Correctness fixes from the internal/ai code review:
- gate reindex commits on the snippet row still existing, so a delete
racing reconcile can't re-insert its embeddings forever
- cap one logical provider request at 90s across all retries, so a
hanging provider can't stall the reply box for minutes
- return proper error envelopes from the copilot message store instead
of raw sqlx errors that surfaced as "Error interface conversion failed"
- reject blank AI replies in the generate-reply and copilot handlers
- build GET tool URLs with url.Parse so a # in the URL doesn't swallow
the query params
- let a blank api_key clear the stored key (masked keeps it), restoring
a way to disable AI
Cleanups and hardening:
- share one SSRF transport between tool and provider clients instead of
building a new transport per AI call
- copy the provider config struct instead of listing every field
- backfill temperature 0.7 for configured providers upgrading from
releases that hardcoded it
- only pass user/assistant roles from copilot history to the provider
- reject enc:-prefixed secrets that would be stored raw and fail decrypt
- rune-safe truncation of test errors, one-line doc comments
The DefaultTransport type-assert guard hardcoded the stdlib pool defaults, the exact duplication cloning was meant to avoid, to defend against a case that never happens in this codebase. Back to the plain clone.
Type-asserting http.DefaultTransport.(*http.Transport) panics if a dependency swaps the global for a different RoundTripper. Fall back to a fresh transport with the standard defaults so boot can't crash.
handleAiPromptSelected now captures the conversation uuid and bails if it changed while the completion was in flight, and uses convertTextToHtml with a null guard, matching handleGenerateReply. This stops a generated prompt from landing in the wrong conversation's draft or crashing on an empty response.
Add ssrf.NewTransport that clones http.DefaultTransport (keeping proxy and connection-pool defaults) and applies the dial guard, then use it for the OIDC, AI, OpenAI, and webhook clients. The bare transports were dropping HTTP_PROXY handling and pool defaults.
The reply-box AI prompts used a legacy inline dialog to set the OpenAI key, hitting a separate PUT /api/v1/ai/provider endpoint. That is now fully covered by the admin provider settings page, so drop the dialog, the endpoint, the UpdateProvider method, the set-completion-key query, and the orphaned i18n keys. With no provider configured an agent now gets the 'ask your administrator' toast instead.
Also make the inline AI prompts (Make Friendly, etc.) toggle the existing isGenerating state so the editor shows the same generating animation as Generate Reply, with a guard against overlapping runs.
The earlier SSRF work covered the provider, custom-tool, webhook, and OIDC
outbound calls but missed the knowledge base URL import path, which still used
http.DefaultClient. An admin importing a URL could reach loopback, link-local,
or RFC1918 hosts. Renamed the Manager's toolClient to httpClient (it is a
general SSRF-guarded outbound client now) and pointed fetchURL at it. The guard
sits on the dialer's Control, so it also runs on redirect targets, not just the
first hop.
Also fixed the OIDC discovery client. Its hand-built transport dropped two
defaults that http.DefaultTransport sets. Added Proxy: http.ProxyFromEnvironment
so it honors HTTP_PROXY/HTTPS_PROXY (SSO discovery now works behind an egress
proxy) and ForceAttemptHTTP2 so HTTP/2 is negotiated.
Move the webhook SSRF guard into a shared internal/ssrf package and wire it
into every place the server fetches an admin-set URL: webhooks, OIDC discovery,
the AI provider base URL, and custom AI tool calls. Add a global [ssrf] config
block, off by default with an allowed_cidrs bypass, so single-tenant self-hosters
keep reaching internal hosts while multi-tenant or hosted deploys can turn it on.
The old [webhook] allowed_hosts key is still read for backward compat and folds
into the guard.
Fix a reindex race: snippet embedding runs outside the lock, so a slower
job from an older edit could commit stale vectors after a newer edit.
Split Reindex into embed (lock-free) and commit (locked), and gate the
commit with a per-snippet generation counter so only the latest edit
wins. Delete drops the counter so an in-flight job can't re-insert
vectors for a deleted snippet.
Clear the DB avatar reference before deleting the avatar media file. The
old order deleted the file first, so a failed DB update left the DB
pointing at a missing file and a broken image.
Also handle unchecked errors flagged by errcheck (tx.Rollback,
html.Render, fmt.Fprintf), guard capToTokens against a negative limit
that would panic, and stop logging the full import URL since it can
carry credentials.
Embedding:
- Persist embedding_max_tokens. It was dropped when saving the provider config
and always reset to the 8192 default, so self-hosted models with a smaller
limit kept getting rejected.
- Split embedding requests to stay under the provider's item and token caps. A
large snippet or URL import used to go in one request, hit the cap, fail, and
get retried every minute forever.
- Skip blank chunks. An empty input string makes the API reject the whole batch.
- Add the tiktoken tokenizer (cl100k_base) with tests for token counting.
Agent prompt:
- Move contact fields, the conversation subject, and custom attributes out of
the system prompt and into a delimited user-role block, so a crafted name or
subject can't act as an instruction.
Frontend:
- Switch the primary and sidebar colors to a green theme.
- Show a "summarizing" info toast while the AI summary runs, and add an info
toast variant.
- Make the snippet content box taller and the snippet dialog wider.
- Drop a hardcoded dark hover color on the scroll-to-bottom button.
Snippet import: new "Import from URL" flow fetches a page and stores its
readable content as a snippet. Extraction uses the mackee/go-readability
library (Mozilla Readability port) and outputs Markdown, so nav/footer
boilerplate is dropped. The snippet list shows the source, and the edit
dialog is now wider with a taller content box.
Summarize: new "Summarize with AI" action on a conversation calls the AI and
adds the result as a private note. It shows an info toast right away so the
user knows it started, since the call can take a few seconds. This adds an
"info" toast variant that any feature can use.
Assistant languages: assistants can be given a list of allowed reply
languages. The assistant replies in the customer's language when it is one of
them, otherwise it falls back to the first. The preview also lists the
knowledge sources it used.
Cleanup: replace hardcoded gray/zinc/white colors with theme tokens
(text-muted-foreground, text-foreground, bg-accent) across several components.
Thread context through provider calls so cancelled requests stop retrying, make snippet delete and FAQ review transitions atomic, cap provider response reads, guard stale AI replies and copilot responses from overwriting newer conversation state, and stop logging raw search queries and chunk content.
Reasoning models like gpt-5 reject max_tokens and non-default temperature.
The client now reads the structured 400, renames max_tokens to
max_completion_tokens or drops the bad tuning param, and retries. This
removes the old rule that only sent max_completion_tokens when reasoning
effort was set.
Provider form gets a "Test connection" button that makes one live call
with the form values and shows the provider's real error. Embedding test
also checks the returned vector length against the Dimensions field.
Add an Anthropic preset (their OpenAI-compatible endpoint) and stop
pre-filling temperature since blank is safe on every model.
Fix the shared Button loader so the spinner is visible on non-filled
variants, and polish the copilot panel: bot avatars on messages, dot
loader while thinking, cleaner input box, and tab slide-in animation.
The model marks its trailing confirmation question with a [[confirm]] line.
We split that off and send it as its own chat bubble so the widget reads like
a real conversation. The customer never sees the marker. Email is left as one
combined message since separate bubbles only suit the widget.
Confirmation messages are tagged with is_confirmation in meta so they no
longer inflate the reply count in assistant stats.
Keep the widget typing indicator alive during long agent runs by
re-broadcasting every 3s, under the widget's 5s typing expiry.
Fix the AI assistant and tool edit forms to stay on the page after save.
They now only go back to the list when creating, matching the rest of the
admin forms.
Custom tools are HTTP calls with no read-only guarantee, so a mutating tool
could fire from copilot chat or while drafting a reply. Copilot and
generate-reply now get only the built-in knowledge search, and custom tools
stay exclusive to assistants where admins pick them explicitly. This matches
Intercom, Chatwoot, and Freshdesk. The contact identity headers now flow only
on the assistant path.
Also:
- new "Offer handoff to a human" switch on assistants (default on). When off,
the hand_off_to_human tool is not registered and the prompt tells the
assistant to say it cannot help instead of offering a human. Safety exits
(error, max turns, other participant) still unassign as before.
- workspace admin instructions from the AI config no longer leak into the
customer-facing assistant prompt.
- copilot and reply-draft prompts now treat conversation text and tool
outputs as untrusted data.
Replies from the AI agent are now converted from markdown to HTML with
goldmark before queueing, so bold, links, and lists render properly in the
widget, agent app, and email. The prompt now allows simple markdown. Raw
HTML in model output is escaped by goldmark, and both frontends sanitize
on render anyway.
Other fixes bundled in:
- validate avatar type and size before creating or updating an assistant,
and roll back the assistant if the avatar upload fails after create
- return 404 from agent update and API key endpoints for AI assistant
identity users, and hide assistants from mention and SLA user pickers
- reserve the autonomous assistant's built-in tool names so custom tools
cannot shadow them
- apply resolve after the reply is posted so the CSAT survey follows the
answer instead of preceding it
- unassign the assistant on handoff even when the fallback team is the
same team
- count reopens by status category instead of status name, and exclude
CSAT messages from the turn cap
- split oversized wrapper divs into child blocks when chunking KB HTML
instead of truncating them
- return 404 when soft-deleting an already-deleted agent, and keep AI
assistants (which have no email) visible in the compact users list
Pass the conversation subject to the AI agent, show the AI's own expectation in the chat widget while it handles a chat, and label AI messages in continuity emails.
Stop dropping customer follow-ups sent mid-response, reset the turn cap only on reassignment, send max_tokens for non-reasoning models, and make FAQ approval atomic. Also exclude CSAT surveys from assistant stats and remove unused scoped-search code.
A single-file page to test the livechat widget's JWT contact auth locally. Signs an HS256 token in the browser from an editable payload and loads the widget with it.
Show an animated border on the reply box while a reply is generating, and hide generate-reply for private notes. Link an AI assistant message author to the assistants settings page for admins who can manage AI.
Reply drafts are now written in first person as the agent and never offer to escalate, since a human is already handling the conversation. Add quick preset buttons to the copilot panel and reuse a shared transcript helper.
Add a reasoning effort field to the completion provider config, sent as-is to the model. Reasoning models such as GPT-5.x need it set to "none" to use tools. Also switch to max_completion_tokens.
Limit custom tools to GET and POST and validate the method on save. GET args now go on the query string and POST args in the body. Also stop sending the raw stored secret when its decryption fails, which would leak ciphertext and fail auth confusingly.
CSAT was only sent when a human resolved from the UI. AI and automation resolves go straight through UpdateConversationStatus and skipped it. Moved the send there so all paths trigger it. It is idempotent and no-ops when the contact has no email.
The agent now only replies to a turn the primary contact authored, so a CC'd or plus-address participant hands off to a human instead of driving tool calls under the contact's identity. It also asks the customer to confirm before resolving, and can read the contact's other recent conversations for context.
Medium (title + short body):
add WIP autonomous AI agent
New internal/aiagent package runs AI assistants that reply to customers
on conversations assigned to them, grounded on a knowledge base. Also mines
resolved conversations for FAQ suggestions. Adds admin UI and the v2.7.0
schema. Still work in progress.
Adds two agent-facing AI features: a copilot chat panel in the conversation
sidebar and a generate-reply button in the reply box. Both run an agentic
tool-calling loop whose first tool searches the knowledge base.
Snippets are chunked and embedded on save, then searched in memory with
brute-force cosine similarity (no pgvector). Providers are split into completion
and embedding types. Both are OpenAI-compatible and the API key is encrypted at
rest. Admins can also register custom HTTP tools the model can call.
A new admin AI settings page covers provider config, snippets, and tools. The
v2.6.0 migration and schema add the ai_knowledge_base, embeddings, and ai_tools
tables plus the ai_providers type column.
Draft attachments now keep url and disposition, so a restored draft can render its attachments correctly instead of losing them.
The prechat phone field now picks a country code from the browser locale when the field is required (falling back to US), so users are not forced to pick one manually. Non-required phone fields still start blank.
Also soften the widget iframe and launcher bubble box shadows to a lighter, tinted look.
CreateContact treated any SetExternalUserID failure as "ext_id taken" and
fell through to the upsert, so a transient DB error created a second
contact with the same email. Now only a unique violation or a contact
deleted mid-flight falls through, other errors are returned.
SetExternalUserID also reports whether a row was actually updated, and
dbutil error checks use errors.As so wrapped errors match. Widget JWTs
with no email now store NULL instead of an empty string.
Abhinav Raut