CHANGELOG.md

Changelog

0.10.0 - 2026-08-11

• Advanced the gateway to oida/gateway/v0.6 and aligned the complete core with AKOÚŌ 0.9.2, Earworm/Akousma 0.7.0/spec 1.6, and Akousmata 0.7.0.
• Added separate linked human listening accounts to REST, MCP, dashboard, and shared memory. Machine accounts remain immutable; human changes are additive revisions, and notes never become heard without an attributable human pass.
• Preserved separately declared human participants and passes at the host boundary, while demoting heard when its pass does not resolve to a human listener.

0.9.2 - 2026-08-03

• Reserved heard for separately attributable human reports. Local and host model transcripts, events, captions, and observations now remain inferred or interpreted, while DSP remains measured.
• Demoted machine-supplied heard input at the gateway boundary and kept the original basis visible for audit.
• Aligned with AKOÚŌ 0.9.1 and Earworm/Akousma 0.6.1 revision integrity, expanded environment-file ignore coverage, and added regression tests.

0.9.1 - 2026-08-03

• Documented a dated exception for two PyTorch findings limited to local APIs Oída does not call, while retaining the validated MOSS-Audio 2.10 runtime.
• Corrected dependency-audit comments to distinguish the unpatched .pt2 loader finding from the TorchScript finding fixed only in unvalidated 2.13.
• Pinned and validated the macOS Keychain executable and account arguments, bounded covenant numeric parsing, and removed backend exception details from public status, retention, memory, and reasoning responses.
• Documented the protected local-audio path boundary and aligned the public stack table with GERM 0.3.1, Algophony 0.5.1, and ORAM 0.4.1.

0.9.0 - 2026-07-27

• Decision-first gateway: oida/gateway/v0.5 and oida/route-outcome/v0.1 make a pre-listening refusal a complete result with an attributable route decision and receipt. No empty listening event, acoustic claim, or audio asset is fabricated; its decision-only Akousma is durable only after an explicit permitted Remember request.
• Accountable listening event v0.3: every completed hearing now carries AKOÚŌ context v2, temporal listening passes, source/cut/corpus provenance, and route decisions. Host perception advances to v0.4 while v0.3/context-v1 inputs remain readable for compatibility.
• Stack contract refresh: Oída now targets AKOÚŌ 0.9, Earworm/Akousma 0.6 and spec 1.5, and Akousmata 0.6. The harness adds corpus listening and no longer treats prompts, transcripts, captions, or contextual notes as heard evidence. undetermined remains a claim category, never an honest absence.
• Plural listening without false consensus: listening passes keep participants, influences, disagreements, permissions, and revisions attributable. Multiple routes from one host are not presented as an ear swarm without an explicit ensemble contract.
• GERM as an optional capability: discovery reports configuration and enablement without probing the network, and host skills require an explicit enabled handoff. There is no implicit localhost target; unconfigured endpoints return 409 and dashboard handoff controls remain absent. Core listening, memory, and release validation do not depend on a GERM installation.

0.8.0 - 2026-07-22

• Accountable listening events: every local and host-supplied result now declares oida/listening-event/v0.2 plus the shared akouo/listening-context/v1 block: position, actual apertures, auditory scales, sources of listening, participants, honest absences, and a distinct action-authority boundary. Capture metadata no longer opens a measurement aperture by itself, and host model observations remain separate from DSP.
• Gateway v0.4 / host perception v0.3: hosts may submit their situated listening declaration, which OÍDA retains as attributed input. OÍDA still computes the effective runtime context and clamps listening to observe-only; capability declarations cannot grant retention, revelation, or execution.
• Earworm 0.5 auditums: explicit memory writes preserve every route report under one honest OÍDA listener identity, carry context and apparatus refs, retain attributable disagreement and honest absence, and record scoped action receipts. Earworm traces add listening-report and listening-action events rather than reducing listening to a generic analysis frame.
• Stack contract refresh: AKOÚŌ advances to v0.8, Akousma to spec v1.4, Earworm and Akousmata integrations to 0.5, and all bundled host skills teach the same context, authority, disagreement, and explicit-memory rules.
• Reproducible release harness: CI and distribution builds resolve AKOÚŌ, Earworm, and Akousmata from their exact v0.8/v0.5 release tags so package verification cannot silently fall back to an older default branch.

0.7.0 - 2026-07-17

• Explicit listening identity: Oída now creates an empty, local LISTENING.md in its data directory and exposes it in the shared desktop dashboard under Settings → Listening. Its bounded perspective is carried by interpretive perception and every model-backed grounded conversation while remaining subordinate to evidence, privacy, routes, and Covenants. DSP and exact transcription remain literal. The freeform document has an optional conceptual scaffold for position, conditions, intentions, attention, relations, open questions, voice, and possibilities.
• Identity provenance across the stack: normalized events, private Earworm context, shared akousma extensions, and conversation audits now preserve a content-free oida/listening-identity/v0.1 revision block. The oida/host-perception/v0.2 contract lets a hosted ear declare the revision it applied; missing and changed revisions remain visible without becoming evidence. Host integrations preflight the active Covenant before direct perception, so perspective cannot be mistaken for permission.
• Release contract and editing safety: the gateway advances to oida/gateway/v0.3, advertises the listening-identity contract explicitly, and exposes 14 structured MCP tools. Dashboard edits use revision-aware atomic saves, preserving a newer external or agent edit instead of silently overwriting it. Capability discovery reports only identity metadata; reading the text is an explicit action. The release is covered by 316 tests plus package, macOS, dependency, schema, and JavaScript release gates.
• Covenant coverage closed around secondary hearings: live/background analysis, route reruns, and generation re-listening now apply the active Covenant's input, content, output, and memory-retention gates before their identity-shaped event is emitted. A new perspective can never become a path around an existing refusal.

0.6.5 - 2026-07-16

• Stricter API contracts: JSON request bodies now reject coercions such as integer values for booleans, closing inconsistencies found through generated OpenAPI request testing. Malformed GERM handoffs are rejected at the request boundary instead of reaching integration processing.
• Accurate service discovery: GET /api derives its endpoint list from the routes actually mounted by the running service, including sessions, covenants, remote capture, and optional integrations without a stale manual inventory. The OpenAPI document now describes shared error responses and HTML surfaces accurately, while the dashboard and embedded Akousmata navigator share a working root favicon route.
• Release and package hardening: distribution metadata uses current SPDX and license-file declarations, source archives exclude property-test state, and CI checks lint, both browser entry points, and built-package metadata. Obsolete development dependencies and old GitHub Action runtimes were removed or updated.
• Secure optional model runtime: the embedded MOSS dependency set now uses the compatible PyTorch 2.10, TorchCodec 0.10, and Transformers 5.14 releases, removing every fixable finding from the all-extras dependency graph. The loader now requires Safetensors and preserves the official MOSS-Audio API across the patched Transformers runtime.

0.6.0 - 2026-07-14

• Installable Python release chain: the public distribution is now sonicfield-oida because the shorter PyPI name belongs to an unrelated project; imports and CLIs remain oida. A tag-gated Trusted Publishing workflow builds and isolates the canonical akouo-contract, akousma, akousmata, and Oída distributions without repository tokens. Publishing is enabled only after all four PyPI Trusted Publishers are configured.

• Oída-owned reasoning layer: event conversation now has a fixed prompt hierarchy, a whitelist-built evidence packet, and a strict response contract. Listening results remain immutable; answer blocks and hypotheses cite stable evidence refs, uncertainty stays explicit, and chain-of-thought is never part of the contract or stored audit data. Conversation records move to v0.2 with read compatibility for v0.1, one primary event anchor, and up to three explicitly selected comparison events.

• Choose what each model does: Reasoning settings assign fast perception, deep perception, transcription, music analysis, conversation, and targeted re-listening independently. The shared dashboard manages provider/model selection and conversation profiles for tone, depth, initiative, focus, language, and bounded custom instructions. Profile text cannot override evidence, covenant, privacy, or output rules.

• Capability-aware audio model catalog: embedded/local-host presets cover MOSS-Audio 8B, MOSS-Music 8B, MOSS Transcribe + Diarize, MiDashengLM, MiMo-Audio, Qwen3-Omni, Gemma 3n, and experimental Mellow. API presets cover Gemini 3.5 Flash, Alibaba Qwen Omni, NVIDIA Nemotron, and the OpenRouter free Nemotron route. Large Qwen/Nemotron targets are configuration-only and were not executed on this machine.

• RAM-aware routing: settings report physical RAM, estimated peak local model memory, residency mode, platform requirements, gated/experimental status, and explicit warnings when a selection exceeds the machine. Models are not downloaded or loaded to calculate this assessment.

• Separate external-audio consent: conversation packets still never contain audio. Cloud perception receives audio bytes only when its provider/model is assigned and enabled and the new default-off permission is on; incognito and raw-audio covenant rules still block it. Requests never expose a local path; larger NVIDIA inputs use a temporary NVCF asset deleted after the call. Targeted re-listening remains local.

• Local, host, and explicit cloud providers: deterministic local reasoning remains the no-model fallback; existing Ollama and OpenAI-compatible endpoints can be added without model downloads. Codex, Claude, Hermes, OpenClaw, and OpenCode can reason through their existing host authentication, while OpenRouter supports BYOK and a localhost PKCE flow. All host and network providers are disabled until the operator enables them.

• Evidence stays smaller than the event: reasoning packets contain only covenant-filtered derived evidence. Raw audio, local paths, source URIs, credentials, and arbitrary event fields are excluded. Transcript and memory content have separate, default-off permissions. Incognito forces local-only reasoning and prevents conversation persistence.

• One disclosed local re-listen: a model may request at most one focused covenant-compliant MOSS/audio pass per turn. Its observation is added as new derived evidence before one final reasoning pass; it never modifies the original listening event. Oída never downloads or pulls a local model to satisfy the request.

• Predictable failure behavior: malformed provider output gets one repair attempt with the same provider. A failure then returns a visible deterministic local answer; Oída does not silently send the packet to a different external service.

• Host turn handoff: /conversation/prepare and /conversation/commit, mirrored as oida_prepare_turn and oida_commit_turn, let an active host run the supplied system prompt and schema without recursively invoking its own CLI. oida_ask remains the daemon-managed path. OpenCode and OpenClaw join oida integrate and doctor.

• Credential boundary: provider secrets use the macOS Keychain or an available cross-platform keyring; an environment-only store is the fallback when secure persistence is unavailable. Secrets are excluded from settings, logs, URLs, and process arguments.

0.5.0 - 2026-07-13

• Listening sessions: one listening session now holds many results, and the session is daemon state — the dashboard, the floating listener, the hotkeys, MCP, and agent calls all file into the same place. GET/POST /sessions, activate/rename/archive/restore/delete, per-result rename and delete (raw audio is never touched by history deletion), and POST /sessions/{id}/remember to remember a whole session. Sessions persist with recent history and survive restarts; pre-session history appears as the legacy "Earlier listens" group with the same batch actions.
• The session dashboard: the center column is now a session feed (every result a card with inline rename, copy, and an action menu), the left rail groups history by session with Archive below, and tag chips filter the feed (#music-id, presets, session tags — combinable). "Related" memory entries open their trace inline.
• Measured, drawn: DSP results now carry a compact normalized log-frequency spectrogram (features.spectrogram, 144×64), rendered as a sonogram with a high-definition modal view plus a frequency-energy band chart and a metric grid (LUFS, peak, RMS, crest, flatness, centroid). Silence stays visually empty instead of normalizing the noise floor into a bright box.
• Music ID, opt-in per listen: a song_id flag on /listen-event, /background/capture, /background/capture-request, live capture, and the native analyze path runs ShazamIO recognition only in Music mode; matches land as music_id on the event (tag music-id) and lead the result summary. Under a covenant that withholds song identity the answer is attributed absence, never a lookup. /health reports provider availability.
• Memory is the shared store: Remember now writes the compatibility trace and the shared Akousmata record in one act (/memory/remember, session remember); the Memory rail lists the shared store with rename and forget (PATCH/DELETE /akousmata/records/{id} — forgetting never deletes audio), and each memory card links back to its session and result.
• Rules in the footer: the covenant layer is surfaced as a Rules popover — a toggle, a document picker, and a plain-text editor; the daemon keeps its covenant vocabulary and file format. Still empty by default.
• Settings, consolidated: appearance (light/dark, persisted, native shell follows), interface reset, capture-permission and open-in-browser actions, Engine, Reasoning, and Path — one modal. A collapsible activity console under the result box records phases, requests, and errors.
• Native shell, chromed: titlebar accessories (sidebar toggles, source popovers with buffer/direction/mic level, floating listener, settings) via a window-chrome bridge; a real listening lifecycle (capturing → processing → result/failed) drives the dashboard, menu bar, and floating listener through one state sync; the floating listener gains an editable title (renames the session result), share/copy actions, the concentric listening glyph, and session-scoped readings (a fresh launch no longer resurrects old history). The supervisor launches the daemon through run_oida_mps.sh (self-syncing, offline-tolerant), terminates its managed daemon on quit, and the mac-mps engine monitor no longer reports a stale "ready" after models unload; the shell requests a launch prewarm when it attaches to a cold daemon. ScreenCaptureKit access is preflighted with an actionable message, and the dev build is signed with a stable designated requirement so privacy consent survives rebuilds.
• Directed capture everywhere: capture requests carry direction, source, skills, and Music ID; the native analyze path records capture and enforces covenant gates (423 refusals, pass filtering, perception/claim redaction, memory-retention refusal) like every other listen surface. 176 tests.

0.4.0 - 2026-07-12

• The sovereignty layer (spec v1.3 / AKOÚŌ v0.7): oída can now listen under a listening covenant — a small, human-written declaration of what this ear will not listen to, will release after hearing, will not reveal, will not retain, will blur, or will refuse at certain hours, and why. oida/covenant.py parses the easy text format (English and Spanish verbs; ## rules, ## commitments, ## because; unknown rule lines become commitments — the bridge, not the cage) and enforces the executable subset at four gates: input (refuse sources, quiet hours, max window — a refusal answers 423 with the covenant and rule named), content (ignored classes have their perception passes never run and their report traces dropped), output (withhold transcript / speaker-identity / affect / location / song-identity / events / spectral-detail, or coarsen location to a declared radius), and retention (raw audio released, memory writes refused — reported, never silent).
• Honest absence, on the record: every hearing under a covenant carries a covenant block (identity, lineage, rules applied, withheld — counted and attributed, never described — and commitment count) on the event, in memory traces, and in the shared-store akousma; the akousmata navigator can filter "everything listened under this covenant".
• Surfaces: GET/PUT /covenant, POST /covenant/activate, GET|DELETE /covenant/{name}; a Covenant panel on the dashboard (list, activate, plain-text editor); the oida_covenant MCP tool (agents may propose covenants; activation stays the operator's act); a per-listen covenant pin on /listen-event. The remote ear reports "under " and withholding counts on its result card. Empty by default: sovereignty is opted into, never imposed.
• Contracts: AKOÚŌ pin moves to akouo/v0.7 (sovereign-listening mode + /covenant command mirrored in the harness registries), components to earworm/v0.4 + akousmata/v0.4; gateway contract stays oida/gateway/v0.2 (all additive). 163 tests (13 new).

0.3.0 - 2026-07-11

• The remote ear (/remote): a phone-first capture surface served by the daemon itself. Wherever, whenever — the phone records (PCM in the page, WAV encoded on-device), attaches its GPS fix when granted, and posts to the new POST /remote/listen; the server runs the full listening pipeline, keeps the WAV, writes the akousma (the sound + its listening file) into the shared store with location and capture, and answers with the listening event rendered in the remote UI. Network publication remains operator-managed and uses the same host/origin guards as the dashboard.
• Future / past listening (spec v1.2 capture): every listen surface can now declare its temporal direction. Past slices the ring buffer that was already listening when the trigger fired; future records the window after it. The floating listener gains a ⟲/⟳ direction pill (persisted), Settings gains the default, the native mic ring grows from 30 s to 120 s with real bufferedSeconds accounting, the remote ear implements both modes with an on-phone ring buffer that overwrites itself (nothing accumulates), and /listen-event accepts capture_direction / capture_seconds / capture_trigger — carried on the event, its segment metadata, memory traces, and akousmata records.
• Geolocated listening (spec v1.2 location): /listen-event, /gateway/listen, the germ handoff, and oida_listen (MCP) accept an optional consent-scoped location {lat, lon, accuracy_m, altitude_m, label, source}; it rides the listening event and lands in the akousma, where the akousmata navigator's new listening map plots it.
• MCP: oida_live gains the capture action (slice the last N seconds from a live ring — the past direction made callable, with optional analysis); oida_listen gains location.
• Contracts: gateway manifest components move to earworm/v0.3 and akousmata/v0.3 (spec v1.2); the gateway contract itself stays oida/gateway/v0.2 — every addition is optional and additive. New remote_ear transport advertised.
• 150 tests (5 new: capture/location validation and plumbing, the complete remote-ear flow into a temp shared store, bridge-level spec v1.2 blocks).

0.2.0 - 2026-07-10

• Oída is now the complete listening distribution and gateway: installing it resolves AKOÚŌ (akouo-contract), Earworm's akousma, and the Akousmata navigator. One FastAPI process exposes the agent/dashboard, REST gateway, official streamable HTTP MCP at /mcp, and full library at /library/.

• Provider-neutral host perception: oida/gateway/v0.2 accepts either Oída-owned audio (/gateway/listen) or a declared report from an audio-capable host (/gateway/harness). Host-model observations cannot be silently promoted to DSP measurements; apparatus limits and unknowns remain explicit, while MOSS keeps its 16 kHz mono restrictions.

• Gateway lifecycle and local adapters: start, stop, status, doctor, agent, gateway, and integrate commands manage a singleton daemon and install bundled Hermes, Codex, and Claude integrations. The remote capture page is served by the daemon without configuring network access. MCP processes ensure or reuse the gateway and pin the active Python runtime.

• Earworm on every pass, durable memory only by consent: session-scoped Earworm context is emitted for all gateway listens and host harness passes; Akousmata writes still require an explicit remember action.

• Akousmata history embedded in the dashboard: a new left-side "Akousmata" section browses the shared store natively inside oída (search, compact cards, and a detail modal with listenings, lineage, kinship, audio playback, and the three germ buttons) — the same library the standalone akousmata navigator (github.com/sonicfieldlabs/akousmata) serves, without launching the external app. Backed by new optional /akousmata/* routes (oida/akousmata_view.py) over py-akousma; card shapes stay compatible with the navigator. Read-only by design: oída writes through its listen flow and the germ bridge; edits belong to the navigator.

• Preset ids aligned with AKOÚŌ v0.6's portable vocabulary: environment→field, speech→voice, memory→recall (with a LEGACY_PRESET_ALIASES map so saved configs, sessions, and older clients keep resolving; the manifest exposes preset_aliases). oída's preset id set is now a strict subset of the upstream presets/presets.json vocabulary, enforced by a test.

• AKOÚŌ v0.6 contract adopted (akouo_contract_version: v0.6): new /remember command with a remember route preset (memory comparison plus registration into the akousmata) alongside the read-only memory preset; memory-lineage-listening added to the harness mode set; /fiction now grants declared speculative permission and /forensic keeps its interpreted/speculative suppression, both matching the published command_permission_overrides. Listening outputs now carry akouo_version, an apparatus declaration (hybrid MOSS+DSP substrate, perception sources, known blind spots, model ids), and a listener block; claims carry source (dsp / model / context) and, for events, time_range anchors. A new harness/akouo/manifest.py loads the upstream machine-readable contract (akouo.manifest.json + presets/presets.json) and the test suite includes a drift check so the harness fallback tables cannot silently diverge.

• Earworm v0.2 / akousma spec v1.1 adopted in the bridge: listen records now carry a skimmable summary, listening entries are wrapped in the v1.1 envelope (contract/created_at/summary/payload, with akouo.* entries pinned to akouo/v0.6), and registration links recurrences — when the same audio content hash already exists in the store, the new record gets a same_source_as relation to the most recent holder.

• Floating listener redesigned: the macOS floating listener is no longer a window. It is a borderless, transparent, non-activating panel whose visible body is the listening-result box itself — the reading is always shown, with a small reactive waveform in its corner that animates only while listening. Every control floats just outside the box and fades in on hover: listening mode at the top-left, close at the top-right, and source / Listen / control center in a row below. The box and its type are larger than the prior card, and it drags from the box. The box stays empty until something is actually heard, and the mode (preset) control matches the size of the row below it. (An earlier take on this rebuild used a Metal shader orb; that was dropped in favor of the box-centric layout.)

• Dashboard restyled flat and grey, then re-laid-out: one continuous surface instead of white cards (sections separated by hairlines, no shadows, flush to the width); no green anywhere. The Listen surface is now two lines — source + its config (capture / input / file) on one line with the Skills / Engine / Path controls as segmented tabs at the top-right (each toggling a panel), and the grey Listen button beside the route presets on the next line. Presets and the result actions (Ask, Remember, Wiki, JSON, and the germ handoff) are borderless icon+word buttons. The claim groups (Hypotheses, Heard, Interpreted, Measured, Undetermined, Memory) are tabs below the reading. Source descriptions, the preset hint, and the evidence-level chip were removed; the header is just the wordmark and the daemon address sits in the footer as host:port. Skills / Engine / Path open as modal dialogs (not inline). The source row is one line — source tabs plus their config; in System the buffer length reads Buffer: 10sec (no route text), and in Mic the input list loads by default (no Devices button) with a small icon monitor toggle. Selects use the app font, not a mono face.

• Native top bar: the mac window is just the embedded dashboard with one top-right bar of icon buttons, split into two groups of three — Skills / Engine / Path (each opens its modal in the page via evaluateJavaScript) and the shell actions floating-listener / reload / browser (or Start daemon when offline). Icons get a subtle hover wash. The shell injects window.__oidaNative so the page hides its own browser-only panel icons.

• Dashboard reorganized into an Obsidian-style three-rail shell: collapse toggles live in the top corners (state persists). The left rail carries the source tabs (System / Mic / File) and their config at its top, then Recent and Memory; the right rail carries the tool icons at its top — Skills / Engine / Path, plus (inside the native app) floating-listener / reload / browser — then the claim Breakdown as a vertical nav. The center is the reading, usable with either rail collapsed, and the rails stay rails at native-window widths instead of stacking full-width. The foot reads like an agent chatbox: Remember / Wiki / JSON / germ as icon+word actions at the bottom-left (germ opens a Sound / Prompt / Lineage menu), the listening mode as a quiet borderless selector, and Listen as a round dark send-style button at the bottom-right. The Ask button, page footer, and wordmark are gone — the daemon address, audio path, and API/health links moved into the Engine modal — and opening any panel modal closes whichever was open.

• Native shell is chrome-free: the capsule top bar was removed; the page's right rail hosts the shell actions, which post back over a oidaShell script-message bridge (reload reloads the WebView directly). The only native overlay left is a Start-daemon card when the daemon is offline.

• Floating listener: the status dot in the box footer was removed (the status word and corner waveform remain), and the listening-mode button is a rectangular frost tile with the same rounding and margins as the buttons under the box.

• Shell refinements: the rails are drag-resizable (widths persist) with Obsidian-style panel toggles in the top corners; the sources are icons whose config always fits the rail (including the mic meter); memory search is an icon. The claim Breakdown is stacked collapsible sections, all expanded by default. The result actions condensed into one export menu at the box's bottom-right — Remember, Expand on Wiki, Export JSON, Generate derived sound, Convert listening to prompt (Lineage removed) — next to the Skills icon, the quiet mode selector, and Listen. Two corner buttons float over everything: a configuration menu bottom-left (Engine / Path / Reload / Open in browser) and, inside the native app, the floating-listener toggle bottom-right.

• App icon: a minimal abstract listening mark — a single bold two-hump sine wave on a charcoal squircle, transparent corners. Sourced from apps/macos/Resources/AppIcon.svg, shipped as AppIcon.icns, and staged into the bundle by both packaging scripts.

• germ handoff fixed and surfaced: /germ/handoff was unusable (with from __future__ import annotations FastAPI could not resolve the function-local request model and degraded the JSON body to a query param); the model is now module-scoped. The dashboard gained the three germ buttons (Sound / Prompt / Lineage) that persist an akousma to the shared store and deep-link germ; origins are normalized to the akousma vocabulary; opt-in song identification (OIDA_SONGID=1) now enriches handoff records.

• Dashboard hardening: SSE events from other surfaces no longer stomp a local mic recording (orphaned hot mic); mic streams are released when MediaRecorder fails and on page hide; duplicate render/refresh after each listen removed; engine fold no longer rebuilds every 20 s (dropdown-closing); phase guards on drop/file/path during a running listen; capture-cancel sends the request id; the 15 s "no capture yet" hint verifies with the daemon before flipping to an error; SSE reconnects after a permanently closed stream; recording shows elapsed time; JSON view gained copy; Ask cannot double-submit via Enter.

• Dashboard accessibility and polish: visible focus rings, AA-passing text tones, radio semantics + arrow-key navigation for source/preset groups, a live daemon status label, meter semantics, reduced-motion-aware scrolling, evidence-level chip colors for all levels, dsp only no longer styled as an error, dead CSS removed, tags overflow as +n.

• Backend correctness: memory list/limit now ordered by createdAt (was filename-random); OIDA_MOSS_PREWARM precedence fixed (legacy HMM_/AEAR_ could override it); OIDA_REQUIRE_MODEL accepts true/yes/on; per-listen DSP inspection memoized (a listen decoded and hashed the same file 2–3×); ffmpeg conversion gets a timeout; /sonicfield/reveal containment can no longer be bypassed by sibling directories; corrupt memory/generation/conversation records return 4xx instead of 500; /health legacy_name reports hmm, aear; /api lists the capture-request, engine-model, status, and germ routes; double prewarm race closed.

• macOS shell: native capture filenames had a corrupted date pattern from the rename (HOidassSSS); the shell now reads OIDA_AUTH_TOKEN first; the managed daemon is stopped on app quit (its pipes died with the app and the orphan would crash on SIGPIPE); transient refresh errors keep recent/pinned history; hotkey signature renamed HMMK → OIDA.

• MCP tools renamed to canonical oida_* with hmm_*/aear_*/ear_* aliases kept and the previously undeclared aear_live_* aliases declared.

• Environment/docs unification: project scripts export OIDA_* names (legacy still honored), release smoke reads the *_SERVER_URL chain and cleans its prompt record from the daemon data dir, README/SECURITY/shell docs are rewritten OIDA-first and match current behavior, CI/local checks run the suite once via pytest, and the server-integration tests isolate ambient OIDA_*/HMM_*/AEAR_* variables.

• Renamed the project hmm → oída (part of the Sonic Field "sonic evolution"). Python package aear → oida; primary CLI oida (oida-daemon), with hmm/aear kept as aliases. Environment variables use the OIDA_* prefix first, falling back to HMM_*/AEAR_*. macOS shell target renamed HmmMacOS → OidaMacOS (oida-macos). Status route /hmm/status → /oida/status; server identity, journal headers, and UI titles now read oída. Data directory prefers ~/Library/Application Support/oida, falling back to a pre-rename hmm directory so existing listening memory is not orphaned. The Sonic Field bridge now resolves a configured or sibling checkout.

• Made mac-mps the default engine everywhere (daemon, macOS shell, env examples); the daemon prewarms MOSS-Audio-4B-Instruct in the background and reports readiness at /engine/status (HMM_MOSS_PREWARM=0 disables).

• Added the deterministic signal listener (oida/signal_listener.py): DSP-only classification (silence / speech-like / music-like / tonal / percussive / noise / ambient), honest captions, and inferred hypotheses with numeric bases. Perception never falls back to placeholder text; without a model the evidence level honestly stays measured_signal.

• Route presets now scope MOSS passes (moss_passes): Basic runs one caption pass (seconds, not minutes), Signal is DSP-only, Speech runs transcribe+speech, and the new Deep preset runs the full report.

• Chunked mac-mps inference at 45 s per pass (HMM_MOSS_CHUNK_SECONDS), with a decode-degeneracy guard that discards token-soup output instead of letting it become claims, and a tokenizer-safe decode fallback.

• Added Sonic Field bridge (oida/sonicfield.py, /sonicfield/*): after a listen, "Explore in the wiki" searches the wiki/lexicon, topics, journal, 93k-item library, paths, research, notes, and labs for related concepts, with taxonomy alias normalization and Finder reveal.

• Added /events/stream (SSE) so every surface mirrors one daemon state, and /background/capture-request so the web dashboard can ask the native shell to capture system audio.

• Default audio directory moved to ~/Documents/oida/audio (HMM_AUDIO_DIR); captures, uploads, and fixtures land there.

• Rebuilt the dashboard as one Listen surface (source + presets + skill manager), claims rendered by AKOUO category, memory and history cards, and a soft-minimal visual system; removed the floating spectral agent, its settings, RMS/peak meters, and the task tabs.

• Rebuilt the macOS shell: the main window embeds the dashboard (WKWebView) so web and app never diverge; the floating listener is a minimal always-on-top panel with live signal, latest reading, and one Listen action; global hotkeys default to ⌃⌥L (listen) and ⌃⌥H (show/hide listener); the shell auto-starts the daemon and system tap and claims dashboard capture requests.

0.1.0 - Unreleased

• Added platform data directory support through HMM_DATA_DIR / AEAR_DATA_DIR with macOS default ~/Library/Application Support/oida.
• Disabled implicit Hugging Face hub model fallback unless explicitly enabled.
• Added bearer-token enforcement for wildcard/LAN daemon binds.
• Added raw-audio status and wipe endpoints for local upload/live-buffer retention, including pre-data-dir checkout uploads/ recordings via the include_legacy wipe option (the dashboard Wipe raw audio button sets it).
• Bounded in-memory stopped live sessions to the most recent few.
• Added Earworm-compatible Akousmata trace/session/context metadata.
• Updated AKOUO harness routing for the v0.5 public command contract and derived evidence levels.
• Reduced long-audio memory pressure by bounded DSP inspection and per-chunk source reads.
• Added release/security/contribution/citation metadata.