chore: rewrite changelog

packages/agent/CHANGELOG.md

@@ -4,8 +4,8 @@
 
 ### Fixed
 
-- Fixed response-only usage records being treated as authoritative context anchors while preserving prompt and total-only provider telemetry ([#7163](https://github.com/can1357/oh-my-pi/pull/7163) by [@harshav167](https://github.com/harshav167)).
-- Fixed compaction summaries being allowed to grow with the context window: the summary output budget is now capped at `MAX_SUMMARY_TOKENS` (16384), so a large window no longer authorizes a summary big enough to copy the conversation instead of compressing it.
+- Fixed an issue where response-only usage records were incorrectly treated as authoritative context anchors, while ensuring prompt and total-only provider telemetry remains preserved.
+- Fixed context compaction summaries growing excessively with large context windows by capping the summary output budget to 16,384 tokens, ensuring conversations are properly compressed rather than duplicated.
 
 ## [17.2.0] - 2026-07-30
 

packages/ai/CHANGELOG.md

@@ -4,21 +4,18 @@
 
 ### Added
 
-- Added the `gmi-cloud` provider registry definition with an API-key paste login (validates against `https://api.gmi-serving.com/v1/models`) and wired it into the provider registry, pairing with the catalog entry in `@oh-my-pi/pi-catalog`.
-### Changed
-
-- `AuthStorage.redeemResetCredit` now spends the account's soonest-expiring available saved reset (via the new `pickSoonestExpiringCredit`) instead of the first credit the backend lists, and reports a distinct retryable `credit_list_failed` outcome when the credit listing itself fails in transport instead of conflating it with a genuine `no_credit`.
-### Fixed
+- Added support for the `gmi-cloud` provider registry, including API-key paste login validation and integration with `@oh-my-pi/pi-catalog`.
 
 ### Changed
 
-- Exported `SENSITIVE_TOKEN_RE` from `providers/transform-messages` so hosts can route the same credential shapes through reversible obfuscation instead of the irreversible redaction fallback ([#6968](https://github.com/can1357/oh-my-pi/issues/6968)).
+- Updated `AuthStorage.redeemResetCredit` to prioritize spending the soonest-expiring available saved reset credit, and improved error handling to distinguish between transport failures (`credit_list_failed`) and a genuine lack of credits.
+- Exported `SENSITIVE_TOKEN_RE` from `providers/transform-messages` to allow hosts to route credential shapes through reversible obfuscation instead of irreversible redaction.
 
 ### Fixed
 
-- Fixed Cursor conversation checkpoints being recorded as billable output tokens; authoritative context occupancy is now tracked separately so usage totals remain accurate while the context gauge retains the server-reported value ([#7163](https://github.com/can1357/oh-my-pi/pull/7163) by [@harshav167](https://github.com/harshav167)).
-- Fixed `AuthStorage.refreshStoredOAuthCredential` returning an expired-but-refreshable OAuth credential without refreshing it when the caller's observed credential mismatched the stored row. A concurrent usage-fetch cycle could rotate the in-memory selected credential out from under a request (e.g. plan finalization); the observed-mismatch guard then adopted the stored copy verbatim, and if that copy was itself expired it flowed into `getOAuthApiKey`, which refused it — surfacing as a misleading `No API key found for <provider>`. The mismatch guard now only short-circuits when the stored credential is still fresh; expired stored copies fall through to a normal refresh ([#7179](https://github.com/can1357/oh-my-pi/issues/7179)).
-- Fixed Cursor history replay flattening assistant tool calls/results and dropping same-model Kimi K3 thinking, preserving Cursor's structured message order and rejecting unsafe mid-session switches to K3 ([#7184](https://github.com/can1357/oh-my-pi/issues/7184)).
+- Fixed an issue where Cursor conversation checkpoints were incorrectly recorded as billable output tokens, ensuring accurate usage totals.
+- Fixed an issue in `AuthStorage.refreshStoredOAuthCredential` where expired OAuth credentials were returned without being refreshed when a credential mismatch occurred, which previously resulted in misleading "No API key found" errors.
+- Fixed Cursor history replay issues by preserving structured message order for assistant tool calls/results, retaining Kimi K3 thinking blocks, and preventing unsafe mid-session switches to K3.
 
 ## [17.2.1] - 2026-07-30
 

packages/catalog/CHANGELOG.md

@@ -4,16 +4,16 @@
 
 ### Added
 
-- Added the GMI Cloud provider (`gmi-cloud`), an OpenAI-compatible inference gateway at `https://api.gmi-serving.com/v1` with dynamic model discovery via `/v1/models` and an API-key paste login. Authenticates with the `GMI_API_KEY` environment variable. The default model (`deepseek-ai/DeepSeek-V4-Flash`) is seeded into the bundled catalog so a fresh install resolves it before the first discovery pass.
-- Added optional authoritative context occupancy to usage records for providers whose checkpoint telemetry is separate from billable token buckets ([#7163](https://github.com/can1357/oh-my-pi/pull/7163) by [@harshav167](https://github.com/harshav167)).
+- Added support for the GMI Cloud provider (`gmi-cloud`), an OpenAI-compatible inference gateway with dynamic model discovery and API-key authentication via the `GMI_API_KEY` environment variable.
+- Added optional authoritative context occupancy to usage records for providers with separate checkpoint telemetry and billable token buckets.
 
 ### Fixed
 
-- Fixed dynamically discovered Cursor Kimi K3 effort variants being classified as non-reasoning models when Cursor omits `thinkingDetails` ([#7184](https://github.com/can1357/oh-my-pi/issues/7184)).
+- Fixed classification of dynamically discovered Cursor Kimi K3 effort variants as non-reasoning models when `thinkingDetails` is omitted.
 - Fixed Google AI Studio OpenAI-compatible requests failing with HTTP 400 by omitting the unsupported `store` field.
-- Fixed Synthetic models losing their thinking selector, vision input, output cap and pricing: the discovery mapper read `supports_reasoning`, `supports_vision` and `max_tokens`, none of which Synthetic sends. It advertises `supported_features`, `reasoning_parameters.efforts`, `input_modalities`, `max_output_length` and `$`-prefixed `pricing`, so any route without a bundled reference (the `syn:*` router aliases, newly added routes such as `hf:moonshotai/Kimi-K3`) resolved to `reasoning: false` — hiding the effort dial and dropping `reasoning_effort` from every request — plus text-only input, zero cost, and an 8k output cap low enough to end turns on `length` and trigger recovery compaction. Effort ladders now come from the per-model wire vocabulary, with the router's `none` tier mapped onto `minimal`; a route whose wire vocabulary contains no known tier reports non-reasoning rather than fabricating an unadvertised ladder, the wire vocabulary overrides stale bundled reference ladders, and a populated `supported_features` list without `tools` no longer downgrades routes a bundled reference already marked tool-capable. An advertised effort vocabulary is authoritative over the bundled reference's reasoning flag, a route with at least one named tier reasons even if that tier stands alone (only a `none`-only vocabulary is the off-switch), and a present-but-empty `supported_features` list now marks the route tool-less rather than falling back to the reference default. The production model-manager merge (`mergeDynamicModel`) now treats Synthetic's discovered `reasoning` as authoritative instead of OR-ing the stale bundled flag back over a wire-advertised off-state — matching the existing Copilot `input` precedent — so a `none`-only route stays non-reasoning through `createModelManager`, not just in the raw fetcher.
-- Fixed Cursor 1M-context model discovery to expose the 1M-token context window for models Cursor labels with a "1M" display name (Claude and GPT families), natively 1M families Cursor serves unlabeled (Kimi K3, GLM 5.2+), and max-mode Claude/Gemini variants, instead of pinning every discovered model to the 200k default. Bumped the Cursor model-cache namespace so windows cached before this fix are refetched. ([#4798](https://github.com/can1357/oh-my-pi/issues/4798))
-- Fixed GitHub Copilot routing `grok-4.5` through the unsupported Chat Completions endpoint instead of the Responses endpoint ([#7096](https://github.com/can1357/oh-my-pi/issues/7096)).
+- Fixed Synthetic models losing capabilities (such as reasoning/thinking selectors, vision input, output limits, and pricing) by correcting how the discovery mapper parses Synthetic's advertised features, effort vocabularies, and pricing structures.
+- Fixed Cursor model discovery to correctly expose the 1M-token context window for supported models (including Claude, GPT, Kimi K3, and GLM 5.2+ families) instead of defaulting to 200k.
+- Fixed GitHub Copilot routing for `grok-4.5` to use the correct Responses endpoint instead of the unsupported Chat Completions endpoint.
 
 ## [17.2.1] - 2026-07-30
 

packages/coding-agent/CHANGELOG.md

@@ -4,51 +4,48 @@
 
 ### Added
 
-- Added an `app.live.toggle` keybinding (default `Ctrl+L`) that starts or stops live voice mode, same as `/live`.
+- Added an app.live.toggle keybinding (default Ctrl+L) to start or stop live voice mode.
+- Added ctx.invokeTool(params, options?) to extension contexts, allowing wrappers to run native tools while inheriting context, abort signals, and progress updates.
 
 ### Changed
 
-- Moved the display-reset default chord (`app.display.reset`) from `Ctrl+L` to `Alt+L` to make room for the live-mode toggle.
-- Updated the hashline edit tool, streaming preview, and plan-mode guidance for the unified `PUT`/`CUT` grammar, `.=` ranges, and named registers.
-- Moved subagent model-registry refresh and session-file opening off the launch critical path. Registry refresh now runs in the background while model resolution uses the registry's static/provider lookup, and `SessionManager.open` starts before prewalk/output-schema/session-option assembly so session creation waits only at its actual input boundary.
-- Batched same-turn file-session appends behind the session writer's microtask boundary. Render and tool paths can now record several transcript entries without paying one synchronous file write per entry, while explicit `flush()` and `close()` still make queued lines durable in order.
-### Changed
-
-- Rewrote the Codex saved-reset auto-redeem algorithm to be pool-wide, window-exact, and expiry-aware. A new salvage sweep piggybacks on the 5-minute usage refresh and spends any account's saved reset that would otherwise expire within `codexResets.salvageHorizonHours` (default 12, `0` disables) while either chat window has meaningful usage to restore; the `keepCredits` reserve never holds back a dying credit. The blocked-turn path considers ALL stored accounts and builds eligibility from the exact exhausted 5h/weekly windows — a banked reset also clears a 5h-only block (openai/codex#28525) — with the natural unblock taken as the latest reset among them, gated by `codexResets.minBlockedMinutes`. Because the usage-report layer can serve a pre-block snapshot even after invalidation (in-flight fetch adoption, last-good-on-failure under `/wham/usage` IP throttling), the live 429's parsed unblock timestamp is authoritative for the active account, synthesizing the candidate when no usable report survives; per-account credit counts are overlaid from the dedicated live credits route on every blocked pass, since a stale `/wham/usage` zero is never corrected upstream. Non-terminal consume outcomes (`nothing_to_reset`, `credit_list_failed`, thrown transport failures) release the attempt key and defer the episode 30 minutes instead of permanently burying a still-banked credit. Previously the reactive-only, active-account-only, weekly-only check was unreachable in multi-account setups — sibling rotation always preempted it — so banked resets expired unused.
-### Fixed
+- Moved the display-reset default keybinding (app.display.reset) from Ctrl+L to Alt+L to accommodate the new live-mode toggle.
+- Updated the hashline edit tool, streaming preview, and plan-mode guidance to support the unified PUT/CUT grammar, .= ranges, and named registers.
+- Improved startup performance by moving subagent model-registry refresh and session-file opening off the launch critical path.
+- Optimized session file writing performance by batching same-turn file-session appends.
+- Rewrote the Codex saved-reset auto-redeem algorithm to be pool-wide, window-exact, and expiry-aware, ensuring banked resets are automatically and reliably redeemed across multi-account setups before they expire.
 
 ### Fixed
 
-- Fixed Kitty terminals crashing while rendering live or restored non-PNG tool-result images when the runtime throws synchronously during PNG conversion ([#7160](https://github.com/can1357/oh-my-pi/issues/7160)).
-- Fixed a subagent's eval `reset: true` wiping the shared kernel it inherits from its parent, destroying every co-owner's interpreter state mid-session. A reset from a non-exclusive owner now forks into a private per-owner kernel (sticky for that owner and reaped on its teardown) across the Python, JavaScript, Ruby, and Julia executors, while an exclusive owner still resets in place.
-- Fixed the copy selector and ask dialog rendering raw key IDs instead of human-readable keybinding labels ([#7164](https://github.com/can1357/oh-my-pi/issues/7164)).
-- Fixed CLI positional initial messages bypassing automatic session-title generation, which left shell-launched sessions unnamed until a later editor submission ([#7166](https://github.com/can1357/oh-my-pi/issues/7166)).
-- Fixed the environment-variable reference omitting the Kitty Unicode placeholder controls and tmux placement caveat ([#7172](https://github.com/can1357/oh-my-pi/issues/7172)).
-- Fixed `omp plugin install` failing extension validation for pi extensions that import `compact` from `@earendil-works/pi-coding-agent` (e.g. pi-claude-bridge) — the legacy `@oh-my-pi/pi-coding-agent` shim did not forward `compact` (it lives in `@oh-my-pi/pi-agent-core/compaction`, the same module as the already-bridged `estimateTokens`), so a named import threw Bun's static "Export named 'compact' not found" error. Added the missing `compact` re-export ([#7174](https://github.com/can1357/oh-my-pi/issues/7174)).
-- Fixed Bash interceptor rules matching only the complete command input, so an anchored rule such as `^\s*git\s+commit\b` missed a later command in a flat compound command. Rules now also inspect unquoted/unescaped `&&`, `||`, `;`, `|`, `&`, and newline-separated command fragments, including forms with leading environment-variable assignments, while retaining the original whole-input match and conservatively skipping complex shell syntax.
-- Fixed `ExtensionContext.cwd` staying pinned to the directory a session started in, even after it moved: `ExtensionRunner` cached `cwd` from its constructor argument instead of reading the owning session's live directory, so extensions watching `ctx.cwd` (e.g. for git-worktree tracking) kept observing a stale path for the rest of the session — reachable via the interactive `/move` command, a programmatic `AgentSession.moveSession()`/`SessionManager.moveTo()` call, or an SDK/ACP session opened with a `cwd` different from the process's own. `ExtensionRunner.cwd` is now a getter over `this.sessionManager.getCwd()`, the same session-scoped value `moveTo()` updates directly, instead of the process-global project directory.
-- Fixed the web-search provider picker describing xAI/Grok as requiring `XAI_API_KEY`, which hid that the `xai` search provider already accepts a SuperGrok/X Premium+ `xai-oauth` sign-in (via `/login xai-oauth` or `XAI_OAUTH_TOKEN`). The description now matches the OAuth-aware wording used by the Anthropic, OpenAI, and Gemini rows.
-- Fixed `/reload-plugins` not reconnecting MCP servers or refreshing the session's MCP tool and prompt-command registries despite listing MCP in its documented reload scope, so `.mcp.json` edits stayed inactive and removed prompt commands remained invocable until restart. The TUI reload pipeline now clears stale MCP prompt commands and runs the same disconnect/rediscover/`refreshMCPTools` path as `/mcp reload`, deriving discovery filters from settings so the reload keeps honoring `mcp.enableProjectConfig: false` instead of starting opted-out project servers ([#7189](https://github.com/can1357/oh-my-pi/issues/7189)).
-- Removed the `read` exemption from the centralized artifact spill wrapper. Oversized read results now persist to a recoverable session artifact and return the configured inline head/tail instead of bypassing the threshold shared by other tools.
-- Fixed DuckDuckGo web search under-returning requests above the first-page result count by submitting the returned continuation form until the requested limit is reached ([#7116](https://github.com/can1357/oh-my-pi/issues/7116)).
-- Fixed DuckDuckGo web search silently ignoring `after:`/`before:` date bounds: DuckDuckGo does not parse date operators, so the provider relies on the shared post-filter to enforce them, but `parseHtmlResults()` discarded the ISO timestamps DuckDuckGo now emits in each result row — every source was treated as undated and passed the filter. The parser now extracts those timestamps into `publishedDate`/`ageSeconds` so date bounds are honored ([#7115](https://github.com/can1357/oh-my-pi/issues/7115)).
-- Fixed env-driven OTLP trace export ignoring `OTEL_RESOURCE_ATTRIBUTES`: the exported resource only carried `service.name`, so backends that attribute spans via resource attributes never received them. The resource now merges the vendored `envDetector`, which parses `OTEL_RESOURCE_ATTRIBUTES` (percent-decoded, per spec) with `OTEL_SERVICE_NAME` taking precedence for `service.name` ([#7134](https://github.com/can1357/oh-my-pi/issues/7134)).
-- Fixed a fresh session with deferred MCP discovery injecting the newly mounted `xd://` tool catalog twice into the first model request: the post-discovery prompt rebuild already renders the full device catalog, but the pre-user `xdev-mount-notice` re-listed the same names because the rebuild never updated announcement tracking. The rebuild now reports the catalog it exposed and those devices are folded into the announced-mount baseline, so the notice is suppressed for anything the prompt already carries while non-rebuild mount changes and unmount notices still fire ([#7139](https://github.com/can1357/oh-my-pi/issues/7139)).
-- Fixed the bash tool failing every command with `EACCES: permission denied, open '/tmp/omp-shell-snapshots/snapshot-bash-<uuid>.sh'` on machines where more than one Unix account runs `omp`. The snapshot directory was a single fixed name under the shared `os.tmpdir()` created 0700, so the first account to run `omp` owned it and every other account's pre-create write was denied; the throw escaped `getOrCreateSnapshot` into `executeBash`, and nothing was cached, so the failure repeated for every command. The directory is now scoped per uid (`omp-shell-snapshots-<uid>`), and an unusable snapshot directory degrades to running without a snapshot instead of failing the command.
+- Fixed a crash in Kitty terminals when rendering non-PNG tool-result images if PNG conversion fails.
+- Fixed subagent evaluation resets (reset: true) wiping the shared kernel inherited from the parent session; resets from non-exclusive owners now fork into a private per-owner kernel.
+- Fixed the copy selector and ask dialog rendering raw key IDs instead of human-readable keybinding labels.
+- Fixed CLI positional initial messages bypassing automatic session-title generation.
+- Fixed the environment-variable reference omitting Kitty Unicode placeholder controls and tmux placement caveats.
+- Fixed extension validation failures during omp plugin install for extensions importing compact from @earendil-works/pi-coding-agent by adding the missing re-export.
+- Fixed Bash interceptor rules to inspect unquoted/unescaped compound command fragments (e.g., &&, ||, ;, |, &, and newlines) instead of only matching the complete command input.
+- Fixed ExtensionContext.cwd staying pinned to the initial session directory; it now dynamically tracks the active session's current working directory.
+- Fixed the web-search provider picker description for xAI/Grok to clarify that it supports SuperGrok/X Premium+ OAuth sign-ins.
+- Fixed /reload-plugins failing to reconnect MCP servers or refresh MCP tool and prompt-command registries.
+- Enforced the centralized artifact spill threshold on oversized read results, persisting them as recoverable session artifacts.
+- Fixed DuckDuckGo web search under-returning results above the first-page limit by automatically submitting continuation forms.
+- Fixed DuckDuckGo web search ignoring after: and before: date bounds by correctly parsing and filtering result timestamps.
+- Fixed env-driven OTLP trace export ignoring OTEL_RESOURCE_ATTRIBUTES.
+- Fixed a fresh session with deferred MCP discovery injecting the newly mounted xd:// tool catalog twice into the first model request.
+- Fixed the bash tool failing with EACCES permission errors on multi-user machines by scoping the snapshot directory per user ID.
 - Fixed LSP write batching replaying stale whole-file snapshots over newer external changes made before the batch flushed.
-- Fixed `ctx.ui.editor()` in ACP mode always resolving to `undefined`: it now routes through the same elicitation bridge as `select`/`confirm`/`input`, so `/review`'s custom-instructions prompt and the `ask` tool's custom-input path reach the ACP client instead of silently no-opping.
-- Fixed `omp commit` failing to resolve extension-provided models in both the agentic and legacy pipelines ([#7099](https://github.com/can1357/oh-my-pi/issues/7099)).
-- Fixed RPC hosts receiving no subagent lifecycle or progress frames when an IRC message revives an idle or parked keep-alive subagent ([#7105](https://github.com/can1357/oh-my-pi/issues/7105)).
-- Fixed copied fenced-code body rows in assistant messages retaining component and container margins ([#7055](https://github.com/can1357/oh-my-pi/pull/7055) by [@GratefulDave](https://github.com/GratefulDave)).
-- Fixed mid-turn auto-compaction repeating dead-end rescue work and warnings at every tool boundary within one oversized turn ([#7151](https://github.com/can1357/oh-my-pi/issues/7151)).
-- Fixed automatic terminal appearance changes clearing native scrollback and snapping readers away from their current scroll position; the active output grid now repaints non-destructively, while Ctrl+L remains the explicit full-history recolor.
-- Fixed exact-match edits failing on files containing credential-shaped tokens when `secrets.enabled` (Hide Secrets) is on: unconfigured API keys (GitHub/GitLab/OpenAI-shaped) in tool results now get reversible keyed placeholders from the secret obfuscator — restored byte-exact in tool-call arguments before execution — instead of pi-ai's irreversible `[*_token_redacted]` rewrite, so the model's `old_text` matches the real file bytes while credentials still never reach the provider ([#6968](https://github.com/can1357/oh-my-pi/issues/6968)).
-- Fixed context usage collapsing to the latest response size for Cursor models that omit prompt-token usage. Output-only assistant turns no longer become context anchors, so active and resumed sessions fall back to full local context reconstruction while compaction retains its conservative stored-context floor ([#7163](https://github.com/can1357/oh-my-pi/pull/7163) by [@harshav167](https://github.com/harshav167)).
-- Fixed the browser tool crashing OMP with an unhandled `EBUSY: resource busy or locked, rm '<TEMP>\puppeteer_dev_chrome_profile-<random>'` rejection when a headless Chromium profile was still locked during cleanup on Windows. `launchHeadlessBrowser` now owns the profile directory via an explicit `--user-data-dir` (disabling puppeteer's unretried temp cleanup) and removes it on dispose with lock-tolerant retry, warning and leaving the directory in place if it stays busy rather than crashing ([#7058](https://github.com/can1357/oh-my-pi/issues/7058)).
-- Fixed the Python RPC client dropping current context, compaction, OAuth URL, and terminal-settlement fields, and made additive notification variants observable without stopping the stdout reader.
-- Fixed the browser tool silently ignoring `url` when opening a new tab on an attached browser (`app.cdp_url` or `app.path`), so the tab now navigates on open exactly as it already did on reuse and in headless mode.
-- Fixed browser automation disrupting a browser it attached to over `app.cdp_url`: the tool now adopts the tab the user actually has in the foreground and no longer raises its own tab when taking a screenshot. Owned and headless browsers keep activating the target before capture.
-- Added `ctx.invokeTool(params, options?)` to a re-registered built-in's extension context, letting a wrapper run the native tool of the same name instead of reimplementing it, and inheriting the caller's context, abort signal, and progress updates.
+- Fixed ctx.ui.editor() in ACP mode always resolving to undefined by routing it through the elicitation bridge.
+- Fixed omp commit failing to resolve extension-provided models in both agentic and legacy pipelines.
+- Fixed RPC hosts receiving no subagent lifecycle or progress frames when an IRC message revives an idle or parked keep-alive subagent.
+- Fixed copied fenced-code body rows in assistant messages retaining component and container margins.
+- Fixed mid-turn auto-compaction repeating dead-end rescue work and warnings at every tool boundary within a single oversized turn.
+- Fixed automatic terminal appearance changes clearing native scrollback and snapping readers away from their current scroll position.
+- Fixed exact-match edits failing on files containing credential-shaped tokens when secrets.enabled is active by using reversible placeholders instead of irreversible redactions.
+- Fixed context usage collapsing to the latest response size for Cursor models that omit prompt-token usage.
+- Fixed the browser tool crashing with EBUSY errors on Windows when a headless Chromium profile is locked during cleanup.
+- Fixed the Python RPC client dropping context, compaction, OAuth URL, and terminal-settlement fields.
+- Fixed the browser tool ignoring the url parameter when opening a new tab on an attached browser.
+- Fixed browser automation disrupting attached browsers by adopting the active foreground tab and avoiding raising new tabs during screenshots.
 
 ## [17.2.1] - 2026-07-30
 

packages/collab-web/CHANGELOG.md

@@ -4,7 +4,7 @@
 
 ### Fixed
 
-- Fixed the guest rendering idle while the host agent is running (spinner vanishing mid-turn after a reconnect dropped `agent_start`): host `state` frames are now authoritative for liveness in both directions, and an idle report clears tool cards pinned by a dropped `tool_execution_end` ([#6990](https://github.com/can1357/oh-my-pi/pull/6990) by [@metaphorics](https://github.com/metaphorics)).
+- Fixed an issue where the guest UI could incorrectly appear idle (such as the loading spinner disappearing) while the host agent was still running after a reconnection, and ensured tool cards are properly cleared if a connection drop occurs.
 
 ## [17.2.0] - 2026-07-30
 

packages/hashline/CHANGELOG.md

@@ -4,21 +4,21 @@
 
 ### Breaking Changes
 
-- Replaced legacy `SWAP`, `INS`, and `PASTE` syntax with unified `PUT` and `CUT` hunks
+- Replaced legacy SWAP, INS, and PASTE syntax with unified PUT and CUT hunks
 
 ### Added
 
-- Added named register support (`@reg`) and span paste capabilities to clipboard operations
-- Added conservative recovery for uniformly omitted replacement indents that would otherwise escape an immediately preceding brace opener, while preserving intentional indentation-only edits
+- Added named register support (@reg) and span paste capabilities to clipboard operations
+- Added conservative recovery for uniformly omitted replacement indents near brace openers, preserving intentional indentation-only edits
 
 ### Changed
 
-- Made `.=` the canonical inclusive range separator while retaining legacy separator variants as lenient input
-- Unified replacement, insertion, register paste, block, head/tail, move, and removal headers under the composable `PUT`, `CUT`, `MV`, and `REM` grammar
+- Made .= the canonical inclusive range separator while retaining legacy separator variants as lenient input
+- Unified replacement, insertion, register paste, block, head/tail, move, and removal headers under a composable PUT, CUT, MV, and REM grammar
 
 ### Fixed
 
-- Recovered common model output mistakes including numbered read rows, summarized ranges, diff-style old/new rows, empty `PUT` deletes, harmless `CUT` colons, and single-line span shorthand without weakening overlap or stale-snapshot guards
+- Improved resilience against common model output formatting errors, including numbered read rows, summarized ranges, diff-style old/new rows, empty PUT deletes, harmless CUT colons, and single-line span shorthand
 
 ## [17.2.0] - 2026-07-30
 

packages/mnemopi/CHANGELOG.md

@@ -4,7 +4,7 @@
 
 ### Fixed
 
-- Fixed one-shot prepared statements never being released, which kept the SQLite connection alive after `close()`. Every `db.prepare(...)` in `beam/store.ts`, `annotations.ts`, `memory.ts`, `beam/consolidate.ts`, `beam/helpers.ts`, and the e6 migration now either goes through `db.run()` or is bound with `using`, so the statement is finalized when it leaves scope. On Windows this left the memory DB and its `-wal`/`-shm` sidecars locked, so the file could not be deleted, moved, or rotated after mnemopi closed it; on POSIX the handle leaked silently because open files can be unlinked.
+- Fixed a resource leak where SQLite prepared statements were not properly released, keeping the database connection alive after calling close(). This resolves file locking issues on Windows (which prevented deleting, moving, or rotating database files) and silent file handle leaks on POSIX systems.
 
 ## [17.0.8] - 2026-07-22
 

packages/natives/CHANGELOG.md

@@ -4,14 +4,14 @@
 
 ### Changed
 
-- Updated native HTML-to-Markdown rendering to html-to-markdown-rs 3.9.2 defaults; Markdown formatting can differ from 2.30.0, including fenced code blocks and cycling nested-list bullets.
+- Updated native HTML-to-Markdown rendering to html-to-markdown-rs 3.9.2 defaults, which may result in formatting differences (such as fenced code blocks and cycling nested-list bullets) compared to version 2.30.0.
 
 ### Fixed
 
-- Fixed `/live` corrupting the heap when opening PulseAudio on Linux ARM64 by shipping target-specific miniaudio Rust layouts for GNU and musl native addons ([#7138](https://github.com/can1357/oh-my-pi/pull/7138) by [@olegpulatov](https://github.com/olegpulatov)).
-- Fixed local Bazel addon builds on NixOS by exposing system CMake tools to sandboxed build scripts and forcing bundled Opus into the library directory expected by `audiopus_sys` ([#7136](https://github.com/can1357/oh-my-pi/pull/7136) by [@olegpulatov](https://github.com/olegpulatov)).
-- Fixed workspace native addon loads preferring an installed leaf package over the workspace build ([#7059](https://github.com/can1357/oh-my-pi/pull/7059) by [@GratefulDave](https://github.com/GratefulDave)).
-- Fixed pathological HTML inputs crashing the process; conversions that reach the native-stack DOM depth limit now reject instead of returning silently truncated Markdown.
+- Fixed a heap corruption crash when opening PulseAudio on Linux ARM64 by shipping target-specific miniaudio Rust layouts for GNU and musl native addons.
+- Fixed local Bazel addon builds on NixOS by exposing system CMake tools to sandboxed build scripts and correctly bundling Opus.
+- Fixed workspace native addon loading to correctly prefer the workspace build over an installed leaf package.
+- Fixed process crashes caused by pathological HTML inputs; conversions that exceed the native-stack DOM depth limit now reject instead of returning silently truncated Markdown.
 
 ## [17.2.1] - 2026-07-30
 

packages/tui/CHANGELOG.md

@@ -4,12 +4,12 @@
 
 ### Added
 
-- Added request tokens to explicit OSC 11 appearance refreshes so consumers can correlate responses across queued and coalesced terminal probes.
+- Added request tokens to explicit OSC 11 appearance refreshes to allow consumers to correlate responses across queued and coalesced terminal probes.
 
 ### Fixed
 
-- Fixed the event-loop watchdog reporting system sleep as a synchronous `ui.loop-blocked` stall. A long missed probe interval is now treated as process suspension and suppressed, while shorter real stalls continue to report with their phase attribution.
-- Fixed terminal copies of literal fenced-code body rows retaining component, list, or blockquote margins in coding-agent assistant messages ([#7055](https://github.com/can1357/oh-my-pi/pull/7055) by [@GratefulDave](https://github.com/GratefulDave)).
+- Fixed the event-loop watchdog incorrectly reporting system sleep or suspension as a synchronous ui.loop-blocked stall.
+- Fixed terminal copies of fenced-code blocks retaining margins from components, lists, or blockquotes in assistant messages (#7055 by @GratefulDave).
 
 ## [17.2.0] - 2026-07-30