Introduced the xd:// virtual device protocol for mounting tools as URLs readable/writable via read/write tools, configurable via the new tools.xdev setting (defaults to true). Added the hub tool, consolidating agent peer messaging, background job control, and supervised long-running processes. Added the edit.enforceSeenLines configuration setting (defaults to false) to optionally reject edits on lines that have not been fully displayed. Added the ToolLoadMode type and an optional satisfies predicate to SoftToolRequirement to support compliance checks against specific invocation shapes (such as writing to a virtual device path). Added `xd://` virtual tool devices (controlled by the `tools.xdev` setting, default on), allowing mounted tools to be discovered via `read xd://`, documented via `read xd://`, and executed via `write xd://`. Added the `edit.enforceSeenLines` setting (default off) to gate the hashline seen-line guard. When enabled, edits anchored on lines that a prior `read` or `grep` never displayed are rejected. Added per-agent prewalk for subagents, featuring a `prewalk` frontmatter field, a `task.agentPrewalk` settings override toggled from the `/agents` dashboard, and a `task.prewalk` boolean (default off) to arm the bundled generic `task` agent. Added `enforceSeenLines` option to `PatcherOptions` (defaulting to `true`) to control whether seen-line validation is enforced on anchored edits, allowing tags to validate on content hash alone when disabled. Improved LaTeX rendering for \underbrace, \overbrace, \overset, \underset, and \stackrel to use drawn horizontal braces with centered labels and stacked annotations instead of flat inline glyphs. Improved LaTeX rendering of multi-letter subscripts and superscripts by displaying them as raised or lowered blocks instead of ragged per-character Unicode glyphs. Added an opt-in Editor.setImeSafeCursorLayout() method to protect macOS IME preedit while retaining the compact bordered layout by default. Added GPT-5.6 Luna, Sol, and Terra entries for Amazon Bedrock, Azure, and Cloudflare Added KAT-Coder Air/Pro V2.5 entries across Kilo, OpenRouter, NanoGPT, and Vercel Added Inkling model entries for Baseten and Vercel AI Gateway Added Umans DeepSeek V4 Pro DSpark as an experimental model listing Added Claude Opus 4.7 Fast and 4.8 Fast on Vercel AI Gateway Added Workers AI GLM-5.2, Muse Spark 1.1, Stealth GPT-5.6 Sol, and nano-gpt-help entries Added native cmux notification delivery targeted to the current terminal surface. 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. Added a fullscreen overlay mouse-tracking opt-out to allow selection-first dialogs to preserve native terminal text selection. Added `Terminal.refreshAppearance()` to allow consumers to manually trigger a refresh of the detected dark/light terminal appearance without periodic polling. Added a structured log sink API (`registerLogSink`, `LogEvent`, `LogLevel`) to the centralized logger, enabling out-of-band consumers (such as OpenTelemetry) to receive log events without affecting local file or console logging. Added a per-message token estimation cache to optimize performance by reusing token counts for settled message history, with automatic cache invalidation on message mutation. Added an Anthropic compatibility flag to allow non-official OAuth endpoints to opt into configured Claude Code fingerprint header overrides. Added support for Codex (ChatGPT subscription) in `generate_image` via the `providers.image: "openai-codex"` option, including automatic subscription detection and fallback logic. Added an optional `provider` parameter to `generate_image` to override the global image provider setting for a single request. Added OpenTelemetry log and metric export capabilities alongside existing trace exports, supporting standard OTLP environment variables. Added support for id-prefixed targets and keys in `retry.fallbackChains` wildcards (e.g., `"openrouter/google/*"`). Added support for `Shift+Enter` in the session tree selector (`/tree`, `/branch`) to summarize and switch branches in a single step. Added the `PI_CONFIG_FILES` environment variable to load settings overlays before `--config` overlays. Added optional PTY start callbacks that report the spawned child PID before command completion. Added static fallback seed for Devin's `swe-1-7` model so it is bundled even when catalog generation runs without a Devin session token. Added native Warp CLI-agent events for rich session status, tool approvals, and completion notifications ([#5592](https://github.com/can1357/oh-my-pi/pull/5592) by [@metaphorics](https://github.com/metaphorics)). Added Codex (ChatGPT subscription) support to `generate_image`. The tool now resolves a connected `openai-codex` OAuth credential and drives OpenAI's hosted `image_generation` tool through the ChatGPT backend (`chatgpt.com/backend-api/codex/responses`, `chatgpt-account-id` header) **independent of the active chat model** — so image generation works on a ChatGPT/Codex subscription with no metered `OPENAI_API_KEY`, even when the active model is Claude/Gemini/etc. A new `providers.image: "openai-codex"` option forces it; `auto` now auto-detects a connected subscription (priority: active GPT image tool > Codex subscription > Antigravity > xAI > OpenRouter > Gemini), and the `openai` preference falls back to it when no `OPENAI_API_KEY`/active GPT model is present. Added an optional `provider` parameter to `generate_image` (`auto` | `openai` | `openai-codex` | `antigravity` | `xai` | `gemini` | `openrouter`) that overrides the `providers.image` setting **for a single request** — so "generate this using gemini / codex / xai" routes per-call without changing the global setting. Absent → the `providers.image` setting applies, unchanged; the named provider uses the same resolution semantics (falls back to auto-detect if it has no credentials). File: `tools/image-gen.ts` (`imageProviderSchema`, `findImageApiKey` `preference` arg). Added OpenTelemetry log and metric export alongside the existing trace export. When `OTEL_EXPORTER_OTLP_LOGS_ENDPOINT` (or the shared `OTEL_EXPORTER_OTLP_ENDPOINT`) is set, `omp` registers a `LoggerProvider` and forwards every centralized-logger event as an OTLP log record (severity + attributes + active span context for log↔trace correlation, min level via `OTEL_LOG_LEVEL`, plus a structured `agent run completed` summary event). When `OTEL_EXPORTER_OTLP_METRICS_ENDPOINT` (or the shared endpoint) is set, it registers a `MeterProvider` with a `PeriodicExportingMetricReader` and records GenAI-semconv `gen_ai.client.token.usage` plus `pi.omp.agent.*` counters/histograms (runs, steps, chat/tool calls by name+status+finish reason, latencies, estimated cost, errors) from the agent run summary and per-chat usage hooks. Each signal honors its own `OTEL_*_EXPORTER=none` kill switch, the global `OTEL_SDK_DISABLED`, and declines non-`http/protobuf` protocols independently ([#4604](https://github.com/can1357/oh-my-pi/issues/4604)). `retry.fallbackChains` wildcards now support id-prefixed targets and keys: a chain entry like `"openrouter/google/*"` re-prefixes the failing model's bare id (`google-antigravity/gemini-x` → `openrouter/google/gemini-x`), a plain `"provider/*"` entry falling back *from* an aggregator strips the vendor prefix when the target provider only knows the bare id (`openrouter/google/x` → `google-vertex/x`), and an id-prefixed key (`"openrouter/google/*"`) scopes a chain to that provider's ids under the prefix. The session tree selector (`/tree`, `/branch`) now supports Shift+Enter to summarize-and-switch in one step: it forks from the selected entry with a branch summary, with no extra prompt and regardless of `branchSummary.enabled`. Plain Enter keeps the current behavior (direct switch by default; the summary prompt only when `branchSummary.enabled` is on). ([#5152](https://github.com/can1357/oh-my-pi/issues/5152)) Added the turn's local timestamp (`YYYY-MM-DD HH:mm:ss`, down to the second) to the per-turn token-usage row shown under assistant messages when `display.showTokenUsage` is enabled. Added support for several new models across multiple providers, including MiniMax M3, Gemini 3.5 Flash Lite, Gemini 3.6 Flash (with thinking support), Hy3, Doubao-Seed-Character, LongCat 2.0, Laguna S 2.1 (free and paid tiers), Qwen 3.6 35B A3B, SWE-1.6 Slow (devin agent catalog), and XiaomiMiMo/MiMo-V2.5. Added a `/tree` re-answer option for past `ask` tool results, allowing users to re-open the picker with original questions and branch the new answer as a sibling while keeping the original branch reachable. Added configurable Hindsight client request deadlines via `hindsight.requestTimeoutMs`, `reflectTimeoutMs`, `recallTimeoutMs`, and `retainTimeoutMs` settings (and matching `HINDSIGHT_*_TIMEOUT_MS` environment variables). Added `omp-linux-musl-x64` and `omp-linux-musl-arm64` release binaries for Alpine and other musl-based Linux distributions, with automatic musl selection in the installer and self-updater. Added jsdiff-compatible native diff exports: `diffLines`, `diffWords`, `diffLineRuns`, and `structuredPatchHunks`. Added batch vector kernels for mnemopi recall paths: `cosineSimilarityPairs`, `vectorIndexTopK`, and `mmrRerankIndices`. Added Synthetic (synthetic.new) usage provider: `/usage` now reports the rolling 5-hour request limit and weekly credit quota via `GET /v2/quotas`, including per-tick regeneration rates in the window labels. Added optional `UsageWindow.resetLabel` so rolling windows can render their countdown with an accurate verb (e.g. "tick in 12m" / "regen in 51m" instead of "resets in") — both quota windows on Synthetic regenerate incrementally rather than hard-resetting. Added per-call `model` selection to the `task` tool, including per-item batch selectors, fallback chains, and explicit reasoning suffixes. Added Firecrawl keyless mode: explicitly selecting `firecrawl` as the web-search provider now works without `FIRECRAWL_API_KEY` by calling the Firecrawl REST API without an `Authorization` header; the automatic provider chain remains credential-gated (#4332). Added `mcp.renderMarkdownResults` (enabled by default): non-JSON MCP text results render as Markdown in the terminal transcript; set it to `false` to keep raw text. Added `SelectList.setMaxVisible()` so hosts can refit the visible row budget to available height after construction. Added support for tracking Cloudflare AI Gateway cache status (hit, miss, bypass, unknown) on chat spans. Added support for caller-owned `cachedContent` on Google Generative AI and Google Vertex AI `GenerateContent` options, allowing passing of opaque cache resource names. Added Anthropic extra-usage reporting across CLI, interactive, and ACP usage endpoints, normalizing the authoritative `spend` payload into a 'Claude Extra Usage' USD row with accurate limit, remaining, and status details. Added process-scoped OAuth account pools for trusted auth-broker clients via `OMP_AUTH_BROKER_ACCOUNT_POOL_FILE` to filter snapshots, updates, refreshes, and usage reports to selected OAuth identities. Added opt-in Vercel AI Gateway automatic prompt caching for OpenAI Chat Completions, including support for cache anchors and cache lifetimes. Added opt-in explicit prompt-cache controls for OpenAI GPT-5.6+ Responses and Chat Completions, supporting stable boundary selection, stateful Responses markers, and future GPT-5.x/6.x models. Added support for forwarding `statefulResponses` through `streamSimple` to allow diagnostic callers to explicitly disable OpenAI Responses `previous_response_id` chaining. Added native QwenCloud Token Plan support, including API-key login, model discovery, and an optional interactive console-Cookie prompt for quota reporting. Added interactive Meta Model API key login and support for `MODEL_API_KEY` and `META_API_KEY` environment variables. Added model-scoped usage health tracking and same-provider reselection for native coding-plan credential pools. Added Bedrock Converse prompt-cache compatibility limits, including 5-minute checkpoint support for Nova models (Lite, Micro, Pro, Premier, Nova 2 Lite) and 1-hour Claude retention. Added native Meta Model API provider and Muse Spark 1.1 support, featuring Responses API reasoning replay, image input, and reasoning-effort controls. Added Vercel AI Gateway integration features, including opt-in automatic prompt-cache compatibility, provider routing preferences, and Responses cache-anchor and cache-lifetime controls. Added prompt-cache breakpoint capability metadata for OpenAI GPT-5.6, with opt-in support for older models and compatible endpoints. Added native alibaba-token-plan provider with QwenCloud Token Plan Individual discovery and a curated chat-model fallback catalog. Added dynamic multi-root workspace context support, allowing users to manage multiple workspace directories mid-session via `/add-dir`, `/remove-dir`, and `/dirs` slash commands, or seed them at launch using the `--add-dir` CLI flag. Added `/live`, a Codex-authenticated real-time voice interface that streams microphone audio over WebRTC and routes coding tasks through the active agent session. Added opt-in usage-aware model fallback for rationed coding plans, including a `/usage` command to view live quantitative usage data and automatic fallback chain traversal. Added `error.notify` configuration to allow failed model turns to trigger distinct terminal or desktop notifications. Added auto-following light and dark themes to HTML session exports, with a `/export --themes` option to bundle selected TUI themes. Added owner-routed asynchronous job delivery, ensuring background bash and task results are injected directly into the owning subagent or agent session rather than the top-level session. Added background-on-steer capability for auto-backgrounded bash commands, allowing incoming user or peer messages to immediately background running commands. Added `friendlyName` support for hidden secrets, allowing model-visible placeholders to carry sanitized semantic labels, hashes, and case hints. Added support for Jujutsu (`jj`) repositories in the statusline `git` segment, displaying the nearest bookmark or change ID and retrieving working-copy change counts. Added `block` and `unblock` operations for tasks, introducing a `blocked` status for tasks waiting on external input to exclude them from incomplete-todo reminders. Added a toggle-list editor in `/settings` for managing array-of-enum settings like search and image provider orders. Added `models.yml` Bedrock Converse prompt-cache capability overrides for bundled and opaque inference profiles. Added `getServiceTiers()` and `setServiceTier()` extension APIs to read and modify the live per-family service tier for session requests. Added opt-in `omp bench --cache` for independent cold/warm prompt-cache benchmarking with stable-prefix controls. Added `tools.xdevDocs` prompt-doc modes and the `tools.xdevInlineDevices` glob allowlist to control which mounted device documentation is inlined into the system prompt. Added the opt-in `read.renderMarkdown` setting for formatted Markdown read previews. Added an `includeThinking` serialization option (defaulting to `true`) to allow excluding assistant reasoning (`¶think:` sections) from archived transcripts. Added Text.setStyleFn() to apply foreground stylers at render time, allowing components to dynamically re-resolve colors after invalidation instead of baking in the palette active at construction. Added the provider-neutral native computer-call lifecycle, preserving observation outputs and input actions across pending and acknowledged tool results. Added `setCodexAttestationProvider` API for injecting `x-oai-attestation` headers in ChatGPT-OAuth Codex requests Added OAuth account session pinning and active status tracking in storage Added OpenAI Responses native computer-use transport, including batched actions and exact `computer_call`/`computer_call_output` replay with pending/acknowledged safety checks and `image_url`/`file_id` output references. Models without native support receive the same action surface as a regular function tool; provider-specific tool-choice forcing is used where supported. Added `PI_CODEX_RESPONSES_LITE` to override the catalog-selected Codex Responses transport for diagnostics (`1`/`true` forces Lite; `0`/`false` forces the standard body). Added caller-owned `cachedContent` on `google-generative-ai` and `google-vertex` GenerateContent options: pass an opaque cache resource name through the shared builder (blank values rejected); no create/refresh/delete lifecycle and no guessed model/project/location validation; existing `cachedContentTokenCount` → `Usage.cacheRead` normalization is unchanged. Added Anthropic extra-usage reporting across `omp usage`, interactive `/usage`, and ACP `/usage`: the OAuth usage endpoint's authoritative `spend` payload (or legacy `extra_usage` fallback when absent) is normalized into a `Claude Extra Usage` USD row; capped accounts show limit/remaining/fractions and status, while uncapped spend exposes only its absolute used amount—rendered as `$… used` in CLI/TUI and `123.45 usd used` in ACP—without a fabricated cap, percentage, or status. ([#5575](https://github.com/can1357/oh-my-pi/issues/5575)) Added process-scoped OAuth account pools for trusted auth-broker clients via `OMP_AUTH_BROKER_ACCOUNT_POOL_FILE`, consistently filtering snapshots, streaming updates, refreshes, and usage reports to selected OAuth identities while leaving API-key credentials and the shared encrypted snapshot cache unrestricted. Added opt-in Vercel AI Gateway automatic prompt caching for OpenAI Chat Completions while preserving `only` and `order` routing preferences. Added Vercel AI Gateway Responses cache anchors and cache lifetimes, emitted only with automatic caching. Added opt-in OpenAI GPT-5.6 explicit prompt-cache controls for Responses and Chat Completions. Existing requests remain implicit; the policy marks at most one existing stable-history block and is rejected locally on unsupported explicit routes. Forwarded `statefulResponses` through `streamSimple`, so diagnostic callers can explicitly disable OpenAI Responses `previous_response_id` chaining. Added native QwenCloud Token Plan API-key login, model discovery, and an optional interactive console-Cookie prompt for 5-hour and 7-day quota reporting ([#6151](https://github.com/can1357/oh-my-pi/issues/6151)). Added model-scoped usage health and same-provider reselection for native coding-plan credential pools, preserving OAuth/login-pool precedence, scoped broker blocks, sibling rotation state, and conservative unknown-account handling while excluding ordinary configured API keys ([#5018](https://github.com/can1357/oh-my-pi/issues/5018)). Added catalog metadata for models that support native computer-use requests. Added resolved Bedrock Converse prompt-cache compatibility limits, including explicit 5-minute checkpoint support for bundled Nova Lite, Micro, Pro, Premier, and Nova 2 Lite models plus their documented in-region, regional, and global IDs, and model-specific 1-hour Claude retention. Added resolved Bedrock Converse prompt-cache compatibility limits, including explicit 5-minute checkpoint support for bundled Nova Lite, Micro, Pro, and Premier models plus Nova Premier's documented in-region model ID, and model-specific 1-hour Claude retention. Added the native Meta Model API provider and Muse Spark 1.1 with Responses API reasoning replay, image input, and the full supported reasoning-effort ladder ([#4941](https://github.com/can1357/oh-my-pi/issues/4941)). Added an opt-in Vercel AI Gateway automatic prompt-cache compatibility option alongside provider routing preferences. Added Vercel AI Gateway Responses cache-anchor and cache-lifetime compatibility controls. Added resolved OpenAI GPT-5.6 prompt-cache breakpoint capability metadata, keeping older models and compatible endpoints opt-in only. Added the native `alibaba-token-plan` provider with QwenCloud Token Plan Individual discovery and a curated chat-model fallback catalog ([#6151](https://github.com/can1357/oh-my-pi/issues/6151)). Added the `/session pin` subcommand and account picker to pin provider OAuth accounts for the current session Added the disabled-by-default `computer` essential tool with configurable enablement, backend, display, and maximum width/height settings. Native desktop execution runs through a `DesktopSession` worker; observation uses read approval, input uses exec approval, and provider checks always prompt and fail closed. Added the `/computer` slash command (`on`/`off`/`status`/toggle) to enable or disable the computer tool for the current session without persisting settings. Exposed `computer` to models without native OpenAI computer-use support as a regular function tool with a typed GA action schema; the same native desktop backend and approval policy apply on both paths. Hardened computer action ingress: action-specific fields, modifier/key arrays, coordinates, drag points, and scroll deltas fail closed before native input; numeric fields must be signed 32-bit integers and coordinates must be non-negative. Added native `AudioCapture`, `AudioPlayback`, and `LiveWebRtcPeer` classes for low-latency microphone capture, gapless speaker playback, and WebRTC offer/answer sessions with Opus media and `oai-events` data-channel delivery. Added a macOS `deviceCheckGenerateToken` export that generates Apple DeviceCheck attestation tokens natively: it drives `DCDevice.generateToken` through raw Objective-C runtime FFI with a hand-built completion block literal and a bounded one-second wait, resolving `{ supported, tokenBase64, error, latencyMs }` to mirror the ChatGPT desktop app's `devicecheck.node` addon contract. Non-macOS builds resolve `supported: false` without touching the network. Added a genuine native desktop backend for computer use, bundled in the core addon on every published platform: macOS Quartz/CGEvent, Windows Win32/`SendInput`, and a pure-Rust Linux X11 backend (`x11rb` capture over the display socket, XTest input with keysym mapping) that links no GUI system libraries — so Linux x64/arm64, glibc and musl are all supported and headless hosts are unaffected. Wayland sessions work through XWayland. Execute batches enforce a 60-second native deadline (`DESKTOP_DEADLINE_EXCEEDED`) and never emit input after it expires; unsupported pure-Wayland capture and out-of-XTest-range or negative-origin coordinate layouts fail closed. Added `resolveFallbackTool` option to allow routing unadvertised tool calls to host-side transports (e.g., device mounts) Added `GET /v1/usage/history` to the auth broker (recorded usage-limit snapshots with `sinceMs`/`provider` filters) and `AuthBrokerClient.fetchUsageHistory` — in broker deployments the broker host performs every upstream usage fetch, so its durable history is the only complete utilization record Added per-client burn tracking to the auth broker: clients batch observed request usage per (provider, model) and flush it to `POST /v1/usage/observed` every 10 seconds (install id as client key, hostname as display name); the broker persists 5-minute buckets in `client_usage`/`clients` and serves aggregates from `GET /v1/usage/clients`. Brokers without the endpoint disable reporting for the process lifetime Added in-process moreutils-style shell builtins to the bash tool's embedded shell: `ts` (timestamp lines; `-i`/`-s` elapsed modes, `-m` monotonic clock, `-r` relative rewriting of RFC3339/syslog timestamps, `%.S`/`%.s`/`%.T` subsecond extensions), `sponge` (soak stdin fully before atomically writing the target, so `foo file | ... | sponge file` works; `-a` appends), `ifne` (run a command only when stdin is non-empty; `-n` inverts and passes non-empty stdin through), `isutf8` (streaming UTF-8 validation with line/char/byte diagnostics; `-q`, `-l`, `-i`), `combine` (boolean `and`/`not`/`or`/`xor` on the lines of two files, `-` for stdin), and `errno` (errno name/number/description lookup with `-l` list and `-s` search; unix only). Like the uutils-backed builtins, they run in-process against the command's own stdio, resolve paths against the shell working directory, honor cancellation, and are disabled by `PI_DISABLE_UUTILS_BUILTINS`. The bash tool prompt now lists the available shell builtins (`mkdir` through `jq`, `rm`/`mv`/`ln`, and the moreutils set) so the model relies on them without existence checks; the line is dropped when `PI_DISABLE_UUTILS_BUILTINS` disables the builtins and omits unix-only `errno` on Windows. Sessions using a broker-backed auth store now report each completed request's token usage and cost to the auth broker (batched, 10s cadence) so the broker can track actual token burn per client install Added a per-spawn `effort` parameter to the `task` tool (`"lo"` | `"med"` | `"hi"`): each selector maps onto the resolved model's supported thinking range (lowest, middle, and highest level — whatever the model tops out at, e.g. `high`, `xhigh`, or `max`) and overrides the agent's default selector, including `auto`. Omitting `effort` keeps the existing automatic per-prompt thinking classification. Added `searxng.engines` setting for the SearXNG web search provider: a comma-separated list of engine names or bang shortcuts (e.g. `ddg, br, startpage`) sent as the API's `engines=` parameter. Shortcuts are resolved to canonical engine names via the instance's `/config` endpoint (cached per endpoint; entries pass through verbatim if `/config` is unreachable). Bang syntax in queries (`!ddg foo`) continues to pass through to the instance, and external bangs (`!!g`) are now stripped client-side since SearXNG answers them with an HTTP redirect even for JSON requests. Web search queries now understand Google-style directives on every provider: `site:`/`-site:` (plus `domain:`/`host:` aliases), `after:`/`before:`/`since:`/`until:` date bounds, `inurl:`/`intitle:`/`intext:`/`allin*:`, `filetype:`/`ext:`, `lang:`, quoted phrases (including smart quotes), `+term`, `-`/`NOT` exclusions, and `OR`/`|` groups. A shared parser (`web/search/query.ts`) structures the query once per request; each provider maps constraints onto native API filters where the upstream supports them (Perplexity domain/date/language filters on both the API-key and ask paths, Tavily `include_domains`/`exclude_domains` + `start_date`/`end_date`, Exa domain lists + published-date bounds on API and MCP paths, Anthropic `allowed_domains`/`blocked_domains`, xAI `filters.allowed_domains`/`excluded_domains`, Parallel `source_policy`, Brave absolute `freshness` ranges, Firecrawl `tbs=cdr` date ranges, Jina `X-Site`, SearXNG `language`) and otherwise re-emits only the operator syntax its engine parses (full Google syntax for Gemini grounding, OpenAI, Kagi, and the credential-free scrapers — with scraper-hostile path-`site:`/`inurl:` operators demoted to plain keywords; conservative subsets for DuckDuckGo, Mojeek, Kimi, Z.AI, TinyFish, and Synthetic). The pipeline then applies a lenient post-filter to every response: constraints the engine ignored are enforced on the returned sources, and any constraint dimension that would eliminate every result is relaxed and reported to the model (`Note: no results matched \`site:...\`; the constraint was relaxed`) instead of returning nothing. Directive-free queries are passed through byte-identical everywhere. Eval and browser `run` code previews are now prettified for display: minified/compact agent-written cells are expanded by conservative streaming formatters (Python, JavaScript, Ruby, Julia) that indent block structure, split statements, normalize operator spacing (JS), and expand object literals wider than 100 columns onto one property per line — purely for rendering, the executed source is untouched. The formatters handle in-flight prefixes (unterminated strings/comments/blocks) without inventing closers, and layout decisions are prefix-stable so already-rendered lines never reflow while a call streams. Added a Providers dashboard section: per-provider totals, stacked token/cost burn over time, peak-burn-hours histogram, subscription-window insights (windows burned, estimated tokens per window, peak concurrent utilization, ideal account count, exhaustion events), and latest window utilization per account — window analytics read the auth broker's `/v1/usage/history` when a broker is configured (falling back to the local agent DB), since broker deployments record usage history on the broker host MiniMax Token Plan accounts now report quota in `omp usage`. `GET /v1/token_plan/remains` returns one bucket per plan quota, each carrying a rolling interval window and a weekly window, so `minimax-code` surfaces real remaining percentages instead of an empty report. A model the plan does not include comes back looking like an untouched quota; those buckets are dropped from the report and named in its metadata. The mainland id `minimax-code-cn` is untouched. OAuth logins now stamp `authorizedAt` (epoch ms of the interactive login) on the stored credential, and every refresh-persist path preserves it. Anthropic expires the whole OAuth grant family ~30 days after authorization regardless of refresh-token rotation (observed as `invalid_grant: "Refresh token expired"` on the latest rotated token, exactly 30 days after login, across four production accounts), so the login anchor is what makes re-login deadlines computable. Exported `ANTHROPIC_OAUTH_GRANT_TTL_MS` alongside the anthropic OAuth flow. Added `GET /v1/credentials/disabled` to the auth broker and `AuthBrokerClient.listDisabledCredentials`: disabled-credential tombstones (`DisabledCredentialSummary` — identity, verbatim disable cause, disable timestamp; never token material) so auto-disabled accounts stay visible to clients instead of silently vanishing from the snapshot. `AuthStorage.listDisabledCredentials` serves the same data locally from SQLite; clients of brokers predating the endpoint get an empty list (404 mapped, no error). Added `AuthStorage.revalidateCredentials()` and the optional `AuthCredentialStore.refreshSnapshot` hook: remote broker stores re-fetch `GET /v1/snapshot` on demand so callers pairing live per-credential data with stored identities (`omp usage`) never render against the up-to-an-hour-stale disk-cached snapshot; local SQLite stores are always current and only reload. Added an optional per-request `codexSseMaxAttempts` stream option to bound Codex SSE pre-response retries while preserving the six-attempt default when omitted. Fixed Cursor requests failing with `Connect error internal: Unable to parse image: ...` whenever the session history contained an image: `rootPromptMessagesJson` image parts now embed a `data:;base64,` URI instead of bare base64, matching the convention used by the OpenAI-completions provider ([#6564](https://github.com/can1357/oh-my-pi/pull/6564)). Added Claude Opus 5 model entries for Amazon Bedrock: `anthropic.claude-opus-5` plus its `us.`, `eu.`, `au.`, and `global.` regional/geo IDs. `omp usage` now surfaces auto-disabled credentials as red `✗` tombstone rows (identity, how long ago, the shortened upstream cause — e.g. `Refresh token expired` — and a re-login hint), including a provider section when no active credential remains. User-driven tombstones (`replaced by newer credential`, `deleted by user`) and API-key rows stay hidden. Requires a broker with `GET /v1/credentials/disabled`; older brokers degrade to no tombstone rows. `omp usage` warns about Anthropic's ~30-day OAuth grant lifetime: accounts whose interactive login (`authorizedAt`) is within a week of the deadline get a yellow `⚠ re-login within