chore: rewritten changelogs

.bazelrc

@@ -10,6 +10,19 @@ common --enable_platform_specific_config
 # PATH through crate annotations in MODULE.bazel.
 build --incompatible_strict_action_env
 
+# Third-party crate resolution (rules_rust crate_universe) shells out to
+# `cargo-bazel splice`, which runs `cargo metadata` + fetches every git/registry
+# dep. Two knobs matter on cold caches:
+#   1. CARGO_BAZEL_ISOLATED=0 — reuse the host's ~/.cargo registry/index instead
+#      of building a throwaway cargo home per invocation. Without this, every
+#      splice re-clones the entire crates.io index (~1 GiB, minutes on cold DNS).
+#   2. CARGO_BAZEL_TIMEOUT=1800 — cargo-bazel honors this to raise Bazel's
+#      default `repository_ctx.execute` 600s cap. Fresh machines / ephemeral
+#      runners with git-hosted forks (brush, uutils) blow past 600s and abort
+#      with `Timed out` (rules_rust common_utils.bzl:61) before we can help.
+common --repo_env=CARGO_BAZEL_ISOLATED=0
+common --repo_env=CARGO_BAZEL_TIMEOUT=1800
+
 # NOTE: rust pipelined_compilation stays OFF: handing dependents rmeta-only
 # crates breaks `rust_test(crate = ...)` harness compiles whose deps export
 # macro_rules! ("can't find crate" at macro expansion).

packages/agent/CHANGELOG.md

@@ -4,12 +4,11 @@
 
 ### Added
 
-- Added an opt-in `retry` option to `instrumentedCompleteSimple` (`InstrumentedChatSpanOptions.retry`) that re-issues a oneshot completion on transient Anthropic failures. Opt-in rather than default-on because `oneshotKind` is free-form and callers may pass arbitrary `ctx.tools`, so the funnel cannot itself prove a request is replay-safe. Response headers are captured per attempt and cleared between attempts, so a stale `retry-after` can never be reused for a later failure.
+- Added automatic retry support for transient provider failures during one-shot completions, allowing callers such as compaction to opt in to resilient request handling.
 
 ### Fixed
 
-- Fixed `/handoff`, branch summarization and manual `/compact` aborting on the first transient provider blip. Each is a single, side-effect-free completion whose result is parsed after it resolves, so an Anthropic `overloaded_error` / 429 / 529 now retries instead of failing the whole operation — previously one blip left the user's context full.
-- Added `SummaryOptions.oneshotRetry` so the compaction summarization oneshots can be retried by the caller that needs it without inflating the caller that does not. Manual `/compact` has no outer loop and gets retry by default; auto-compaction passes `false` because `session-maintenance` already retries the whole attempt, and nesting would multiply the request budget (10 outer x 3 inner) while stacking each outer wait on an inner backoff.
+- Fixed /handoff, branch summarization, and manual /compact failing outright on transient provider errors (e.g. Anthropic overloaded/429/529 responses); these operations now retry automatically instead of leaving the user's context full.
 
 ## [17.3.4] - 2026-08-14
 

packages/ai/CHANGELOG.md

@@ -4,29 +4,25 @@
 
 ### Added
 
-- Added `retryTransientCompletion` for oneshot (non-agent-loop) LLM calls. `completeSimple` reports a transient provider failure — Anthropic `overloaded_error` / `rate_limit_error`, HTTP 429/500/502/503/529 — by **resolving** with `stopReason: "error"` rather than throwing, so callers that wrapped it in a try/catch-based retry never actually retried. The helper handles both shapes (resolved error-stop and thrown error), classifies with the existing `AIError` predicates so the retryable set stays defined in one place, honors `retry-after` / `x-ratelimit-reset*` from response headers (supplied via `getResponseHeaders`, since an `AssistantMessage` carries none) or from the error text, and returns/rethrows the failure unchanged once attempts are exhausted so existing caller fallbacks keep working. Aborts surface immediately.
+- Added retryable oneshot completion support (`retryTransientCompletion`) so non-agent LLM calls correctly retry on transient provider failures (Anthropic overload/rate-limit errors, HTTP 429/500/502/503/529), honoring provider-supplied retry-after timing before giving up.
 
 ### Fixed
 
-- Stopped treating `XAI_API_KEY` as SuperGrok (`xai-oauth`) sign-in for availability, so paid-key-only setups default to `xai/grok-4.5` instead of the zero-cost SuperGrok catalog path. Explicit `xai-oauth/…` selectors still accept the paid key via the existing env fallback.
-- Omitted unsupported `reasoning.summary` on paid xAI Responses requests (`xai/grok-4.5`), matching SuperGrok, so a thinking level no longer serializes `summary: "auto"`.
-- Omitted presence/frequency penalties on all first-party xAI Responses models, including non-reasoning ids such as `xai/grok-2`.
-- Fixed Umans usage reporting computing utilization from raw request counts instead of the model-weighted "effective requests", falsely reporting the quota as exhausted while the account still had weighted headroom. The request limit is now split into a weighted soft-cap row (warns, never exhausts) and a raw burst-ceiling row (exhausted only where throttling actually starts), and the rolling 5h window's absolute `resets_at` is surfaced as a countdown ([#7858](https://github.com/can1357/oh-my-pi/issues/7858)). Payloads without a reported burst ceiling collapse to a single weighted `umans:requests` row that can exhaust, so request exhaustion is never unreportable.
-- Fixed `omp usage invalidate` to discard stale OAuth and API-key usage snapshots, then force a cache-bypassing, per-provider serialized refresh so upgraded subscriptions do not silently retain pre-change quota data.
-- Classified Cursor HTTP/2 `NGHTTP2_INTERNAL_ERROR` / `NGHTTP2_REFUSED_STREAM` stream closes as transient so session recovery can continue instead of treating the RST as a hard stop.
-- Fixed OpenAI-compatible streams that close without a `finish_reason` chunk (connection dropped mid-generation, e.g. DeepSeek's `insufficient_system_resource` interruption) being silently finalized as a clean `stop`; the truncated turn now surfaces as a retryable incomplete-stream error instead of halting mid-sentence.
-- Fixed DeepSeek's `insufficient_system_resource` finish reason mapping to a non-retryable error; it now matches the session retry classifier's transient-transport pattern so the turn is auto-retried.
-- Preserved opaque Chat Completions tool-call IDs during same-model replay so custom gateways retain provider correlation state ([#8641](https://github.com/can1357/oh-my-pi/issues/8641)).
-- Fixed Kimi Code multi-account routing to rank OAuth credentials by 5-hour and 7-day quota headroom, retain usage-limit blocks until reset, and preserve JWT account identity across token refreshes for stable usage labels and history ([#8630](https://github.com/can1357/oh-my-pi/issues/8630)).
-- Fixed Anthropic custom signing-proxy continuations dropping native tool-search call/result blocks from signed assistant history, preserving interleaved thinking for byte-identical replay ([#8559](https://github.com/can1357/oh-my-pi/issues/8559)).
-- Stopped runaway exact response cycles across model providers and kept persistent loops fail-closed after bounded guarded retries ([#8669](https://github.com/can1357/oh-my-pi/pull/8669) by [@pstarkgit](https://github.com/pstarkgit)).
-- Fixed xAI 400ing the whole turn on MCP schemas whose tool root is an object plus a typeless exclusive-required `anyOf` (e.g. codebase-memory `check_index_coverage`). Flatten only that root fragment — nested unions and branch property/`additionalProperties` constraints stay intact. Leftover object-root unions quarantine that one tool on paid xAI Completions and xAI OAuth Responses, not on OpenAI/Azure/Codex.
-
-### Removed
-
-- Fixed Alibaba DashScope/Bailian per-minute token throttle (`429 Throttling.AllocationQuota`, "You exceeded your current quota, please check your plan and billing details. … error-code#token-limit", `type=insufficient_quota`) being misclassified as `QUOTA_EXHAUSTED`. The OpenAI-compatible billing wording (and the `insufficient_quota` payload code) matched the usage-limit classifier, so a transient TPM/TPS cap — which Bailian's docs document as clearing within the minute window — blocked the credential and stalled the session for the 30-minute quota backoff instead of retrying with a short backoff on the same credential. Bodies linking the `error-code#token-limit` anchor now classify as `RATE_LIMIT_EXCEEDED` and stay in the transient lane; the identical wording without the anchor (OpenAI's real account-quota error) is unchanged.
-- Fixed Anthropic-compatible streams dropping thinking bytes supplied by `content_block_start`, which invalidated signed-thinking replay ([#8319](https://github.com/can1357/oh-my-pi/pull/8319) by [@max12525k](https://github.com/max12525k)).
-- Updated the Alibaba Coding Plan China login to open Bailian API-key management after the DashScope console was retired ([#8691](https://github.com/can1357/oh-my-pi/issues/8691)).
+- Fixed xAI availability detection so paid-key-only setups correctly default to `xai/grok-4.5` instead of the free SuperGrok catalog; explicit `xai-oauth/…` selectors still work as before.
+- Fixed xAI Responses requests sending unsupported parameters (reasoning summary, presence/frequency penalties) that some models rejected.
+- Fixed Umans usage reporting incorrectly marking quota as exhausted based on raw request counts instead of actual weighted usage, and improved the usage display to show both a soft-cap warning and a hard exhaustion limit with an accurate countdown to reset.
+- Fixed `omp usage invalidate` to fully clear stale usage data and force a fresh refresh, so upgraded subscriptions no longer show outdated quota information.
+- Improved session recovery to correctly treat certain Cursor HTTP/2 connection errors as transient instead of ending the session.
+- Fixed OpenAI-compatible streams (e.g. DeepSeek) that are cut off mid-generation being silently treated as a completed response instead of being retried.
+- Fixed DeepSeek resource-exhaustion interruptions not being automatically retried.
+- Fixed tool-call IDs being lost during same-model replay, which could break correlation with custom gateways.
+- Fixed Kimi Code multi-account routing to prefer accounts with more available quota, respect usage-limit cooldowns, and keep consistent usage history across token refreshes.
+- Fixed Anthropic custom signing-proxy conversations losing tool-search results and thinking content during replay.
+- Fixed rare runaway response loops across model providers so they now fail gracefully instead of repeating indefinitely.
+- Fixed xAI rejecting entire turns due to certain MCP tool schema shapes, restoring compatibility while isolating any remaining incompatible tools rather than failing the whole request.
+- Fixed Alibaba DashScope/Bailian transient per-minute rate limits being misclassified as full quota exhaustion, causing unnecessary long backoffs instead of quick retries.
+- Fixed Anthropic-compatible streams dropping thinking content, which broke replay of prior reasoning.
+- Updated the Alibaba Coding Plan China login flow to point to the current Bailian API-key management console.
 
 ## [17.3.4] - 2026-08-14
 

packages/catalog/CHANGELOG.md

@@ -4,30 +4,20 @@
 
 ### Added
 
-- Added support for GLM-5.3 on the z.AI provider. GLM-5.3 introduces a uniform wire-exact `low`/`high`/`max` reasoning-effort ladder on every host (replacing GLM-5.2's host-specific dialects), makes thinking mandatory (`thinking.type` must always be `enabled`; disabling is no longer supported), and defaults to `max` effort. The model is pinned to 1M context and set as the z.AI provider default.
+- Added support for GLM-5.3 on the z.AI provider, featuring a unified low/high/max reasoning-effort ladder across all hosts, mandatory thinking mode, 1M context, and default-model status for the z.AI provider.
 
 ### Changed
 
-- Switched the paid xAI provider (`xai` / `XAI_API_KEY`) from Chat Completions to the OpenAI Responses API (`POST https://api.x.ai/v1/responses`), matching SuperGrok `xai-oauth`. Prompt-cache affinity (`x-grok-conv-id`), reasoning-effort allowlisting, and encrypted-reasoning replay rules are now shared across both first-party xAI hosts.
-- Changed the paid xAI (`XAI_API_KEY`) default model from `grok-4-fast-non-reasoning` to `grok-4.5`.
-- Changed the SuperGrok (`xai-oauth`) default model from `grok-4.3` to `grok-4.5`.
-- Requested `reasoning.encrypted_content` on first-party xAI Responses calls (`xai` and `xai-oauth`) via the `include` parameter.
-- Replayed xAI encrypted reasoning items on later Responses turns instead of stripping `type: "reasoning"` history.
+- Switched the paid xAI provider (xai / XAI_API_KEY) from Chat Completions to the OpenAI Responses API, aligning it with SuperGrok (xai-oauth) for prompt-cache affinity, reasoning-effort handling, and encrypted-reasoning replay.
+- Changed the paid xAI (XAI_API_KEY) default model to grok-4.5.
+- Changed the SuperGrok (xai-oauth) default model to grok-4.5.
+- Improved reasoning continuity for xAI models by requesting and replaying encrypted reasoning content across multi-turn Responses API calls.
 
 ### Fixed
 
-- Fixed Codex Daybreak Blue and Red model discovery reporting zero token prices, which made the model picker label the models as free.
-- Fixed Baseten's `moonshotai/Kimi-K3` catalog metadata so its `low`/`high`/`max` thinking levels are available.
-- Invalidated stale paid-xAI model-cache rows written under Chat Completions so the Responses migration takes effect immediately instead of waiting for TTL expiry.
-- Clamped paid xAI Responses `minimal` reasoning effort to `low` (same wire map as SuperGrok) so `xai/grok-4.5` does not 400.
-- Suppressed presence/frequency penalties and stop sequences on xAI reasoning models so a configured `presencePenalty` does not 400 after the `grok-4.5` default change.
-- Stopped emitting stale `thinking.efforts` dials on paid xAI Responses catalog rows that reject `reasoning.effort` (`grok-code-fast-1`, `grok-build-0.1`, `grok-4.20-0309-reasoning`, and other off-allowlist reasoners).
-- Marked first-party xAI Responses hosts (`xai` and `xai-oauth`) as not supporting `reasoning.summary`, so paid `xai/grok-4.5` effort requests omit the unsupported field instead of sending `summary: "auto"`.
-- Removed unsupported `xhigh` (and `max`) thinking tiers from first-party Grok 4.5 / 4.3 / 3-mini Responses rows; leftover `xhigh`/`max` requests clamp to `high`. `grok-4.6*` and `grok-4.20-multi-agent*` advertise unmapped `xhigh`.
-- Stopped baking `reasoningEffortMap` on first-party xAI catalog rows that omit `reasoning.effort` (`omitReasoningEffort: true`).
-- Suppressed presence/frequency penalties on every first-party xAI Responses model, including non-reasoning ids such as `grok-2`; xAI's `/v1/responses` marks those fields unsupported.
-- Routed `grok-4.6` (added on main) through first-party xAI Responses and advertised its documented `xhigh` effort tier (4.5 stays 4-tier).
-- Fixed `opencode-go/deepseek-v4-flash` Responses requests sending forced named `tool_choice` selectors that Console Go rejects while thinking mode is active.
+- Fixed Codex Daybreak Blue and Red model discovery reporting zero token prices, which incorrectly labeled the models as free in the model picker.
+- Fixed Baseten's moonshotai/Kimi-K3 catalog metadata so its low/high/max thinking levels are available.
+- Fixed opencode-go/deepseek-v4-flash Responses requests sending forced named tool_choice selectors that are rejected while thinking mode is active.
 
 ## [17.3.4] - 2026-08-14
 

packages/coding-agent/CHANGELOG.md

@@ -6,66 +6,60 @@
 
 ### Changed
 
-- Routed paid xAI models (`XAI_API_KEY` / `xai/…`) through the Responses API used by SuperGrok OAuth instead of Chat Completions.
-- Changed the default model for `XAI_API_KEY` (`xai`) from `grok-4-fast-non-reasoning` to `grok-4.5`.
-- Changed the default model for SuperGrok OAuth (`xai-oauth`) from `grok-4.3` to `grok-4.5`.
-- Included `reasoning.encrypted_content` in Responses `include` for paid xAI and SuperGrok OAuth models.
-- Replayed encrypted xAI reasoning on follow-up Responses turns for `xai` and `xai-oauth`.
-- Kept automatic model selection on paid `xai/grok-4.5` when only `XAI_API_KEY` is set, instead of preferring SuperGrok `xai-oauth/grok-4.5`. Explicit `xai-oauth/grok-4.5` still works with that paid key.
-- Stopped sending presence/frequency penalties and stop sequences to xAI reasoning models such as `grok-4.5`, which reject them.
+- Routed paid xAI models (XAI_API_KEY / xai/…) through the Responses API used by SuperGrok OAuth instead of Chat Completions, including reliable replay of encrypted reasoning content on follow-up turns.
+- Updated the default model for XAI_API_KEY (xai) to grok-4.5, and the default SuperGrok OAuth (xai-oauth) model to grok-4.5. Automatic model selection continues to prefer paid xai/grok-4.5 when only XAI_API_KEY is set, with xai-oauth/grok-4.5 still available explicitly.
+- Stopped sending presence/frequency penalties and stop sequences to xAI reasoning models such as grok-4.5, which reject them.
 
 ### Fixed
 
-- Fixed prompt caching on open-weight providers (DeepSeek, Qwen, GLM, …) by moving the per-request date/cwd line out of the system prompt into a `<system-reminder>` block on the first user turn, so the tool schemas rendered after the system content stay cached across directory changes and midnight rollovers ([#7404](https://github.com/can1357/oh-my-pi/issues/7404)).
-- Kept completed edit previews within their collapsed line limit for oversized single hunks and cached header change statistics across terminal repaints.
-- Fixed `omp --fork` omitting the source session's artifact directory, so CLI-created forks now preserve `artifact://` references like interactive `/fork`.
-- Fixed long `ask` option labels being hard-truncated at the terminal width with no way to read the clipped tail; the rich ask dialog now wraps labels onto indented continuation lines, matching the legacy selector ([#1243](https://github.com/can1357/oh-my-pi/issues/1243), [#8594](https://github.com/can1357/oh-my-pi/issues/8594)).
-- Fixed toggling `display.showTokenUsage` from `/settings` leaving existing token-usage rows stale until the transcript was rebuilt.
-- Fixed mid-run auto-compaction waiting on `auto_compaction_end` / `session_compact` extension handlers before the next provider call, which could hang the live loop after a snapcompact or context-full pass. Mid-run those handlers now run concurrently with the next turn; `auto_compaction_start` is still awaited.
-- Reduced peak memory for persisted subagent revival probes by streaming large file-backed session journals instead of loading the complete journal ([#8117](https://github.com/can1357/oh-my-pi/issues/8117)).
-- Kept streaming edit previews responsive for large diffs by selecting and highlighting only the visible tail instead of scanning the discarded prefix on every update.
-- Fixed repeated `/btw` panels committing transient frames to native scrollback and replaying conversation history after dismissal.
+- Fixed llama.cpp model discovery producing a baseUrl without the /v1 prefix for non-Qwen models, causing 404 errors on OpenAI-compatible endpoints.
+- Fixed prompt caching on open-weight providers (DeepSeek, Qwen, GLM, …) so tool schemas stay cached across directory changes and midnight rollovers.
+- Fixed omp --fork omitting the source session's artifact directory, so CLI-created forks now preserve artifact:// references like interactive /fork.
+- Fixed long ask option labels being hard-truncated at the terminal width; labels now wrap onto indented continuation lines.
+- Fixed toggling display.showTokenUsage from /settings leaving existing token-usage rows stale until the transcript was rebuilt.
+- Fixed mid-run auto-compaction blocking the live loop while waiting on extension handlers, which could hang after a snapcompact or context-full pass.
+- Reduced peak memory for persisted subagent revival probes by streaming large file-backed session journals instead of loading them fully.
+- Improved responsiveness of streaming edit previews for large diffs by rendering only the visible tail.
+- Fixed repeated /btw panels committing transient frames to native scrollback and replaying conversation history after dismissal.
 - Clarified that closing browser tool sessions releases managed handles without closing pages in spawned, CDP-connected, or relay browsers.
-- Fixed interrupted `vibe_wait` calls being reported as elapsed timeout windows while preserving the per-call wait timeout.
-- Optimized checkpoint/rewind prompt rendering: the rewind instruction moved from the permanent checkpoint tool result (stale after rewind) into a transient `<system-notice>` that is branch-cut away on rewind, leaving only the goal; the rewind-report prompt is now forward-looking with no negation guard; the rewind tool description collapsed to one line. ([#8499](https://github.com/can1357/oh-my-pi/issues/8499))
-- Continued Cursor turns that died with `NGHTTP2_INTERNAL_ERROR` / `NGHTTP2_REFUSED_STREAM` after tool calls already had results, instead of leaving the agent idle until the user typed "continue". HTTP/2 stream resets now use the same preserve-and-continue path as idle stream stalls, without requiring the Cursor exec-resolved marker that MCP/todo blocks never carry.
-- Stopped mixed-case plugin tool names from being lowercased during tool-set refresh, which unmounted them from `xd://` whenever MCP tools connected.
-- Kept Exa MCP servers mounted when their config explicitly requests tools the native Exa integration does not provide (`web_fetch_exa`, `web_search_advanced_exa`), instead of filtering them out so `/mcp reconnect exa` could never connect.
-- Fixed Claude Code custom tool discovery attempting to import non-module files from `.claude/tools` ([#8471](https://github.com/can1357/oh-my-pi/pull/8471) by [@Kigbnajd](https://github.com/Kigbnajd)).
-- Fixed Agent Hub parking a mid-spawn child JSONL (title + session header only) so `task` then fails with `already owned by another session generation` and the row cannot be revived.
-- Fixed the welcome banner displaying a stale model name when the session's active model changes after startup (e.g. after a delayed config load or an explicit `/model` switch). The banner now subscribes to `model_changed` events and calls `WelcomeComponent.setModel()`, which already existed but was never wired up.
+- Fixed interrupted vibe_wait calls being reported as elapsed timeout windows.
+- Improved checkpoint/rewind prompt rendering to stay accurate after a rewind and be more concise.
+- Fixed Cursor turns dying with HTTP/2 stream errors (NGHTTP2_INTERNAL_ERROR / NGHTTP2_REFUSED_STREAM) after tool calls already had results, instead of leaving the agent idle until the user typed "continue".
+- Fixed mixed-case plugin tool names being lowercased during tool-set refresh, which unmounted them from xd:// whenever MCP tools connected.
+- Fixed Exa MCP servers being unmounted when their config explicitly requests tools the native Exa integration does not provide, breaking /mcp reconnect exa.
+- Fixed Claude Code custom tool discovery attempting to import non-module files from .claude/tools.
+- Fixed Agent Hub parking a mid-spawn child session so subsequent task calls failed with an ownership error and the row could never be revived.
+- Fixed the welcome banner displaying a stale model name when the session's active model changes after startup (e.g. after a delayed config load or an explicit /model switch).
 - Fixed Nix standalone binaries retaining Bun's build-time package in their runtime closure.
-- Fixed birch user/custom message card contrast on dark terminals: `userMessageText` and `customMessageText` were empty and inherited the terminal default foreground over light paper card backgrounds, so chat bubbles could render light-on-light. Those tokens now use `forestFloor`. Selection chrome and shared `toolTitle` stay adaptive (terminal-default) so light terminals and unpainted inline tool rows are not regressed.
-- Fixed hidden tool snapshots preventing long streamed assistant responses from entering terminal scrollback ([#8285](https://github.com/can1357/oh-my-pi/pull/8285) by [@dannyboy-ai](https://github.com/dannyboy-ai)).
-- Prevented `omp models` from loading ambient hook factories while preserving extension-contributed providers ([#8414](https://github.com/can1357/oh-my-pi/pull/8414) by [@starlink-awaken](https://github.com/starlink-awaken)).
-- Fixed the ask dialog's multi-select mode dead-ending on Enter: Space toggles options and Enter submits the current selection (matching single-select), instead of Enter silently toggling the focused option with the only submit path hidden behind the Submit tab ([#8252](https://github.com/can1357/oh-my-pi/issues/8252)).
-- Fixed workspace diagnostics reporting a clean workspace when its checker crashed without producing output ([#8386](https://github.com/can1357/oh-my-pi/issues/8386)).
-- Fixed manual `/compact` failing outright when a summarization request hit a transient provider overload. The auto-compaction path is unchanged: it keeps its own retry loop and now explicitly opts out of the inner one.
-- Fixed transient Anthropic failures (`overloaded_error`, `rate_limit_error`, 429/500/502/503/529) aborting or silently degrading side-effect-free oneshot LLM calls. Session title generation, TTS speech enhancement, commit-message generation, the auto-thinking and unexpected-stop classifiers, memory extraction/consolidation, the commit analysis/summary/changelog/map/reduce passes, and the mnemopi LLM callback now retry with backoff that honors `retry-after`, instead of failing on the first blip or returning `null` — which made a transient overload indistinguishable from a legitimate empty result.
-- Fixed the commit analysis, summary, changelog and reduce passes feeding a provider error message straight into their response parsers: they never checked `stopReason`, so a failed request produced garbage output instead of surfacing the error.
-- Fixed the commit map phase never retrying transient failures: its retry helper only caught thrown errors, so a `stopReason: "error"` response bypassed it entirely.
-- Fixed the shared headless browser daemon launching from the macOS system Google Chrome bundle (`com.google.Chrome`), which let macOS LaunchServices route the user's link clicks to the automation daemon and silently swallow them; the daemon now prefers the isolated Chrome for Testing binary (`com.google.chrome.for.testing`) on macOS and falls back to system Chrome only when Chrome for Testing cannot be obtained ([#8673](https://github.com/can1357/oh-my-pi/issues/8673)).
-- Reclaimed abandoned daemon runtime directories under `~/.omp/run/daemons/`: each broker now prunes sibling scopes with a dead/absent `broker.pid`, no live client presence, and past a short stale grace on startup, removing the leftover Chromium profiles and broker state that grew unbounded from short-lived project directories ([#8674](https://github.com/can1357/oh-my-pi/issues/8674)).
-- Kept the welcome screen's Tips, LSP Servers, and Recent sessions visible when a long model name still leaves enough terminal width for both columns ([#8657](https://github.com/can1357/oh-my-pi/issues/8657)).
-- Fixed focused `ultrathink`, `orchestrate`, and `workflowz` shimmer frames repainting the full TUI every 70 ms, causing high CPU usage while composing prompts on WSL2 ([#8646](https://github.com/can1357/oh-my-pi/issues/8646)).
-- Fixed the `/debug` report bundle including unrelated historic sessions: subagent transcripts were collected by scanning the entire sessions directory, leaking other sessions' `.jsonl` files and bloating archives to tens of MB. The bundle now captures only the current session's own artifacts subtree ([#8648](https://github.com/can1357/oh-my-pi/issues/8648)).
-- Fixed adopted keep-alive agents remaining `running` in the registry after deferred turn settlement, and prevented stale detached or non-streaming refs from sustaining bare `hub wait` calls indefinitely ([#8634](https://github.com/can1357/oh-my-pi/issues/8634)).
-- Expanded home-relative marketplace catalog paths before cache access, preventing updates from writing into a literal `~` directory under the process working directory ([#8627](https://github.com/can1357/oh-my-pi/issues/8627)).
-- Fixed broker-owned headless Chromium opening and retaining an unowned blank foreground window on Windows ([#8615](https://github.com/can1357/oh-my-pi/issues/8615)).
-- Fixed the `auto` thinking classifier failing every turn on Anthropic models served through LiteLLM/Vertex with `max_tokens must be greater than thinking.budget_tokens`. The classifier's disabled-reasoning request is downgraded to the lowest reasoning effort on the `openai-completions` transport, which the proxy translates to an Anthropic thinking budget of at least 1024 tokens; the classifier now reserves enough output room (4096) to clear that budget instead of capping at exactly 1024 ([#8610](https://github.com/can1357/oh-my-pi/issues/8610)).
-- Fixed `always-ask` approval prompts bypassing edit preview readiness when a built-in tool executes under its wire-level alias, such as `edit` running as `apply_patch` ([#8607](https://github.com/can1357/oh-my-pi/issues/8607)).
-- Fixed `lsp reload` crashing non-rust-analyzer language servers (e.g. Roslyn/`roslyn-language-server`) by sending the rust-analyzer-specific `rust-analyzer/reloadWorkspace` request to every server; the request is now gated on the server being rust-analyzer, and all other servers reload via `workspace/didChangeConfiguration` directly ([#8571](https://github.com/can1357/oh-my-pi/issues/8571)).
-- Fixed `browser open` failing with "Shared browser daemon unavailable" when `HTTP_PROXY`/`HTTPS_PROXY` is set (e.g. a local Clash proxy), because the shared-browser CDP liveness probes routed loopback `127.0.0.1` requests through the proxy, which 502'd them and killed the healthy daemon. The probes now talk to the endpoint over raw TCP and never touch a proxy ([#8567](https://github.com/can1357/oh-my-pi/issues/8567)).
-- Fixed `defaultThinkingLevel: auto` skipping classification for user-invoked `/skill:<name>` turns, which left the effort stuck on pending `auto`; user-attributed skill prompts now classify like any user turn while agent/autoload injections stay excluded ([#8554](https://github.com/can1357/oh-my-pi/issues/8554)).
-- Fixed custom-tool and other directory discovery recursing into subtrees despite a non-recursive default: `loadFilesFromDir` built a top-level-only glob pattern but never forwarded its `recursive` flag to the native glob (which defaults to recursive), so `~/.codex/tools` scans descended into a Python venv's `site-packages` and imported browser-only frontend assets as tools, crashing startup on an unhandled `window` rejection ([#8552](https://github.com/can1357/oh-my-pi/issues/8552)).
-- Repaired torn session JSONL appends after disk-write failures, rewrote malformed resumed files before their next append, retried transient persistence failures, and surfaced failures in the TUI ([#8596](https://github.com/can1357/oh-my-pi/issues/8596)).
-- Prevented Anthropic model fallback from replaying model-bound thinking blocks across models, and surfaced immutable-thinking `invalid_request_error` responses without retrying the unchanged invalid turn ([#8558](https://github.com/can1357/oh-my-pi/issues/8558)).
-- Fixed the capped empty-stop failure always naming the context/`/shake images` hint even when the provider billed output tokens. A zero-block `stop` (no content blocks) with billed output beyond any provider-reported reasoning usage means content was generated and dropped downstream (a filter/refusal flattened to `finish_reason: "stop"` by a proxy, or a lossy API translation), so the message now reports the billed output-token count and points at a provider-side filter/translation instead of a context problem, and logs `outputTokens` alongside the existing warning fields. Thinking-only and known reasoning-only stops retain the context hint ([#8511](https://github.com/can1357/oh-my-pi/issues/8511)).
-- Fixed a parked, session-less agent-registry entry with no reviver permanently poisoning its agent id for the process lifetime: a fresh subagent spawn reusing that id died post-registration with `already owned by another session generation`, and messages to it failed with `is parked and cannot be revived`. Such dead corpses (left by an isolated run's park or an interrupted construction) are now reclaimed on a fresh-spawn collision so the id becomes reusable; the corpse's transcript remains readable at `history://<id>` ([#8490](https://github.com/can1357/oh-my-pi/issues/8490)).
+- Fixed birch user/custom message card contrast on dark terminals, where chat bubbles could render light-on-light.
+- Fixed hidden tool snapshots preventing long streamed assistant responses from entering terminal scrollback.
+- Prevented omp models from loading ambient hook factories while preserving extension-contributed providers.
+- Fixed the ask dialog's multi-select mode dead-ending on Enter; Space now toggles options and Enter submits the current selection.
+- Fixed workspace diagnostics reporting a clean workspace when its checker crashed without producing output.
+- Fixed manual /compact failing outright when a summarization request hit a transient provider overload.
+- Fixed transient Anthropic failures (overloaded_error, rate_limit_error, 429/500/502/503/529) aborting or silently degrading side-effect-free background LLM calls such as session title generation, TTS speech enhancement, commit-message generation, thinking/stop classifiers, memory extraction/consolidation, and commit analysis/summary/changelog passes; these now retry with backoff honoring retry-after instead of failing or returning an indistinguishable empty result.
+- Fixed the shared headless browser daemon launching from the macOS system Google Chrome bundle, which could cause macOS to route the user's link clicks to the automation daemon and silently swallow them; the daemon now prefers an isolated Chrome for Testing binary on macOS.
+- Reclaimed abandoned daemon runtime directories under ~/.omp/run/daemons/, preventing unbounded growth of leftover Chromium profiles and broker state.
+- Kept the welcome screen's Tips, LSP Servers, and Recent sessions visible when a long model name still leaves enough terminal width for both columns.
+- Fixed focused shimmer animation frames (ultrathink, orchestrate, workflowz) repainting the full TUI too frequently, causing high CPU usage while composing prompts on WSL2.
+- Fixed the /debug report bundle including unrelated historic sessions, leaking other sessions' files and bloating archives.
+- Fixed adopted keep-alive agents remaining stuck in a running state in the registry after deferred turn settlement, and prevented stale refs from sustaining bare hub wait calls indefinitely.
+- Fixed home-relative marketplace catalog paths not being expanded before cache access, preventing updates from writing into a literal ~ directory.
+- Fixed broker-owned headless Chromium opening and retaining an unowned blank foreground window on Windows.
+- Fixed the auto thinking classifier failing every turn on Anthropic models served through LiteLLM/Vertex due to a thinking-budget mismatch.
+- Fixed always-ask approval prompts bypassing edit preview readiness when a built-in tool executes under its wire-level alias, such as edit running as apply_patch.
+- Fixed lsp reload crashing non-rust-analyzer language servers by sending them a rust-analyzer-specific request; that request is now gated to rust-analyzer only.
+- Fixed browser open failing with "Shared browser daemon unavailable" when HTTP_PROXY/HTTPS_PROXY is set, because liveness probes were incorrectly routed through the proxy.
+- Fixed defaultThinkingLevel: auto skipping classification for user-invoked /skill:<name> turns, leaving the effort stuck on pending auto.
+- Fixed custom-tool directory discovery recursing into subtrees despite a non-recursive default, which could crash startup when scanning large dependency directories such as Python venvs.
+- Repaired torn session JSONL appends after disk-write failures, rewrote malformed resumed files before their next append, retried transient persistence failures, and surfaced failures in the TUI.
+- Prevented Anthropic model fallback from replaying model-bound thinking blocks across models, and surfaced immutable-thinking errors without retrying the unchanged invalid turn.
+- Fixed empty-stop failure messages always suggesting a context problem even when the provider billed output tokens; the message now reports the billed token count and points at a provider-side filter/translation issue when appropriate.
+- Fixed a parked, session-less agent-registry entry with no reviver permanently poisoning its agent id, preventing fresh subagent spawns from reusing that id.
 - Made extension tool-call timeouts configurable and paused them during user dialogs.
-- Fixed `/vibe` cancellation leaving an in-flight model turn unaware that Vibe mode and its tools were removed ([#8326](https://github.com/can1357/oh-my-pi/issues/8326)).
-- Fixed empty local-model stops lingering on the persisted active branch after retries; discarded turns now durably select their parent, preserve safe metadata children, and cannot resurface after reload or a mid-retry process kill. ([#5179](https://github.com/can1357/oh-my-pi/issues/5179))
-- Fixed the Biome linter client silently dropping every diagnostic because `#parseJsonOutput` expected an outdated `--reporter=json` schema (`location.path.file`, byte-offset `span`, `description`); it now reads Biome 2.x's string `location.path`, 1-indexed `location.start`/`location.end`, and `message`, and warns when a non-empty diagnostics array has no recognizable location ([#8694](https://github.com/can1357/oh-my-pi/issues/8694)).
+- Fixed /vibe cancellation leaving an in-flight model turn unaware that Vibe mode and its tools were removed.
+- Fixed empty local-model stops lingering on the persisted active branch after retries, preventing them from resurfacing after reload or a mid-retry process kill.
+- Fixed the Biome linter client silently dropping every diagnostic due to an outdated JSON output schema; it now supports Biome 2.x's diagnostic format.
 
 ## [17.3.4] - 2026-08-14
 

packages/mnemopi/CHANGELOG.md

@@ -4,7 +4,7 @@
 
 ### Fixed
 
-- Fixed transient provider failures collapsing to `null` in the configured-completion and remote-LLM paths, which made an Anthropic overload or rate limit indistinguishable from "the model returned nothing". Both now retry transient failures before falling back.
+- Fixed an issue where transient provider failures (such as Anthropic overload or rate limit errors) were incorrectly treated as empty responses; these failures are now retried automatically before falling back.
 
 ## [17.3.4] - 2026-08-14
 

packages/natives/CHANGELOG.md

@@ -4,8 +4,8 @@
 
 ### Fixed
 
-- Fixed the native `xargs` builtin panicking in `-I`/`-i` replace mode when stdin is empty; it now exits successfully without running the command, matching GNU behavior ([#8595](https://github.com/can1357/oh-my-pi/issues/8595)).
-- Fixed inline-code foreground color carrying into plain text when a Markdown codespan ended exactly at a soft-wrap boundary ([#8582](https://github.com/can1357/oh-my-pi/issues/8582)).
+- Fixed the native `xargs` builtin panicking in `-I`/`-i` replace mode when stdin is empty; it now exits successfully without running the command, matching GNU behavior.
+- Fixed inline-code foreground color incorrectly carrying into plain text when a Markdown codespan ended exactly at a soft-wrap boundary.
 
 ## [17.3.4] - 2026-08-14
 

packages/tui/CHANGELOG.md

@@ -4,8 +4,8 @@
 
 ### Fixed
 
-- Fixed long CPU-bound event-loop stalls being misclassified as system sleep and omitted from `ui.loop-blocked` diagnostics ([#8442](https://github.com/can1357/oh-my-pi/pull/8442) by [@Mustaqeem66](https://github.com/Mustaqeem66)).
-- Fixed marker-bearing focused components falling back from direct row writes to full TUI renders, while preserving hardware-cursor position and native scrollback across marker changes ([#8646](https://github.com/can1357/oh-my-pi/issues/8646)).
+- Fixed long CPU-bound event-loop stalls being misclassified as system sleep and omitted from loop-blocked diagnostics.
+- Fixed focused components with markers falling back to full-screen redraws instead of direct row updates, preserving cursor position and native scrollback across marker changes.
 
 ## [17.3.4] - 2026-08-14
 

packages/utils/CHANGELOG.md

@@ -4,7 +4,7 @@
 
 ### Fixed
 
-- Fixed the vendored Markdown lexer letting an indented code block interrupt a paragraph: a 4-space-indented line directly attached to paragraph text (e.g. a box-drawing tree child under a `└──` branch) now stays a lazy continuation of that paragraph, matching marked/CommonMark, instead of rendering the rest of the message section as a raw code block ([#8557](https://github.com/can1357/oh-my-pi/pull/8557) by [@chan1103](https://github.com/chan1103)).
+- Fixed the Markdown renderer incorrectly breaking into a raw code block when a 4-space-indented line (such as a box-drawing tree child under a └── branch) directly followed paragraph text; it now correctly stays part of the paragraph, matching standard Markdown behavior.
 
 ## [17.3.2] - 2026-08-13