Luigit
repositories / pi-ext

pi-ext

bugabingas pi extensions

owned by admin

CHANGELOG.md

Raw
Rendered preview

Changelog

Unreleased

  • Label Pi's working loader with live Angel/Devil progress, such as Devil · Investigating with grep, for tool calls and automatic consultations; restore the default label when they finish.

  • Treat each persisted session as its own usage ledger: PiPi no longer reports persisted child usage on its tool result or writes pipi-usage-receipt entries, and Usage and Cost drop all PiPi-specific accounting. Parent footer and /session totals now exclude delegated work; PiPi results still report child usage. History recorded before this change counts PiPi child work twice in Usage and Cost.

  • Usage and Cost scan the same session stores as pi --resume: every project of Pi's default store, or only the configured flat sessionDir, instead of everything beside it. SessionManager.listAll itself took about 14 s and 465 MB for 5,368 sessions, so both enumerate the stores directly.

  • Name new and forked PiPi child sessions PiPi · <agent, tier, or model> · <prompt> so pi --resume shows them recognisably.

  • Remove Handoff, Open, and uv, their generated packages, workspace registrations, and Handoff demo verifier. Preserve historical records.

  • Remove Gratis, its generated package, workspace registrations, and Memory live-task dependency. Preserve historical records and generic virtual-provider support; Memory live tasks now use Pi's configured model or an explicit --model.

  • List git_clone_safe in codemode with its guidance instead of deferring it. Deferred exposure hid "Use git_clone_safe instead of raw git clone" from the model since 2026-10-12, and agents fell back to raw clones.

  • Bound every pipi call with a wall-clock deadline (pipi.timeoutMs setting or PIPI_TIMEOUT_MS, default 4 hours). An expired deadline aborts the child, preserves its resumable session, and bounds child teardown at 5 seconds. Per-call timeoutMs overrides the setting; null opts out.

  • Remove Relop's runtime, build/settings registrations, and workspace dependency; remove remaining Relop/Ultra build outputs. Preserve historical specifications, research, session records, and user settings; ordinary PiPi delegation remains available while the dedicated peer-review skill is planned.

  • Modernize against Pi 1.5.1 while preserving minimum Pi 1.4.0: refresh workspace Pi dependencies, honor saved distrust in context measurements, resolve slash-containing provider IDs exactly, and distinguish requested versus served models in accounting. Use canonical edited context for handoffs, report focused custom dialogs as waiting for input, accept CR-only MCP SSE, and enforce RTK/Git-probe output bounds independently of host Pi version. Settle fixture tools before shutdown and launch generated browser helpers outside project cwd for deep WSL paths. Retain specialized extension workflows; idle cache-warming footer refresh remains unchanged because no public extension completion event exists.

  • Deliver Ephemlet choice and text answers usefully: steer a working agent before its next model call instead of queueing them for the next user prompt, and while idle hold them in the extension so only the latest revision joins the next run as one message. Inline text answers and the editor now wrap inside themed field rows instead of truncating or scrolling on the terminal background; the end-of-text cursor survives row breaks.

  • Make the Activity widget an attention surface: show active work plus unacknowledged failures, hide finished successes, and disappear when nothing needs attention. Acknowledge failures by closing their inspector or with d/D in the /activities picker; hidden records stay inspectable and agent-readable. Replace the static inspect: /activities hint with a padded Nerd Font header icon coloured by state, and align rows under the header text.

  • Give Relop native compact/expanded tool views with grouped findings, inherited theme backgrounds, safe actionable agent text, explicit finding pagination, ordinary status retrieval of complete finding fields/limitations, truthful immutable-scope recovery, safe domain-error explanations, and actual scope/selection/round/snapshot-read progress. Verify registered tool behavior, both native themes and remapped expand bindings, terminal-control injection, background inheritance including ellipsis, native-container inspection, and regular-file-to-FIFO acquisition races without live model spend.

  • Tier MiniMax M Plan spending: image generation, understanding, speech, and MiniMax search run without dialogs; H3 submission asks once or for the session unless user/env minimax.videoApproval is always. Replace minimax.allowCreditsOverflow and web.minimax.mPlanCreditsApproved without compatibility shims; Web's automatic chain includes MiniMax again for valid M Plan keys. Render MiniMax tools like built-in/web rows, send concise model text, and return structuredContent for codemode. Report remaining shared M Plan allowance after consuming operations and emit quota:provider-activity, so Quota shows MiniMax beside another active chat provider. Harden MiniMax quota parsing for reported window lengths, unlimited and not-in-plan pools, weekly boosts, and ambiguous count direction.

  • Implement MiniMax M Plan capabilities (PX-SPEC-1RR4LU8M): preserve built-in chat while adding native image generation, explicit image understanding, speech artifacts, local capability discovery, and session-owned resumable H3 jobs. Require human per-operation consent or user/env spending authorization, reject PAYG/ambiguous or mismatched regional credentials, prevent automatic consuming retries, and keep unverified conditional APIs unavailable. Bound HTTPS artifact downloads with public-address pinning, redirect rejection, signed-URL redaction, and exclusive output creation; preserve partial image estimates and distinguish cancellation from terminal-record deletion. Add explicit MiniMax-only Web search without fallback, replace arbitrary host overrides with actionable removal guidance, and retire unverified CN quota routing instead of retaining legacy endpoints. Include native media usage in cost/history surfaces and persist PiPi contribution receipts so mixed codemode/media/child aggregates retain media estimates without recounting child-session spending. Verify relocated MiniMax execution against installed Pi 1.4.0 and 1.5.0 with deterministic HTTP; live extension acceptance, account tier, Credits attribution, and conditional capabilities remain unverified.

  • Use the verified current international MiniMax M Plan quota endpoint without legacy-route fallback. Preserve reported percentages when absolute counts are zero placeholders, and treat the general pool as account-wide so the most-consumed shared window drives the footer. Verify API-key credential resolution, reset visibility, truthful unknown values, and preserved measured/scoped allowances through parser and real Pi command regressions.

  • Repair Windows suite fixtures: isolate home discovery, close fixture-owned SQLite handles, use portable path assertions, and match Canvas startup diagnostics to the selected browser executable. Exercise native LSP navigation through a supported TypeScript launcher, preserve Go peer coverage and POSIX escalation checks, and verify bounded Windows process exit without assuming POSIX signal reporting. Separate settings-precedence cases from short-deadline timing and pass release-version input through Mise's parsed environment value instead of POSIX shell quoting; verify literal POSIX/cmd injection payloads leave package metadata untouched.

  • Implement Relop Independent Review Loop (PX-SPEC-R8L4Q2NV): agent-callable scoped snapshot review through isolated parent-linked Pi children, ordered independent reviewer pools, free-only OpenRouter, structured stable findings, explicit readiness limits, and caller-owned fixes/delivery. Persist settings, deadlines, consumed invocation reservations, findings, and substitutions across fixes, waits, retries, tree navigation, and reload; abort active work on cancellation, deadline, and shutdown. Register standalone codegen/settings/debug, Mise tasks, package relocation, README configuration, and bounded-context measurement; verify alias equivalence, adversarial scoped tools, artifact races/FIFOs, persistent crash reservations, and terminal outcomes. Reject short regular-file acquisitions without claiming prefix coverage; immediately exhaust nonclean final-round results while retaining failure causes, coverage limitations, and outstanding findings, without overriding clean reviews or cancellation. Recover persisted loops in one pass without intermediate history arrays or spread arguments; preserve latest matching state, maximum consumed rounds across branches, and snapshot isolation.

  • Remove Ultra, its workflow runtime, bundled authoring guidance, tasks, and build artifacts. Retain PiPi as the child-session delegation primitive without migrating Ultra's workflow engine.

  • Report Devdocs refresh-slot saturation as a transient busy source reason instead of silently skipping uncached local documentation. Retain the two-job bound and verify immediate local recovery after both remote refreshes complete, without adding an artificial retry cooldown.

  • Preserve typed upstream HTTP failures and share status/Retry-After decoding with fixed-provider networking so ordinary retrieval honors server delays without exposing response bodies. Verify one-hour 429 delta-seconds and 503 HTTP-date deferrals, exact-deadline resumption, short-delay minimums, and missing/invalid-header fallback against the real adapter and SQLite-backed engine.

  • Reject Devdocs retrievals whose version differs from the discovered cache identity, including compatible-but-different releases and lost version evidence; preserve prior content/refs and release refresh leases on rejection. Add explicit/implicit-version drift regressions, document the existing ten-hit search cap in the tool and README, and verify narrower queries can reach omitted retained sections.

  • Reduce Devdocs cached-read allocations and repeated JSON serialization with an array-free Unicode counter, a safe penultimate-prefix lower bound, and no duplicate partial-content join. Preserve original budget/continuation decisions; verify 1,683 differential boundary/error/Unicode cases and add 1/2/20-ref aggregate-boundary regressions. Add //extensions/devdocs:bench using disposable SQLite caches and optional native CPU profiles. On Linux/Node 26.13.0, five alternating processes per revision with ten warmups and thirty timed reads each reduce 20-ref ASCII/Unicode medians from 32.84/65.09 ms to 6.98/12.32 ms; paired bootstrap 95% speedup intervals are 4.68–4.75×/5.24–5.34× after within-process IQR filtering, with similar unfiltered gains.

  • Implement Devdocs Technical Documentation (PX-SPEC-D7V3KQ9M): add native Pi search/read/status/refresh with version evidence, provenance, serialized-output budgets, continuation refs, and a portable personal database-path override. Isolate SQLite/FTS work in an extension-owned native child, publish revisions atomically, coordinate refresh leases, retain offline discovery, enforce source/context namespaces, and preserve prior content across failed updates. Add replaceable local, DevDocs, upstream, and Context7 adapters, platform-neutral source catalog, registered files/trees/HTTPS sources, bounded static and archive readers, and index-to-body resolution. Verify real Pi settings/trust, query-isolated persistent caches, corruption recovery, capacity and lease fencing, live anonymous source retrieval/offline restarts, and relocated package/worker execution.

  • Share LSP location/symbol byte-budget admission and count each item's encoded JSON once instead of serializing every growing result prefix. Preserve exact UTF-8 envelope/comma costs, result order, truncation flags and final output validation; cover exact-boundary admission with Unicode/escaping and both initial truncation states.

  • Complete Agent LSP Navigation (PX-SPEC-LSP4N8Q2, PX-PLAN-LSPCP6R4): add types, implementations, document/workspace symbols, and one-hop incoming/outgoing calls with explicit ambiguous-target selection and stale-source rejection. Curate Go, TypeScript and C++ profiles through an explicit pinned-Helix refresh with provenance, licensed source, reviewed overrides, conflicts, atomic publication and last-good preservation; runtime never fetches catalogs or installs servers. Route TypeScript semantic requests through the semantic server rather than its partial-semantic startup fallback; require an external SDK containing tsserver.js. Resolve supported Windows npm shims through validated adjacent metadata and direct Node launch without a shell; retain unknown server completeness/freshness, bounded results and owned-process cleanup.

  • Retry Web fetches refused with 403, 429, or 5xx through Z.AI's web reader when a Z.AI key exists, for Markdown and text formats only. Gate the fallback on a status-tagged refusal so safety-rejected, oversized, aborted, and redirect-failed URLs never reach Z.AI; share the fetch deadline, enforce maxBytes on reader output, report both failures, and mark results with source: "zai". Share Z.AI key resolution, MCP error mapping, and nested-JSON decoding between search and reader; search order is unchanged.

  • Declare Loop's loop_begin and loop_control tools model-only so codemode scripts cannot start, pause, or resume runs; the model still calls them directly.

  • Roll generated extension debug logs at 2 MiB without dropping the triggering event; retain the newest three segments plus protected active writers, using closed-file markers and conservative PID checks across concurrent sessions. Preserve private JSONL files, schema-bound metadata, nonfatal diagnostics, and standalone extension packaging without handwritten extension changes.

  • Add LSP MVP for bounded Go definition, references, and hover using installed gopls, typed protocol connections, reviewed pinned Helix provenance, trusted timeout settings, disk synchronization, explicit uncertainty, and owned-process cleanup. Verify native Pi tool use against relocated built output and gopls 0.24, including external managed/unopened dependency edits; keep registry refresh and remaining operations deferred. Reduce evidence-path allocation and I/O by sharing bounded line extraction and reusing one last-readable source snapshot within each normalization, never across queries. Add reproducible profiled first/late-line benchmarks and regressions for duplicate freshness, mixed files, CRLF/Unicode boundaries, and cancellation during repeated hits. Verify actual Pi TUI/model/gopls tool traces through external edits, Unicode correction, truncation, expected failures, reload, and shutdown.

  • Move Activity status above Pi's editor using a native non-focusable widget with bounded task rows, active-work prioritization, failure and overflow counts, and compact retained history. Suppress output-only redraws, preserve RPC exclusion and session cleanup, sanitize title controls, and cover rendering widths plus extension-level debouncing and unchanged-count state transitions. Add selected-task stopping in the picker and non-terminal inspectors with truthful pending/error feedback; preserve attached-terminal input and reserve narrow-row space for kind/exit metadata while retaining both title ends. Keep retained failures from hiding active jobs, expose a terminal-safe stop path, and show process/terminal elapsed time that freezes at the final event. Preserve unexpected process termination signals in failure details, distinguish hidden-row counts from state totals, mark clipped non-terminal output, and use readable native text roles for essential metadata and actions. Reserve a separate failure-attention row without evicting three active jobs, mark clipped captions and omitted output tails, and keep terminal-safe stop guidance readable at narrow widths.

  • Bundle System's browser module graph into one self-contained entry through an extension-local build opt-out, eliminating lazy Mermaid fetches during navigation without lifecycle queues or delays. Preserve strict rendering-error reporting; cover initialized lifecycle transitions, entry-load failures, cancellation, diagram families, and unchanged lazy behavior in other build targets.

  • Scope System browser-health cancellation handling to pre-navigation requests and distinct confirmed replacement documents; preserve active errors, failed navigations, ambiguous contexts, and original health assertions, with native browser and old-document lifecycle regression coverage.

  • Retain failure-only Canvas/System browser diagnostics in CI logs: bounded request/navigation timelines, allowlisted resource paths, range status, and numeric media state; exclude headers, bodies, query strings, DOM, and screenshots while preserving original assertions and failure status.

  • Surface bounded, scrubbed Canvas Chromium startup exit evidence, including executable, effective UID, exit code/signal, and stderr head/tail, without weakening sandbox flags or changing launch behavior; preserve initial explanations and final markers across multi-chunk output exceeding the rolling buffer, with credential/profile redaction.

  • Revalidate npm registry metadata when baking exact-version external dependencies so a stale shared cache cannot reject newly locked versions; retain exact pins, installation safeguards, and a real cached-registry regression fixture.

  • Repair clean-checkout CI tests: create release archive fixtures under OS temp rather than assuming .pi/tmp exists, and wait for Good-job processing-claim cleanup before asserting completed queue state while retaining runtime-only authentication assertions.

  • Add Mockup multipage flows, keyboard navigation, page-scoped annotations, and capped headless previews; ship declared Playwright runtime files, expose browser provisioning through Mise, block preview fragment networking, preserve cancellation, avoid cache filename collisions, and correct scrolled pin geometry and stale note bubbles.

  • Add Keysmash prompt-editor list/task/quote continuation, conditional nesting, indentation, and bracket pairing; preserve native completion, app bindings, paste streams, fenced code, and foreign editor ownership, with stale-completion and asynchronous redraw protection.

  • Return Ephemlet waits immediately for already-answered requested fields while preserving selected-field filtering; cover an answered target alongside an unrelated open question, and remove unnecessary list-before-update guidance.

  • Refine PiPi helper routing and reusable-agent lifecycle guidance: inspect preset purpose and authority, keep task overrides local, validate authorized adaptation on fresh calls and continuations, and preserve one-off isolation; add concrete native-result caching, text/provider-error projection, and recovery without regeneration, plus explicit model-family selection and neutral role instructions for review panels.

  • Add the compact PiPi Round Table skill for independent, budgeted multi-model judging with anchored criteria, primary-evidence checks, explicit uncertainty, and preserved dissent; include optional research limitations and refine PiPi workflow handoff, recovery, and parent-verification guidance without changing native result contracts.

  • Clarify Activity agent guidance: requested follow-ups for long jobs use notifications, bounded waits serve immediate dependencies and redraws, monitoring prefers suitable events, and owned-work cleanup uses stop plus final-status verification; preserve silent defaults and permission boundaries, and verify guidance delivery through inline and discoverable codemode paths.

  • Isolate release snapshot filesystem-mock tests so shared-worker module caching cannot bypass symlink-read and stage-cleanup observations; retain all security assertions.

  • Prefer host-injected renderer keybinding hints, including lazy, live-preview, error, and shared-helper paths, while retaining Pi 1.4 fallbacks; inline question, consent, picker, and inspection flows now declare native waiting status.

  • Preserve provider-billed cost provenance through PiPi, Ultra, Cost, and Usage, distinguish estimated or mixed totals, honor billed zeroes, and avoid labeling empty aggregates as billed; retain catalog-priced model evaluation.

  • Add explicit /prompt-autopsy export for native historical HTML with prompts and tool updates, isolated from caller settings in private temporary storage; retain the current-prompt editor command and shortcut.

  • Package optional root consumer RELEASE_NOTES.md for one native collection update notice; disable fallback changelog lookup when absent, without publishing internal engineering history or creating extension-local changelogs.

  • Bound RTK version/rewrite and Loop Git-ignore helper output, reject truncated results, and reconcile Ephemlets demo reading pauses with the six-second contract without weakening assertions.

  • Track latest dependency releases with wildcard ranges and reproducible locks; refresh Pi to 1.5.0, stable TypeScript to 7.0.2, root development TypeBox to 1.4.4, Mermaid converter to 3.6.0, xterm headless to 6.0.0, and PTY to 1.2.0-beta.18. Keep Canvas React runtime and type packages within major 18 because Excalidraw's pinned Radix subtree excludes React 19; record remaining constraints in DEPENDENCIES.md.

  • Replace only typed-rest-client 2.3.1 with implementation-identical 2.3.0 through a version-scoped override, permitting patched qs 6.17.1 within Stryker's supported dependency range without accepting audit's incompatible Excalidraw and remark-math downgrades.

  • Adapt enum inference and renderer test contexts to Pi 1.5, preserve local debug export types under stable TypeScript's isolated declarations, and remove the converter's obsolete startOnLoad option. Update package-discovery expectations for Pi 1.5 and count cached prompt input in Dilemma's faux-provider accounting test.

  • Autopsy adopts the accepted transcript-first Nugu browser layout: compact scope totals, hidden Filters/Insights/Session panels, theme selection, and no primary-view helper-label clutter. Replace corpus-wide preview and 10,000-entry omission caps with bounded server pages, stable ID cursors, bidirectional infinite scrolling, viewport-adaptive row windowing, retained active readers, and direct graph-to-entry lookup. Search complete readable text and JSON; aggregate cheap accounting over the entire matching history and load detailed metric streams and runtime captures only on demand. Generate extension-owned browser CSS from the canonical Nugu source; verify real 300-entry relocated history through an initially underfilled viewport, page pruning/recovery, complete-on-click content, unchanged scope totals, and light/dark desktop/mobile screenshots.

  • Autopsy now shows bounded inline previews and streams complete stored text and JSON on entry/disclosure open, with bounded chunks, backpressure, explicit completion/error events, lazy original-index image references, retry, and cancellation. Track actual preview clipping instead of scanning content for truncation-marker strings; remove per-entry warning spam and present the session-history caveat as informational text. Reuse visible rows without discarding open content, scroll, or selection; update derived measurements separately and block queued disclosure events from starting detached-row requests. Verify large text/signatures, deep/wide JSON, Unicode boundaries, HTTP disconnects, interrupted streams, stale-toggle disposal, changed derived metrics, and relocated Chromium flows.

  • Add Autopsy's usage dashboard with stored/current-catalog cost comparisons, token/category breakdowns, measured response/tool durations, original text/image sizes, provider/model/date filters, and selectable actual-time graphs with keyboard-accessible cross-page entry drill-down. Preserve absent quantities as unknown with subtotal coverage, separate timestamp spans/gaps from latency, apply request-wide price tiers and one-hour cache-write rates, and keep original image counts/previews independent of truncated JSON previews. Bound metric scans and chart sampling; verify nullable usage, pricing thresholds, size limits, shared filters, graph selection, desktop/mobile layouts, and relocated Chromium workflows.

  • Highlight Autopsy's expanded entry JSON with theme-aware colors for keys, strings, numbers, and boolean/null literals while preserving exact text and treating all source content as inert text.

  • Autopsy renders session image blocks through authenticated on-demand local image routes, with byte and declared-dimension limits, native-size toggles, and preview teardown on collapse. Preserve provider signatures and signature-named tool fields in bounded entry JSON; collapse JSON/provider metadata by default instead of stripping it. Verify image delivery, deferred browser loading, native-size inspection, collapse, and oversized rejection through HTTP and the relocated-package Chromium workflow.

  • Rename prompt-autopsy to autopsy: extension directory, package identity, slash command /autopsy, browser title, Mise tasks, and debug environment namespace (PI_AUTOPSY_DEBUG). Keep Ctrl+Alt+A; do not retain the old command alias.

  • Supersede prompt-autopsy's external-editor workflow and native historical HTML export with a token-protected loopback browser inspector for the current Pi session: search, role/type/branch filters, reported-usage bars and totals, readable entries, sanitized JSON, and exact/grouped system-prompt captures. Preserve the post-before-agent-start capture boundary, retain bounded runtime-local prompt history with explicit expiry/truncation notices, separate session history from provider request context, and stop the inspector on shutdown or reload. Add real-Pi capture-to-endpoint coverage and a relocated-package Chromium browser check; remove superseded editor-handoff automation.

  • Deepen the loop vixen's animation craft with classic principles: anticipation before pose changes, arced gestures with follow-through, breathing with chest rise, counterbalanced hip-bust sway, heavier hair inertia, grounded foot contact, and soft elbow clearance that keeps arms from clipping her torso. Smooth camera wrap and kiss recovery, fix a reachable peek elbow flip, and keep her tasteful bikini pin-up style, palette, exports, and render budget unchanged.

  • Render /loop view goals as normal-weight Markdown instead of one bold heading, with readable lists, code, tables, links, and contained wrapping in both Nugu themes. Escape raw HTML, suppress automatic image loads, allow only explicit HTTP(S)/mailto links, cache the last rendered goal per observer, and preserve text selection across live gate updates.

  • Add a safe-for-work loop kitten selected whenever PI_LOOP_ANIM is present, including empty, 0, false, or arbitrary values; resolve trusted loop.anim through the extension-local settings helper and select the kitten with a warning for invalid configuration. Draw an original copper-and-cream plush tabby with loop-shaped yarn and gate-watching, working, puzzled, curled, and proud reactions, without runtime imports from AFK or constructing vixen frames on the safe-for-work path. Preserve bottom-right portrait geometry, progress, keyboard focus, animation visibility toggles, and timer/overlay teardown across resume and reload.

  • Anchor the loop vixen bottom-right and enlarge painted pixels 1.5× before half-block packing, producing a tightly cropped portrait up to 48 columns by 23 rows without changing its camera motion. Leave nine rows clear for progress, editor, and footer; hide rather than shrink the portrait below 60 columns or 32 rows.

  • Move the loop vixen from inline layout into a tightly cropped, keyboard-noncapturing top-right portrait overlay, retaining natural raster scale and camera movement while reserving zero transcript or editor rows. Keep progress and counters in a separate inline widget of at most three rows; preserve /loop vixen [on|off] and remove owned overlays and timers on toggle, stop, afterglow, resize, and teardown without dismissing other overlays. Verify native regular/fullscreen compositors and keyboard focus; document native rectangular opacity and fullscreen mouse interception instead of patching Pi's compositor.

  • Combine loop progress and vixen into one aspect-aware above-editor widget: use a cropped natural-scale portrait beside status on wide terminals, above it on narrow terminals, and status only when a usable portrait cannot fit. Bound height to eight rows and normally one quarter of terminal height; keep attempts and turns only in the widget, remove duplicate loop footer status, and add /loop vixen [on|off] as a runtime-local animation toggle. Stop refresh timers for hidden or expired-success figures; document that terminal-background transparency cannot reveal Pi text underneath an inline widget.

  • Show persistent loop progress above Pi's editor: baseline or gate-check phase, animated activity indicator, gate position, elapsed time during silent gates, sanitized output tail, and cancellation hint. Transition to agent waiting/working, passed, paused, or stopped without implying a percentage or completion estimate; dispose refresh timers on widget removal and session teardown.

  • Restore loop runs by Pi session identity as paused diagnostics, preserving goals, gates, budgets, counters, last activity, and completed results without automatic execution. Fence run ownership across Pi processes, detect abandoned owners and surviving gate processes, and require explicit offline confirmation before recovering legacy runs without ownership metadata. Run command-started baselines in the background so status, view, and pause remain responsive; show running-gate elapsed time and bounded live output, drain cancellation before session teardown, and fence stale callbacks. Match interrupted gates by attempt and gate index, and publish durable restore events so other observers see recovered state without rewriting the original activity timestamp.

  • Nushell keeps bash and powershell displaced on resumed and reloaded sessions: when another extension (e.g. horst's per-host +powershell) re-adds a shell on top of the restored loadout before nushell's session_start, nushell now removes it again and records it for /nushell off, instead of returning early because nu was already active.

  • Load the whole collection from one Pi package entry, ~/Workspace/pi-ext/dist: every //:build-ext and //:build-all regenerates dist/package.json, listing built extension packages with pfui and footer first and lifting declared skills, prompts, and themes under <name>/ (Pi reads nested manifests only for extensions). New extensions load after a build and /reload without a settings edit; narrow the collection with package filters such as "extensions": ["!ultra/**"]. //:settings-entries is removed. Spec PX-SPEC-DZLVGRW4 and plan PX-PLAN-RMJ8BIOX are amended accordingly. READMEs say /reload instead of restarting Pi after a rebuild.

  • Add the loop companion: a red-haired pin-up in a pastel orange bikini rendered as supersampled truecolor half-blocks from hand-picked hue-shifted ramps below the editor in the TUI. She holds slow pin-up poses with breathing and a hip roll, two-bone IK keeps limb lengths fixed, and a slow camera glides between close-ups of face, bust and hips with zoom-gated soft-tissue follow-through; she waits with fingertips at her collarbone while gates run, looks unimpressed after failures, lounges while paused, stretches overhead and then blows kisses on success, and carries one subtle line of attempts, turns, and the failing gate. //extensions/loop:preview shows her without Pi.

  • Tighten loop internals: SQLite statements are prepared once per session instead of per write, the run page's poll rebuilds a snapshot only when the store changed (live controller changes still push immediately), gate-pass debug spans live in the controller and also cover rejected passes, and duplicated tool rendering, error formatting, and run lookup code is shared.

  • Loop TUI surfaces use Nerd Font Material Design icons: the status line shows autorenew with gate attempts, forum turns, and a timer while gates run, or pause/check/stop circles for paused, succeeded, and stopped runs; loop prompts, loop_begin, loop_control, and run notifications carry matching icons.

  • Let a blocked loop agent pause and resume its run with the loop_control tool: the loop prompt asks it to work autonomously first and pause only when it truly needs the human; paused turns are ordinary conversation and spend no loop budget, and resume reruns the gates before looping again.

  • Local //:e2e and //:e2e-extension trust the checkout's Mise config again: the isolated per-run MISE_STATE_DIR holds no trust records, so the runner now sets MISE_TRUSTED_CONFIG_PATHS to the repository root (appended to any inherited value).

  • Biome skips .system/**/assets; spec assets are frozen human-review mockups, not shipped code.

  • Add the loop extension (PX-SPEC-IJGV4GGK): /loop start|guide|resume|pause|stop|status|view|list runs an agent in a goal–gate loop until every configured gate (raw executable with args, env, cwd) exits 0 in one ordered pass; one attempt stops at the first failing gate, the injected loop prompt keeps a stable goal/gate/anti-cheat prefix with bounded output tails and older prompts collapse to stubs in context; minutes/turns/attempts budgets (none uncapped) pause the run, as do Escape during the baseline, typed stop/pause, aborted turns, and session switch/fork/tree/shutdown, and nothing auto-resumes; /loop resume keeps consumed turns/attempts and restarts the time budget; runs, gate attempts, and events persist in project-local SQLite (.pi/loop/state.db, loop.statePath, PI_LOOP_STATE_PATH) with control events outranking telemetry on queue overflow; /loop guide arms a brainstorm session whose loop_begin tool is rejected unless armed and advises on an unignored state file; /loop view serves a token-protected 127.0.0.1 page with SSE updates, Nugu light/dark palettes, run card, gate output drawer, timeline, events, history, and succeeded/paused banners. Root //:e2e now includes //extensions/loop:e2e.

  • Harden clipboard image reads against malformed payloads and decode bombs: PNG inputs must carry a well-formed IHDR and IEND trailer, inputs and converted outputs are bounded to 33554432 pixels (16384 px per side) before passthrough or return, and oversized or malformed images fail explicitly without partial data.

  • Bound clipboard backend calls with deadlines: native text, image, and file-reference reads fail after 3 s and copies after 10 s instead of hanging the sequential tool, and aborting mid-read now rejects promptly; losing timeout or abort paths never surface as unhandled rejections.

  • Rewrite ephemlet agent guidance around when to create a panel and when not to, one panel per task updated at milestones, mentioning open questions in chat before ending a turn, and wait only for questions a panel already holds; the tool description now covers text fields, wait, and using the revision returned by put.

  • Add ephemlet text fields edited inline in the panel and a wait action that blocks only the agent's tool call until awaited fields are answered (default every open question), times out (default 5 minutes, at most 30), or is aborted; waited fields show agent waiting, the human is notified, and the answer returns as the tool result without a duplicate answer message.

  • Render ephemlet panels with theme background surfaces instead of box-drawing characters: header band with checklist progress, read-receipt and Nerd Font icon controls, banded question headers, focus and hover row tints blended in OKLCH, confirm answers as Yes/No segments, and struck-through human vetoes; palette-indexed theme colors are resolved through OSC 4 with plain-token fallbacks. Add the ephemlets.icons setting (nerd default, unicode fallback). Open questions scrolled out of view get a pinned ↓ N needs you marker that jumps to them, the first tab focuses an open question, and new agent revisions start at the top. Question headers show their delivery before you answer: instant on a warm-tinted band for now, a muted next turn for next.

  • Add ephemlet choice and confirm fields with per-field delivery: ambient answers wait for get, next answers join the next turn, now answers steer a running agent or start a turn; rapid answers coalesce into one visible ephemlet-answer message carrying its answer revision (Pi may deliver a queued older message after a newer one), open questions show needs you in the panel and indicator, confirmations notify once, and the read receipt moves from sent to read when the model context contains the answers.

  • Add ephemlet checklist fields: the agent declares items and defaults, the human ticks them in the panel without starting a turn, human answers survive agent updates for unchanged IDs, list/get flag unseen changes, get with include: ["answers"] returns values with by: agent|user, and the panel header shows a read receipt.

  • Remove the ephemlet panel's alt+tab button selection, which operating systems usually reserve; ctrl+s, ctrl+w and header clicks remain, and the panel hint now lists them.

  • Dilemma's parent_session context view now follows Pi's context edits (omitted entries hidden, replacements shown) and the branch view marks them; system-role messages, which leaked the parent system prompt, plus context-edit and usage entries are no longer retrievable; Dilemma's own advice and status messages become retrievable.

  • Add Dilemma prior_consultations: advisors inspect same-role earlier child sessions on the active parent branch, with read results marked stale when the file changed or vanished, and a count hint in the assignment.

  • Rework memory into an agent-owned knowledge store: the memory tool takes raw SQLite (sql, params, readonly, timeoutMs) against one local database whose schema agents create and evolve; a seeded _guide and agent-maintained _schema_docs carry conventions; every call lands in an extension-owned audit journal with pre-write undo snapshots; ATTACH, VACUUM INTO, and ownership pragmas are blocked; oversized results spill to temp NDJSON. Removes the v1 review gate, pending inbox, action enum, and memory.reviewShortcut; /memory becomes a five-tab viewer (Usage, Size, Health, Browse, Journal) with journaled human edits and undo; v1 databases are refused instead of migrated, so delete the old file or point memory.databasePath at a new one.

  • Reduce ephemlets to session-persistent Markdown reports: remove interactive native components, choices/explorer templates, background builds, patch, state, Submit and the jiti dependency; load the runtime eagerly. Restored legacy interactive records show a removal notice while their session data stays untouched. purpose now defaults to title on creation and include: [] is accepted, removing the two most common first-call failures seen in archived sessions.

  • Strip native child background colors from PiPi live previews, including the edit renderer's self-painted header, so the parent card background remains continuous.

  • Route Good Job child analysis through the parent model registry so virtual selections such as grts/auto and runtime-only auth work without synthetic-provider credentials; retain thinking settings and bounded output.

  • Explain context-extension net removal effects with actual request codemode catalog additions/removals; distinguish empty/unavailable catalogs and withhold savings on catalog-only control drift.

  • Resolve managed Pi installations for context measurements, preferring an explicit managed root, then PATH, then conventional HOME fallbacks.

  • Trim memory tool/schema descriptions and System board tool descriptions.

  • Gratis routes to a built-in list of verified free-tier coding models (Kilo, OpenRouter pool, Mistral ministral, NVIDIA nemotron) plus user gratis.models patterns instead of zero-priced catalog entries; registers the Kilo gateway provider with a live free-model catalog; skips models with permanent failures for the session. Google Gemma and Mistral labs models are no longer built in.

  • Pad component widgets above the editor (notes, dilemma, voice, voice setup) by one column on each side, matching Pi's padding of string-array widgets.

  • Disable nested SDK tool shells inside PiPi through the public renderShell setting, preserving native content renderers and the parent card background.

  • Keep PiPi's last useful preview across empty assistant frames instead of replacing it with running status; reserve a stable collapsed expansion-hint row through completion and synchronize thinking visibility on retained messages.

  • Quota footer shows the account-wide allowance closest to exhaustion instead of always preferring the five-hour window; scoped limits and balances stay in the widget.

  • Add OpenRouter quota (GET /api/v1/key): free-model daily requests with reset at next UTC midnight; paid keys add credit cap and daily/weekly/monthly spend.

  • Add best-effort quota for Groq (requests per day from response headers), DeepSeek and Moonshot (balance), and Zai CN; xAI, Mistral, and NVIDIA expose no key-readable quota.

  • Quota follows virtual model selections such as grts/auto to the physical provider of the latest successful response since the selection; response headers are ignored while routed.

  • Render PiPi's current child assistant and tool activity with native SDK components, limiting collapsed previews to three lines and preserving expanded content, thinking visibility, child renderers, and Markdown transformers; isolate live UI state, filter unsafe terminal commands, and recover from broken preview components.

  • Align quota widget bars, show multi-day resets in days, and drop units already named by the bucket label.

  • Ignore empty Codex response-header window placeholders so absent secondary allowance does not reappear as an unused quota bucket.

  • Replace quota's fixed five-hour/weekly schema with provider-owned dynamic buckets, isolated provider modules, shared detail/footer rendering, independent Zai measurements, scoped Anthropic limits, and Codex balance/spend-cap support; remove Google quota support.

  • Preserve PiPi collapsed-header ellipsis backgrounds by clipping plain text before styling; verify native container backgrounds with terminal-cell assertions for pending, success, and error states.

  • Keep collapsed PiPi call headers on one line with terminal-width-aware ellipsis truncation; retain complete prompts in expanded view.

  • Show sanitized child-provider failure reasons in expanded PiPi results, preserving partial text and structured submissions.

  • Display PiPi as π² and stream child tool names with bounded, expanded-only arguments; keep collapsed activity compact and reset previews between tool execution and answers.

  • Classify PiPi child tools after extension startup: inherited hidden tools appear only in expanded Config; explicit hidden requests, inactive tools, and missing tools retain diagnostics.

  • Parse Codex rate-limit windows by their declared duration, preserve unavailable windows, and render a weekly-only known quota without fabricating a five-hour window.

  • Allow the activity demo up to 30 seconds to observe terminal cleanup after stop, preserving the required stopped lifecycle assertion.

  • Bundle task-neutral PiPi workflow guidance for delegation, evidence-based handoffs, and capability-driven persistent progress views.

  • Isolate Canvas Chromium-discovery module mocks in the serial test pool instead of the shared unit pool.

  • Preserve PFUI prompt-section mutability for later extensions while retaining lazy rendering until explicit overrides; verify PFUI/PiPi instruction delivery and inherited-section precedence through the real SDK.

  • Include actual child-extension error messages in PiPi diagnostics instead of reporting only hook identity.

  • Keep Mockup annotation bubbles inside narrow and scrolled fragments; preserve pin anchoring, note editing, and placement across option changes and viewport resizing.

  • Discover cached Playwright Chromium for Canvas layout in CI; start System video playback before requiring decoded frames and retain media diagnostics on failure.

  • Add manual fail-closed Luci release preparation with version-lock, digest provenance, and leased release branches.

  • Add pipi: blocking native child-session tasks, optional reusable JSON/Markdown presets, discoverable dynamic extension files, model-tier recommendations, continuation/forking, human-only streaming, and structured result submission with per-call native usage accounting.

  • Unify System book, board, and document views behind one authenticated per-project localhost application with shared actions, navigation, live refresh, last-good recovery, and shutdown cleanup.

  • Verify System book workflows in Chromium and Firefox across desktop/mobile and light/dark views; fix clipped search, repeated highlights, encoded previews, disclosure retention, proposal/readiness races, generated footnotes and accessibility references, fixed-header anchor offsets, and navigation-time Mermaid diagnostics while preserving authored HTML/SVG ownership.

  • Bundle System Mermaid and MiniSearch as browser modules instead of copying upstream prebundles; refresh compatible locked dependencies and verify DOMPurify 3.4.17 in rebuilt Canvas and System assets.

  • Protect good-job claim acquisition and recovery across same-process module reloads; reject stale quota/provider/cache results and out-of-order VCS refresh publication with controlled concurrency regressions.

  • Validate Chrome command arguments and complete Chrome/IntelliJ subcommands; correct quoted-Python and Windows executable handling, runtime home-boundary discovery, portable resource URLs, and broken focused relocation task selectors.

  • Remove unused Firefox BiDi bookkeeping; verify priority-provider request dispatch and TPS event/status behavior; read the configured usage store for in-memory sessions and include cached reads in the displayed token formula.

  • Add real-Pi activity, AFK, ephemlet, usage, and synthetic quota demonstrations; refresh stale recordings and capture real canvas/mockup browser examples.

  • Verify handoff replacement by native session IDs and reviewed editor text; add negative-control hook-safety and post-editor input checks so demos cannot pass from typed echoes or missing tools.

  • Make Ultra's orchestrator responsible for subagent knowledge, instructions, skills, tools, and session continuity instead of minimizing context length; preserve risk-based verification separately.

  • Increase Ultra action-summary output allowance from 40 to 256 tokens and Voice response allowance from 768 to 4,096 tokens without changing concise output instructions or timeout bounds.

  • Reject failed or empty handoff generation before review or session replacement; distinguish failures from cancellation and ignore late provider settlement after the loader is cancelled.

  • Disable backup storage on invalid explicit destination settings and automatic uploads on invalid opt-out settings instead of falling back; preserve manual backups with valid storage.

  • Align cost session totals with Pi's full-session native usage, scan monthly responses by recorded time, retain distinct usage records, and suppress identifiable copied fork history without double-counting parent tool aggregates.

  • Fail closed on failed or malformed Git tree inspection, verify unborn repositories before skipping checkout, cap combined subprocess output at 16 MiB, and retain active-target overlap protection across queued clone replacements.

  • Rewrite the memory README around remember → review → recall and add a reproducible real-Pi demo showing proposal approval and scoped search with isolated synthetic data.

  • Preserve authored final reading pauses in asciicast endpoints and GIF exports instead of silently using the renderer's three-second default; force an isolated memory DB even when the caller exports a user database override, with live isolation verification behind build-backed demo-check rather than normal unit tests.

  • Add an interactive read-only memory search browser with live keyword queries, scoped approved results, complete detail scrolling, empty/error recovery, and no new transcript dumps; reuse review layout and preserve non-TUI search output.

  • Replace memory review's raw-ID picker and ASCII inspector with a themed, searchable inbox and content-first preview; preserve revision-bound inspection, adapt selection and detail panes to short/narrow terminals, and add a disposable real-Pi preview task.

  • Consolidate extension history into this internal changelog; remove extension-local changelogs and README links, and restrict changelog automation and validation to the root file.

  • Restore configured codemode, tool search, and MCP capabilities in Ultra SDK children; builders retain child-default tools instead of requiring the parent's full tool registry.

0.1.10

  • Refresh dependencies and require Pi 1.0.2 with updated extension contracts
  • Add /handoff for moving context into a focused new session
  • Run interactive Nushell bang commands by suspending Pi's terminal UI
  • Allow manual masks to select tools and system-prompt instructions
  • Refresh workspace branch status after external checkouts and update quota from provider response headers
  • Harden Ultra workflow trust checks, inactive-tool preflight, and prompt interpolation
  • Ship Klaus SDK dependencies per platform and load its provider lazily
  • Rewrite all 41 extension guides around first use, with generated configuration references and deterministic terminal demos
  • Modernize TypeScript tooling, scoped typechecking, structural checks, and test execution

0.1.9

  • Publish release archives under the package server's public /releases/ route

0.1.8

  • Keep live board controls disabled until script binding and initial state synchronization complete
  • Cover delayed state initialization in Chromium and Firefox without timing-based waits

0.1.7

  • Expose the release checksum manifest at a flat CI artifact path
  • Keep release archives gated by the complete CI suite and relocated package validation

0.1.6

  • Keep FFmpeg active inside the nested Chrome E2E task
  • Wait for live board script binding after document reloads before metadata interactions in browser tests

0.1.5

  • Declare FFmpeg 9.0.2 as an E2E tool instead of relying on a host encoder
  • Diagnose System board status failures at HTTP, disk, and rendered-site boundaries

0.1.4

  • Refresh the System board's static output before acknowledging plan status changes
  • Run headed Chrome acceptance tests under Xvfb in CI, with sandbox isolation confined to the test executable wrapper
  • Reuse explicitly configured CI browser and download caches while keeping E2E HOME, agent data, and trust state isolated

0.1.3

  • Persist Mise toolchains, downloads, Go build caches, and npm downloads across failed CI runs using durable cache mounts
  • Keep ephemeral Mise trust state outside the source checkout and cached toolchain directories
  • Remove Node startup latency from RTK timeout test fixtures

0.1.2

  • Wait for bak's background initialization instead of assuming a fixed startup delay in settings tests
  • Keep Firefox fake-child ownership tests independent of a host Firefox installation

0.1.1

  • Exclude cached toolchains and temporary release output from Taplo checks
  • Replace failed v0.1.0 CI candidate without moving its tag

0.1.0

  • Build all 40 extensions as minified, self-contained packages with one repository version
  • Validate relocated packages and exact archive bytes; publish release archives with SHA-256 manifests through Luci
  • Ship one root installation package containing every extension, plus individual extension archives
  • Exclude cached CI toolchains from source linting
  • Raise web fetch and map visible-output limit to 100,000 characters and improve invalid-argument errors
  • Optimize uv spawn-hook command checks and PATH rewriting; add benchmark, profiling, and comparison tasks
  • Add uv blockers for pip3 and Poetry commands with uv guidance
  • Remove RTK find -path compatibility probing and guard; find rewrites now apply directly
  • Cache successful RTK command rewrites per session, with benchmark and profiling tasks
  • Harden dynamic-extension validator protocol handling and share workflow preparation across command and tool runs
  • Identify Ultra child assignments by workflow, phase, and agent; attribute interpolated prior-agent output to its immediate producers, preserve provenance through fanout transformations, and invalidate checkpoints when attribution changes
  • Precompile uv spawn-hook matchers and guidance, optimize POSIX PATH handling, and add spawn-hook benchmarks and profiling tasks
  • Add uv spawn-hook microbenchmarks and optimize command matching and POSIX PATH rewriting, reducing measured benchmark time by 89%

0.0.3

  • Rename autoresearch to tune; add jj backend, state migration, timeout clamps
  • Rework diff-review to SSR reviews; add headless Vitest coverage
  • Rework q extraction; add benchmark and 112 parser tests
  • Retire superpowers; move workspace to npm, Vitest, dprint, Biome, tsc, mise

0.0.2

  • Bump package.json from 0.0.1 to 0.0.2
  • Update README install examples to git:github.com/bugabinga/pi-ext@v0.0.2
  • Leave extension sources, READMEs, tests unchanged from v0.0.1

0.0.1

  • Ship initial pi-ext set: goal, tldr, leaks, tps, quota, vcs-status
  • Add goal loop: autocomplete, progress UI, standing prompts, judge-gated updates
  • Add tldr live summaries; model footer; quota parsers; jj-aware workspace status
  • Fix session startup races, stale telemetry, aux LLM calls, provider secret leaks

Legacy extension history

Consolidated from former extension-local changelogs; original version labels and claims are preserved as historical notes, not verified release records.

context-size

context-size: 0.0.3

  • Release @bugabinga/pi-ext-context-size 0.0.3
  • Resolve pi-ext conflict stack; retire superpowers references
  • Replace Bun test wiring with Vitest workspace tooling
  • Isolate extension harness env in tests
  • Tighten Biome lint rules; remove low-risk explicit any

context-size: 0.0.2

  • Release contains no context-size commits since v0.0.1
  • Runtime still publishes footer:segment; status fallback remains context-size
  • Harness test still checks extension load in Pi runtime

context-size: 0.1.0 - 2026-05-21

  • a40a427 prepare extensions for npm release
  • 6a97080 add asciinema demo workflow
  • c1e9f71 fix(context-size): show visible usage bar
  • 517ba54 pi: add extension footer fallbacks
  • 133cb7d chore(pi): migrate extensions to earendil packages
  • f144b06 fix(pi): restore extension notifications
  • 5ca1296 Rework Pi agent extensions

cost

cost: Unreleased

  • Estimate zero-cost messages from Pi's model catalog through calculateCost instead of a hand-maintained zai/minimax price table; cache writes are now charged.

cost: 0.0.3

  • Resolve pi-ext conflict stack; retire superpowers.
  • Replace Bun test runner with Vitest workspace tooling.
  • Isolate extension harness env; fix import.meta.dirname and chai matchers.
  • Tighten Biome rules; replace low-risk explicit any usage.

cost: 0.0.2

  • Source unchanged from v0.0.1
  • README unchanged from v0.0.1
  • Harness tests unchanged from v0.0.1

cost: 0.1.0 - 2026-05-21

  • a40a427 prepare extensions for npm release
  • 6a97080 add asciinema demo workflow
  • 517ba54 pi: add extension footer fallbacks
  • 133cb7d chore(pi): migrate extensions to earendil packages
  • 5ca1296 Rework Pi agent extensions
  • b87a61a feat(pi): monorepo workspace — all extensions are proper packages
  • b408bba pi(ext): footer segment producers (context, timing, cost, runtime, git-status, pr-status, thinking)

dilemma

dilemma: 0.0.3

  • Replace Bun test harness with Vitest workspace setup
  • Format markdown and code fences with dprint
  • Tighten lint rules; remove low-risk explicit any usage
  • Update extension guides and demos

dilemma: 0.0.2

  • Extension has no commits after v0.0.1; release-only cut
  • README.md and index.ts unchanged
  • /angel, /angel-policy, hooks, and bash policy gate unchanged

dilemma: 0.1.0 - 2026-05-21

  • a40a427 prepare extensions for npm release
  • 133cb7d chore(pi): migrate extensions to earendil packages
  • 5ca1296 Rework Pi agent extensions
  • b87a61a feat(pi): monorepo workspace — all extensions are proper packages
  • 5fbf178 pi(ext): add remaining extensions (angel, animations, code-actions, files-widget, ghost, git-safe, neko, prune, web, etc.)

fast

fast: 0.0.3

  • Own fast status in @bugabinga/pi-ext-fast
  • Replace Bun test runner with Vitest across workspace
  • Fix extension harness tests for Vitest deps and Chai matchers
  • Migrate workspace tooling to mise; keep root Bun workspace

fast: 0.0.2

  • Git history: no extensions/fast commits since v0.0.1
  • Keep fast.enable settings contract unchanged
  • Keep serviceTier: "priority" wrapper for openai-codex unchanged

footer

footer: Unreleased

  • Emit footer:branch-change with { branch } when Pi reports a git branch change, replacing the no-op rerender subscription
  • Forward branch changes while hidden as well; add branch.change debug event

footer: 0.0.3

  • Replace Bun test runner with Vitest
  • Migrate workspace tooling to mise and root Bun workspace
  • Fix footer harness tests; isolate env, deps, and Chai matchers
  • Tighten Biome lint; remove explicit any and non-null assertions

footer: 0.0.2

  • Ship unchanged footer package; no source diff since v0.0.1
  • Keep /footer and ctrl+alt+f toggle behavior unchanged
  • Keep footer:segment event contract and 4-line footer layout unchanged

footer: 0.1.0 - 2026-05-21

  • a40a427 prepare extensions for npm release
  • 6a97080 add asciinema demo workflow
  • 517ba54 pi: add extension footer fallbacks
  • 133cb7d chore(pi): migrate extensions to earendil packages
  • f144b06 fix(pi): restore extension notifications
  • 5ca1296 Rework Pi agent extensions
  • b87a61a feat(pi): monorepo workspace — all extensions are proper packages
  • b18f0e2 refactor(pi/footer): clean sweep — kill dead code, extract helpers, name constants, type properly
  • 972978f fix(pi/footer): force full redraw (requestRender(true)) to clear old footer on hide
  • 0ea30fd fix(pi/footer): force render after swap to immediately clear old lines
  • a49441f fix(pi/footer): swap footer via setFooter() on toggle for true 0-height when hidden
  • 1cfa6a1 fix(pi/footer): import Key from pi-tui, not pi-coding-agent
  • 933bcb6 pi(ext/footer): semantic zones, flow layout, 4-line stable height, /footer toggle

git-safe

git-safe: Unreleased

  • Remove the Bash Git editor policy: it injected Bash source into every shell tool and broke non-Bash shells such as Nushell
  • Stop intercepting tool_call; git-safe no longer rewrites agent-issued shell commands
  • Keep safe-clone behavior unchanged
  • Format byte sizes with Pi's formatSize instead of two local formatters; size-limit errors and the clone result now use binary units (KB = 1024 B) and cap at MB

git-safe: 0.0.3

  • Add sparse safe clone support.
  • Recover tracked paths on session start.
  • Replace Bun test runner with Vitest.
  • Isolate extension harness environment in tests.

git-safe: 0.0.2

  • Release 0.0.2 with no commits touching extensions/git-safe
  • Keep git_clone_safe(url, path, branch?) behavior unchanged
  • Keep README, harness test, and package metadata unchanged

git-safe: 0.1.0 - 2026-05-21

  • a40a427 prepare extensions for npm release
  • 133cb7d chore(pi): migrate extensions to earendil packages
  • 5ca1296 Rework Pi agent extensions
  • b045a6b fix(pi/mux): cooldown exhausted keys on provider usage-limit errors
  • b87a61a feat(pi): monorepo workspace — all extensions are proper packages
  • 5fbf178 pi(ext): add remaining extensions (angel, animations, code-actions, files-widget, ghost, git-safe, neko, prune, web, etc.)

good-job

good-job: Unreleased

  • Steering input (event.streamingBehavior === "steer") no longer waits for runtime startup and queue writes before continuing to the running turn; the capture runs in the background and failures still notify.

handoff

handoff: Unreleased

  • Add /handoff <goal>: compaction-aware branch summary, editable generated prompt, new child session with the prompt in the editor.
  • Generation prompt references artifacts by path or URL, redacts secrets and PII, and adds a Suggested skills section.

masks

masks: Unreleased

  • Add optional tools and instructions to mask settings, validated like other fields and reported when invalid
  • Apply mask tools on equip, warning on unknown names and reusing the prompt-mask snapshot and rollback path
  • Inject mask instructions as the masks system-prompt section via before_agent_start
  • Let prompt frontmatter tools override mask tools for one prompt
  • Show tools and a truncated instructions preview in picker descriptions; non-TUI selection now shows shortcuts too

masks: 0.0.3

  • Add /masks, configurable picker shortcut, and per-mask shortcuts
  • Fuzzy-match masks by name, model, thinking level, and shortcut
  • Apply configured model and thinking level together
  • Let prompt templates equip a mask for one agent run through mask frontmatter
  • Restore the previous model and thinking level after masked prompt runs
  • Validate global masks settings and report invalid entries

mockup

mockup: Unreleased

  • Extract the generic browser mockup tool from The System.
  • Show browser URLs and opening failures in Pi's native expandable tool UI.
  • Reopen disconnected reviews and recover abandoned iteration chains with preserved history.

model-info

model-info: Unreleased

  • Keep the model version in short labels: gpt-6-astra becomes astra-6, claude-opus-5-5 becomes opus-5.5
  • Read the thinking level from pi.getThinkingLevel() instead of scanning the session branch for thinking_level_change entries

model-info: 0.0.3

  • Own fast status rendering in model-info
  • Switch extension tests from Bun test to Vitest
  • Isolate extension harness environment in tests
  • Tighten Biome rules and remove low-risk explicit any usage

model-info: 0.0.2

  • Mark 0.0.2; source, README, tests, lockfile match v0.0.1
  • Keep model footer segment plus model-info status fallback
  • Keep fast.enable badge for openai-codex
  • Keep harness tests for extension load and label format

nu-bang

nu-bang: Unreleased

  • Run interactive commands (!i <command>, editors, pagers, editor-opening Git subcommands, known terminal UIs) with the terminal handed over while the TUI is suspended; outside the TUI they are refused with an explanatory result.
  • Remove the unused standalone executor, output control stripping, and pi-nu-bang-output-*.log files; Pi sanitizes, truncates, and persists bang output itself.
  • Correct the README behavior section to match the runtime path.

nu-bang: 0.0.3

  • Fix win32 release readiness
  • Replace shell scripts with cross-platform Lua scripts
  • Add docx-to-md.lua, EPUB, Typst, Beamer, and pdf-report.yaml
  • Isolate harness env and temp-dir cleanup tests

nu-bang: 0.0.2

  • Source unchanged under extensions/nu-bang since v0.0.1.
  • README lists hooks: user_bash, session_start, session_shutdown.
  • Tests cover Nu exec, completion, cwd errors, truncation, temp-dir cleanup.

nu-bang: 0.1.0 - 2026-05-21

  • a40a427 prepare extensions for npm release
  • 133cb7d chore(pi): migrate extensions to earendil packages
  • 3b453ee feat(pi): add nushell bang extension

os-notifier

os-notifier: Unreleased

  • Rename terminal-notifier to os-notifier.
  • Prefer native Linux, macOS, and Windows notifications with OSC fallback.
  • Notify only for direct interactive TUI runs completed while unfocused.
  • Preserve rich metadata with natural urgency and bounded sanitized text.
  • Remove the unused extension notification event bridge.
  • Publish pane-scoped Pi identity, session name, and project through OSC 1337 during TUI sessions.
  • Subscribe to ui_prompt_start and ui_prompt_end through Pi's typed pi.on overloads instead of a local shim type and cast.

os-notifier: 0.0.3

  • Replace Bun test runner with Vitest in the workspace
  • Add isolated extension harness env for tests
  • Fix harness tests for import.meta.dirname, Vitest deps, and Chai matchers
  • Migrate project tooling to mise

os-notifier: 0.0.2

  • Release only: no commits touched extensions/terminal-notifier since v0.0.1
  • Source unchanged: agent_end still sends OSC/Windows notify
  • Tests unchanged: harness still verifies extension runtime load

os-notifier: 0.1.0 - 2026-05-21

  • a40a427 prepare extensions for npm release
  • 133cb7d chore(pi): migrate extensions to earendil packages
  • f144b06 fix(pi): restore extension notifications

prompt-autopsy

prompt-autopsy: Unreleased

  • Honor Pi's externalEditor setting before $VISUAL and $EDITOR when opening the prompt file, via SettingsManager.getExternalEditorCommand()
  • Fall back to Pi's default editor (nano, notepad on win32) instead of vi when nothing is configured

prompt-autopsy: 0.0.3

  • Replace Bun test runner with Vitest in the workspace
  • Isolate extension harness environment for tests
  • Fix harness tests for import.meta.dirname, Vitest deps, and Chai matchers
  • Format prompt-autopsy markdown docs

prompt-autopsy: 0.0.2

  • Release @bugabinga/pi-ext-prompt-autopsy with no code delta from v0.0.1
  • Keep /prompt-autopsy and Ctrl+Alt+A behavior unchanged
  • Leave README, harness test, bench, and changelog unchanged

prompt-autopsy: 0.1.0 - 2026-05-21

  • a40a427 prepare extensions for npm release
  • 133cb7d chore(pi): migrate extensions to earendil packages
  • 5ca1296 Rework Pi agent extensions

quota

quota: Unreleased

  • Update Anthropic and OpenAI Codex quota from provider response headers via after_provider_response instead of waiting for the poll.
  • Read the Codex account id through Pi's typed readStoredCredential export.

quota: 0.0.3

  • Display quota reset in parentheses.
  • Replace shell scripts with cross-platform Pandoc Lua.
  • Use --highlight-style for Pandoc 3.6.
  • Add EPUB, Typst, Beamer, and Sandbox mode report output.
  • Move workspace tests from Bun to Vitest.

quota: 0.0.2

  • Source: no extension-local commits after v0.0.1
  • Keep /quota widget, footer segment, and status fallback unchanged
  • Keep provider quota parsers and harness tests unchanged

quota: 0.1.0 - 2026-05-21

  • a40a427 prepare extensions for npm release
  • 6a97080 add asciinema demo workflow
  • 517ba54 pi: add extension footer fallbacks
  • b924295 config: add kimi coding model
  • 5ca1296 Rework Pi agent extensions
  • b87a61a feat(pi): monorepo workspace — all extensions are proper packages

rtk

rtk: Unreleased

  • add minimal RTK bash rewrite hook
  • warn once on load when RTK is missing or too old

runtime

runtime: 0.0.3

  • Switch tests from Bun to Vitest; fix harness imports and Chai matchers
  • Isolate extension harness env for runtime tests
  • Migrate workspace tooling to mise with a root Bun workspace
  • Replace suite demos

runtime: 0.0.2

  • Keep index.ts unchanged from v0.0.1
  • Keep README, package metadata, changelog, and harness test unchanged
  • Ship no command, tool, setting, or hook contract changes

runtime: 0.1.0 - 2026-05-21

  • a40a427 prepare extensions for npm release
  • 6a97080 add asciinema demo workflow
  • 517ba54 pi: add extension footer fallbacks
  • 133cb7d chore(pi): migrate extensions to earendil packages
  • 5ca1296 Rework Pi agent extensions
  • b87a61a feat(pi): monorepo workspace — all extensions are proper packages
  • 0bf9702 pi(ext/runtime): update runtime extension
  • b408bba pi(ext): footer segment producers (context, timing, cost, runtime, git-status, pr-status, thinking)

session-rename

session-rename: Unreleased

  • Read the current session name through pi.getSessionName() instead of a guarded ctx.sessionManager.getSessionName() wrapper

session-rename: 0.0.3

  • Resolve pi-ext conflict stack; remove superpowers path
  • Isolate extension harness env in tests
  • Remove non-null assertions under stricter Biome rules
  • Replace Bun test runner with Vitest across workspace

session-rename: 0.0.2

  • Ship release-only version bump; no session-rename source changes
  • Keep /session-rename [name] command behavior unchanged
  • Keep session naming hooks and harness load test unchanged

session-rename: 0.1.0 - 2026-05-21

  • a40a427 prepare extensions for npm release
  • 133cb7d chore(pi): migrate extensions to earendil packages
  • 5ca1296 Rework Pi agent extensions
  • b87a61a feat(pi): monorepo workspace — all extensions are proper packages
  • 5fbf178 pi(ext): add remaining extensions (angel, animations, code-actions, files-widget, ghost, git-safe, neko, prune, web, etc.)

the-system

the-system: Unreleased

  • Add reversible dropped board state for specs, plans, and issues, with DROP/RESTORE controls, inactive counts, and dropped-last ordering; dropped plans no longer block spec completion.

  • Add an EDIT action to board cards that opens the indexed document in the configured editor and reports launch failures.

  • Show specs in one board column, copy document IDs on click, and start /system exec <ID> from a separate card action in the current idle Pi session.

  • Remove the board's behavioral subtitle and order specs unfinished first, then most recently modified first with stable ID ties.

  • Register deferred System tools upfront without lifecycle-stage activation or consumption gates.

  • Prevent board polling from interrupting pending approval requests or applying stale pre-write snapshots.

  • Move the document, Markdown, board, and exec contracts out of every /system stage message into project-scoped skills (the-system-documents, the-system-markdown, the-system-board, the-system-exec), so repeated stages no longer duplicate several kilobytes of identical instructions into the session.

  • Reshape the stage message to a heading with the stage and the human's words, then the board summary, then the stage prompt; the Human instructions header is gone.

  • Cut /system index and render cost: frontmatter extraction no longer full-parses document bodies (index load -94%), and the renderer harvests search text from the same parse as the HTML pipeline instead of re-parsing every document (cold board render -7%).

  • Resolve the active System root as the nearest .system directory at or above cwd; nested systems override parents, /system init initializes at cwd, and split Git/JJ roots no longer conflict.

  • Use rich Markdown bundles with YAML graph metadata, config-only machine state, doctor validation, and file-level spec ownership.

  • Add /system board with tracked schema-0 operational state, nested plan cards, independent issues, responsive accordions, and human-reviewed /system assess proposals.

  • Rename plan status done to approved; remove issue status.

  • Open rendered pages through the native Win32 URL handler without flashing a command window.

  • Add turn-scoped system_bundle_slot allocation with title-derived slugs and repeatable ID reservation.

  • Redefine phases as project-wide operating epochs selected through config, injected into agent context, and shown in Pi and book navigation.

  • Reset the unreleased config schema to version 0; /system init updates machine schema markers mechanically, then offers agent-led bundle migration.

  • Add initialized-project-only GFM, TeX, and Mermaid skills for already-selected System document formats.

  • Keep generated /system lifecycle turns alive in print and JSON modes.

  • Align bundle-slot and board-proposal prompt metadata and tool rendering with Pi defaults.

  • Keep board proposals and repeatable bundle slots available across clarification turns and compaction.

  • Add a generally invocable vertical-slicing skill and require it for /system exec.

  • Treat plans as required outcome rather than implementation order, and remove slice boundaries as human approval points.

  • Make visual models the System Markdown default, require visuals in specs and plans, and decompose plans by behavior instead of architectural layer.

  • Let /system exec apply verified board states directly through system_board_update with source: agent and an evidence note, keeping /system assess proposal-only.

  • Inject a board state block with counts and unfinished items into every /system lifecycle prompt.

  • Compact the injected document contract without dropping rules.

  • Replace unconditional core-document blocking with validated one-shot human approval.

  • Let completed /system exec runs propose evidence-backed board updates for human review.

  • Serve /system board pages from a live snapshot that re-renders when .system changes, so refreshes show persisted state and open boards follow disk edits.

  • Toggle plan status and spec board state directly from the rendered metadata box.

the-system: 0.0.3

  • Resolve pi-ext conflict stack; retire superpowers.
  • Replace os.execute wrappers with pandoc.pipe argv lists in conversion scripts.
  • Fix batch-convert patterns and keep web-to-md H1 content out of metadata.
  • Add docx-to-md image extraction to ./media/, Decision Guide, pandoc lua -e examples.
  • Document heading shift, CJK, images, empty titles, Lua filters, PDF engines.

tps

tps: 0.0.3

  • Replace Bun test runner with Vitest for extension tests
  • Isolate extension harness env in tests
  • Fix harness tests for import.meta.dirname, Vitest deps, and Chai matchers
  • Tighten Biome rules and remove low-risk explicit any usage
  • Move workspace tooling to mise

tps: 0.0.2

  • Keep @bugabinga/pi-ext-tps source unchanged from v0.0.1
  • Keep README hooks, footer segment, status fallback unchanged
  • Keep bun tests for TPS math and pi-test-harness load unchanged

ultra

ultra: Unreleased

  • Add the {item.FIELD} interpolation token: one field of the current record fanout item, rendering null or missing fields as empty text in prompts and summaries and as undefined in selectors, with a validation error for non-record items.
  • Treat braces in prompt and summary as interpolation tokens only when they start with args., item, or PHASE.results/PHASE.failures; other braces, such as JSON fragments, TypeScript types, or empty braces, pass through as literal text, while malformed token-shaped spans still fail validation and selectors stay strict.
  • Refuse a workflow before creating any child when a step without flags lists a tool that is registered but inactive in the parent Pi (for example bash under pi --nushell), naming the inactive tools, the child extensions with startup flags that may have deactivated them, and the active nu tool when a shell tool is inactive; builtins Pi leaves inactive by default are ignored.
  • Report a deactivated builtin tool as builtin tool, deactivated at startup in the child-startup error instead of a bare name.
  • Fail the not-yet-started siblings of a fanout phase immediately with attempts: 0 once one child's requested tools were switched off at startup, instead of creating every sibling session to fail the same way.
  • Request strict constrained sampling for structured_output only when the step declares a schema; against the schema-less property-free object schema it forced models to submit {}.
  • Honor project workflows (<cwd>/.pi/workflows/*.json) only in trusted projects: untrusted project workflows are hidden from /ultra run completions, never shadow bundled or user workflows, and run_workflow and /ultra run refuse them with an error naming the file.
  • Label each /ultra run completion with its workflow source (bundled, user, project).

usage

usage: 0.0.3

  • Retire superpowers conflict stack
  • Move workspace tooling from bun to mise and vitest
  • Isolate extension harness env in tests
  • Remove non-null assertions for Biome lint

usage: 0.0.2

  • Keep /usage dashboard unchanged from v0.0.1.
  • Ship no extensions/usage diff: source, README, tests, assets unchanged.
  • Retain harness smoke test for Pi extension runtime load.

usage: 0.1.0 - 2026-05-21

  • a40a427 prepare extensions for npm release
  • 133cb7d chore(pi): migrate extensions to earendil packages
  • 5ca1296 Rework Pi agent extensions
  • b87a61a feat(pi): monorepo workspace — all extensions are proper packages
  • 5fbf178 pi(ext): add remaining extensions (angel, animations, code-actions, files-widget, ghost, git-safe, neko, prune, web, etc.)

vcs-status

vcs-status: Unreleased

  • Refresh the VCS segment on footer:branch-change using the latest hook context, so branch switches in other terminals no longer leave it stale
  • Add branch.change debug event

vcs-status: 0.0.3

  • Move workspace tooling to mise and root Bun workspace
  • Replace Bun test runner with Vitest across workspace
  • Fix harness tests: import.meta.dirname, deps, Chai matchers
  • Isolate extension harness environment in tests
  • Tighten Biome lint rules and fix violations

vcs-status: 0.0.2

  • git diff v0.0.1 -- extensions/vcs-status: empty
  • Source, README, tests: unchanged
  • Runtime behavior: unchanged

vcs-status: 0.1.0 - 2026-05-21

  • a40a427 prepare extensions for npm release
  • 6a97080 add asciinema demo workflow
  • 517ba54 pi: add extension footer fallbacks
  • 133cb7d chore(pi): migrate extensions to earendil packages
  • 5ca1296 Rework Pi agent extensions
  • b87a61a feat(pi): monorepo workspace — all extensions are proper packages
  • b408bba pi(ext): footer segment producers (context, timing, cost, runtime, git-status, pr-status, thinking)

voice

Historical development entries include the retired browser implementation and superseded defaults. See Voice README for current behavior and availability.

voice: Unreleased

  • Refine the call surface with human-approved quiet-luxury styling, a measured-PCM ribbon, visible status, captioned controls, quieter Diagnostics and a labelled Pi activity pill.

  • Preserve control labels and focus across icon updates; honour reduced motion, root text scaling and compact recovery layouts.

  • Clear the Resume action after ASR recovery completes instead of leaving stale controls visible.

  • Add host-native CMake Moonshine builds for Linux, macOS and Windows, portable temporary workspaces, explicit Node/ORT inputs and platform-specific packaging; desktop runtime verification remains deferred.

  • Load Windows native dependencies through UTF-16 paths, preserve ORT library aliases and validate the required ORT API before loading Moonshine.

  • Add opt-in native CPU English and multilingual Nemotron 560/1120 ms, Moonshine Streaming Small/Medium, and Parakeet Realtime EOU choices without changing production defaults.

  • Verify pinned model downloads, bundle Android arm64 Moonshine runtime with notices, and reject missing native assets or unsupported thread overrides before downloading weights.

  • Supply Parakeet Realtime's complete checkpoint vocabulary without runtime conversion; hide EOU control tokens from transcripts without suppressing native predictor updates.

  • Flush Nemotron's complete encoder window at forced EOF, preserving existing VAD, natural endpoint and submission policies.

  • Cancel streaming model downloads on reset and fence stale startup work from restarted sessions.

  • Flush Moonshine directly through its native stop contract and roll bounded stream context without ending long idle calls.

  • Release acquired microphone tracks on failed worklet startup, clear recovered startup errors, and keep technical provider failures inside Diagnostics.

  • Enforce long-reply handoff during actual playback, retain full accepted transcript text, and isolate duration/completion accounting per turn.

  • Recover bounded playback acknowledgements, including stalled progress, rejected boundaries, silent notifications and PCM delivered before its receipt.

  • Deliver idle executor completions at acknowledged boundaries; support deliberate silent skips, correlated result navigation and stable keyboard focus.

  • Map real word timings to ordered transcript positions, preserve result actions while words highlight, and expose unavailable timing reasons without estimates.

  • Preserve manual transcript scrollback and verify layered views on desktop, mobile, short landscape and enlarged text.

  • Use human-approved paper, bubble, mechanical-loop and ambient cues as defaults after production-level comfort audition; preserve saved overrides and gains.

  • Initialize native TTS and offline ASR asynchronously, preserving responsive controls and fencing cancelled startup results, retries, progress, and temporary vocabulary ownership.

  • Begin eased cue ducking before speech playback instead of adding an unconditional 120 ms cue delay; preserve audio-clock state and ordered playback.

  • Replace the audition shortlist with twenty level-checked CC0 candidates from five contrasting source packs; preserve legacy choices and production defaults, support authored continuous loops, and ease short transient attacks.

  • Add browser-controlled voice replay and cue selection, gesture-driven audio resume, bounded PCM/audio-clock diagnostics, and production-level cue audition without a preview boost.

  • Clear previous recognition hotwords when a voice reply fails its structured output contract.

  • Fence speech submission against early native activity and pending decodes without weakening confirmed barge-in or adding another silence wait; retain deferred transcript fragments.

  • Credit validated natural endpoint silence against existing submission budgets, retain conservative forced boundaries, and skip classifier requests after the deadline.

  • Add explicit native classifier selection and opt-in host Codex Luna chat labels, with visible non-native identity and no invented confidence; retain the silence fallback and reject late or cancelled answers.

  • Add pinned decoder-only and full-FP32 streaming controls without changing production speech models or concealing recognition regressions.

  • Add opt-in native streaming recognition with bounded in-process workers and lifecycle fencing; retain Parakeet defaults because controlled corpus tests expose opening-word loss in the experimental model.

  • Add native acoustic word alignment, late correlated timing delivery and callback-chunk playback offsets without delaying speech audio; reject unsupported or mismatched alignment.

  • Resolve packaged speech-worker entry points and verify native bundled-server ASR/TTS, relocated alignment assets, and late browser timing without replay or stale highlighting.

  • Add descriptive five-engine recorded batches, original-recording Whisper-reference reuse and atomic checkpoints without invented acceptance thresholds.

  • Add native Whisper references and checkpointed human-corpus ASR measurements with explicit raw versus production paths, empty-result reporting and no invented acceptance thresholds.

  • Remove voice-owned WASM runtimes, fallback, backend settings, worker paths, packaging and comparison controls; speech uses native Sherpa only.

  • Add guided sixteen-prompt microphone corpus recording with replay, retakes, bounded PCM upload and exclusive WAV/manifest output.

  • Apply native Parakeet per-turn hotwords using official BPE configuration.

  • Release queued executor results after confirmed empty recognition; fence old decodes from consuming restarted audio.

  • Add five licensed cue candidates per state, terminal-driven WebAudio audition, and read-only persisted production choices.

  • Add portable fixed-key speech-engine shootout with gated browser microphone relay, ordered PCM playback and cancellation fencing.

  • Preserve unresolved executor requests and queue result announcements behind actual conversation boundaries.

  • Bound stalled voice replies, speak initialization fallbacks, and retire cancelled child sessions before replacement.

  • Keep recognition and corrections inside a stable transcript drawer, with real-timing-only highlighting and explicit unsupported capability labels.

  • Speak long-reply handoff notices before acknowledging playback completion; fence stale synthesis and interrupted turns.

  • Separate accessible voice diagnostics from executor detail, including actual runtime identity and observed latency intervals.

  • Add CPU-headroom desktop thread defaults and a native model-backed batch runner with truthful verification gaps.

  • Add selectable offline Parakeet TDT asset validation and VAD-segment recognition, while rejecting unsupported Nemotron streaming selection clearly.

  • Add bounded local WebRTC inbound bytes and audio-energy diagnostics alongside media-element state; these distinguish receiver flow from scheduled PCM and unknown physical audibility.

  • Add hands-free local CALL flow with automatic speech turns, interruption, mute, end, and streamed local speech output.

  • Preserve fast GPT automatic voice-model selection.

  • Default local TTS setup to verified Piper vits-piper-en_US-ljspeech-medium assets.

  • Fence voice finals at server boundary so reset, close, and duplicate submissions cannot dispatch stale speech.

  • Replace browser energy turn control with pinned neural Silero VAD in persistent local ASR worker.

  • Add a reduced-motion-safe voice waveform, focused executor depth panel, accessible transcript back handling, and local scanner waiting audio that fades for Piper PCM playback.

  • Debounce local ASR finals into bounded continuation batches, holding resumed speech for at most two seconds while preserving authenticated generation and request-ID validation.

  • Add explicit selected local TTS identity with verified Sherpa Kokoro af_nicole speaker maps, while retaining Piper LJSpeech default.

  • Keep scanner progress through thinking and synthesis, with distinct modes and a 120 ms fade before PCM playback.

web

web: 0.0.3

  • Persist capped fetch results.
  • Show local image preview in understand_image.
  • Fix win32 release readiness.
  • Retire superpowers conflict stack.

web: 0.0.2

  • Source unchanged; no extensions/web commits since v0.0.1
  • Tests unchanged; bun test ./__tests__/*.test.ts still applies
  • Docs unchanged; README still lists web_fetch, web_search, understand_image

web: 0.1.0 - 2026-05-21

  • a40a427 prepare extensions for npm release
  • 133cb7d chore(pi): migrate extensions to earendil packages
  • 2161a15 fix(web): use Z.AI MCP search
  • db7b1c5 Update agent and terminal configs
  • 5ca1296 Rework Pi agent extensions
  • 2157199 test(pi): add mux harness and MiniMax/web tool tests
  • b045a6b fix(pi/mux): cooldown exhausted keys on provider usage-limit errors
  • b87a61a feat(pi): monorepo workspace — all extensions are proper packages
  • 5fbf178 pi(ext): add remaining extensions (angel, animations, code-actions, files-widget, ghost, git-safe, neko, prune, web, etc.)
# Changelog

## Unreleased

- Label Pi's working loader with live Angel/Devil progress, such as `Devil · Investigating with grep`, for tool calls and automatic consultations; restore the default label when they finish.

- Treat each persisted session as its own usage ledger: PiPi no longer reports persisted child usage on its tool result or writes `pipi-usage-receipt` entries, and Usage and Cost drop all PiPi-specific accounting.
  Parent footer and `/session` totals now exclude delegated work; PiPi results still report child usage.
  History recorded before this change counts PiPi child work twice in Usage and Cost.
- Usage and Cost scan the same session stores as `pi --resume`: every project of Pi's default store, or only the configured flat `sessionDir`, instead of everything beside it.
  `SessionManager.listAll` itself took about 14 s and 465 MB for 5,368 sessions, so both enumerate the stores directly.
- Name new and forked PiPi child sessions `PiPi · <agent, tier, or model> · <prompt>` so `pi --resume` shows them recognisably.

- Remove Handoff, Open, and uv, their generated packages, workspace registrations, and Handoff demo verifier.
  Preserve historical records.

- Remove Gratis, its generated package, workspace registrations, and Memory live-task dependency.
  Preserve historical records and generic virtual-provider support; Memory live tasks now use Pi's configured model or an explicit `--model`.

- List `git_clone_safe` in codemode with its guidance instead of deferring it.
  Deferred exposure hid "Use git_clone_safe instead of raw git clone" from the model since 2026-10-12, and agents fell back to raw clones.

- Bound every pipi call with a wall-clock deadline (`pipi.timeoutMs` setting or `PIPI_TIMEOUT_MS`, default 4 hours).
  An expired deadline aborts the child, preserves its resumable session, and bounds child teardown at 5 seconds.
  Per-call `timeoutMs` overrides the setting; `null` opts out.

- Remove Relop's runtime, build/settings registrations, and workspace dependency; remove remaining Relop/Ultra build outputs.
  Preserve historical specifications, research, session records, and user settings; ordinary PiPi delegation remains available while the dedicated peer-review skill is planned.

- Modernize against Pi 1.5.1 while preserving minimum Pi 1.4.0: refresh workspace Pi dependencies, honor saved distrust in context measurements, resolve slash-containing provider IDs exactly, and distinguish requested versus served models in accounting.
  Use canonical edited context for handoffs, report focused custom dialogs as waiting for input, accept CR-only MCP SSE, and enforce RTK/Git-probe output bounds independently of host Pi version.
  Settle fixture tools before shutdown and launch generated browser helpers outside project cwd for deep WSL paths.
  Retain specialized extension workflows; idle cache-warming footer refresh remains unchanged because no public extension completion event exists.

- Deliver Ephemlet choice and text answers usefully: steer a working agent before its next model call instead of queueing them for the next user prompt, and while idle hold them in the extension so only the latest revision joins the next run as one message.
  Inline text answers and the editor now wrap inside themed field rows instead of truncating or scrolling on the terminal background; the end-of-text cursor survives row breaks.

- Make the Activity widget an attention surface: show active work plus unacknowledged failures, hide finished successes, and disappear when nothing needs attention.
  Acknowledge failures by closing their inspector or with `d`/`D` in the `/activities` picker; hidden records stay inspectable and agent-readable.
  Replace the static `inspect: /activities` hint with a padded Nerd Font header icon coloured by state, and align rows under the header text.

- Give Relop native compact/expanded tool views with grouped findings, inherited theme backgrounds, safe actionable agent text, explicit finding pagination, ordinary status retrieval of complete finding fields/limitations, truthful immutable-scope recovery, safe domain-error explanations, and actual scope/selection/round/snapshot-read progress.
  Verify registered tool behavior, both native themes and remapped expand bindings, terminal-control injection, background inheritance including ellipsis, native-container inspection, and regular-file-to-FIFO acquisition races without live model spend.

- Tier MiniMax M Plan spending: image generation, understanding, speech, and MiniMax search run without dialogs; H3 submission asks once or for the session unless user/env `minimax.videoApproval` is `always`.
  Replace `minimax.allowCreditsOverflow` and `web.minimax.mPlanCreditsApproved` without compatibility shims; Web's automatic chain includes MiniMax again for valid M Plan keys.
  Render MiniMax tools like built-in/web rows, send concise model text, and return `structuredContent` for codemode.
  Report remaining shared M Plan allowance after consuming operations and emit `quota:provider-activity`, so Quota shows MiniMax beside another active chat provider.
  Harden MiniMax quota parsing for reported window lengths, unlimited and not-in-plan pools, weekly boosts, and ambiguous count direction.

- Implement MiniMax M Plan capabilities (PX-SPEC-1RR4LU8M): preserve built-in chat while adding native image generation, explicit image understanding, speech artifacts, local capability discovery, and session-owned resumable H3 jobs.
  Require human per-operation consent or user/env spending authorization, reject PAYG/ambiguous or mismatched regional credentials, prevent automatic consuming retries, and keep unverified conditional APIs unavailable.
  Bound HTTPS artifact downloads with public-address pinning, redirect rejection, signed-URL redaction, and exclusive output creation; preserve partial image estimates and distinguish cancellation from terminal-record deletion.
  Add explicit MiniMax-only Web search without fallback, replace arbitrary host overrides with actionable removal guidance, and retire unverified CN quota routing instead of retaining legacy endpoints.
  Include native media usage in cost/history surfaces and persist PiPi contribution receipts so mixed codemode/media/child aggregates retain media estimates without recounting child-session spending.
  Verify relocated MiniMax execution against installed Pi 1.4.0 and 1.5.0 with deterministic HTTP; live extension acceptance, account tier, Credits attribution, and conditional capabilities remain unverified.

- Use the verified current international MiniMax M Plan quota endpoint without legacy-route fallback.
  Preserve reported percentages when absolute counts are zero placeholders, and treat the general pool as account-wide so the most-consumed shared window drives the footer.
  Verify API-key credential resolution, reset visibility, truthful unknown values, and preserved measured/scoped allowances through parser and real Pi command regressions.

- Repair Windows suite fixtures: isolate home discovery, close fixture-owned SQLite handles, use portable path assertions, and match Canvas startup diagnostics to the selected browser executable.
  Exercise native LSP navigation through a supported TypeScript launcher, preserve Go peer coverage and POSIX escalation checks, and verify bounded Windows process exit without assuming POSIX signal reporting.
  Separate settings-precedence cases from short-deadline timing and pass release-version input through Mise's parsed environment value instead of POSIX shell quoting; verify literal POSIX/cmd injection payloads leave package metadata untouched.

- Implement Relop Independent Review Loop (PX-SPEC-R8L4Q2NV): agent-callable scoped snapshot review through isolated parent-linked Pi children, ordered independent reviewer pools, free-only OpenRouter, structured stable findings, explicit readiness limits, and caller-owned fixes/delivery.
  Persist settings, deadlines, consumed invocation reservations, findings, and substitutions across fixes, waits, retries, tree navigation, and reload; abort active work on cancellation, deadline, and shutdown.
  Register standalone codegen/settings/debug, Mise tasks, package relocation, README configuration, and bounded-context measurement; verify alias equivalence, adversarial scoped tools, artifact races/FIFOs, persistent crash reservations, and terminal outcomes.
  Reject short regular-file acquisitions without claiming prefix coverage; immediately exhaust nonclean final-round results while retaining failure causes, coverage limitations, and outstanding findings, without overriding clean reviews or cancellation.
  Recover persisted loops in one pass without intermediate history arrays or spread arguments; preserve latest matching state, maximum consumed rounds across branches, and snapshot isolation.

- Remove Ultra, its workflow runtime, bundled authoring guidance, tasks, and build artifacts.
  Retain PiPi as the child-session delegation primitive without migrating Ultra's workflow engine.

- Report Devdocs refresh-slot saturation as a transient `busy` source reason instead of silently skipping uncached local documentation.
  Retain the two-job bound and verify immediate local recovery after both remote refreshes complete, without adding an artificial retry cooldown.

- Preserve typed upstream HTTP failures and share status/`Retry-After` decoding with fixed-provider networking so ordinary retrieval honors server delays without exposing response bodies.
  Verify one-hour 429 delta-seconds and 503 HTTP-date deferrals, exact-deadline resumption, short-delay minimums, and missing/invalid-header fallback against the real adapter and SQLite-backed engine.

- Reject Devdocs retrievals whose version differs from the discovered cache identity, including compatible-but-different releases and lost version evidence; preserve prior content/refs and release refresh leases on rejection.
  Add explicit/implicit-version drift regressions, document the existing ten-hit search cap in the tool and README, and verify narrower queries can reach omitted retained sections.

- Reduce Devdocs cached-read allocations and repeated JSON serialization with an array-free Unicode counter, a safe penultimate-prefix lower bound, and no duplicate partial-content join.
  Preserve original budget/continuation decisions; verify 1,683 differential boundary/error/Unicode cases and add 1/2/20-ref aggregate-boundary regressions.
  Add `//extensions/devdocs:bench` using disposable SQLite caches and optional native CPU profiles.
  On Linux/Node 26.13.0, five alternating processes per revision with ten warmups and thirty timed reads each reduce 20-ref ASCII/Unicode medians from 32.84/65.09 ms to 6.98/12.32 ms; paired bootstrap 95% speedup intervals are 4.68–4.75×/5.24–5.34× after within-process IQR filtering, with similar unfiltered gains.

- Implement Devdocs Technical Documentation (PX-SPEC-D7V3KQ9M): add native Pi search/read/status/refresh with version evidence, provenance, serialized-output budgets, continuation refs, and a portable personal database-path override.
  Isolate SQLite/FTS work in an extension-owned native child, publish revisions atomically, coordinate refresh leases, retain offline discovery, enforce source/context namespaces, and preserve prior content across failed updates.
  Add replaceable local, DevDocs, upstream, and Context7 adapters, platform-neutral source catalog, registered files/trees/HTTPS sources, bounded static and archive readers, and index-to-body resolution.
  Verify real Pi settings/trust, query-isolated persistent caches, corruption recovery, capacity and lease fencing, live anonymous source retrieval/offline restarts, and relocated package/worker execution.

- Share LSP location/symbol byte-budget admission and count each item's encoded JSON once instead of serializing every growing result prefix.
  Preserve exact UTF-8 envelope/comma costs, result order, truncation flags and final output validation; cover exact-boundary admission with Unicode/escaping and both initial truncation states.

- Complete Agent LSP Navigation (PX-SPEC-LSP4N8Q2, PX-PLAN-LSPCP6R4): add types, implementations, document/workspace symbols, and one-hop incoming/outgoing calls with explicit ambiguous-target selection and stale-source rejection.
  Curate Go, TypeScript and C++ profiles through an explicit pinned-Helix refresh with provenance, licensed source, reviewed overrides, conflicts, atomic publication and last-good preservation; runtime never fetches catalogs or installs servers.
  Route TypeScript semantic requests through the semantic server rather than its partial-semantic startup fallback; require an external SDK containing tsserver.js.
  Resolve supported Windows npm shims through validated adjacent metadata and direct Node launch without a shell; retain unknown server completeness/freshness, bounded results and owned-process cleanup.

- Retry Web fetches refused with 403, 429, or 5xx through Z.AI's web reader when a Z.AI key exists, for Markdown and text formats only.
  Gate the fallback on a status-tagged refusal so safety-rejected, oversized, aborted, and redirect-failed URLs never reach Z.AI; share the fetch deadline, enforce `maxBytes` on reader output, report both failures, and mark results with `source: "zai"`.
  Share Z.AI key resolution, MCP error mapping, and nested-JSON decoding between search and reader; search order is unchanged.

- Declare Loop's `loop_begin` and `loop_control` tools model-only so codemode scripts cannot start, pause, or resume runs; the model still calls them directly.

- Roll generated extension debug logs at 2 MiB without dropping the triggering event; retain the newest three segments plus protected active writers, using closed-file markers and conservative PID checks across concurrent sessions.
  Preserve private JSONL files, schema-bound metadata, nonfatal diagnostics, and standalone extension packaging without handwritten extension changes.

- Add LSP MVP for bounded Go definition, references, and hover using installed gopls, typed protocol connections, reviewed pinned Helix provenance, trusted timeout settings, disk synchronization, explicit uncertainty, and owned-process cleanup.
  Verify native Pi tool use against relocated built output and gopls 0.24, including external managed/unopened dependency edits; keep registry refresh and remaining operations deferred.
  Reduce evidence-path allocation and I/O by sharing bounded line extraction and reusing one last-readable source snapshot within each normalization, never across queries.
  Add reproducible profiled first/late-line benchmarks and regressions for duplicate freshness, mixed files, CRLF/Unicode boundaries, and cancellation during repeated hits.
  Verify actual Pi TUI/model/gopls tool traces through external edits, Unicode correction, truncation, expected failures, reload, and shutdown.

- Move Activity status above Pi's editor using a native non-focusable widget with bounded task rows, active-work prioritization, failure and overflow counts, and compact retained history.
  Suppress output-only redraws, preserve RPC exclusion and session cleanup, sanitize title controls, and cover rendering widths plus extension-level debouncing and unchanged-count state transitions.
  Add selected-task stopping in the picker and non-terminal inspectors with truthful pending/error feedback; preserve attached-terminal input and reserve narrow-row space for kind/exit metadata while retaining both title ends.
  Keep retained failures from hiding active jobs, expose a terminal-safe stop path, and show process/terminal elapsed time that freezes at the final event.
  Preserve unexpected process termination signals in failure details, distinguish hidden-row counts from state totals, mark clipped non-terminal output, and use readable native text roles for essential metadata and actions.
  Reserve a separate failure-attention row without evicting three active jobs, mark clipped captions and omitted output tails, and keep terminal-safe stop guidance readable at narrow widths.

- Bundle System's browser module graph into one self-contained entry through an extension-local build opt-out, eliminating lazy Mermaid fetches during navigation without lifecycle queues or delays.
  Preserve strict rendering-error reporting; cover initialized lifecycle transitions, entry-load failures, cancellation, diagram families, and unchanged lazy behavior in other build targets.
- Scope System browser-health cancellation handling to pre-navigation requests and distinct confirmed replacement documents; preserve active errors, failed navigations, ambiguous contexts, and original health assertions, with native browser and old-document lifecycle regression coverage.
- Retain failure-only Canvas/System browser diagnostics in CI logs: bounded request/navigation timelines, allowlisted resource paths, range status, and numeric media state; exclude headers, bodies, query strings, DOM, and screenshots while preserving original assertions and failure status.
- Surface bounded, scrubbed Canvas Chromium startup exit evidence, including executable, effective UID, exit code/signal, and stderr head/tail, without weakening sandbox flags or changing launch behavior; preserve initial explanations and final markers across multi-chunk output exceeding the rolling buffer, with credential/profile redaction.
- Revalidate npm registry metadata when baking exact-version external dependencies so a stale shared cache cannot reject newly locked versions; retain exact pins, installation safeguards, and a real cached-registry regression fixture.
- Repair clean-checkout CI tests: create release archive fixtures under OS temp rather than assuming `.pi/tmp` exists, and wait for Good-job processing-claim cleanup before asserting completed queue state while retaining runtime-only authentication assertions.
- Add Mockup multipage flows, keyboard navigation, page-scoped annotations, and capped headless previews; ship declared Playwright runtime files, expose browser provisioning through Mise, block preview fragment networking, preserve cancellation, avoid cache filename collisions, and correct scrolled pin geometry and stale note bubbles.
- Add Keysmash prompt-editor list/task/quote continuation, conditional nesting, indentation, and bracket pairing; preserve native completion, app bindings, paste streams, fenced code, and foreign editor ownership, with stale-completion and asynchronous redraw protection.
- Return Ephemlet waits immediately for already-answered requested fields while preserving selected-field filtering; cover an answered target alongside an unrelated open question, and remove unnecessary list-before-update guidance.
- Refine PiPi helper routing and reusable-agent lifecycle guidance: inspect preset purpose and authority, keep task overrides local, validate authorized adaptation on fresh calls and continuations, and preserve one-off isolation; add concrete native-result caching, text/provider-error projection, and recovery without regeneration, plus explicit model-family selection and neutral role instructions for review panels.
- Add the compact PiPi Round Table skill for independent, budgeted multi-model judging with anchored criteria, primary-evidence checks, explicit uncertainty, and preserved dissent; include optional research limitations and refine PiPi workflow handoff, recovery, and parent-verification guidance without changing native result contracts.
- Clarify Activity agent guidance: requested follow-ups for long jobs use notifications, bounded waits serve immediate dependencies and redraws, monitoring prefers suitable events, and owned-work cleanup uses stop plus final-status verification; preserve silent defaults and permission boundaries, and verify guidance delivery through inline and discoverable codemode paths.
- Isolate release snapshot filesystem-mock tests so shared-worker module caching cannot bypass symlink-read and stage-cleanup observations; retain all security assertions.
- Prefer host-injected renderer keybinding hints, including lazy, live-preview, error, and shared-helper paths, while retaining Pi 1.4 fallbacks; inline question, consent, picker, and inspection flows now declare native waiting status.
- Preserve provider-billed cost provenance through PiPi, Ultra, Cost, and Usage, distinguish estimated or mixed totals, honor billed zeroes, and avoid labeling empty aggregates as billed; retain catalog-priced model evaluation.
- Add explicit `/prompt-autopsy export` for native historical HTML with prompts and tool updates, isolated from caller settings in private temporary storage; retain the current-prompt editor command and shortcut.
- Package optional root consumer `RELEASE_NOTES.md` for one native collection update notice; disable fallback changelog lookup when absent, without publishing internal engineering history or creating extension-local changelogs.
- Bound RTK version/rewrite and Loop Git-ignore helper output, reject truncated results, and reconcile Ephemlets demo reading pauses with the six-second contract without weakening assertions.

- Track latest dependency releases with wildcard ranges and reproducible locks; refresh Pi to 1.5.0, stable TypeScript to 7.0.2, root development TypeBox to 1.4.4, Mermaid converter to 3.6.0, xterm headless to 6.0.0, and PTY to 1.2.0-beta.18.
  Keep Canvas React runtime and type packages within major 18 because Excalidraw's pinned Radix subtree excludes React 19; record remaining constraints in DEPENDENCIES.md.
- Replace only typed-rest-client 2.3.1 with implementation-identical 2.3.0 through a version-scoped override, permitting patched qs 6.17.1 within Stryker's supported dependency range without accepting audit's incompatible Excalidraw and remark-math downgrades.
- Adapt enum inference and renderer test contexts to Pi 1.5, preserve local debug export types under stable TypeScript's isolated declarations, and remove the converter's obsolete startOnLoad option.
  Update package-discovery expectations for Pi 1.5 and count cached prompt input in Dilemma's faux-provider accounting test.

- Autopsy adopts the accepted transcript-first Nugu browser layout: compact scope totals, hidden Filters/Insights/Session panels, theme selection, and no primary-view helper-label clutter.
  Replace corpus-wide preview and 10,000-entry omission caps with bounded server pages, stable ID cursors, bidirectional infinite scrolling, viewport-adaptive row windowing, retained active readers, and direct graph-to-entry lookup.
  Search complete readable text and JSON; aggregate cheap accounting over the entire matching history and load detailed metric streams and runtime captures only on demand.
  Generate extension-owned browser CSS from the canonical Nugu source; verify real 300-entry relocated history through an initially underfilled viewport, page pruning/recovery, complete-on-click content, unchanged scope totals, and light/dark desktop/mobile screenshots.

- Autopsy now shows bounded inline previews and streams complete stored text and JSON on entry/disclosure open, with bounded chunks, backpressure, explicit completion/error events, lazy original-index image references, retry, and cancellation.
  Track actual preview clipping instead of scanning content for truncation-marker strings; remove per-entry warning spam and present the session-history caveat as informational text.
  Reuse visible rows without discarding open content, scroll, or selection; update derived measurements separately and block queued disclosure events from starting detached-row requests.
  Verify large text/signatures, deep/wide JSON, Unicode boundaries, HTTP disconnects, interrupted streams, stale-toggle disposal, changed derived metrics, and relocated Chromium flows.

- Add Autopsy's usage dashboard with stored/current-catalog cost comparisons, token/category breakdowns, measured response/tool durations, original text/image sizes, provider/model/date filters, and selectable actual-time graphs with keyboard-accessible cross-page entry drill-down.
  Preserve absent quantities as unknown with subtotal coverage, separate timestamp spans/gaps from latency, apply request-wide price tiers and one-hour cache-write rates, and keep original image counts/previews independent of truncated JSON previews.
  Bound metric scans and chart sampling; verify nullable usage, pricing thresholds, size limits, shared filters, graph selection, desktop/mobile layouts, and relocated Chromium workflows.

- Highlight Autopsy's expanded entry JSON with theme-aware colors for keys, strings, numbers, and boolean/null literals while preserving exact text and treating all source content as inert text.

- Autopsy renders session image blocks through authenticated on-demand local image routes, with byte and declared-dimension limits, native-size toggles, and preview teardown on collapse.
  Preserve provider signatures and signature-named tool fields in bounded entry JSON; collapse JSON/provider metadata by default instead of stripping it.
  Verify image delivery, deferred browser loading, native-size inspection, collapse, and oversized rejection through HTTP and the relocated-package Chromium workflow.

- Rename `prompt-autopsy` to `autopsy`: extension directory, package identity, slash command `/autopsy`, browser title, Mise tasks, and debug environment namespace (`PI_AUTOPSY_DEBUG`).
  Keep `Ctrl+Alt+A`; do not retain the old command alias.

- Supersede prompt-autopsy's external-editor workflow and native historical HTML export with a token-protected loopback browser inspector for the current Pi session: search, role/type/branch filters, reported-usage bars and totals, readable entries, sanitized JSON, and exact/grouped system-prompt captures.
  Preserve the post-before-agent-start capture boundary, retain bounded runtime-local prompt history with explicit expiry/truncation notices, separate session history from provider request context, and stop the inspector on shutdown or reload.
  Add real-Pi capture-to-endpoint coverage and a relocated-package Chromium browser check; remove superseded editor-handoff automation.

- Deepen the loop vixen's animation craft with classic principles: anticipation before pose changes, arced gestures with follow-through, breathing with chest rise, counterbalanced hip-bust sway, heavier hair inertia, grounded foot contact, and soft elbow clearance that keeps arms from clipping her torso.
  Smooth camera wrap and kiss recovery, fix a reachable peek elbow flip, and keep her tasteful bikini pin-up style, palette, exports, and render budget unchanged.

- Render `/loop view` goals as normal-weight Markdown instead of one bold heading, with readable lists, code, tables, links, and contained wrapping in both Nugu themes.
  Escape raw HTML, suppress automatic image loads, allow only explicit HTTP(S)/mailto links, cache the last rendered goal per observer, and preserve text selection across live gate updates.

- Add a safe-for-work loop kitten selected whenever `PI_LOOP_ANIM` is present, including empty, `0`, `false`, or arbitrary values; resolve trusted `loop.anim` through the extension-local settings helper and select the kitten with a warning for invalid configuration.
  Draw an original copper-and-cream plush tabby with loop-shaped yarn and gate-watching, working, puzzled, curled, and proud reactions, without runtime imports from AFK or constructing vixen frames on the safe-for-work path.
  Preserve bottom-right portrait geometry, progress, keyboard focus, animation visibility toggles, and timer/overlay teardown across resume and reload.

- Anchor the loop vixen bottom-right and enlarge painted pixels 1.5× before half-block packing, producing a tightly cropped portrait up to 48 columns by 23 rows without changing its camera motion.
  Leave nine rows clear for progress, editor, and footer; hide rather than shrink the portrait below 60 columns or 32 rows.

- Move the loop vixen from inline layout into a tightly cropped, keyboard-noncapturing top-right portrait overlay, retaining natural raster scale and camera movement while reserving zero transcript or editor rows.
  Keep progress and counters in a separate inline widget of at most three rows; preserve `/loop vixen [on|off]` and remove owned overlays and timers on toggle, stop, afterglow, resize, and teardown without dismissing other overlays.
  Verify native regular/fullscreen compositors and keyboard focus; document native rectangular opacity and fullscreen mouse interception instead of patching Pi's compositor.

- Combine loop progress and vixen into one aspect-aware above-editor widget: use a cropped natural-scale portrait beside status on wide terminals, above it on narrow terminals, and status only when a usable portrait cannot fit.
  Bound height to eight rows and normally one quarter of terminal height; keep attempts and turns only in the widget, remove duplicate loop footer status, and add `/loop vixen [on|off]` as a runtime-local animation toggle.
  Stop refresh timers for hidden or expired-success figures; document that terminal-background transparency cannot reveal Pi text underneath an inline widget.

- Show persistent loop progress above Pi's editor: baseline or gate-check phase, animated activity indicator, gate position, elapsed time during silent gates, sanitized output tail, and cancellation hint.
  Transition to agent waiting/working, passed, paused, or stopped without implying a percentage or completion estimate; dispose refresh timers on widget removal and session teardown.

- Restore loop runs by Pi session identity as paused diagnostics, preserving goals, gates, budgets, counters, last activity, and completed results without automatic execution.
  Fence run ownership across Pi processes, detect abandoned owners and surviving gate processes, and require explicit offline confirmation before recovering legacy runs without ownership metadata.
  Run command-started baselines in the background so status, view, and pause remain responsive; show running-gate elapsed time and bounded live output, drain cancellation before session teardown, and fence stale callbacks.
  Match interrupted gates by attempt and gate index, and publish durable restore events so other observers see recovered state without rewriting the original activity timestamp.

- Nushell keeps `bash` and `powershell` displaced on resumed and reloaded sessions: when another extension (e.g. horst's per-host `+powershell`) re-adds a shell on top of the restored loadout before nushell's `session_start`, nushell now removes it again and records it for `/nushell off`, instead of returning early because `nu` was already active.

- Load the whole collection from one Pi package entry, `~/Workspace/pi-ext/dist`: every `//:build-ext` and `//:build-all` regenerates `dist/package.json`, listing built extension packages with `pfui` and `footer` first and lifting declared skills, prompts, and themes under `<name>/` (Pi reads nested manifests only for extensions). New extensions load after a build and `/reload` without a settings edit; narrow the collection with package filters such as `"extensions": ["!ultra/**"]`. `//:settings-entries` is removed. Spec PX-SPEC-DZLVGRW4 and plan PX-PLAN-RMJ8BIOX are amended accordingly. READMEs say `/reload` instead of restarting Pi after a rebuild.

- Add the loop companion: a red-haired pin-up in a pastel orange bikini rendered as supersampled truecolor half-blocks from hand-picked hue-shifted ramps below the editor in the TUI. She holds slow pin-up poses with breathing and a hip roll, two-bone IK keeps limb lengths fixed, and a slow camera glides between close-ups of face, bust and hips with zoom-gated soft-tissue follow-through; she waits with fingertips at her collarbone while gates run, looks unimpressed after failures, lounges while paused, stretches overhead and then blows kisses on success, and carries one subtle line of attempts, turns, and the failing gate. `//extensions/loop:preview` shows her without Pi.
- Tighten loop internals: SQLite statements are prepared once per session instead of per write, the run page's poll rebuilds a snapshot only when the store changed (live controller changes still push immediately), gate-pass debug spans live in the controller and also cover rejected passes, and duplicated tool rendering, error formatting, and run lookup code is shared.
- Loop TUI surfaces use Nerd Font Material Design icons: the status line shows autorenew with gate attempts, forum turns, and a timer while gates run, or pause/check/stop circles for paused, succeeded, and stopped runs; loop prompts, `loop_begin`, `loop_control`, and run notifications carry matching icons.
- Let a blocked loop agent pause and resume its run with the `loop_control` tool: the loop prompt asks it to work autonomously first and pause only when it truly needs the human; paused turns are ordinary conversation and spend no loop budget, and resume reruns the gates before looping again.

- Local `//:e2e` and `//:e2e-extension` trust the checkout's Mise config again: the isolated per-run `MISE_STATE_DIR` holds no trust records, so the runner now sets `MISE_TRUSTED_CONFIG_PATHS` to the repository root (appended to any inherited value).
- Biome skips `.system/**/assets`; spec assets are frozen human-review mockups, not shipped code.

- Add the `loop` extension (PX-SPEC-IJGV4GGK): `/loop start|guide|resume|pause|stop|status|view|list` runs an agent in a goal–gate loop until every configured gate (raw executable with args, env, cwd) exits 0 in one ordered pass; one attempt stops at the first failing gate, the injected loop prompt keeps a stable goal/gate/anti-cheat prefix with bounded output tails and older prompts collapse to stubs in context; minutes/turns/attempts budgets (`none` uncapped) pause the run, as do Escape during the baseline, typed `stop`/`pause`, aborted turns, and session switch/fork/tree/shutdown, and nothing auto-resumes; `/loop resume` keeps consumed turns/attempts and restarts the time budget; runs, gate attempts, and events persist in project-local SQLite (`.pi/loop/state.db`, `loop.statePath`, `PI_LOOP_STATE_PATH`) with control events outranking telemetry on queue overflow; `/loop guide` arms a brainstorm session whose `loop_begin` tool is rejected unless armed and advises on an unignored state file; `/loop view` serves a token-protected 127.0.0.1 page with SSE updates, Nugu light/dark palettes, run card, gate output drawer, timeline, events, history, and succeeded/paused banners. Root `//:e2e` now includes `//extensions/loop:e2e`.

- Harden clipboard image reads against malformed payloads and decode bombs: PNG inputs must carry a well-formed IHDR and IEND trailer, inputs and converted outputs are bounded to 33554432 pixels (16384 px per side) before passthrough or return, and oversized or malformed images fail explicitly without partial data.
- Bound clipboard backend calls with deadlines: native text, image, and file-reference reads fail after 3 s and copies after 10 s instead of hanging the sequential tool, and aborting mid-read now rejects promptly; losing timeout or abort paths never surface as unhandled rejections.

- Rewrite ephemlet agent guidance around when to create a panel and when not to, one panel per task updated at milestones, mentioning open questions in chat before ending a turn, and `wait` only for questions a panel already holds; the tool description now covers text fields, `wait`, and using the revision returned by `put`.

- Add ephemlet `text` fields edited inline in the panel and a `wait` action that blocks only the agent's tool call until awaited fields are answered (default every open question), times out (default 5 minutes, at most 30), or is aborted; waited fields show `agent waiting`, the human is notified, and the answer returns as the tool result without a duplicate answer message.

- Render ephemlet panels with theme background surfaces instead of box-drawing characters: header band with checklist progress, read-receipt and Nerd Font icon controls, banded question headers, focus and hover row tints blended in OKLCH, confirm answers as Yes/No segments, and struck-through human vetoes; palette-indexed theme colors are resolved through OSC 4 with plain-token fallbacks. Add the `ephemlets.icons` setting (`nerd` default, `unicode` fallback).
  Open questions scrolled out of view get a pinned `↓ N needs you` marker that jumps to them, the first `tab` focuses an open question, and new agent revisions start at the top.
  Question headers show their delivery before you answer: `instant` on a warm-tinted band for `now`, a muted `next turn` for `next`.

- Add ephemlet `choice` and `confirm` fields with per-field delivery: `ambient` answers wait for `get`, `next` answers join the next turn, `now` answers steer a running agent or start a turn; rapid answers coalesce into one visible `ephemlet-answer` message carrying its answer revision (Pi may deliver a queued older message after a newer one), open questions show `needs you` in the panel and indicator, confirmations notify once, and the read receipt moves from sent to read when the model context contains the answers.

- Add ephemlet checklist fields: the agent declares items and defaults, the human ticks them in the panel without starting a turn, human answers survive agent updates for unchanged IDs, `list`/`get` flag `unseen` changes, `get` with `include: ["answers"]` returns values with `by: agent|user`, and the panel header shows a read receipt.

- Remove the ephemlet panel's `alt+tab` button selection, which operating systems usually reserve; `ctrl+s`, `ctrl+w` and header clicks remain, and the panel hint now lists them.

- Dilemma's `parent_session` context view now follows Pi's context edits (omitted entries hidden, replacements shown) and the branch view marks them; system-role messages, which leaked the parent system prompt, plus context-edit and usage entries are no longer retrievable; Dilemma's own advice and status messages become retrievable.
- Add Dilemma `prior_consultations`: advisors inspect same-role earlier child sessions on the active parent branch, with `read` results marked stale when the file changed or vanished, and a count hint in the assignment.

- Rework memory into an agent-owned knowledge store: the `memory` tool takes raw SQLite (`sql`, `params`, `readonly`, `timeoutMs`) against one local database whose schema agents create and evolve; a seeded `_guide` and agent-maintained `_schema_docs` carry conventions; every call lands in an extension-owned audit journal with pre-write undo snapshots; `ATTACH`, `VACUUM INTO`, and ownership pragmas are blocked; oversized results spill to temp NDJSON. Removes the v1 review gate, pending inbox, action enum, and `memory.reviewShortcut`; `/memory` becomes a five-tab viewer (Usage, Size, Health, Browse, Journal) with journaled human edits and undo; v1 databases are refused instead of migrated, so delete the old file or point `memory.databasePath` at a new one.

- Reduce ephemlets to session-persistent Markdown reports: remove interactive native components, `choices`/`explorer` templates, background builds, `patch`, `state`, Submit and the `jiti` dependency; load the runtime eagerly.
  Restored legacy interactive records show a removal notice while their session data stays untouched.
  `purpose` now defaults to `title` on creation and `include: []` is accepted, removing the two most common first-call failures seen in archived sessions.

- Strip native child background colors from PiPi live previews, including the edit renderer's self-painted header, so the parent card background remains continuous.

- Route Good Job child analysis through the parent model registry so virtual selections such as `grts/auto` and runtime-only auth work without synthetic-provider credentials; retain thinking settings and bounded output.

- Explain context-extension net removal effects with actual request codemode catalog additions/removals; distinguish empty/unavailable catalogs and withhold savings on catalog-only control drift.
- Resolve managed Pi installations for context measurements, preferring an explicit managed root, then PATH, then conventional HOME fallbacks.

- Trim memory tool/schema descriptions and System board tool descriptions.

- Gratis routes to a built-in list of verified free-tier coding models (Kilo, OpenRouter pool, Mistral ministral, NVIDIA nemotron) plus user `gratis.models` patterns instead of zero-priced catalog entries; registers the Kilo gateway provider with a live free-model catalog; skips models with permanent failures for the session. Google Gemma and Mistral labs models are no longer built in.

- Pad component widgets above the editor (notes, dilemma, voice, voice setup) by one column on each side, matching Pi's padding of string-array widgets.

- Disable nested SDK tool shells inside PiPi through the public renderShell setting, preserving native content renderers and the parent card background.

- Keep PiPi's last useful preview across empty assistant frames instead of replacing it with running status; reserve a stable collapsed expansion-hint row through completion and synchronize thinking visibility on retained messages.

- Quota footer shows the account-wide allowance closest to exhaustion instead of always preferring the five-hour window; scoped limits and balances stay in the widget.
- Add OpenRouter quota (`GET /api/v1/key`): free-model daily requests with reset at next UTC midnight; paid keys add credit cap and daily/weekly/monthly spend.
- Add best-effort quota for Groq (requests per day from response headers), DeepSeek and Moonshot (balance), and Zai CN; xAI, Mistral, and NVIDIA expose no key-readable quota.
- Quota follows virtual model selections such as `grts/auto` to the physical provider of the latest successful response since the selection; response headers are ignored while routed.

- Render PiPi's current child assistant and tool activity with native SDK components, limiting collapsed previews to three lines and preserving expanded content, thinking visibility, child renderers, and Markdown transformers; isolate live UI state, filter unsafe terminal commands, and recover from broken preview components.

- Align quota widget bars, show multi-day resets in days, and drop units already named by the bucket label.

- Ignore empty Codex response-header window placeholders so absent secondary allowance does not reappear as an unused quota bucket.

- Replace quota's fixed five-hour/weekly schema with provider-owned dynamic buckets, isolated provider modules, shared detail/footer rendering, independent Zai measurements, scoped Anthropic limits, and Codex balance/spend-cap support; remove Google quota support.

- Preserve PiPi collapsed-header ellipsis backgrounds by clipping plain text before styling; verify native container backgrounds with terminal-cell assertions for pending, success, and error states.

- Keep collapsed PiPi call headers on one line with terminal-width-aware ellipsis truncation; retain complete prompts in expanded view.

- Show sanitized child-provider failure reasons in expanded PiPi results, preserving partial text and structured submissions.

- Display PiPi as π² and stream child tool names with bounded, expanded-only arguments; keep collapsed activity compact and reset previews between tool execution and answers.

- Classify PiPi child tools after extension startup: inherited hidden tools appear only in expanded Config; explicit hidden requests, inactive tools, and missing tools retain diagnostics.

- Parse Codex rate-limit windows by their declared duration, preserve unavailable windows, and render a weekly-only known quota without fabricating a five-hour window.

- Allow the activity demo up to 30 seconds to observe terminal cleanup after stop, preserving the required `stopped` lifecycle assertion.

- Bundle task-neutral PiPi workflow guidance for delegation, evidence-based handoffs, and capability-driven persistent progress views.

- Isolate Canvas Chromium-discovery module mocks in the serial test pool instead of the shared unit pool.

- Preserve PFUI prompt-section mutability for later extensions while retaining lazy rendering until explicit overrides; verify PFUI/PiPi instruction delivery and inherited-section precedence through the real SDK.
- Include actual child-extension error messages in PiPi diagnostics instead of reporting only hook identity.
- Keep Mockup annotation bubbles inside narrow and scrolled fragments; preserve pin anchoring, note editing, and placement across option changes and viewport resizing.

- Discover cached Playwright Chromium for Canvas layout in CI; start System video playback before requiring decoded frames and retain media diagnostics on failure.

- Add manual fail-closed Luci release preparation with version-lock, digest provenance, and leased release branches.

- Add pipi: blocking native child-session tasks, optional reusable JSON/Markdown presets, discoverable dynamic extension files, model-tier recommendations, continuation/forking, human-only streaming, and structured result submission with per-call native usage accounting.

- Unify System book, board, and document views behind one authenticated per-project localhost application with shared actions, navigation, live refresh, last-good recovery, and shutdown cleanup.
- Verify System book workflows in Chromium and Firefox across desktop/mobile and light/dark views; fix clipped search, repeated highlights, encoded previews, disclosure retention, proposal/readiness races, generated footnotes and accessibility references, fixed-header anchor offsets, and navigation-time Mermaid diagnostics while preserving authored HTML/SVG ownership.
- Bundle System Mermaid and MiniSearch as browser modules instead of copying upstream prebundles; refresh compatible locked dependencies and verify DOMPurify 3.4.17 in rebuilt Canvas and System assets.

- Protect good-job claim acquisition and recovery across same-process module reloads; reject stale quota/provider/cache results and out-of-order VCS refresh publication with controlled concurrency regressions.
- Validate Chrome command arguments and complete Chrome/IntelliJ subcommands; correct quoted-Python and Windows executable handling, runtime home-boundary discovery, portable resource URLs, and broken focused relocation task selectors.
- Remove unused Firefox BiDi bookkeeping; verify priority-provider request dispatch and TPS event/status behavior; read the configured usage store for in-memory sessions and include cached reads in the displayed token formula.
- Add real-Pi activity, AFK, ephemlet, usage, and synthetic quota demonstrations; refresh stale recordings and capture real canvas/mockup browser examples.
- Verify handoff replacement by native session IDs and reviewed editor text; add negative-control hook-safety and post-editor input checks so demos cannot pass from typed echoes or missing tools.

- Make Ultra's orchestrator responsible for subagent knowledge, instructions, skills, tools, and session continuity instead of minimizing context length; preserve risk-based verification separately.
- Increase Ultra action-summary output allowance from 40 to 256 tokens and Voice response allowance from 768 to 4,096 tokens without changing concise output instructions or timeout bounds.

- Reject failed or empty handoff generation before review or session replacement; distinguish failures from cancellation and ignore late provider settlement after the loader is cancelled.

- Disable backup storage on invalid explicit destination settings and automatic uploads on invalid opt-out settings instead of falling back; preserve manual backups with valid storage.
- Align cost session totals with Pi's full-session native usage, scan monthly responses by recorded time, retain distinct usage records, and suppress identifiable copied fork history without double-counting parent tool aggregates.
- Fail closed on failed or malformed Git tree inspection, verify unborn repositories before skipping checkout, cap combined subprocess output at 16 MiB, and retain active-target overlap protection across queued clone replacements.

- Rewrite the memory README around remember → review → recall and add a reproducible real-Pi demo showing proposal approval and scoped search with isolated synthetic data.
- Preserve authored final reading pauses in asciicast endpoints and GIF exports instead of silently using the renderer's three-second default; force an isolated memory DB even when the caller exports a user database override, with live isolation verification behind build-backed `demo-check` rather than normal unit tests.

- Add an interactive read-only memory search browser with live keyword queries, scoped approved results, complete detail scrolling, empty/error recovery, and no new transcript dumps; reuse review layout and preserve non-TUI search output.

- Replace memory review's raw-ID picker and ASCII inspector with a themed, searchable inbox and content-first preview; preserve revision-bound inspection, adapt selection and detail panes to short/narrow terminals, and add a disposable real-Pi preview task.

- Consolidate extension history into this internal changelog; remove extension-local changelogs and README links, and restrict changelog automation and validation to the root file.

- Restore configured codemode, tool search, and MCP capabilities in Ultra SDK children; builders retain child-default tools instead of requiring the parent's full tool registry.

## 0.1.10

- Refresh dependencies and require Pi 1.0.2 with updated extension contracts
- Add `/handoff` for moving context into a focused new session
- Run interactive Nushell bang commands by suspending Pi's terminal UI
- Allow manual masks to select tools and system-prompt instructions
- Refresh workspace branch status after external checkouts and update quota from provider response headers
- Harden Ultra workflow trust checks, inactive-tool preflight, and prompt interpolation
- Ship Klaus SDK dependencies per platform and load its provider lazily
- Rewrite all 41 extension guides around first use, with generated configuration references and deterministic terminal demos
- Modernize TypeScript tooling, scoped typechecking, structural checks, and test execution

## 0.1.9

- Publish release archives under the package server's public `/releases/` route

## 0.1.8

- Keep live board controls disabled until script binding and initial state synchronization complete
- Cover delayed state initialization in Chromium and Firefox without timing-based waits

## 0.1.7

- Expose the release checksum manifest at a flat CI artifact path
- Keep release archives gated by the complete CI suite and relocated package validation

## 0.1.6

- Keep FFmpeg active inside the nested Chrome E2E task
- Wait for live board script binding after document reloads before metadata interactions in browser tests

## 0.1.5

- Declare FFmpeg 9.0.2 as an E2E tool instead of relying on a host encoder
- Diagnose System board status failures at HTTP, disk, and rendered-site boundaries

## 0.1.4

- Refresh the System board's static output before acknowledging plan status changes
- Run headed Chrome acceptance tests under Xvfb in CI, with sandbox isolation confined to the test executable wrapper
- Reuse explicitly configured CI browser and download caches while keeping E2E HOME, agent data, and trust state isolated

## 0.1.3

- Persist Mise toolchains, downloads, Go build caches, and npm downloads across failed CI runs using durable cache mounts
- Keep ephemeral Mise trust state outside the source checkout and cached toolchain directories
- Remove Node startup latency from RTK timeout test fixtures

## 0.1.2

- Wait for bak's background initialization instead of assuming a fixed startup delay in settings tests
- Keep Firefox fake-child ownership tests independent of a host Firefox installation

## 0.1.1

- Exclude cached toolchains and temporary release output from Taplo checks
- Replace failed v0.1.0 CI candidate without moving its tag

## 0.1.0

- Build all 40 extensions as minified, self-contained packages with one repository version
- Validate relocated packages and exact archive bytes; publish release archives with SHA-256 manifests through Luci
- Ship one root installation package containing every extension, plus individual extension archives
- Exclude cached CI toolchains from source linting
- Raise web fetch and map visible-output limit to 100,000 characters and improve invalid-argument errors
- Optimize uv spawn-hook command checks and PATH rewriting; add benchmark, profiling, and comparison tasks
- Add uv blockers for `pip3` and Poetry commands with uv guidance
- Remove RTK `find -path` compatibility probing and guard; `find` rewrites now apply directly
- Cache successful RTK command rewrites per session, with benchmark and profiling tasks
- Harden dynamic-extension validator protocol handling and share workflow preparation across command and tool runs
- Identify Ultra child assignments by workflow, phase, and agent; attribute interpolated prior-agent output to its immediate producers, preserve provenance through fanout transformations, and invalidate checkpoints when attribution changes
- Precompile uv spawn-hook matchers and guidance, optimize POSIX PATH handling, and add spawn-hook benchmarks and profiling tasks
- Add uv spawn-hook microbenchmarks and optimize command matching and POSIX PATH rewriting, reducing measured benchmark time by 89%

## 0.0.3

- Rename autoresearch to tune; add jj backend, state migration, timeout clamps
- Rework diff-review to SSR reviews; add headless Vitest coverage
- Rework q extraction; add benchmark and 112 parser tests
- Retire superpowers; move workspace to npm, Vitest, dprint, Biome, tsc, mise

## 0.0.2

- Bump package.json from 0.0.1 to 0.0.2
- Update README install examples to git:github.com/bugabinga/pi-ext@v0.0.2
- Leave extension sources, READMEs, tests unchanged from v0.0.1

## 0.0.1

- Ship initial pi-ext set: goal, tldr, leaks, tps, quota, vcs-status
- Add goal loop: autocomplete, progress UI, standing prompts, judge-gated updates
- Add tldr live summaries; model footer; quota parsers; jj-aware workspace status
- Fix session startup races, stale telemetry, aux LLM calls, provider secret leaks

## Legacy extension history

Consolidated from former extension-local changelogs; original version labels and claims are preserved as historical notes, not verified release records.

### context-size

#### context-size: 0.0.3

- Release @bugabinga/pi-ext-context-size 0.0.3
- Resolve pi-ext conflict stack; retire superpowers references
- Replace Bun test wiring with Vitest workspace tooling
- Isolate extension harness env in tests
- Tighten Biome lint rules; remove low-risk explicit any

#### context-size: 0.0.2

- Release contains no context-size commits since v0.0.1
- Runtime still publishes `footer:segment`; status fallback remains
  `context-size`
- Harness test still checks extension load in Pi runtime

#### context-size: 0.1.0 - 2026-05-21

- a40a427 prepare extensions for npm release
- 6a97080 add asciinema demo workflow
- c1e9f71 fix(context-size):
  show visible usage bar
- 517ba54 pi:
  add extension footer fallbacks
- 133cb7d chore(pi):
  migrate extensions to earendil packages
- f144b06 fix(pi):
  restore extension notifications
- 5ca1296 Rework Pi agent extensions

### cost

#### cost: Unreleased

- Estimate zero-cost messages from Pi's model catalog through `calculateCost` instead of a hand-maintained `zai`/`minimax` price table; cache writes are now charged.

#### cost: 0.0.3

- Resolve pi-ext conflict stack; retire superpowers.
- Replace Bun test runner with Vitest workspace tooling.
- Isolate extension harness env; fix import.meta.dirname and chai matchers.
- Tighten Biome rules; replace low-risk explicit any usage.

#### cost: 0.0.2

- Source unchanged from v0.0.1
- README unchanged from v0.0.1
- Harness tests unchanged from v0.0.1

#### cost: 0.1.0 - 2026-05-21

- a40a427 prepare extensions for npm release
- 6a97080 add asciinema demo workflow
- 517ba54 pi:
  add extension footer fallbacks
- 133cb7d chore(pi):
  migrate extensions to earendil packages
- 5ca1296 Rework Pi agent extensions
- b87a61a feat(pi):
  monorepo workspace — all extensions are proper packages
- b408bba pi(ext):
  footer segment producers (context, timing, cost, runtime, git-status,
  pr-status, thinking)

### dilemma

#### dilemma: 0.0.3

- Replace Bun test harness with Vitest workspace setup
- Format markdown and code fences with dprint
- Tighten lint rules; remove low-risk explicit `any` usage
- Update extension guides and demos

#### dilemma: 0.0.2

- Extension has no commits after v0.0.1; release-only cut
- README.md and index.ts unchanged
- /angel, /angel-policy, hooks, and bash policy gate unchanged

#### dilemma: 0.1.0 - 2026-05-21

- a40a427 prepare extensions for npm release
- 133cb7d chore(pi):
  migrate extensions to earendil packages
- 5ca1296 Rework Pi agent extensions
- b87a61a feat(pi):
  monorepo workspace — all extensions are proper packages
- 5fbf178 pi(ext):
  add remaining extensions (angel, animations, code-actions, files-widget,
  ghost, git-safe, neko, prune, web, etc.)

### fast

#### fast: 0.0.3

- Own fast status in @bugabinga/pi-ext-fast
- Replace Bun test runner with Vitest across workspace
- Fix extension harness tests for Vitest deps and Chai matchers
- Migrate workspace tooling to mise; keep root Bun workspace

#### fast: 0.0.2

- Git history:
  no `extensions/fast` commits since `v0.0.1`
- Keep `fast.enable` settings contract unchanged
- Keep `serviceTier:
  "priority"` wrapper for `openai-codex` unchanged

### footer

#### footer: Unreleased

- Emit `footer:branch-change` with `{ branch }` when Pi reports a git branch change, replacing the no-op rerender subscription
- Forward branch changes while hidden as well; add `branch.change` debug event

#### footer: 0.0.3

- Replace Bun test runner with Vitest
- Migrate workspace tooling to mise and root Bun workspace
- Fix footer harness tests; isolate env, deps, and Chai matchers
- Tighten Biome lint; remove explicit any and non-null assertions

#### footer: 0.0.2

- Ship unchanged footer package; no source diff since v0.0.1
- Keep `/footer` and `ctrl+alt+f` toggle behavior unchanged
- Keep `footer:segment` event contract and 4-line footer layout unchanged

#### footer: 0.1.0 - 2026-05-21

- a40a427 prepare extensions for npm release
- 6a97080 add asciinema demo workflow
- 517ba54 pi:
  add extension footer fallbacks
- 133cb7d chore(pi):
  migrate extensions to earendil packages
- f144b06 fix(pi):
  restore extension notifications
- 5ca1296 Rework Pi agent extensions
- b87a61a feat(pi):
  monorepo workspace — all extensions are proper packages
- b18f0e2 refactor(pi/footer):
  clean sweep — kill dead code, extract helpers, name constants, type properly
- 972978f fix(pi/footer):
  force full redraw (requestRender(true)) to clear old footer on hide
- 0ea30fd fix(pi/footer):
  force render after swap to immediately clear old lines
- a49441f fix(pi/footer):
  swap footer via setFooter() on toggle for true 0-height when hidden
- 1cfa6a1 fix(pi/footer):
  import Key from pi-tui, not pi-coding-agent
- 933bcb6 pi(ext/footer):
  semantic zones, flow layout, 4-line stable height, /footer toggle

### git-safe

#### git-safe: Unreleased

- Remove the Bash Git editor policy: it injected Bash source into every shell tool and broke non-Bash shells such as Nushell
- Stop intercepting `tool_call`; git-safe no longer rewrites agent-issued shell commands
- Keep safe-clone behavior unchanged
- Format byte sizes with Pi's `formatSize` instead of two local formatters; size-limit errors and the clone result now use binary units (KB = 1024 B) and cap at MB

#### git-safe: 0.0.3

- Add sparse safe clone support.
- Recover tracked paths on session start.
- Replace Bun test runner with Vitest.
- Isolate extension harness environment in tests.

#### git-safe: 0.0.2

- Release 0.0.2 with no commits touching `extensions/git-safe`
- Keep `git_clone_safe(url, path, branch?)` behavior unchanged
- Keep README, harness test, and package metadata unchanged

#### git-safe: 0.1.0 - 2026-05-21

- a40a427 prepare extensions for npm release
- 133cb7d chore(pi):
  migrate extensions to earendil packages
- 5ca1296 Rework Pi agent extensions
- b045a6b fix(pi/mux):
  cooldown exhausted keys on provider usage-limit errors
- b87a61a feat(pi):
  monorepo workspace — all extensions are proper packages
- 5fbf178 pi(ext):
  add remaining extensions (angel, animations, code-actions, files-widget,
  ghost, git-safe, neko, prune, web, etc.)

### good-job

#### good-job: Unreleased

- Steering input (`event.streamingBehavior === "steer"`) no longer waits for runtime startup and queue writes before continuing to the running turn; the capture runs in the background and failures still notify.

### handoff

#### handoff: Unreleased

- Add `/handoff <goal>`: compaction-aware branch summary, editable generated prompt, new child session with the prompt in the editor.
- Generation prompt references artifacts by path or URL, redacts secrets and PII, and adds a `Suggested skills` section.

### masks

#### masks: Unreleased

- Add optional `tools` and `instructions` to mask settings, validated like other fields and reported when invalid
- Apply mask `tools` on equip, warning on unknown names and reusing the prompt-mask snapshot and rollback path
- Inject mask `instructions` as the `masks` system-prompt section via `before_agent_start`
- Let prompt frontmatter `tools` override mask `tools` for one prompt
- Show tools and a truncated instructions preview in picker descriptions; non-TUI selection now shows shortcuts too

#### masks: 0.0.3

- Add `/masks`, configurable picker shortcut, and per-mask shortcuts
- Fuzzy-match masks by name, model, thinking level, and shortcut
- Apply configured model and thinking level together
- Let prompt templates equip a mask for one agent run through `mask` frontmatter
- Restore the previous model and thinking level after masked prompt runs
- Validate global `masks` settings and report invalid entries

### mockup

#### mockup: Unreleased

- Extract the generic browser mockup tool from The System.
- Show browser URLs and opening failures in Pi's native expandable tool UI.
- Reopen disconnected reviews and recover abandoned iteration chains with preserved history.

### model-info

#### model-info: Unreleased

- Keep the model version in short labels: `gpt-6-astra` becomes `astra-6`, `claude-opus-5-5` becomes `opus-5.5`
- Read the thinking level from `pi.getThinkingLevel()` instead of scanning the session branch for `thinking_level_change` entries

#### model-info: 0.0.3

- Own `fast` status rendering in `model-info`
- Switch extension tests from Bun test to Vitest
- Isolate extension harness environment in tests
- Tighten Biome rules and remove low-risk explicit `any` usage

#### model-info: 0.0.2

- Mark 0.0.2; source, README, tests, lockfile match `v0.0.1`
- Keep `model` footer segment plus `model-info` status fallback
- Keep `fast.enable` badge for `openai-codex`
- Keep harness tests for extension load and label format

### nu-bang

#### nu-bang: Unreleased

- Run interactive commands (`!i <command>`, editors, pagers, editor-opening Git subcommands, known terminal UIs) with the terminal handed over while the TUI is suspended; outside the TUI they are refused with an explanatory result.
- Remove the unused standalone executor, output control stripping, and `pi-nu-bang-output-*.log` files; Pi sanitizes, truncates, and persists bang output itself.
- Correct the README behavior section to match the runtime path.

#### nu-bang: 0.0.3

- Fix win32 release readiness
- Replace shell scripts with cross-platform Lua scripts
- Add docx-to-md.lua, EPUB, Typst, Beamer, and pdf-report.yaml
- Isolate harness env and temp-dir cleanup tests

#### nu-bang: 0.0.2

- Source unchanged under `extensions/nu-bang` since `v0.0.1`.
- README lists hooks:
  `user_bash`, `session_start`, `session_shutdown`.
- Tests cover Nu exec, completion, cwd errors, truncation, temp-dir cleanup.

#### nu-bang: 0.1.0 - 2026-05-21

- a40a427 prepare extensions for npm release
- 133cb7d chore(pi):
  migrate extensions to earendil packages
- 3b453ee feat(pi):
  add nushell bang extension

### os-notifier

#### os-notifier: Unreleased

- Rename `terminal-notifier` to `os-notifier`.
- Prefer native Linux, macOS, and Windows notifications with OSC fallback.
- Notify only for direct interactive TUI runs completed while unfocused.
- Preserve rich metadata with natural urgency and bounded sanitized text.
- Remove the unused extension notification event bridge.
- Publish pane-scoped Pi identity, session name, and project through OSC 1337 during TUI sessions.
- Subscribe to `ui_prompt_start` and `ui_prompt_end` through Pi's typed `pi.on` overloads instead of a local shim type and cast.

#### os-notifier: 0.0.3

- Replace Bun test runner with Vitest in the workspace
- Add isolated extension harness env for tests
- Fix harness tests for import.meta.dirname, Vitest deps, and Chai matchers
- Migrate project tooling to mise

#### os-notifier: 0.0.2

- Release only:
  no commits touched `extensions/terminal-notifier` since `v0.0.1`
- Source unchanged:
  `agent_end` still sends OSC/Windows notify
- Tests unchanged:
  harness still verifies extension runtime load

#### os-notifier: 0.1.0 - 2026-05-21

- a40a427 prepare extensions for npm release
- 133cb7d chore(pi):
  migrate extensions to earendil packages
- f144b06 fix(pi):
  restore extension notifications

### prompt-autopsy

#### prompt-autopsy: Unreleased

- Honor Pi's `externalEditor` setting before `$VISUAL` and `$EDITOR` when opening the prompt file, via `SettingsManager.getExternalEditorCommand()`
- Fall back to Pi's default editor (`nano`, `notepad` on win32) instead of `vi` when nothing is configured

#### prompt-autopsy: 0.0.3

- Replace Bun test runner with Vitest in the workspace
- Isolate extension harness environment for tests
- Fix harness tests for import.meta.dirname, Vitest deps, and Chai matchers
- Format prompt-autopsy markdown docs

#### prompt-autopsy: 0.0.2

- Release @bugabinga/pi-ext-prompt-autopsy with no code delta from v0.0.1
- Keep `/prompt-autopsy` and `Ctrl+Alt+A` behavior unchanged
- Leave README, harness test, bench, and changelog unchanged

#### prompt-autopsy: 0.1.0 - 2026-05-21

- a40a427 prepare extensions for npm release
- 133cb7d chore(pi):
  migrate extensions to earendil packages
- 5ca1296 Rework Pi agent extensions

### quota

#### quota: Unreleased

- Update Anthropic and OpenAI Codex quota from provider response headers via `after_provider_response` instead of waiting for the poll.
- Read the Codex account id through Pi's typed `readStoredCredential` export.

#### quota: 0.0.3

- Display quota reset in parentheses.
- Replace shell scripts with cross-platform Pandoc Lua.
- Use `--highlight-style` for Pandoc 3.6.
- Add EPUB, Typst, Beamer, and Sandbox mode report output.
- Move workspace tests from Bun to Vitest.

#### quota: 0.0.2

- Source:
  no extension-local commits after v0.0.1
- Keep `/quota` widget, footer segment, and status fallback unchanged
- Keep provider quota parsers and harness tests unchanged

#### quota: 0.1.0 - 2026-05-21

- a40a427 prepare extensions for npm release
- 6a97080 add asciinema demo workflow
- 517ba54 pi:
  add extension footer fallbacks
- b924295 config:
  add kimi coding model
- 5ca1296 Rework Pi agent extensions
- b87a61a feat(pi):
  monorepo workspace — all extensions are proper packages

### rtk

#### rtk: Unreleased

- add minimal RTK bash rewrite hook
- warn once on load when RTK is missing or too old

### runtime

#### runtime: 0.0.3

- Switch tests from Bun to Vitest; fix harness imports and Chai matchers
- Isolate extension harness env for runtime tests
- Migrate workspace tooling to mise with a root Bun workspace
- Replace suite demos

#### runtime: 0.0.2

- Keep `index.ts` unchanged from `v0.0.1`
- Keep README, package metadata, changelog, and harness test unchanged
- Ship no command, tool, setting, or hook contract changes

#### runtime: 0.1.0 - 2026-05-21

- a40a427 prepare extensions for npm release
- 6a97080 add asciinema demo workflow
- 517ba54 pi:
  add extension footer fallbacks
- 133cb7d chore(pi):
  migrate extensions to earendil packages
- 5ca1296 Rework Pi agent extensions
- b87a61a feat(pi):
  monorepo workspace — all extensions are proper packages
- 0bf9702 pi(ext/runtime):
  update runtime extension
- b408bba pi(ext):
  footer segment producers (context, timing, cost, runtime, git-status,
  pr-status, thinking)

### session-rename

#### session-rename: Unreleased

- Read the current session name through `pi.getSessionName()` instead of a guarded `ctx.sessionManager.getSessionName()` wrapper

#### session-rename: 0.0.3

- Resolve pi-ext conflict stack; remove superpowers path
- Isolate extension harness env in tests
- Remove non-null assertions under stricter Biome rules
- Replace Bun test runner with Vitest across workspace

#### session-rename: 0.0.2

- Ship release-only version bump; no session-rename source changes
- Keep `/session-rename [name]` command behavior unchanged
- Keep session naming hooks and harness load test unchanged

#### session-rename: 0.1.0 - 2026-05-21

- a40a427 prepare extensions for npm release
- 133cb7d chore(pi):
  migrate extensions to earendil packages
- 5ca1296 Rework Pi agent extensions
- b87a61a feat(pi):
  monorepo workspace — all extensions are proper packages
- 5fbf178 pi(ext):
  add remaining extensions (angel, animations, code-actions, files-widget,
  ghost, git-safe, neko, prune, web, etc.)

### the-system

#### the-system: Unreleased

- Add reversible dropped board state for specs, plans, and issues, with DROP/RESTORE controls, inactive counts, and dropped-last ordering; dropped plans no longer block spec completion.

- Add an EDIT action to board cards that opens the indexed document in the configured editor and reports launch failures.

- Show specs in one board column, copy document IDs on click, and start `/system exec <ID>` from a separate card action in the current idle Pi session.
- Remove the board's behavioral subtitle and order specs unfinished first, then most recently modified first with stable ID ties.
- Register deferred System tools upfront without lifecycle-stage activation or consumption gates.
- Prevent board polling from interrupting pending approval requests or applying stale pre-write snapshots.

- Move the document, Markdown, board, and exec contracts out of every `/system` stage message into project-scoped skills (`the-system-documents`, `the-system-markdown`, `the-system-board`, `the-system-exec`), so repeated stages no longer duplicate several kilobytes of identical instructions into the session.
- Reshape the stage message to a heading with the stage and the human's words, then the board summary, then the stage prompt; the `Human instructions` header is gone.

- Cut `/system` index and render cost: frontmatter extraction no longer full-parses document bodies (index load -94%), and the renderer harvests search text from the same parse as the HTML pipeline instead of re-parsing every document (cold board render -7%).

- Resolve the active System root as the nearest `.system` directory at or above cwd; nested systems override parents, `/system init` initializes at cwd, and split Git/JJ roots no longer conflict.

- Use rich Markdown bundles with YAML graph metadata, config-only machine state,
  doctor validation, and file-level spec ownership.
- Add `/system board` with tracked schema-0 operational state, nested plan cards, independent issues, responsive accordions, and human-reviewed `/system assess` proposals.
- Rename plan status `done` to `approved`; remove issue status.
- Open rendered pages through the native Win32 URL handler without flashing a command window.
- Add turn-scoped `system_bundle_slot` allocation with title-derived slugs and repeatable ID reservation.
- Redefine phases as project-wide operating epochs selected through config, injected into agent context, and shown in Pi and book navigation.
- Reset the unreleased config schema to version 0; `/system init` updates machine schema markers mechanically, then offers agent-led bundle migration.
- Add initialized-project-only GFM, TeX, and Mermaid skills for already-selected System document formats.
- Keep generated `/system` lifecycle turns alive in print and JSON modes.
- Align bundle-slot and board-proposal prompt metadata and tool rendering with Pi defaults.
- Keep board proposals and repeatable bundle slots available across clarification turns and compaction.
- Add a generally invocable vertical-slicing skill and require it for `/system exec`.
- Treat plans as required outcome rather than implementation order, and remove slice boundaries as human approval points.
- Make visual models the System Markdown default, require visuals in specs and plans, and decompose plans by behavior instead of architectural layer.
- Let `/system exec` apply verified board states directly through `system_board_update` with `source: agent` and an evidence note, keeping `/system assess` proposal-only.
- Inject a board state block with counts and unfinished items into every `/system` lifecycle prompt.
- Compact the injected document contract without dropping rules.
- Replace unconditional core-document blocking with validated one-shot human approval.
- Let completed `/system exec` runs propose evidence-backed board updates for human review.
- Serve `/system board` pages from a live snapshot that re-renders when `.system` changes, so refreshes show persisted state and open boards follow disk edits.
- Toggle plan `status` and spec board state directly from the rendered metadata box.

#### the-system: 0.0.3

- Resolve pi-ext conflict stack; retire superpowers.
- Replace os.execute wrappers with pandoc.pipe argv lists in conversion scripts.
- Fix batch-convert patterns and keep web-to-md H1 content out of metadata.
- Add docx-to-md image extraction to ./media/, Decision Guide, pandoc lua -e
  examples.
- Document heading shift, CJK, images, empty titles, Lua filters, PDF engines.

### tps

#### tps: 0.0.3

- Replace Bun test runner with Vitest for extension tests
- Isolate extension harness env in tests
- Fix harness tests for import.meta.dirname, Vitest deps, and Chai matchers
- Tighten Biome rules and remove low-risk explicit `any` usage
- Move workspace tooling to mise

#### tps: 0.0.2

- Keep @bugabinga/pi-ext-tps source unchanged from v0.0.1
- Keep README hooks, footer segment, status fallback unchanged
- Keep bun tests for TPS math and pi-test-harness load unchanged

### ultra

#### ultra: Unreleased

- Add the `{item.FIELD}` interpolation token: one field of the current record fanout item, rendering null or missing fields as empty text in prompts and summaries and as `undefined` in selectors, with a validation error for non-record items.
- Treat braces in `prompt` and `summary` as interpolation tokens only when they start with `args.`, `item`, or `PHASE.results`/`PHASE.failures`; other braces, such as JSON fragments, TypeScript types, or empty braces, pass through as literal text, while malformed token-shaped spans still fail validation and selectors stay strict.
- Refuse a workflow before creating any child when a step without `flags` lists a tool that is registered but inactive in the parent Pi (for example `bash` under `pi --nushell`), naming the inactive tools, the child extensions with startup flags that may have deactivated them, and the active `nu` tool when a shell tool is inactive; builtins Pi leaves inactive by default are ignored.
- Report a deactivated builtin tool as `builtin tool, deactivated at startup` in the child-startup error instead of a bare name.
- Fail the not-yet-started siblings of a fanout phase immediately with `attempts: 0` once one child's requested tools were switched off at startup, instead of creating every sibling session to fail the same way.
- Request strict constrained sampling for `structured_output` only when the step declares a `schema`; against the schema-less property-free object schema it forced models to submit `{}`.
- Honor project workflows (`<cwd>/.pi/workflows/*.json`) only in trusted projects: untrusted project workflows are hidden from `/ultra run` completions, never shadow bundled or user workflows, and `run_workflow` and `/ultra run` refuse them with an error naming the file.
- Label each `/ultra run` completion with its workflow source (`bundled`, `user`, `project`).

### usage

#### usage: 0.0.3

- Retire superpowers conflict stack
- Move workspace tooling from bun to mise and vitest
- Isolate extension harness env in tests
- Remove non-null assertions for Biome lint

#### usage: 0.0.2

- Keep `/usage` dashboard unchanged from v0.0.1.
- Ship no `extensions/usage` diff:
  source, README, tests, assets unchanged.
- Retain harness smoke test for Pi extension runtime load.

#### usage: 0.1.0 - 2026-05-21

- a40a427 prepare extensions for npm release
- 133cb7d chore(pi):
  migrate extensions to earendil packages
- 5ca1296 Rework Pi agent extensions
- b87a61a feat(pi):
  monorepo workspace — all extensions are proper packages
- 5fbf178 pi(ext):
  add remaining extensions (angel, animations, code-actions, files-widget,
  ghost, git-safe, neko, prune, web, etc.)

### vcs-status

#### vcs-status: Unreleased

- Refresh the VCS segment on `footer:branch-change` using the latest hook context, so branch switches in other terminals no longer leave it stale
- Add `branch.change` debug event

#### vcs-status: 0.0.3

- Move workspace tooling to mise and root Bun workspace
- Replace Bun test runner with Vitest across workspace
- Fix harness tests: import.meta.dirname, deps, Chai matchers
- Isolate extension harness environment in tests
- Tighten Biome lint rules and fix violations

#### vcs-status: 0.0.2

- git diff v0.0.1 -- extensions/vcs-status:
  empty
- Source, README, tests:
  unchanged
- Runtime behavior:
  unchanged

#### vcs-status: 0.1.0 - 2026-05-21

- a40a427 prepare extensions for npm release
- 6a97080 add asciinema demo workflow
- 517ba54 pi:
  add extension footer fallbacks
- 133cb7d chore(pi):
  migrate extensions to earendil packages
- 5ca1296 Rework Pi agent extensions
- b87a61a feat(pi):
  monorepo workspace — all extensions are proper packages
- b408bba pi(ext):
  footer segment producers (context, timing, cost, runtime, git-status,
  pr-status, thinking)

### voice

Historical development entries include the retired browser implementation and superseded defaults.
See [Voice README](extensions/voice/README.md) for current behavior and availability.

#### voice: Unreleased

- Refine the call surface with human-approved quiet-luxury styling, a measured-PCM ribbon, visible status, captioned controls, quieter Diagnostics and a labelled Pi activity pill.
- Preserve control labels and focus across icon updates; honour reduced motion, root text scaling and compact recovery layouts.
- Clear the Resume action after ASR recovery completes instead of leaving stale controls visible.

- Add host-native CMake Moonshine builds for Linux, macOS and Windows, portable temporary workspaces, explicit Node/ORT inputs and platform-specific packaging; desktop runtime verification remains deferred.
- Load Windows native dependencies through UTF-16 paths, preserve ORT library aliases and validate the required ORT API before loading Moonshine.

- Add opt-in native CPU English and multilingual Nemotron 560/1120 ms, Moonshine Streaming Small/Medium, and Parakeet Realtime EOU choices without changing production defaults.
- Verify pinned model downloads, bundle Android arm64 Moonshine runtime with notices, and reject missing native assets or unsupported thread overrides before downloading weights.
- Supply Parakeet Realtime's complete checkpoint vocabulary without runtime conversion; hide EOU control tokens from transcripts without suppressing native predictor updates.
- Flush Nemotron's complete encoder window at forced EOF, preserving existing VAD, natural endpoint and submission policies.
- Cancel streaming model downloads on reset and fence stale startup work from restarted sessions.
- Flush Moonshine directly through its native stop contract and roll bounded stream context without ending long idle calls.

- Release acquired microphone tracks on failed worklet startup, clear recovered startup errors, and keep technical provider failures inside Diagnostics.
- Enforce long-reply handoff during actual playback, retain full accepted transcript text, and isolate duration/completion accounting per turn.
- Recover bounded playback acknowledgements, including stalled progress, rejected boundaries, silent notifications and PCM delivered before its receipt.
- Deliver idle executor completions at acknowledged boundaries; support deliberate silent skips, correlated result navigation and stable keyboard focus.
- Map real word timings to ordered transcript positions, preserve result actions while words highlight, and expose unavailable timing reasons without estimates.
- Preserve manual transcript scrollback and verify layered views on desktop, mobile, short landscape and enlarged text.

- Use human-approved paper, bubble, mechanical-loop and ambient cues as defaults after production-level comfort audition; preserve saved overrides and gains.

- Initialize native TTS and offline ASR asynchronously, preserving responsive controls and fencing cancelled startup results, retries, progress, and temporary vocabulary ownership.
- Begin eased cue ducking before speech playback instead of adding an unconditional 120 ms cue delay; preserve audio-clock state and ordered playback.
- Replace the audition shortlist with twenty level-checked CC0 candidates from five contrasting source packs; preserve legacy choices and production defaults, support authored continuous loops, and ease short transient attacks.
- Add browser-controlled voice replay and cue selection, gesture-driven audio resume, bounded PCM/audio-clock diagnostics, and production-level cue audition without a preview boost.
- Clear previous recognition hotwords when a voice reply fails its structured output contract.
- Fence speech submission against early native activity and pending decodes without weakening confirmed barge-in or adding another silence wait; retain deferred transcript fragments.
- Credit validated natural endpoint silence against existing submission budgets, retain conservative forced boundaries, and skip classifier requests after the deadline.
- Add explicit native classifier selection and opt-in host Codex Luna chat labels, with visible non-native identity and no invented confidence; retain the silence fallback and reject late or cancelled answers.
- Add pinned decoder-only and full-FP32 streaming controls without changing production speech models or concealing recognition regressions.
- Add opt-in native streaming recognition with bounded in-process workers and lifecycle fencing; retain Parakeet defaults because controlled corpus tests expose opening-word loss in the experimental model.
- Add native acoustic word alignment, late correlated timing delivery and callback-chunk playback offsets without delaying speech audio; reject unsupported or mismatched alignment.
- Resolve packaged speech-worker entry points and verify native bundled-server ASR/TTS, relocated alignment assets, and late browser timing without replay or stale highlighting.
- Add descriptive five-engine recorded batches, original-recording Whisper-reference reuse and atomic checkpoints without invented acceptance thresholds.

- Add native Whisper references and checkpointed human-corpus ASR measurements with explicit raw versus production paths, empty-result reporting and no invented acceptance thresholds.
- Remove voice-owned WASM runtimes, fallback, backend settings, worker paths, packaging and comparison controls; speech uses native Sherpa only.

- Add guided sixteen-prompt microphone corpus recording with replay, retakes, bounded PCM upload and exclusive WAV/manifest output.

- Apply native Parakeet per-turn hotwords using official BPE configuration.
- Release queued executor results after confirmed empty recognition; fence old decodes from consuming restarted audio.
- Add five licensed cue candidates per state, terminal-driven WebAudio audition, and read-only persisted production choices.
- Add portable fixed-key speech-engine shootout with gated browser microphone relay, ordered PCM playback and cancellation fencing.

- Preserve unresolved executor requests and queue result announcements behind actual conversation boundaries.
- Bound stalled voice replies, speak initialization fallbacks, and retire cancelled child sessions before replacement.
- Keep recognition and corrections inside a stable transcript drawer, with real-timing-only highlighting and explicit unsupported capability labels.
- Speak long-reply handoff notices before acknowledging playback completion; fence stale synthesis and interrupted turns.
- Separate accessible voice diagnostics from executor detail, including actual runtime identity and observed latency intervals.
- Add CPU-headroom desktop thread defaults and a native model-backed batch runner with truthful verification gaps.

- Add selectable offline Parakeet TDT asset validation and VAD-segment recognition, while rejecting unsupported Nemotron streaming selection clearly.
- Add bounded local WebRTC inbound bytes and audio-energy diagnostics alongside media-element state; these distinguish receiver flow from scheduled PCM and unknown physical audibility.

- Add hands-free local CALL flow with automatic speech turns, interruption, mute, end, and streamed local speech output.
- Preserve fast GPT automatic voice-model selection.
- Default local TTS setup to verified Piper `vits-piper-en_US-ljspeech-medium` assets.
- Fence voice finals at server boundary so reset, close, and duplicate submissions cannot dispatch stale speech.
- Replace browser energy turn control with pinned neural Silero VAD in persistent local ASR worker.
- Add a reduced-motion-safe voice waveform, focused executor depth panel, accessible transcript back handling, and local scanner waiting audio that fades for Piper PCM playback.
- Debounce local ASR finals into bounded continuation batches, holding resumed speech for at most two seconds while preserving authenticated generation and request-ID validation.
- Add explicit selected local TTS identity with verified Sherpa Kokoro `af_nicole` speaker maps, while retaining Piper LJSpeech default.
- Keep scanner progress through thinking and synthesis, with distinct modes and a 120 ms fade before PCM playback.

### web

#### web: 0.0.3

- Persist capped fetch results.
- Show local image preview in `understand_image`.
- Fix win32 release readiness.
- Retire superpowers conflict stack.

#### web: 0.0.2

- Source unchanged; no `extensions/web` commits since `v0.0.1`
- Tests unchanged; `bun test ./__tests__/*.test.ts` still applies
- Docs unchanged; README still lists `web_fetch`, `web_search`,
  `understand_image`

#### web: 0.1.0 - 2026-05-21

- a40a427 prepare extensions for npm release
- 133cb7d chore(pi):
  migrate extensions to earendil packages
- 2161a15 fix(web):
  use Z.AI MCP search
- db7b1c5 Update agent and terminal configs
- 5ca1296 Rework Pi agent extensions
- 2157199 test(pi):
  add mux harness and MiniMax/web tool tests
- b045a6b fix(pi/mux):
  cooldown exhausted keys on provider usage-limit errors
- b87a61a feat(pi):
  monorepo workspace — all extensions are proper packages
- 5fbf178 pi(ext):
  add remaining extensions (angel, animations, code-actions, files-widget,
  ghost, git-safe, neko, prune, web, etc.)