What's new

All notable changes to ai-chatbot-widget.

The format is loosely based on Keep a Changelog.
The repository carries no git tags and package.json has stayed at 0.1.6 since
2026-07-07, so releases below are grouped by the date the work landed on develop
rather than by version number.

2026-09-25 — Agent targeting, input padding fix

Fixed

The message input was clipped by the window's rounded bottom corners. In the default
messageInputPosition="bottom" layout the input container is the last child of the chat
surface, so its bottom edge is the surface's bottom edge — and that surface is
rounded-4xl (2rem radius) with overflow-hidden. The container carried px-4 pt-2 and no
bottom padding, which put the textarea's bottom-left corner (and the send button's
bottom-right) roughly 36px from the corner arc's centre, i.e. outside a 32px radius: the
corners were cut off flat. The container is now px-4 pt-2 pb-4, so the bottom inset matches
the sides and the controls clear the curve by ~9px. A longstanding bug, visible on every
default-configured widget.

Added

agentId widget option. An optional agent (UUID) that answers this widget's questions,
forwarded as the agentId query parameter on every GET /v1/post/q. It is the intended
replacement for the retired model prop: the API picks an agent per query, so agentId
is deliberately not sent when opening a thread or reading history.

Legacy setups are untouched by design:
- When no agent is configured the parameter is omitted from the query string entirely
  rather than sent empty, which is how the API is told to use the organization's default
  agent — byte-for-byte the request the widget made before this option existed.
- Blank, whitespace-only and non-string values all count as "not set" (host pages are plain
  JavaScript, and templated snippets such as agentId: "{{AGENT_ID}}" can render empty).
  mountChatbotWidget warns — [ChatbotWidget] "agentId" must be a non-empty string — and
  still mounts, because an optional option must not break a page. Values are trimmed.
- agentId joins the thread-cookie scope, so pointing a widget at a different agent starts
  a new conversation instead of continuing one another agent answered. The agent segment is
  appended to the cookie hash only when an agent is set, so a widget without agentId
  resolves the exact cookie it already had and shipping this build interrupts nobody. A test
  pins the pre-feature hash (verbatim-session-b09i4r for the fixture config) to keep it
  that way.
- The dev harness reads an optional VITE_TEST_AGENT_ID from .env.local.

2026-09-09 — Thread API migration, conversation reset, CDN move

Added

Conversation reset. A reset button next to the "Online" status in the header,
backed by a new resetThread(config) in src/utils/api.ts. It opens a fresh thread
and overwrites the verbatim-session-* cookie in place — the browser session itself is
never touched. Three deliberate behaviours: the message list is cleared only after the
new thread resolves (a failed reset leaves the visitor on the conversation they can still
see); the greeting and quick-reply prompts survive, since they are configuration rather
than history; and a double-click joins the inflight create instead of stranding a second
thread.
CLAUDE.md — repo guidance covering the two entry points, the intentional double
import of widget.css, the split vitest.config.ts, the module-level thread caches, and
the four coordinated edits a new theme token requires.
README.md → "How to develop, run and debug" — requirements, three-command start,
annotated src/ tree, config-file table, npm scripts, built-bundle smoke testing, and the
lint noise to expect.

Changed

/v1/session/ → /v1/thread/. Conversations now open with POST /v1/thread/, and the
session → thread rename is carried through the types (ThreadCreateResponse). The cookie
keeps its verbatim-session- prefix on purpose: it is an opaque hash key, and renaming it
would drop every visitor's thread the moment a new build hits the CDN.
Deprecated spec fields dropped. sessionId query param → threadId on /v1/post/ and
/v1/post/q; model is no longer sent in the thread-create body (the prop is kept and still
feeds the cookie hash, so existing embed snippets and cookies keep working); sessionId
removed from the response types.
getPreviewUrls batches. pages is required and capped at 10 values per request, so
longer page lists are split into parallel calls and concatenated. An empty list
short-circuits to [] instead of issuing a request that cannot succeed.
PreviewSize is SMALL | LARGE (was SMALL | MEDIUM). The wrong value happened to work
because unknown sizes fell through to the default, but it was handed to integrators as API.
CI deploys to Scaleway S3. cloudbuild.yaml replaced gcloud storage cp dist/* gs://bmu-verbatim-cdn/widget/chatbot/v1/ with aws s3 sync ./dist/ --delete to
s3://cdn.verbatim-ai.com/widget/chatbot/v1/ (endpoint s3.fr-par.scw.cloud, region
fr-par, credentials via _AWS_ACCESS_KEY_ID / _AWS_SECRET_ACCESS_KEY substitutions).
Explicit waitFor chains were removed; steps run in declaration order. Tests still gate
build and deploy.

Fixed

History displayed newest-first. GET /v1/post/ defaults to order=DESC, so the restored
conversation was reversed. getDialog now sends order: "ASC" explicitly.
Source thumbnails were labelled one page too high. The + 1 is gone from all three places
it appeared — the visible caption, the button title, and the image alt (the last two mattered
for screen readers and hover text). The API echoes the one-based numbering it was sent; the
OpenAPI schema documents this reply as zero-based and is wrong, verified against the live API.
The test fixture encoded the same off-by-one, which is why the bug survived review; it now pins
the real request/reply shape.
The header reset button was invisible. .widget-header__reset { color: … } has
specificity (0,1,0) and lost to the existing
.ai-chatbot button, input, textarea { color: inherit } reset at (0,1,1), painting the button near-black on a near-black header.
The rule is now scoped to .ai-chatbot .widget-header__reset. Icon-only buttons dodge this
because a descendant svg rule colours them — the trap is documented in CLAUDE.md.
Broken-image icons on preview tiles. Previews render asynchronously and URLs are issued
without an existence check, so a tile can 404. A failing thumbnail now removes itself on
onError.

Tests

Suite grew from 129 to 148 tests, including guards that assert model is never sent and that
threadId (not sessionId) is used.

2026-08-01 → 2026-08-04 — Required props and staging

Added

Required props are validated at the mount boundary. mountChatbotWidget enforces a
non-empty accessToken and at least one non-empty corpusIds entry, with console.error +
throw. ChatbotWidget itself no longer re-checks. Covered by src/index.test.tsx.
Staging section in docs/widget-installation-guide.md, documenting
https://staging-api.verbatim-ai.com as the baseUrl for non-production embeds.

2026-07-31 — Source attachments, test suite, CI

Added

Source attachments. Bot answers backed by documents show an "N sources" button that lazily
loads each document's summary, filename and size, plus per-page thumbnails with a click-to-zoom
lightbox and an open/collapse control (src/components/chatbot-sources.tsx, new attachment
endpoints in src/utils/api.ts, ~130 lines of widget.css).
docs/widget-installation-guide.md — the public embedding reference.
Vitest + Testing Library suite across components, utils and both entry points, with
TESTING.md recording the coverage snapshot and known gaps.
Cloud Build pipelines. cloudbuild.yaml (main: install → test → build → deploy) and
cloudbuild-dev.yaml (branches: install → test → build, no deploy). Lint is not run in CI.

Fixed

Thumbnail and message sizing in full-page / fullscreen mode.