" marker by organization so only the active subscription is flagged. The OAuth login success message names the account and organization that was stored — a login landing on an unintended subscription is visible immediately.
+- `/logout` labels Anthropic accounts with their organization and marks only the credential of the active organization as active; `omp token --list` shows the organization next to each account. Two subscriptions sharing one email are distinguishable when selecting which to remove or mint a token for.
+- `omp auth-broker migrate --from-local` dedupes Anthropic OAuth identities per organization, so a Team seat already on the broker no longer blocks uploading the personal plan under the same email.
+- The status line invalidates its cached usage when the session rotates to a different Anthropic organization (previously the old subscription's quota could linger for the cache TTL), and `omp auth-gateway check` labels each credential with its organization so a failing row says which subscription needs re-login.
+- `omp usage` "no usage data" attribution is org-decisive whenever either the stored account or a report carries an organization: an org-less legacy credential whose own fetch failed is no longer hidden by an org-attributed sibling report sharing the same email.
+- Active-account matching for `/usage`, `/logout`, and `omp token --list` now treats a shared organization as a qualifier rather than a match: two Anthropic Team seats in one org (same org id, per-user pools) no longer flag each other's rows or reports as "in use by this session" — the base identity (account/email/project) is still required, with org-only sessions matching on the org alone.
+- `omp usage` "no usage data" coverage now requires the member's own identity within a shared organization: a sibling Team member's same-org report no longer counts as coverage for an account whose own report is missing, while an org-only account remains covered by any same-org report.
+- `omp auth-broker migrate --from-local` reruns now recognize an already-migrated org-only Anthropic row (login recovered neither email nor account) by its organization id instead of re-uploading it, which could overwrite the broker's newer refresh token with the stale local one.
+- Updated tangential agent forks to ignore parent session history and focus exclusively on the new request
+- Hardened `/tan` fork isolation: the clone's inherited todo list is cleared at fork (parent todo reminders no longer drag the tan back onto the parent's task), the fork notice warns that the parent is concurrently editing the same working directory, and the notice is re-injected after each compaction so the fork boundary survives summarization
+- Added visual markers in the transcript for elided tool calls that have no corresponding result
+- Updated status event log to prioritize the most recent entries in the display window
+- Updated the snapcompact shape preview transcript to use the compact scope format shown to models during compaction.
+
+### Fixed
+
+- Fixed `xd://` mount notices triggering unsolicited model turns by deferring hidden notices until the next user prompt.
+- Fixed `xd://` device tools appearing in the direct tool inventory and prompting invalid function calls ([#5797](https://github.com/can1357/oh-my-pi/issues/5797)).
+- Fixed `history://` read selectors being treated as part of the agent id instead of paging the transcript ([#5806](https://github.com/can1357/oh-my-pi/issues/5806)).
+- Narrowed the `history://` contract in the system prompt to match the implementation: it serves registered agents process-wide plus persisted subagents discoverable from their artifact trees, but does not discover unregistered top-level sessions solely from persisted session files ([#5839](https://github.com/can1357/oh-my-pi/issues/5839)).
+- Fixed expanded `!` bash and `eval` output keeping a stale `… N more lines (ctrl+o to expand)` footer after Ctrl+O revealed every line ([#5842](https://github.com/can1357/oh-my-pi/issues/5842)).
+- Fixed MCP reauthentication continuing to an authorization URL without `client_id` after dynamic client registration fails; the registration error now blocks the flow with the provider response details ([#5852](https://github.com/can1357/oh-my-pi/issues/5852)).
+- Fixed collapsed todo views hiding the in-progress task in large phases. Both the transient `Todo` tool result and the sticky `Todos` HUD now share one walking-viewport policy: completed/abandoned tasks are omitted, every active task (the in-progress one, or a pending task a live subagent is executing) is pulled to the head in todo order, remaining rows fill with the following pending tasks, and an explicit `… N more active todos` summary is shown when active work alone exceeds the preview cap ([#5873](https://github.com/can1357/oh-my-pi/issues/5873)).
+- Fixed legacy provider extensions failing to load when they use the historical synchronous auth-storage surface ([#5879](https://github.com/can1357/oh-my-pi/issues/5879)).
+- Fixed orphaned detached MCP stdio server process trees surviving session dispose by escalating stdin-EOF → group SIGTERM → group SIGKILL on close() (#5578)
+- Fixed `/new` starting an unsolicited old-context provider turn when a hidden steer (e.g. an `xd://` mount notice) was queued: the session transition is now an atomic boundary, so a queued steer/follow-up can no longer auto-resume against the pre-`/new` context while the session is disconnected mid-transition. `/compact` still resumes a steer/follow-up that arrives while it runs, draining the queue once it reconnects ([#5800](https://github.com/can1357/oh-my-pi/issues/5800)).
+- Fixed signed thinking-only Claude stops being discarded and retried as empty responses ([#5881](https://github.com/can1357/oh-my-pi/issues/5881)).
+- Fixed repeated URL reads and URL-backed searches returning stale same-session responses instead of refetching the resource ([#5803](https://github.com/can1357/oh-my-pi/issues/5803)).
+- Fixed `read`/`write` not recognizing ZIP-based `.jar`/`.war`/`.ear`/`.apk` files as archives, so `read lib.jar:META-INF/MANIFEST.MF` failed with path-not-found ([#5808](https://github.com/can1357/oh-my-pi/issues/5808)).
+- Fixed legacy binary `.doc`/`.ppt`/`.xls`/`.rtf` being advertised as convertible in `read`, `fetch`, and CLI `@file` handling despite having no markit converter, which surfaced an `Unsupported format` error instead of falling through to normal file handling ([#5808](https://github.com/can1357/oh-my-pi/issues/5808)).
+- Fixed Kimi Code transport selection to follow live per-model protocol metadata by default while preserving explicit OpenAI and Anthropic overrides ([#5893](https://github.com/can1357/oh-my-pi/issues/5893)).
+- Fixed repeated edit-tool rejections from local models by recovering comma-separated ranges and malformed trailers, while clarifying canonical string input and `.=` syntax ([#5805](https://github.com/can1357/oh-my-pi/issues/5805)).
+- Fixed LSP diagnostics and edit-time diagnostics writethrough for pull-only servers that advertise `textDocument/diagnostic` statically or through dynamic registration ([#5825](https://github.com/can1357/oh-my-pi/issues/5825)).
+- Fixed local `!` command output concatenating carriage-return progress updates by preserving them as readable line boundaries ([#5845](https://github.com/can1357/oh-my-pi/issues/5845)).
+- Fixed `hub`/`irc` peer discovery after process-crash resume by restoring persisted subagents as parked peers before listing the roster ([#5864](https://github.com/can1357/oh-my-pi/issues/5864)).
+
+### Removed
+
+- Removed the unreliable Bing and Yahoo HTML-scraping web search providers
+
+## [17.0.2] - 2026-07-17
+
+### Added
+
+- Added native Warp CLI-agent events for rich session status, tool approvals, and completion notifications.
+- Added support for ChatGPT/Codex subscriptions in the `generate_image` tool, allowing image generation without a metered `OPENAI_API_KEY` even when using other active models.
+- Added an optional `provider` parameter to `generate_image` to override the global `providers.image` setting for a single request.
+- Added OpenTelemetry log and metric export support alongside existing trace exports, enabling forwarding of centralized-logger events and GenAI-semconv metrics when configured.
+- Enhanced `retry.fallbackChains` wildcards to support id-prefixed targets and keys, allowing more flexible model fallback routing across different providers.
+- Added an opt-in per-project model role storage mode with global fallback from the model selector.
+- Added per-advisor on/off toggle (`enabled: false` in `WATCHDOG.yml`): advisors stay in the roster but their runtime is never built — they show `○` in `/advisor status` rather than disappearing. Existing configs are backward-compatible (defaults to `true` when absent).
+- Colored the status line's advisor `++` badge by roster health (green all running, yellow quota-exhausted, red failed, dim paused); per-advisor glyphs (`●`/`○`/`✕`) show in `/advisor status`.
+- Added real provider quota display (usage percent, window, reset timer) to `/advisor status` and the `/advisor configure` preview.
+
+### Changed
+
+- Changed Bash command timeouts to render with a warning (yellow) border instead of an error (red) border, while still indicating to the model that the command did not complete normally.
+- Made the hashline seen-line guard opt-in and off by default via `edit.enforceSeenLines`, and improved handling of column-clipped lines so single-line edits on long lines apply without a full-width re-read.
+- Changed bundled TTSR rules to warn without interrupting generation.
+- Renamed the system prompt's project-context section wrapper from `<context>` to `<repo-rules>` to prevent collisions with the `task` tool's `context` parameter.
+- Enriched `/advisor status` to show per-advisor status glyphs, model, spend breakdown, and quota window for every configured advisor (including disabled ones), replacing the previous single-advisor-only summary.
+
+### Fixed
+
+- Fixed loading issues for linked legacy extensions importing `DefaultPackageManager` or `linkedom`.
+- Fixed the advisor retrying terminal, non-retriable provider failures (e.g., blocked prompts), ensuring they fail immediately while transient failures still retry.
+- Fixed an issue where reassigning the `plan` role model mid-planning did not take effect until the next plan-mode entry; it now applies at the next turn boundary.
+- Added managed timer helpers (`ctx.setInterval`, `ctx.setTimeout`, `ctx.clearTimer`) to the extension context to prevent self-scheduled callbacks from throwing uncaught exceptions and crashing the session.
+- Fixed `/quit` and `/exit` leaving stalled automatic title-generation requests alive during session teardown.
+- Fixed `startup.quiet` still rendering the `xdev: xd://: mounted` status line when MCP tools connect.
+- Fixed command errors in the `hub` tool when using a non-POSIX shell.
+- Fixed xdev-routed checkpoint and rewind writes not tracking checkpoint state.
+- Fixed the built-in advisor silently failing when its model routes through the Cursor provider by adding a Cursor execution bridge scoped to its granted tool set.
+- Fixed the fullscreen plan-review overlay staying visible until the approved execution turn finished; it is now hidden as soon as execution begins.
+- Fixed MCP tools repeatedly unmounting and remounting mid-session due to overlapping sanitized prefixes, and resolved stale tools remaining registered after disconnecting a server with special characters.
+- Fixed the `/usage show` marker and TUI usage panel to display and preserve the active organization suffix (`email (OrgName)`) to distinguish between multiple credentials with the same email.
+- Fixed Windows stdio MCP servers launched through `.cmd` or `.bat` shims failing with `Transport closed` by properly escaping arguments and spawning with `windowsVerbatimArguments`.
+- Fixed a startup crash on Windows when running from a drive root (e.g., `R:\`).
+- Fixed unknown `__omp_worker_*` CLI selectors exiting with code 0 instead of throwing an error.
+- Fixed Plan Review capturing mouse drags as pointer events, which prevented native terminal text selection.
+- Fixed orphaned TUI processes remaining alive after a fatal error and causing high resource consumption.
+- Fixed approved-plan execution looping through filesystem searches when a model rewrites the required plan read path.
+- Fixed Ask dialogs immediately accepting highlighted answers when they appear while the user is typing a space.
+- Stopped post-compaction auto-continue from opening another primary turn after a terminal text answer with no queued work.
+- Fixed the `write` approval gate misclassifying `xd://` device writes as `exec` when the mounted tool declared a function-valued approval.
+- Fixed custom LSP servers (such as `roslyn-language-server`) crashing when requesting unconfigured workspace configuration sections.
+- Fixed late user-initiated bash results being recorded in whichever session or branch was active when execution finished; they now retain their originating transcript.
+- Fixed Claude Code marketplace plugins with `scope: "local"` leaking skills, hooks, tools, commands, and MCP servers into unrelated projects.
+- Fixed headless `omp -p` waiting indefinitely after a completed turn when final consolidation stalls.
+- Fixed explicit-tool sessions bypassing `xd://` presentation for ambient discoverable custom and MCP tools.
+- Fixed `providers.webSearch: kimi` incorrectly sending Moonshot credentials instead of Kimi Code credentials.
+- Fixed `registerTool` demoting essential built-in tools to `discoverable` when a re-registration omitted `loadMode`.
+- Fixed the advisor skipping the next user instruction after auto-learn accepted and pruned a terminal empty assistant stop.
+- Fixed built-in advisors retrying quota- or rate-limited providers instead of applying the matching `retry.fallbackChains` model chain.
+- Updated model selector status messages to use role tags (`SMOL`, `SLOW`) instead of display names (`Fast`, `Thinking`) for consistency.
+- Fixed Cursor models receiving only top-level tools by forwarding mounted `xd://` devices through Cursor's request-context MCP catalog.
+- Fixed Windows bash crashes when a piped command times out while flushing output.
+- Fixed a race condition where hub/IRC `send` and `ensureLive` could inject into a subagent session during disposal.
+- Migrated legacy nested/dotted configuration keys (`dev.autoqa.consent` and `todo.reminders.max`) to flat keys (`dev.autoqaConsent` and `todo.remindersMax`) on settings load.
+- Fixed keyboard input dying after the first keypress when a custom tool module attaches a stdin consumer at import time.
+- Fixed the alignment of the ask tool's custom-input dialog to start at column 0.
+- Fixed advisor context maintenance undercounting the provider context by anchoring compaction decisions on provider-reported context usage.
+- Fixed RPC mode (`--mode rpc`) crashing on non-JSON stdin lines; malformed lines are now reported as errors while the frame loop continues.
+- Fixed local llama.cpp Qwen-family models not honoring the `--thinking off` flag.
+- Documented the `ultrathink`, `orchestrate`, and `workflowz` magic keywords, including their effects, matching rules, and settings.
+- Fixed `/clear` autocomplete selecting `/autoresearch` and updated `/clear` to start a new session as an alias for `/new`.
+- Fixed `/review` aborting entirely when GitHub rejects a pull request's aggregate diff for exceeding the line limit by falling back to the paginated per-file endpoint.
+- Fixed `/q` + Enter running `/queue` instead of `/quit` by adding an explicit `q` alias to `/quit`.
+- Fixed Ctrl+L (`app.display.reset`) not refreshing the dark/light theme on certain terminals by issuing a background re-query before repainting.
+- Fixed a failing advisor stalling the primary agent: the per-turn catch-up gate parked the primary for up to its full 30s budget while a broken advisor (unsupported model, dead endpoint, render bug) retried — and an advisor exception could abort the primary's turn-end outright. A failing advisor now releases parked waiters the moment its turn fails (before any async hook), refuses new parks until a turn succeeds, and the turn-end boundary isolates advisor exceptions completely; a failed render restores the delta cursor so nothing is lost when the advisor recovers.
+- Fixed advisors retrying a permanently rejected request forever (e.g. `invalid_request_error: model not supported with this account`): unlike quota exhaustion — which pauses with a notice until an explicit reset — this class notified once and silently kept re-attempting every turn, re-building heavy context in a shared daemon. The runtime now hard-stops after a permanent rejection or three consecutive backlog-drop cycles, with a visible notice; an explicit reset (`/new`, config rebuild, restart) re-enables it. `waitForCatchup` resolves immediately while halted so the primary agent is never parked on a runtime that cannot drain.
+
 ## [17.0.1] - 2026-07-16
 
-### Changed
+### Removed
 
 - Fixed a crash when a plugin/custom tool renderer returns a component that throws during its later `render()` pass (e.g. `TypeError: th.bold is not a function` from a plugin that styles its header off an object without a `bold` method). `ToolExecutionComponent` now wraps every renderer-returned call/result component so a throwing `render()` degrades to the safe fallback (tool label or raw result text) instead of taking down the transcript ([#4978](https://github.com/can1357/oh-my-pi/issues/4978)).
+- Fixed `/login` for paste-code providers (Codex, Anthropic, Gemini CLI, GitLab Duo, Antigravity, Devin) dropping the pasted fallback redirect URL: the login dialog captured focus but never mounted an input, and the "complete pairing with `/login <url>`" tip pointed at the hidden, unfocused editor. The dialog now mounts a focused input for the manual code/URL paste ([#5339](https://github.com/can1357/oh-my-pi/issues/5339)).
+- Fixed `/tan` and `/fork` clones cold-missing the provider prompt cache: the per-turn supersede/useless-result prune rewrote the live context without persisting it, so file-based forks and resume rebuilt a divergent (un-pruned) prefix and re-wrote the entire cache
+- Fixed `/tan` pinning the clone's prompt-cache key to the parent's session id instead of the parent's effective cache key, dropping shard affinity when the parent was itself a fork or tan
+- Fixed `/resume` and plan approval exposing the previous session while their asynchronous session replacement was still loading by keeping fullscreen overlays mounted until the rebuilt transcript is ready ([#5319](https://github.com/can1357/oh-my-pi/issues/5319)).
+- Fixed inconsistent history rendering when toggling the display setting for compacted items
+- Fixed configured `retry.fallbackChains` never engaging on non-retryable provider errors (e.g. "Cloud Code Assist API returned an empty response"): a hard error on a model covered by a fallback chain now switches to the next candidate instead of failing the turn, while still never backoff-retrying the failing model itself
+- Fixed transcript rebuilds (compaction, `/compact`, and toggling history display) repainting content below stale scrollback when collapsing history; rebuilds now correctly clear the scrollback buffer when history is collapsed
+- Improved auto-compaction to automatically drop images and elide content when context is tight, and added persistent warning badges to the compaction divider when manual intervention is required
+- Fixed backgrounded Bash blocks continuing to repaint with live and final job output; they now freeze with a compact job notice while completion is delivered separately
+- Fixed the downshift plan nudge silently ending the run with no code written when the model answered with a text-only reply (no tool call): the agent loop treats a tool-call-free turn as a natural stop and never prompts again, which the nudge's own "write the plan in your next reply" instruction makes common. The nudge now explicitly tells the model this is a checkpoint, not a final answer, and the session forces one more turn whenever a post-nudge reply lands with zero tool calls
+- Fixed launch tool rendering stacking a stale pending header over a bare `✓ Launch` line and raw text: the tool now uses a merged registry renderer with one per-op status header (op, target, `state · pid · uptime` meta), stripped log cursor suffixes, capped collapsed log/list previews, and a launch tool glyph
+- Fixed confusing launch start/wait results when readiness timed out with the log pattern already matched (readiness needs log AND port): the result printed a contradictory `Ready: <match>` next to `Readiness timed out` without naming the failing condition. Daemon snapshots now carry the unmet conditions (`readyPending`), and start/wait results state exactly what never happened (e.g. `port 3100 on 127.0.0.1 never accepted connections`); the TUI shows a `waiting on port` badge on starting daemons
+- Fixed the in-process `stat` builtin mangling BSD-style invocations like `stat -f "%Sm %N" file` (macOS muscle memory): GNU `-f` means `--file-system`, so the format string was treated as a file operand — printing filesystem info for the real operands and erroring with `cannot read file system information for '%Sm %N'`. A `-f` whose format value contains `%` is now detected as BSD syntax and translated to the GNU equivalent (`%Sm`→`%y`, `%N`→`%n`, `%z`→`%s`, epoch/`S`-form times, owner/group/permission and `H`/`L` sub-field directives, `-L`/`-n`/`-q`/`-F` flag clusters, with `%n`/`%t` as literal newline/tab); directives with no GNU counterpart fail with a clear `unsupported BSD format directive` error
+- Fixed the remaining GNU-flavored shell builtins that broke under macOS/BSD muscle memory, using the same unambiguous-detection approach as the `stat` fix (only invocations that are invalid or nonsensical under GNU semantics are reinterpreted; unsupported BSD forms fail loudly instead of producing wrong output): `date -r <epoch>` formats the epoch when no such file exists (GNU `-r FILE` mtime preserved), signed `date -v±N<unit>` adjustments translate to `-d` relative dates and `-j` is accepted (`-j -f` strptime parse mode and field-set `-v` error clearly); `sed -i '' 's/…/…/' file` drops the BSD empty backup-suffix token instead of treating it as the script; `mktemp -t prefix` without X's creates `$TMPDIR/prefix.XXXXXXXXXX` (the GNU `too few X's` error path); `tail -r` reverses input by delegating to `tac` (with `-n`/`-c`/`-f` combinations erroring clearly); `find -E` maps to `-regextype posix-extended` ahead of the expression; `base64 -D` decodes as an alias of `-d`; and `ln -sfh` works via a `-h` alias of `--no-dereference` (clap's `-h` help short is dropped to match real GNU/BSD ln; `--help` unchanged)
+- Fixed the browser tool crashing the whole process (parent session and every subagent) when a CDP world re-acquire failed mid-navigation: the stealth `puppeteer-core` patch called the bare `debugError` logger, which is `undefined` while the `puppeteer:error` debug channel is disabled (the default), turning a transient acquire failure into a fatal `TypeError` unhandled rejection. The patched `FrameManager`/`WebWorker` acquire paths now use `debugCatchError` ([#5296](https://github.com/can1357/oh-my-pi/issues/5296))
+- Fixed a role with a `:high` thinking suffix resolving to a longer sibling model whose id embeds the tier name (e.g. `kimi-for-coding:high` → `kimi-for-coding-highspeed`). The thinking suffix is now stripped before any fuzzy match, so `provider/model:high` keeps the exact model at high effort ([#5151](https://github.com/can1357/oh-my-pi/issues/5151)).
 
 ### Fixed
 
@@ -20,6 +152,9 @@
 - Fixed Bash internal URLs remaining unresolved when used as unquoted arguments inside command substitutions ([#5535](https://github.com/can1357/oh-my-pi/issues/5535)).
 - Fixed the built-in `fd` printing `fd: Broken pipe (os error 32)` when a downstream pipeline reader exited early (e.g. `fd … | head`); it now exits silently with 141 (128+SIGPIPE), matching real fd.
 - Fixed prewalk repeatedly continuing after a bash-only task such as `commit` had already completed ([#5551](https://github.com/can1357/oh-my-pi/issues/5551)).
+
+### Added
+
 - Fixed the Codex `config.toml` MCP importer dropping `cwd` and leaving relative `command` values unrooted, which broke the bundled Codex Computer Use server (`ENOENT` on spawn); relative `command`/`cwd` now resolve against the Codex config directory like the claude-plugins/omp-plugins providers ([#5561](https://github.com/can1357/oh-my-pi/issues/5561)).
 - Fixed streamed replace-mode edits with `ssh://` paths terminating the active prompt before normal tool dispatch ([#5552](https://github.com/can1357/oh-my-pi/issues/5552)).
 - Fixed concurrent provider OAuth refreshes from invalidating Anthropic's rotating refresh token, and prevented background usage probes from permanently disabling credentials after refresh failures ([#5396](https://github.com/can1357/oh-my-pi/issues/5396)).
@@ -39,25 +174,6 @@
 - Fixed ACP stdio EOF/EPIPE disconnects bypassing awaited session teardown and leaving in-flight tool calls pending in persisted rollouts ([#4788](https://github.com/can1357/oh-my-pi/issues/4788)).
 - Routed the print-mode assistant-error/aborted exit, RPC `pi.shutdown()` and stdin-EOF shutdowns, and the extension command-context `shutdown()` through the awaited, idempotent `session.dispose()` before `process.exit()`, so the bounded browser reaper (`releaseTabsForOwner`) always runs and OMP-owned Chromium no longer outlives the process ([#5643](https://github.com/can1357/oh-my-pi/issues/5643)).
 
-### Removed
-
-- Fixed `/login` for paste-code providers (Codex, Anthropic, Gemini CLI, GitLab Duo, Antigravity, Devin) dropping the pasted fallback redirect URL: the login dialog captured focus but never mounted an input, and the "complete pairing with `/login <url>`" tip pointed at the hidden, unfocused editor. The dialog now mounts a focused input for the manual code/URL paste ([#5339](https://github.com/can1357/oh-my-pi/issues/5339)).
-- Fixed `/tan` and `/fork` clones cold-missing the provider prompt cache: the per-turn supersede/useless-result prune rewrote the live context without persisting it, so file-based forks and resume rebuilt a divergent (un-pruned) prefix and re-wrote the entire cache
-- Fixed `/tan` pinning the clone's prompt-cache key to the parent's session id instead of the parent's effective cache key, dropping shard affinity when the parent was itself a fork or tan
-- Fixed `/resume` and plan approval exposing the previous session while their asynchronous session replacement was still loading by keeping fullscreen overlays mounted until the rebuilt transcript is ready ([#5319](https://github.com/can1357/oh-my-pi/issues/5319)).
-- Fixed inconsistent history rendering when toggling the display setting for compacted items
-- Fixed configured `retry.fallbackChains` never engaging on non-retryable provider errors (e.g. "Cloud Code Assist API returned an empty response"): a hard error on a model covered by a fallback chain now switches to the next candidate instead of failing the turn, while still never backoff-retrying the failing model itself
-- Fixed transcript rebuilds (compaction, `/compact`, and toggling history display) repainting content below stale scrollback when collapsing history; rebuilds now correctly clear the scrollback buffer when history is collapsed
-- Improved auto-compaction to automatically drop images and elide content when context is tight, and added persistent warning badges to the compaction divider when manual intervention is required
-- Fixed backgrounded Bash blocks continuing to repaint with live and final job output; they now freeze with a compact job notice while completion is delivered separately
-- Fixed the downshift plan nudge silently ending the run with no code written when the model answered with a text-only reply (no tool call): the agent loop treats a tool-call-free turn as a natural stop and never prompts again, which the nudge's own "write the plan in your next reply" instruction makes common. The nudge now explicitly tells the model this is a checkpoint, not a final answer, and the session forces one more turn whenever a post-nudge reply lands with zero tool calls
-- Fixed launch tool rendering stacking a stale pending header over a bare `✓ Launch` line and raw text: the tool now uses a merged registry renderer with one per-op status header (op, target, `state · pid · uptime` meta), stripped log cursor suffixes, capped collapsed log/list previews, and a launch tool glyph
-- Fixed confusing launch start/wait results when readiness timed out with the log pattern already matched (readiness needs log AND port): the result printed a contradictory `Ready: <match>` next to `Readiness timed out` without naming the failing condition. Daemon snapshots now carry the unmet conditions (`readyPending`), and start/wait results state exactly what never happened (e.g. `port 3100 on 127.0.0.1 never accepted connections`); the TUI shows a `waiting on port` badge on starting daemons
-- Fixed the in-process `stat` builtin mangling BSD-style invocations like `stat -f "%Sm %N" file` (macOS muscle memory): GNU `-f` means `--file-system`, so the format string was treated as a file operand — printing filesystem info for the real operands and erroring with `cannot read file system information for '%Sm %N'`. A `-f` whose format value contains `%` is now detected as BSD syntax and translated to the GNU equivalent (`%Sm`→`%y`, `%N`→`%n`, `%z`→`%s`, epoch/`S`-form times, owner/group/permission and `H`/`L` sub-field directives, `-L`/`-n`/`-q`/`-F` flag clusters, with `%n`/`%t` as literal newline/tab); directives with no GNU counterpart fail with a clear `unsupported BSD format directive` error
-- Fixed the remaining GNU-flavored shell builtins that broke under macOS/BSD muscle memory, using the same unambiguous-detection approach as the `stat` fix (only invocations that are invalid or nonsensical under GNU semantics are reinterpreted; unsupported BSD forms fail loudly instead of producing wrong output): `date -r <epoch>` formats the epoch when no such file exists (GNU `-r FILE` mtime preserved), signed `date -v±N<unit>` adjustments translate to `-d` relative dates and `-j` is accepted (`-j -f` strptime parse mode and field-set `-v` error clearly); `sed -i '' 's/…/…/' file` drops the BSD empty backup-suffix token instead of treating it as the script; `mktemp -t prefix` without X's creates `$TMPDIR/prefix.XXXXXXXXXX` (the GNU `too few X's` error path); `tail -r` reverses input by delegating to `tac` (with `-n`/`-c`/`-f` combinations erroring clearly); `find -E` maps to `-regextype posix-extended` ahead of the expression; `base64 -D` decodes as an alias of `-d`; and `ln -sfh` works via a `-h` alias of `--no-dereference` (clap's `-h` help short is dropped to match real GNU/BSD ln; `--help` unchanged)
-- Fixed the browser tool crashing the whole process (parent session and every subagent) when a CDP world re-acquire failed mid-navigation: the stealth `puppeteer-core` patch called the bare `debugError` logger, which is `undefined` while the `puppeteer:error` debug channel is disabled (the default), turning a transient acquire failure into a fatal `TypeError` unhandled rejection. The patched `FrameManager`/`WebWorker` acquire paths now use `debugCatchError` ([#5296](https://github.com/can1357/oh-my-pi/issues/5296))
-- Fixed a role with a `:high` thinking suffix resolving to a longer sibling model whose id embeds the tier name (e.g. `kimi-for-coding:high` → `kimi-for-coding-highspeed`). The thinking suffix is now stripped before any fuzzy match, so `provider/model:high` keeps the exact model at high effort ([#5151](https://github.com/can1357/oh-my-pi/issues/5151)).
-
 ## [17.0.0] - 2026-07-15
 
 ### Breaking Changes
@@ -176,26 +292,6 @@
 
 - Enhanced Anthropic credential and usage management to support organization-scoped accounts, including displaying organization names in /usage, /logout, omp token --list, and OAuth login success messages, resolving active-account matching for shared organizations, and deduplicating identities during migration.
 
-### Fixed
-
-- Fixed compatibility of GNU-flavored shell builtins (such as stat, date, sed, mktemp, tail, find, base64, and ln) when invoked with macOS/BSD-style arguments.
-- Fixed subagent model and thinking level resolution to correctly respect the configured modelRoles.task selector instead of intermittently falling back to the parent session's model.
-- Fixed TUI rendering issues, including preventing macOS runtime diagnostics from painting into the viewport, bounding transcript retention in long sessions, and fixing scrollback repainting when collapsing history.
-- Fixed /tan and /fork clones failing to inherit or persist the parent session's prompt cache keys.
-- Fixed Python and JavaScript evaluation kernels suspending the CLI on subprocess foregrounding, deadlocking on non-serializable values, or losing in-flight subagent work during external aborts.
-- Fixed configured retry.fallbackChains failing to engage when encountering non-retryable provider errors.
-- Improved auto-compaction to automatically drop images and elide content when context is tight, and added persistent warning badges to the compaction divider when manual intervention is required.
-- Fixed the downshift plan nudge silently ending runs with no code written when the model answered with a text-only reply.
-- Fixed launch tool rendering and status reporting, including resolving contradictory readiness timeout messages and preventing backgrounded Bash blocks from continuing to repaint.
-- Fixed Advisor containment and timing issues, preventing hallucinated tool calls from contaminating later advice and ensuring late-arriving transcript deltas are coalesced before advisor calls.
-- Fixed omp update on npm-managed Windows installations to prevent downloaded release binaries from overwriting npm launchers.
-- Fixed --max-time duration values (e.g., 5s, 10m, 1h) being ignored instead of setting a session deadline.
-- Fixed omp plugin install --force failing with a dependency loop when replacing an existing pinned Git plugin source.
-- Fixed MCP tools receiving session image attachments as raw local:// URIs instead of resolving them to local filesystem paths.
-- Fixed Pyright LSP semantic requests hanging during startup.
-- Fixed Codex web search requests for GPT-5.6 Responses-Lite models.
-- Fixed custom model/provider configuration discovery to correctly load ~/.omp/agent/models.yaml when models.yml is absent.
-
 ## [16.5.0] - 2026-07-13
 
 ### Breaking Changes
@@ -243,6 +339,32 @@
 
 ### Added
 
+- Added invocation-specific schemas to task subagents and unified task/eval agent execution, including host-enforced read-only plan-mode agents ([#5279](https://github.com/can1357/oh-my-pi/issues/5279))
+
+### Fixed
+
+- Fixed compatibility of GNU-flavored shell builtins (such as stat, date, sed, mktemp, tail, find, base64, and ln) when invoked with macOS/BSD-style arguments.
+- Fixed subagent model and thinking level resolution to correctly respect the configured modelRoles.task selector instead of intermittently falling back to the parent session's model.
+- Fixed TUI rendering issues, including preventing macOS runtime diagnostics from painting into the viewport, bounding transcript retention in long sessions, and fixing scrollback repainting when collapsing history.
+- Fixed /tan and /fork clones failing to inherit or persist the parent session's prompt cache keys.
+- Fixed Python and JavaScript evaluation kernels suspending the CLI on subprocess foregrounding, deadlocking on non-serializable values, or losing in-flight subagent work during external aborts.
+- Fixed configured retry.fallbackChains failing to engage when encountering non-retryable provider errors.
+- Improved auto-compaction to automatically drop images and elide content when context is tight, and added persistent warning badges to the compaction divider when manual intervention is required.
+- Fixed the downshift plan nudge silently ending runs with no code written when the model answered with a text-only reply.
+- Fixed launch tool rendering and status reporting, including resolving contradictory readiness timeout messages and preventing backgrounded Bash blocks from continuing to repaint.
+- Fixed Advisor containment and timing issues, preventing hallucinated tool calls from contaminating later advice and ensuring late-arriving transcript deltas are coalesced before advisor calls.
+- Fixed omp update on npm-managed Windows installations to prevent downloaded release binaries from overwriting npm launchers.
+- Fixed --max-time duration values (e.g., 5s, 10m, 1h) being ignored instead of setting a session deadline.
+- Fixed omp plugin install --force failing with a dependency loop when replacing an existing pinned Git plugin source.
+- Fixed MCP tools receiving session image attachments as raw local:// URIs instead of resolving them to local filesystem paths.
+- Fixed Pyright LSP semantic requests hanging during startup.
+- Fixed Codex web search requests for GPT-5.6 Responses-Lite models.
+- Fixed custom model/provider configuration discovery to correctly load ~/.omp/agent/models.yaml when models.yml is absent.
+
+## [16.4.8] - 2026-07-12
+
+### Added
+
 - Added a predicate form to the browser run's `wait()` helper: `wait(fn, { timeout?, interval? })` polls the function (sync or async) until truthy and resolves with that value, failing with a named timeout error (deadline clamped under the cell budget so it always beats the opaque whole-cell timeout) instead of Bun's `sleep expects a number` or a whole-cell stall from in-page polling Promises; both `wait` forms now register in the stall diagnosis of cell timeouts
 - Added `--reasoning-slide-model` and `--reasoning-slide-turns` to switch a running agent from its initial model after a fixed number of completed assistant turns
 - Added `--reasoning-slide-plan` (with `--reasoning-slide-plan-at`) to steer a hidden deep-planning nudge into the run before the reasoning slide; the switch is held until a substantial plan turn actually lands (bounded by a grace window) and the nudge is scrubbed from the LLM context at the switch so the fast model inherits only the produced plan