codai-sdk 0.7.0__tar.gz → 0.52.1__tar.gz

codai_sdk-0.52.1/CHANGELOG.md

@@ -0,0 +1,731 @@
+# Changelog
+
+All notable changes to `codai-sdk` are documented here. The format is based on
+[Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and this project
+adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+## [Unreleased]
+
+## [0.52.1] - 2026-10-16
+
+- `WalletLedgerEntry.source` literal gains `sandbox_charge`: the agent sandbox exec-time debit (EUR 0.10 per
+  exec-hour after 10 free exec-hours a month), written every 5 minutes once sandbox billing is switched on.
+
+## [0.52.0] - 2026-10-16
+
+- `provenance.check(file, content_type=...)` — public AI-provenance checker (`POST /v1/public/provenance/check`,
+  no key needed, at most 25 MB; `file` is bytes or a path): reports C2PA manifest presence and claimed issuer, the
+  WAV RIFF INFO, MP3 ID3 and MP4 `udta/meta` AI marks and image XMP. `signature_verified` and `synthid_checked`
+  are always `False`.
+
+## [0.51.0] - 2026-10-16
+
+- `client.sandboxes` (beta) — agent sandboxes over REST (`/v1/sandboxes`): `create` / `list` / `retrieve`, `exec`
+  (blocking; `exit_code`, `stdout`, `stderr` and a reproducibility `receipt`), `exec_stream` (generator of event
+  dicts: `start`, `stdout`, `stderr`, `truncated`, `exit`, `error`), `files.put` / `files.get` (returns `SandboxFile`
+  with `data` + `sha256`) / `files.list`, `artifacts` (scanned files from `/workspace/out` with signed 1 h links),
+  `snapshot`, `delete` and one-shot `run`. `exec`, `exec_stream` and `run` accept `idempotency_key=`.
+- Chat `tools` items may now be `{"type": "codai_sandbox", "egress": ..., "template": ...}` (`ChatSandboxTool` in
+  `codai._types`): the gateway runs bash / Python server-side in a sandbox per session and returns the artifacts.
+- REST exec over Environments: `environments.exec(env_id, {"command": ..., "op": ..., "args": ..., "timeout_ms": ...})`
+  (blocking relay op, shell results flattened to `exit_code` / `stdout` / `stderr`), `environments.files.list` /
+  `get` / `put` (UTF-8 text, ≤ 2 MB) and `environments.jobs.start` / `get` (status + output). No WebSocket client
+  needed; 409 `environment_offline`, 504 `host_timeout`.
+
+## [0.50.0] - 2026-10-16
+
+- `abuse_reports.create(share_token, surface, category, explanation, reporter_name, reporter_email, good_faith=True)` —
+  `POST /v1/studio/abuse-reports` (public, no key needed): DSA art. 16 notice about illegal or abusive content on a
+  codai Studio share (`surface="share"`) or review (`"review"`) link — categories `voice_without_consent`,
+  `impersonation`, `fraud`, `election`, `illegal_other`, `other`. Answers 202
+  `{id, object: "studio.abuse_report", status: "open"}`, also for an unknown link (no enumeration). Never retried.
+  `CreateStudioAbuseReportBody` / `CreateStudioAbuseReportResponse` types.
+
+## [0.49.0] - 2026-10-16
+
+- `resolve.list(status=None, limit=None, cursor=None)` — `GET /v1/resolve/jobs`, your Resolve jobs newest first
+  (`{"jobs", "next_cursor"}`; `status` = a string or a list sent as a comma list), and `resolve.iterate(status, limit)`,
+  a generator that follows `next_cursor`. `ResolveJob` gains `issue_url` and `issue_title`; `ResolveJobList` type.
+- `billing` (READ-ONLY, never starts a payment): `billing.get()` — `GET /v1/billing`; `billing.invoices(limit=None)` —
+  `GET /v1/billing/invoices`; `billing.watch_balance(threshold_eur, on_low, interval_s=60, max_checks=None)` polls
+  `get()` and calls `on_low(balance_eur, billing)` once per crossing below the threshold.
+- `referrals`: `referrals.get()` — `GET /v1/referrals`; `referrals.redeem(code)` and `referrals.set_consent(bool)` call
+  `PATCH /v1/account` with `referral_code` / `referral_consent`.
+
+## [0.48.0] - 2026-10-16
+
+- `voice_revocations.create(email, voice_owner_name, details, evidence_url=None)` — `POST /v1/studio/voice-revocations`
+  (public, no key needed): asks for a voice cloned in codai Studio to be taken down; answers 202
+  `{id, object: "studio.voice_revocation", status: "open"}` and never echoes the input (only the sha256 of the e-mail
+  is stored). Never retried. `CreateVoiceRevocationBody` / `CreateVoiceRevocationResponse` types. Custom (cloned) voices
+  themselves are not in the SDK while the feature flag is off.
+
+## [0.47.0] - 2026-10-16
+
+- Lottie export: `creatives.export_lottie(id, format=None, formats=None, seq=None, container=None)` —
+  `GET /v1/studio/creatives/{id}/export.lottie` exports a video creative as a Lottie animation (Bodymovin JSON with
+  embedded images, or a dotLottie `.lottie` archive with `container="dotlottie"`; `formats="all"` for every format),
+  free and synchronous: the export comes back already `succeeded` with a signed `url` and the fidelity report in
+  `warnings`. `creatives.download_lottie(...)` returns the file itself (`{bytes, filename, content_type}`).
+  `StudioExport` gains `kind` (`"fcpxml"` | `"lottie"`) and `container` (Lottie only); `ExportCreativeLottieResponse`.
+
+## [0.46.0] - 2026-10-16
+
+- `ai_label` (`"none"` | `"corner"` | `"end_card"`, EU AI Act art. 50(4)): in the `renders.create` / `renders.render`
+  body, as `creatives.render(id, ai_label=...)` and `batches.create(..., ai_label=...)` / `dry_run`. Burns a visible
+  "AI-generated" label (corner for the whole video, or a ~2 s end card) into the MP4. Omitted = `corner` when the render
+  uses TTS narration or AI-generated images, else `none`; `none` with TTS narration is a 400 `ai_label_required`.
+  Render replies report the resolved `ai_label`, and so do `renders.list` items (`"none"` for renders from before the
+  label existed).
+
+## [0.45.2] - 2026-10-16
+
+- `client.memory.list/search/create/delete` — your persistent memory over REST (`/v1/memory`), the same store
+  the MCP memory tools use. Search and create are metered as `memory:search` / `memory:create`.
+
+## [0.45.1] - 2026-10-15
+
+- `allow_remix` on view share links now defaults to `false` server-side (E-13 legal review): pass
+  `allow_remix: true` to let link holders remix the creative into their own account.
+
+## [0.45.0] - 2026-10-15
+
+- Share links: `creatives.create_share(id, kind, label=None, expires_in_days=None, seq=None, allow_remix=None,
+can_approve=None)` — `POST /v1/studio/creatives/{id}/shares` returns the public `url` and its `token` once;
+  `creatives.list_shares(id)` and `creatives.revoke_share(id, share_id)`.
+- Review: `creatives.review(id)`, `request_review(id)`, `reset_review(id)` (`draft` → `in_review` → `approved` |
+  `changes_requested`, approval pins the version); threads: `review_threads(id, status=None)`,
+  `create_review_thread(id, body, node_id=None, scene_id=None, t_ms=None)`, `reply_review_thread`,
+  `resolve_review_thread`, `reopen_review_thread`, `delete_review_comment`.
+- New `client.shares` (token or whole link): `get(link, format=None, locale=None)` (no key needed),
+  `remix(link, title=None)`, and through review links `comment(link, name, body, …)`, `reply`, `resolve`, `reopen`,
+  `decide(link, name, decision, seq, note=None)`. `codai.share_token_from(value)` extracts the token. All free.
+- Generated types: `StudioShare*`, `StudioReview*`, `StudioPublicShare`, `StudioPublicReview`, `StudioRemix*`.
+
+## [0.44.0] - 2026-10-15
+
+- Localisation: `creatives.locales(id, locale=None, units=None, offset=None, limit=None)` —
+  `GET /v1/studio/creatives/{id}/locales`; `creatives.localize(id, locales, tone=None, glossary=None, only_missing=None,
+voice=None, base_seq=None, force=None)` — `POST /v1/studio/creatives/{id}/localize` translates every text into up to
+  10 BCP-47 locales as per-locale overrides in one new version (token-metered; `voice=True` narrates voice clips again,
+  billed); `creatives.fork_locale(id, locale, title=None, seq=None)` —
+  `POST /v1/studio/creatives/{id}/locales/{locale}/fork` saves a locale as a new creative.
+- `locale=` keyword on `creatives.compiled`, `render`, `frame`, `captions_srt`, `captions_vtt` and `lint`. Lint
+  codes `missing_translation`, `stale_translation`, `text_overflow_locale` and issue `locale`; new ops
+  `setLocaleText`, `setLocale`, `setLocalization`.
+- Batches: new `client.batches` — `create(source, csv=None, sheet=None, delimiter=None, mapping=None, title=None,
+render=None, quality=None, format=None, localize=None)`, `dry_run(...)` (same arguments, nothing stored), `list`,
+  `get(id, rows=None, status=None, offset=None, limit=None)`, `cancel`, `retry`, `export_csv` (status CSV text),
+  `zip` (ZIP of finished renders) and `wait(id, interval=3.0, timeout=600.0)`.
+- Generated types: `CreativeLocaleStatus`, `CreativeLocaleUnit`, `CreativeLocalizeRequest`, `CreativeLocalizeResult`,
+  `CreativeLocaleForkRequest`, `CreativeLocaleForkResult`, `StudioBatch*`.
+
+## [0.43.0] - 2026-10-15
+
+- Auto-captions: `creatives.captions(id, clip_id=None, preset=None, node_id=None, script=None, language=None,
+retranscribe=None, base_seq=None, force=None)` — `POST /v1/studio/creatives/{id}/captions` transcribes a voice clip
+  word by word (words cached on the asset, free next time), corrects the spelling with the script and adds or updates a
+  captions node as one new version (`node_id`, `words`, `cached`, `corrected`, `cost_micro_usd`; a fresh
+  transcription is billed at the speech-to-text rate). `creatives.captions_srt(id, node_id=None, seq=None)` /
+  `captions_vtt(...)` return the captions as SRT / WebVTT text.
+- Platforms: `creatives.adapt(id, platforms, focal=None, base_seq=None, force=None)` —
+  `POST /v1/studio/creatives/{id}/adapt` adds one format per platform (`tiktok`, `reels`, `shorts`, `feed-square`,
+  `feed-portrait`, `youtube`) with text kept inside the platform safe zone; `focal="vision"` (token-metered) finds
+  image focal points. New ops `adaptFormat` and `setControl`.
+- Lint: issues gain `format` and `fixable` and the codes `empty_list`, `off_brand_color`,
+  `outside_platform_safe_zone`, `text_too_small`, `missing_alt`, `missing_captions`, `reading_speed`, `flashing`;
+  the result adds `fixable`. `creatives.lint_fix(id, codes=None, agent=None, base_seq=None, force=None)` —
+  `POST /v1/studio/creatives/{id}/lint/fix`: deterministic fixes are free, `agent=True` adds one billed `codai` turn.
+- Tweak panel: `creatives.controls(id)` — `GET /v1/studio/creatives/{id}/controls`.
+- Render aspect ratio `4:5` (1080×1350).
+- Generated types: `CreativeCaptionsRequest`, `CreativeCaptionsResult`, `CaptionPresetId`, `PlatformId`,
+  `CreativeAdaptRequest`, `CreativeAdaptResult`, `CreativeLintCode`, `CreativeLintFixRequest`,
+  `CreativeLintFixResult`, `Control`, `PanelControl`, `CreativeControls`.
+
+## [0.42.0] - 2026-10-14
+
+- Brand kits as Design Tokens (DTCG 2025.10): kits now carry `dtcg` (the canonical document), `contrast` (WCAG ratio
+  - `aa` for text/background, primary/background, background/primary, accent/background, secondary/background) and
+    `updated_at`; `tokens` also carries the document aliases (`color.bg`, `color.fg`, `color.accent`,
+    `font.heading`, `font.body`), extra colours and `radius`.
+- `brand_kits.create(name, tokens=… | dtcg=…, logo_asset_id=…, source_url=…)` — `POST /v1/studio/brand-kits`
+  (exactly one of `tokens` / `dtcg`, else `ValueError`; 400 `invalid_tokens`), `brand_kits.update(id, name=…,
+tokens=…, dtcg=…, logo_asset_id=…)` — `PATCH /v1/studio/brand-kits/{id}` (`logo_asset_id=None` removes the logo),
+  `brand_kits.export_dtcg(id)` — `GET /v1/studio/brand-kits/{id}/dtcg`. Free.
+- `brand_kits.extract(url=None, name=None, logo_asset_id=None)` — exactly one of `url` / `logo_asset_id` (else
+  `ValueError`); from a logo image one vision turn reads the palette (400 `unsupported_media`).
+- `creatives.restyle(id, brand_kit_id=None, tokens=None, base_seq=None, snap=None, fonts=None, force=None)` —
+  `POST /v1/studio/creatives/{id}/restyle`: re-applies a kit as one new version (literal brand colours become token
+  refs, `snap` maps the rest to the nearest kit colour, fonts by role, locked nodes skipped unless `force`). Free.
+- Generated types: `BrandKitDtcg`, `BrandKitContrastPair`, `BrandKitCreateRequest`, `BrandKitUpdateRequest`,
+  `CreativeRestyleRequest`.
+
+## [0.41.0] - 2026-10-14
+
+- Documents as creatives: `creatives.from_document(document_id)` — `POST /v1/studio/documents/{id}/creative` turns a
+  Studio deck / doc / sheet into an editable creative (slides = scenes with `layout` + `notes`, doc pages = a `flow`
+  group, sheets = a `cell_grid` + chart nodes; nodes carry semantic `role`s). Idempotent: returns the creative plus
+  `document_id` and `existing` (HTTP 201 new / 200 already linked). Free.
+- `creatives.build(creative_id, seq=None, preview=None, filename=None)` — `POST /v1/studio/creatives/{id}/build` builds
+  the creative back into PPTX / DOCX / XLSX as a new version of its document: `{object: "studio.creative_build",
+creative_id, seq, warnings, document}`; 400 `not_supported_for_kind` / `invalid_spec`. Files are free;
+  `preview=True` queues the billed document preview.
+- Regenerated `_types.py`: `StudioCreativeFromDocument`, `CreativeBuildRequest`, `StudioCreativeBuild`,
+  `Document.creative_id`, `CreativeCompiled.kind`, the `createCreativeFromDocument` / `buildCreative` operations.
+
+## [0.40.0] - 2026-10-14
+
+- FCPXML export: `creatives.export_fcpxml(id, format=None, seq=None)` — `GET /v1/studio/creatives/{id}/export.fcpxml`
+  queues a zip with an FCPXML 1.10 project (one clip + marker per scene, audio on its own lanes) and its media
+  (per-scene MP4 segments, voice / music / sfx files) for Final Cut Pro, DaVinci Resolve or Premiere (FCPXML import).
+  Returns `{id, status: "queued", creative_id, seq, format, fps, filename, url, expires_at, bytes, rendered_seconds,
+cost_micro_usd, warnings, error, created_at}`; 400 `unsupported_kind` / `unsupported_format` / `too_long`. Only newly
+  rendered seconds are billed, at the final render price (cached segments are free).
+- `exports.retrieve(id)` — `GET /v1/studio/exports/{id}` (free); `exports.wait(id, interval=3.0, timeout=900.0)` polls
+  until `succeeded` (signed `url`, valid 1 h) or `failed` and returns the final export; `TimeoutError` after `timeout`.
+
+## [0.39.0] - 2026-10-14
+
+- Templates from your own creatives: `creatives.save_template(id, vars=[…], title=…, seq=…)` —
+  `POST /v1/studio/creatives/{id}/template`: each `vars` entry (`{"node_id", "prop": "text" | "image", "key",
+"label"?}`) turns a text node's text or an image node's asset into a template variable (current value = default).
+  Returns `{id, title, kind, source_creative_id, source_seq, vars_schema, thumbnail_url, created_at, doc}`; 400
+  `node_not_found` / `invalid_pick` / `duplicate_key`. Free; templates are private.
+- `templates.list(kind=…, limit=…)`, `.retrieve(id)` (with `doc`), `.delete(id)` — free.
+- `creatives.create({"kind": …, "template": …, "vars": {…}})` starts version 1 from a template with the var values
+  applied (400 `kind_mismatch` / `unknown_var` / `invalid_var_value`, 404 unknown template).
+
+## [0.38.0] - 2026-10-14
+
+- Cards: `cards.product(url, lang=None)` — `POST /v1/cards/product`: a `products` card block with one item read from
+  a public https product page's schema.org `Product` JSON-LD (or OpenGraph product tags): name, image, price +
+  currency (lowest offer), old price, rating (0-5), reviews, seller, availability badge (en/ro). Only what the page
+  states, never sponsored; SSRF-guarded like `unfurl`, cached 6 h, shares its 500/day limit; 422 `product_not_found`.
+
+## [0.37.0] - 2026-10-14
+
+- Brand kits: `brand_kits.extract(url, name=…)` — `POST /v1/studio/brand-kits/extract`: fetches a public https page
+  (SSRF-guarded), one bounded `codai` model turn picks the palette (primary, secondary, accent, background, text) and
+  the heading / body fonts, the logo is stored as an `image` asset. Returns `{id, name, source_url, tokens,
+logo_asset_id, logo_url, logo_mime_type, created_at}`; 400 `url_blocked`, 422 `fetch_failed` / `extract_failed`,
+  504 `timeout`. Token-metered (`codai::brand-kit-extract`).
+- `brand_kits.list(limit=…)`, `.retrieve(id)`, `.delete(id)` — free.
+- `creatives.create({"kind": …, "brand_kit_id": …})` applies a kit to version 1 (`brand = {kitId, tokens}` + logo asset).
+
+## [0.36.0] - 2026-10-14
+
+- A/B variants of a creative: `creatives.variants(id, base_seq=…, count=…, focus=[…], brief=…, render=…, format=…)`
+  — `POST /v1/studio/creatives/{id}/variants`: one bounded `codai` model turn proposes up to 5 variants (focus hook /
+  title / cta / visual / copy), each saved as a fork with the ops applied plus a queued draft render. Returns
+  `{group_id, base_seq, variants: [{creative_id, label, rationale, applied_ops, head_seq, render_id}]}`; 422
+  `no_edit`, 504 `timeout` after 90 s. Token-metered (`codai::creative-variants`), draft renders billed on their poll.
+- `creatives.list_variants(id)` — `GET …/variants`: the latest group with `status` and the stored render state.
+- `creatives.pick_variant(id, variant_id, base_seq=…)` — `POST …/variants/{variantId}/pick`: the winner's ops onto the
+  original (rebased like `edit`, 409 `stale` on overlap), statuses updated.
+
+## [0.35.0] - 2026-10-14
+
+- `SessionControlKind` gains `pause` and `resume`: `POST /v1/sessions/{id}/control` with `kind="pause"`
+  parks the running turn at its next step boundary (after the current tool calls), `kind="resume"` continues the
+  same turn. Editors may send both (same trust as `cancel`). Executors report the state with a `run_state` event
+  `{"state": "paused" | "running"}`.
+
+## [0.34.0] - 2026-10-12
+
+- `creatives.comment(id, t_ms=…, comment=…, base_seq=…, node_id=…, scene_id=…, apply=…)` —
+  `POST /v1/studio/creatives/{id}/comments`: a comment pinned on a frame becomes an edit (one bounded `codai` model
+  turn proposes ops, applied like `edit`). Returns `{head_seq, applied_ops, summary, version}`; 409 `stale` on an
+  overlapping change, 422 `no_edit` when the model made no edit, 504 `timeout` after 60 s. Token-metered
+  (`codai::creative-comment`).
+
+## [0.33.0] - 2026-10-12
+
+- `creatives.frame(id, t_ms=[…], format=…, exact=…, version=…)` — `POST /v1/studio/creatives/{id}/frame` now returns
+  PNG still frames (`{object: "studio.creative.frames", status: "succeeded", frames, billed_micro_usd}`) or a running
+  job (`{object: "studio.creative.frame_job", status: "running", frame_job_id}`) when they are not ready within ~90 s.
+  `exact=True` cuts frames from a draft render of the covering scenes. Cached frames are free; 60 frame jobs per UTC
+  day are free, then billed. The old `frame(id, body)` dict form is replaced by keyword arguments.
+- `creatives.frame_job(id, job_id)` — `GET /v1/studio/creatives/{id}/frames/{jobId}`, polls a frame job.
+- `creatives.render` may now answer 200 with the identical existing render (`cached: True`, no new charge); new
+  renders carry `cached: False`.
+
+## [0.32.0] - 2026-10-12
+
+- `cards.route(from_=…, to=…, origin=…, destination=…, mode=…, depart_at=…)` — `GET /v1/cards/route`, travel time
+  from Google Routes (`route` summary + `map` block); billed $0.010 traffic-aware driving, $0.005 walk / bike / transit.
+- `cards.track(awb, carrier=…)` — `GET /v1/cards/track`, parcel history (Sameday live; FAN Courier, Cargus, DPD with a
+  gateway carrier contract only).
+- `cards.efactura(limit=…)` — `GET /v1/cards/efactura`, your fiscal documents issued through Brivio with their ANAF
+  e-Factura status.
+
+## [0.31.0] - 2026-10-12
+
+- `client.creatives` — codai Studio editable creative documents (ECD) with an append-only version history:
+  `list(kind, before, limit)`, `create(body)`, `retrieve(id, seq=None)`, `outline`, `find(id, query)`,
+  `subtree(id, ids=None, scene_id=None)`, `edit(id, base_seq, ops, intent=None, author=None, force=None)`
+  (`POST /v1/studio/creatives/{id}/ops`; a stale base is rebased, overlapping changes raise `CodaiError` 409
+  `stale`), `versions`, `revert(id, seq)`, `fork`, `delete`, `lint`, `compiled` (preview HTML with signed asset URLs)
+  and `render` (queues a `codai-render`, HTTP 202; poll `renders.retrieve`). `frame` is reserved (501
+  `not_implemented`). Everything except `render` is free.
+- `create_timeline` replies carry `creative_id`; `renders.retrieve` reports `creative_id` for renders linked to a
+  creative.
+- Generated types: `Creative*` TypedDicts and the 15 creatives operation aliases.
+
+## [0.30.0] - 2026-10-12
+
+- `lists` resource — codai Lists + Share to codai: `list`, `create`, `reorder`, `update`, `delete`, `items`, `add`,
+  `update_item`, `delete_item`, `image` (`ListItemImage(data, content_type)`), `export_markdown` (str), `share` (`POST /v1/share`,
+  201 created / 200 dedupe) and `stream()` (generator) (`event: list` frames).
+- `schedules.audit({ schedule_id?, limit? })` — `GET /v1/schedules/audit`, the append-only scheduler audit log.
+- Generated types: 14 new operations (lists, share, schedule audit).
+
+## [0.29.0] - 2026-10-12
+
+- `cards.render(block, format=None, width=None, height=None, theme=None, scale=None)` — `POST /v1/cards/render`:
+  renders a `chart` or `qr` card block to PNG (default) or SVG on the gateway and returns
+  `CardsRendered(data, content_type, width, height)`. Free, 2000 renders per user per day. Mermaid `diagram` / `math`
+  blocks are not supported yet (422 `render_unsupported`).
+- Generated types: operation `cardsRender`.
+
+## [0.28.0] - 2026-10-11
+
+- Studio scene timelines on `client.renders`: `create_timeline(body)` for `POST /v1/studio/timelines` (image/video
+  scenes, optional narration billed now and a music bed; HTTP 202, poll with `retrieve`), `regenerate_scene(render_id,
+scene_id, body)` for `POST /v1/studio/renders/{id}/scenes/{scene_id}/regenerate` (a new render with one scene
+  replaced) and `list(before=None, limit=None, root_render_id=None)` for `GET /v1/studio/renders` (no polling, no URLs).
+- `retrieve(render_id, expand_assets=True)` sends `?expand=assets`; timeline renders carry `timeline`,
+  `parent_render_id`, `root_render_id` and `stage`.
+- Generated types for `StudioRenderList`, `StudioTimelineCreate`, `StudioTimelineAccepted`, `StudioSceneRegenerate`,
+  `StudioSceneRegenerateAccepted`, `StudioTimeline`; operations `listRenders`, `createTimeline`, `regenerateRenderScene`.
+
+## [0.27.0] - 2026-10-11
+
+- `client.cards` — typed card data from `/v1/cards/*`. Every JSON method returns `{"block": ..., "provenance": ...}`
+  where `block` is already a valid `@codai/cards` block: `cards.weather(lat, lng, name=None)` (MET Norway, free),
+  `cards.places(q=None, lat=None, lng=None, radius=None, type=None, limit=None, lang=None, region=None,
+open_now=None)` (Google Places, billed $0.032 per call, 200/day), `cards.geocode(address)` (map block + raw
+  `results`, billed $0.005), `cards.fx(source=None, from_=None, to=None, amount=None, quote=None, symbols=None)`
+  (BNR / ECB, free), `cards.unfurl(url)` (links preview block, 500/day) and `cards.static_map(center=None,
+markers=None, ...)` returning `CardsStaticMap(data, content_type, attribution)` for the PNG (billed $0.002, counted
+  in the 500/day maps limit).
+- Generated types for `CardDataResponse`; operations `cardsWeather`, `cardsPlaces`, `cardsGeocode`,
+  `cardsStaticMap`, `cardsFx`, `cardsUnfurl`.
+
+## [0.26.0] - 2026-10-11
+
+- `schedules.pause/resume` (vacation mode with optional backfill), `schedules.templates/render_template`, and
+  `client.heartbeats` (`list/create/update/delete`) for dead-man-switch monitors of crons codai does not run.
+
+## [0.25.0] - 2026-10-11
+
+- `client.schedules` — the codai Scheduler: `list`, `create`, `preview` (cron or plain words in English or Romanian →
+  next runs + DST warning), `get`, `update`, `delete`, `run_now`, `runs`, `recent_runs`, `decide` (approve / skip a run
+  waiting for you) and `stream`; `schedules.hosts.list/update` (devices as hosts: os, kind, tags); and the host executor
+  protocol `schedules.executor.claim/heartbeat/result/approve_command`.
+- `transcripts.publish_live` mapped to `publishTranscriptLive` in the operation table.
+
+## [0.24.0] - 2026-10-10
+
+- `client.settings` — account-level settings shared by every device of the account (per user, not per key):
+  `client.settings.get()` for `GET /v1/settings` (full object with defaults, `version` 0 / `updatedAt` None until
+  the first write) and `client.settings.update(partial, version=None)` for `PATCH /v1/settings` (deep merge, strict
+  validation; a stale `version` raises a 409 `CodaiError` with code `version_conflict`).
+  `settings_conflict_current(err)` returns the current state carried by that 409.
+- Generated types for `SettingsResponse`, `UserSettings`, `TranscriptsSettings`, `SettingsPatchRequest`,
+  `SettingsConflict`; operations `getSettings`, `patchSettings`.
+
+## [0.23.0] - 2026-10-10
+
+### Added
+
+- `client.documents` — codai Studio documents (`codai-docs`): `create(body)` builds a deck (PPTX), doc (DOCX) or sheet
+  (XLSX) from a JSON `spec` (free; `file` is an asset dict with `kind: "document"`), `retrieve(id)` returns it with the
+  stored `spec`, `preview(id)` queues a `codai-docs-preview` preview (tagged PDF + page PNGs + contact sheet, HTTP 202),
+  `retrieve_preview(preview_id)` polls it and `video(id, body)` turns a deck into an MP4 queued as a `codai-render`
+  render (optional narration). Pollers `create_preview_and_wait(id, poll_interval=, timeout=)` returns the succeeded
+  preview and `video_and_wait(id, body, poll_interval=, timeout=)` returns a `RenderResult`; both raise
+  `RuntimeError` on `failed` and `TimeoutError` on timeout.
+- Generated types: `DocumentCreateRequest`, `Document`, `DocumentPreview`, `DocumentPreviewAccepted`,
+  `DocumentVideoRequest`, `DocumentVideoAccepted`, `DeckSpec`, `DocSpec`, `SheetSpec` and their parts; `AssetKind`
+  gains `document` and `preview`.
+
+## [0.22.0] - 2026-10-10
+
+### Added
+
+- `client.sfx` — the codai Studio sound-effects library (`codai-sfx`, curated CC0 sounds + a Freesound CC0-only
+  fallback): `client.sfx.search(q=, category=, tags=, max_duration=, limit=)` for `GET /v1/studio/sfx` returns
+  `SfxSearchResult` (`data`, `source`, `fallback_unavailable` from `x-codai-sfx-fallback: unavailable`; `tags` accepts
+  a list) and `client.sfx.import_(sfx_id)` for `POST /v1/studio/sfx/{sfx_id}/import` returns the asset dict
+  (`kind: "sfx"`, usable as `asset:<asset_id>` in renders). The method is `import_` because `import` is a Python
+  keyword. Both are free.
+- Generated types: `SfxDto`, `SfxCategory`, `SfxList`, `searchSfxResponse`, `importSfxResponse`; `AssetKind` gains
+  `sfx`.
+
+## [0.21.0] - 2026-10-10
+
+### Added
+
+- `client.audio.speech_json(body)` for `POST /v1/audio/speech` with `timestamps` and/or `store`: returns
+  `SpeechJsonResult` with per-word timings (`words`), decoded `audio` or the stored studio `asset` (kind `speech`,
+  usable as `asset:<id>` in renders), `duration_seconds`, `cost_micro_usd` and `timestamps_unavailable`
+  (`x-codai-timestamps: unavailable`). Defaults to `codai-speech` and `timestamps: True`.
+- Generated types: speech models `codai-speech` (Chirp 3 HD, 30 native Romanian voices) and
+  `codai-speech-expressive` (Gemini 2.5 Flash TTS, honours `instructions`); request fields `language`, `timestamps`,
+  `store`; new TypedDicts `SpeechJson`, `SpeechWord`, `MusicTrackMeta`; music request fields `lyrics`,
+  `instrumental`, `bpm`, `duration_seconds`, `language` and the `codai-music` / `codai-music-pro` /
+  `codai-music-legacy` model enum.
+- `MusicResult.tracks` (per-item `lyrics`, `caption`, `lyrics_match`, `warnings`) and `MusicResult.warnings`
+  (e.g. `lyrics_mismatch` for best-effort non-English lyrics).
+
+### Changed
+
+- `codai-music` is now Lyria 3 Clip (30 s MP3 with vocals); Lyria 2 stays available as `codai-music-legacy`.
+- `audio.speech()` / `speech_detailed()` no longer send the Azure default `voice: "alloy"` for `codai-speech*` models.
+- `music.generate()` client timeout floor raised from 180 s to 300 s (Lyria 3 Pro songs).
+
+## [0.20.0] - 2026-10-10
+
+### Added
+
+- `client.renders` for codai Studio renders (`codai-render`, HyperFrames HTML composition → MP4, up to 180 s at
+  1080p): `create(body)` (`POST /v1/studio/renders`, HTTP 202), `retrieve(id)` (`GET /v1/studio/renders/{id}`, status
+  - `progress` + `qa`) and `render(body, poll_interval=, timeout=)`, which polls until `succeeded` and returns a
+    `RenderResult` (`asset`, `poster_url`, `qa`, `cost_micro_usd`); raises `RuntimeError` on `failed` and
+    `TimeoutError` on timeout.
+- Regenerated `_types.py` (`StudioRender*` schemas, the `createRender` / `getRender` operations, the
+  `asset_not_found`, `too_many_renders` and `renders_unavailable` error codes).
+
+## [0.19.0] - 2026-10-10
+
+### Added
+
+- `client.assets` for codai Studio assets (`/v1/assets`, owner-scoped, kept 90 days): `list(kind=, before=, limit=)`,
+  `iterate()` (follows `next_before`), `retrieve(id)`, `content_url(id)` (fresh 1 h signed URL), `content_path(id)`
+  (authenticated `GET /v1/assets/{id}/content` redirect), `content(id)` (bytes) and `delete(id)`.
+- `response_format: "url"` on `images.generate` / `images.edit` and `music.generate` bodies, and
+  `response_format="url"` on `videos.retrieve` / `videos.generate`: results are stored as assets and returned in
+  `ImagesResult.assets` / `MusicResult.assets` / `VideoResult.asset` (signed URLs) instead of bytes. The default
+  `b64_json` behaviour is unchanged.
+- Regenerated `_types.py` (`AssetDto`, `AssetKind`, `AssetList`, `AssetDeleted`, `MediaResponseFormat`, the four asset
+  operations, `assets_unavailable` error code).
+
+## [0.18.0] - 2026-10-09
+
+### Added
+
+- codai Studio media (US-only, English prompts): `music.generate(body)` -> `MusicResult` (`POST /v1/audio/music`,
+  Lyria 2, 30 s WAV clips, timeout floor 180 s) and `videos.create` / `.retrieve(id)` /
+  `.generate(body, poll_interval=5.0, timeout=300.0)` -> `VideoResult` (`POST /v1/videos` + `GET /v1/videos/{id}`,
+  Veo 3.1 Fast; raises `RuntimeError` on `failed`/`filtered`, `TimeoutError` on timeout).
+
+## [0.17.0] - 2026-10-08
+
+### Added
+
+- `client.tools.ground(body)` — `POST /v1/tools/ground`: locate a UI element described in words on an inline
+  screenshot (codai's self-hosted GUI grounding model).
+
+## [0.16.1] - 2026-10-07
+
+### Added
+
+- Web search result `source` gains `ocw` (MIT OpenCourseWare course records for education queries that name MIT / OCW).
+
+## [0.16.0] - 2026-10-06
+
+### Added
+
+- `transcripts.export(id, "pdf" | "docx")` returns `{bytes, filename, content_type}`; text formats keep returning
+  `{text, ...}`.
+- `transcripts.shares.create(id, expires_in_days=None)` / `.list(id)` / `.revoke(id, share_id)` - read-only share
+  links (`https://hub.codai.ro/t/<token>`; the token is returned once).
+- `transcripts.public_view(token)` - the read-only view behind a share link.
+- Types regenerated from the gateway spec (`TranscriptShare*`, `PublicTranscript`).
+
+## [0.15.3] - 2026-10-06
+
+### Added
+
+- `TranscriptDetail` types gain `processing` (pending job: `kind`, `attempts`, `max_attempts`, `last_error`,
+  `next_attempt_at`) - regenerated from the gateway spec.
+
+## [0.15.2] - 2026-10-05
+
+### Added
+
+- `POST /v1/tools/search` types: request `depth: "fast" | "deep"` (deep = more sources, ~12 s budget); `mode`
+  also accepts `fast` / `deep` as an alias (means `codai` with that depth).
+
+## [0.15.1] - 2026-10-05
+
+### Added
+
+- `POST /v1/tools/search` types: request `fallback: "auto" | "off"` (automatic Grounding with Google Search when
+  codai results are weak, returned as a separate `google` block + `fallback: { used, reason }`), new intents
+  `academic` / `definition` / `education`, and new result sources (`mdn`, `mslearn`, `hn`, `osv`, `endoflife`,
+  `openalex`, `crossref`, `arxiv`, `europepmc`, `doaj`, `core`, `wiktionary`, `openstax`, `dexonline`).
+
+## [0.15.0] - 2026-10-05
+
+### Added
+
+- `client.usage.get(query)` → `GET /v1/usage`: your own requests, errors, tasks, tokens and cost grouped by
+  `day` | `key` | `device` | `app` | `platform` | `model` | `client` | `tag`, filtered by `key`, `device`, `app`,
+  `platform`, `model`, `tag`; `from` / `to` are inclusive `YYYY-MM-DD` (default last 30 days, max 400).
+  `cost_micro_usd` is USD micros as a string. Unknown query keys raise `ValueError`.
+- `client.keys.list(include_revoked=False)` → `GET /v1/keys`: your API keys with `kind`, `device_id`, `platform`,
+  `client_version`, `tags` and `current`.
+- `client.keys.set_tags(id, tags)` → `PATCH /v1/keys/{id}`: replace a key's tags (≤ 16, each ≤ 32 chars).
+- Generated TypedDicts `UsageReport`, `UsageRow`, `ApiKeyInfo` (+ `getUsage` / `listKeys` / `setKeyTags` aliases).
+
+## [0.14.0] - 2026-10-05
+
+### Added
+
+- `client.transcripts` — Transcripts API: `list` (folder / tag / pinned / status, cursor), `create`, `get`,
+  `update`, `delete`, `finalize`, `retranscribe`, `reanalyze`, `add_bookmark`, `rename_speaker`,
+  `edit_utterance`, `audio`, `export` (md / txt / srt / vtt / json → `{text, filename, content_type}`),
+  `accept_prompt`, `usage`, `search` (lexical / hybrid), `ask`, `stream` (SSE generator), plus
+  `transcripts.folders.*` and `transcripts.vocabulary.get|put`.
+- `client.transcripts.upload_audio(data, track=, transcript_id=, chunk_size=, upload_id=, on_progress=)` —
+  resumable tus 1.0 upload (`transcripts.uploads.capabilities|create|offset|append`).
+- `PromptSource` gains `transcript`.
+
+## [0.13.1] - 2026-10-05
+
+### Added
+
+- `POST /v1/tools/search` types: request `mode: "codai" | "google"` (default `codai`) and the
+  `ToolSearchGoogleResponse` shape (Grounding with Google Search: `answer`, `sources`, `queries`,
+  `search_suggestions_html`, `attribution`), plus the `releases` / `registry` / `engine` result sources.
+
+## [0.13.0] - 2026-10-05
+
+### Added
+
+- Pending prompts track delivery: statuses `sent`, `answered`, `verified` and fields `sent_at`, `answered_at`,
+  `verified_at`, `match_ref`, `match_turn_id`, `match_score`, `base_sha`, `head_sha`, `report`, `report_at`.
+- `client.prompts.match(id, body)` → `POST /v1/prompts/:id/match`; `client.prompts.verify(id, body)` →
+  `POST /v1/prompts/:id/verify` (report: done / partial / not started, what changed, what remains);
+  `client.prompts.verify_many(body)` → `POST /v1/prompts/verify`; `client.prompts.stream()` → `GET /v1/prompts/stream`
+  (SSE generator of `event: prompt` upserts / deletes).
+
+## [0.12.0] - 2026-10-05
+
+### Added
+
+- `client.audio.speech_stream(body)` → `POST /v1/audio/speech/stream` (`codai-tts-ro`, Azure neural TTS with
+  visemes): generator of `{"type": "start"|"audio"|"viseme"|"word"|"done"|"error", ...}`; `audio` events carry the
+  decoded `audio_bytes`. `createSpeechStream` in the parity table.
+- `client.system_one.batch(items, questions, timeout_ms=None, fallback=None)` — the same questions about up to 16
+  states in one call; `system_one.create` documents the new `items`, `timeout_ms` and `fallback` fields and the
+  `fallback_used` response flag.
+- Ephemeral token scope `live` (realtime + audio) accepted by `tokens.create`.
+- `_types.py` regenerated (`SpeechStreamRequest`, `SpeechStreamEvent`, batch fields on `SystemOneRequest` /
+  `SystemOneResponse`).
+
+### Not included
+
+- No realtime (WebSocket) helper: the SDK is zero-dependency and the standard library has no WebSocket client, and
+  `/v1/realtime` is outside the OpenAPI parity contract. Use `tokens.create("live")` and any WebSocket client against
+  `wss://ai.codai.ro/v1/realtime?model=codai-transcribe-live&language=ro&key=<token>` (protocol in the docs,
+  gateway → Embeddings, audio & realtime).
+
+## [0.11.0] - 2026-10-02
+
+### Added
+
+- `client.devices.revoke_self_key()` — `POST /v1/keys/self/revoke`: the calling API key revokes itself (sign out).
+
+## [0.10.0] - 2026-09-29
+
+### Added
+
+- `client.prompts.{list, create, get, update, delete, launch}` — pending prompts (`/v1/prompts/*`):
+  ready-to-paste prompts left for a future chat, with `goal_key` supersede, `not_before` reminders
+  (`prompt.due` notifications) and `launch` into a new session on a device. `launch` is never retried.
+  `_types.py` regenerated (`Prompt`, `PromptCreate`, `PromptPatch`, `PromptCreated`, …).
+
+## [0.9.0] - 2026-09-29
+
+### Added
+
+- `client.sessions.dispatch(session_id, *, device_id=..., text=..., mode="send"|"open", turn_id=..., control_id=...)`
+  keyword form. `mode="open"` hands the session off — the target only opens it (no user turn), and
+  `text`/`turn_id` are dropped. The positional `body` dict still works (keywords override its keys); a
+  `send` without `text` now raises `ValueError` client-side instead of a server `400`.
+- Regenerated `src/codai/_types.py`: `SessionDispatchRequest.mode` (`text` now optional),
+  `SessionDispatchResponse.mode`, `SessionControl.mode` (`'open'` on handoff controls) and
+  `DeviceListItem.online` (seen in the last 2 minutes).
+- `client.images.edit({"prompt": ..., "image": {"b64_json": ..., "mime_type": ...} | [...]})` →
+  `POST /v1/images/edits` (JSON body, 1–14 input images; `mask` unsupported). `editImage` in the
+  parity table; `EditImageBody` / `CoreImageEditRequest` types.
+- `ImagesResult.mime_types` and `ImagesResult.data_region` (`x-codai-data-region`, `eu` | `global`).
+- `generate()` types gain `reference_images`, `aspect_ratio`, `image_size` and the
+  `codai-image-fast` / `codai-image-pro` model ids.
+
+### Changed
+
+- `codai-image` is now `gemini-3.1-flash-image` on the Vertex EU endpoint (2.5 is shut down on
+  2026-10-02 and served as `codai-image`). `cost_micro_usd` is usage-based (≈ $0.074 per default
+  1K image) instead of a flat $0.04. The image request floor timeout is 150 s.
+
+## [0.8.0] - 2026-09-29
+
+### Added
+
+- `client.images.generate({"prompt": ..., "n"?, "size"?, "negative_prompt"?, "model"?})` →
+  `POST /v1/images/generations` (base64 only). Backed by `gemini-2.5-flash-image`; `model`
+  defaults to `codai-image` (`imagen-4` / `imagen-4-fast` are legacy aliases). Flat $0.04 per
+  returned image, reported as `cost_micro_usd`.
+- EXPERIMENTAL, unaudited (HIDE protocol) agent-output signatures — provenance only, never
+  authentication:
+  - `EnvironmentTask.output_signature: Optional[EnvironmentTaskOutputSignature]` plus the
+    `AgentOutputKey` / `AgentOutputVerifyRequest` / `AgentOutputVerifyResult` types
+    (regenerated `src/codai/_types.py`).
+  - `client.verify.agent_output_key()` → `GET /v1/verify/agent-output/key` and
+    `client.verify.agent_output(payload, signature)` → `POST /v1/verify/agent-output`
+    (public; never retried, 503 `signing_disabled` when signing is off).
+
+## [0.7.0] - 2026-09-27
+
+### Added
+
+- Resolve pricing experiment (regenerated `src/codai/_resolve_types.py`):
+  - `ResolveQuote` `intro_price_micro_eur` / `list_price_micro_eur` — present while the €1
+    intro price applies to your first paid accept.
+  - `ResolveAccepted` `paid_with` (`ResolvePaidWith`: `intro` | `credit` | `wallet` | `free`)
+    and `held_micro_eur`; `ResolveJob.paid_with` once accepted.
+  - `ResolveInsufficientFundsError.bundle_url` (`Optional[str]`) — fix-bundle checkout
+    (10 t7 fixes for €49), set only for `t7` jobs; also appended to the error message.
+  - `client.resolve.balance()` — `GET /v1/resolve/balance` (operation `getResolveBalance`) →
+    `ResolveBalance {fix_credits, intro_available, wallet_micro_eur}`.
+
+## [0.6.0] - 2026-09-27
+
+### Added
+
+- Signed Resolve attestations: `ResolveAttestation.signature` (DSSE envelope, `None` for
+  pre-signing attestations) and `ResolveAttestationEnvelope` / `ResolveKeys` / `ResolveKey`
+  types (regenerated `codai._resolve_types`).
+- `client.resolve.keys()` — `GET /v1/resolve/keys` (operation `getResolveKeys`).
+- `client.resolve.verify_attestation(att, keys=None)` and offline
+  `verify_attestation_with_keys(att, keys)` → `{"ok": True, "keyid"}` or
+  `{"ok": False, "reason": "UNSIGNED" | "BAD_SIGNATURE" | "STATEMENT_MISMATCH" | "BAD_PAYLOAD_TYPE"}`.
+  The statement check is stdlib; the ECDSA signature check needs the optional
+  `cryptography` package (lazy import, clear `ImportError` otherwise).
+  `attestation_statement(att)` exposes the recomputed statement.
+
+## [0.5.0] - 2026-09-27
+
+### Added
+
+- Resolve job webhooks: `callback_url` on `ResolveIntake`, one-time `callback_secret` on
+  `ResolveQuote` / `ResolveDeclined`, and `ResolveWebhookEvent` / `ResolveWebhookEventName` /
+  `ResolveWebhookJob` TypedDicts (regenerated `_resolve_types.py`).
+- `verify_resolve_webhook(secret, raw_body, signature_header, tolerance_seconds=300, now=None)`
+  — constant-time (`hmac.compare_digest`) check of `x-codai-signature: t=<unix>,v1=<hex>`
+  with a timestamp window. Exported from `codai` and `codai.resources`.
+
+## [0.4.0] - 2026-09-27
+
+### Added
+
+- `client.resolve` — codai Resolve (pay per verified fix, `https://resolve.codai.ro`,
+  override with `resolve_base_url=`): `submit`, `get`, `accept`, `attestation`,
+  `health`, `wait_for`. `ResolveError` / `ResolveInsufficientFundsError` (402 carries
+  `topup_url`, `balance`, `required`). `RESOLVE_OPERATION_METHODS` maps every
+  `resolve.yaml` operationId; `_resolve_types.py` is generated from it.
+
+## [0.3.1] - 2026-09-26
+
+### Changed
+
+- Regenerated `_types.py`: project prebuild pool fields, environment `prebuild`
+  and `prebuild_claimed` on create (CE-0072).
+
+## [0.3.0] - 2026-09-25
+
+### Added
+
+- Projects + environments resources (26 operations), restoring full parity with the gateway
+  OpenAPI (88 operations): `projects.{list,create,get,update,delete}`,
+  `environments.{list,create,enroll,get,update,start,stop,archive,destroy,enroll_token}`,
+  `environments.members.{list,add,set_ssh_key,set_role,remove}`,
+  `environments.ports.{list,set,remove}` and `environments.secrets.{list,set,remove}`.
+  `enroll`, `create` and `enroll_token` are never retried (the enrolment token is single-use).
+- `_types.py` regenerated: `Project*` and `Environment*` TypedDicts.
+- `system_one.create({"state", "questions"})` — typed decisions from codai-s1 (`POST /v1/systemone`,
+  Choice / Score / Noul with calibrated probabilities). Raises on 503 `s1_unavailable`; keep a fallback.
+- `triggers.*`, `notifications.*`, `environments.members.ssh_cert`, `environments.ssh_ca`,
+  `environments.tasks.{list,create,get}` — parity with the TypeScript SDK.
+
+## [0.2.1] - 2026-09-24
+
+### Added
+
+- `WalletLedgerEntry.source` literal gains `compute_charge` — the debit written every 10 minutes
+  for a running managed cloud environment (F2). Additive; regenerated from the gateway OpenAPI.
+
+## [0.2.0] - 2026-09-23
+
+### Added
+
+- Full gateway coverage: one method per `operationId` in `apps/docs/openapi/en/gateway.yaml`
+  (62 operations), grouped like the TypeScript SDK — `client.chat.completions`, `messages`,
+  `responses`, `embeddings`, `audio`, `tokens`, `models`, `health`, `agents` (+ `agents.runs`),
+  `tools`, `tasks`, `sessions` (+ `events`, `controls`, `lease`, `shares`), `devices`, `hosts`,
+  `orgs` (+ `members`), `account`, `receipt`, `feedback`, `phone_models`.
+- `OPERATION_METHODS` (operationId → dotted method path) and `tests/test_parity.py`, which
+  parses the OpenAPI spec and fails on a missing, stale or non-callable mapping and when the
+  Python table diverges from the TypeScript one.
+- Generated `codai._types` (`TypedDict`s for every schema plus `<operationId>Body` /
+  `<operationId>Response`) from `scripts/gen-types.py`; the parity test fails when it is stale.
+- SSE streams: `chat.completions.stream()` / `messages.stream()` / `responses.stream()` return
+  iterables with `.text()`, raw chunks/events and a `.final` aggregate (content, tool calls,
+  usage, finish reason, routing headers); `agents.runs.stream()`, `sessions.stream()` and
+  `hosts.stream()` are generators of `{"event", "data"}` frames.
+- `ext={...}` extension bag on every method covering all documented `X-Codai-*` request
+  headers (`effort`, `thinking`, `thinking_budget`, `cache`, `no_task`, `task_id`, `device`,
+  `compact`, `best_of`, `share_token`, …) plus raw `headers`; client-wide `defaults`, `device`,
+  `device_name`, `device_platform`, `client` constructor options.
+- `CodaiError.code`, `.request_id`, `.retry_after` parsed from the gateway error envelope.
+- Dev extras `pip install -e ".[dev]"` (pyyaml + pytest); an `http.server` stub test suite.
+
+### Changed
+
+- `__version__` now reports the real version (`0.2.0`; it said `0.1.0` in 0.1.1).
+- `chat_stream()` is built on `chat.completions.stream()`; 429/5xx retries add jitter.
+- Runtime stays zero-dependency (stdlib `urllib`), Python 3.9+.
+
+### Backward compatibility
+
+- Every 0.1.x method keeps its name, signature and return shape: `chat()`, `chat_stream()`,
+  `agents_run()`, `feedback()`, `models()`, `mint_token()`, `embeddings()`, `transcribe()`,
+  `speech()`. `chat`, `embeddings`, `models` and `feedback` are callable resource groups, so
+  `client.chat([...])` and `client.chat.completions.create({...})` are the same call.
+
+## [0.1.1] - 2026-09-13
+
+### Changed
+
+- Repository moved to https://github.com/codai-ro/codai-sdk-python (project URLs updated). No code changes.
+
+## [0.1.0] - 2026-06-23
+
+### Added
+
+- Initial public release.
+- `Codai.chat()` — OpenAI-compatible completions with codai extensions
+  (`session_id`, `agent_mode`, `compact`, `best_of`).
+- `Codai.chat_stream()` — streaming via generator.
+- `Codai.agents_run()` — server-side agent loop.
+- `Codai.feedback()` — thumbs rating on a completed request.
+- `Codai.models()` — list models available to the key.
+- Zero dependencies (stdlib only), Python 3.9+.
+
+[Unreleased]: https://github.com/codai-ro/codai-sdk-python/compare/v0.2.0...HEAD
+[0.2.0]: https://github.com/codai-ro/codai-sdk-python/compare/v0.1.1...v0.2.0
+[0.1.1]: https://github.com/codai-ro/codai-sdk-python/compare/v0.1.0...v0.1.1
+[0.1.0]: https://github.com/codai-ro/codai-sdk-python/releases/tag/v0.1.0