feat(chat): chat history — index, admin UI, realtime (Phases 1–3)

[maksymhryzodub-prog]

Replaces mazda-ai's "chats" with durable, browsable chat history across all
channels (web/Bridle, Telegram, Slack). Message content stays in the agents'
append-only S3 JSONL (source of truth); this adds a queryable Postgres index
so chats are listable/filterable without fanning out to agent pods and remain
available when a pod is down.

Architecture

• Postgres = index only (metadata + monotonic counts). Transcript content is read
  live from S3 via IFileGateway.
• Index populated two ways: S3 reconciliation (ChatSyncService) + realtime
  session_activity over the always-on agent↔hub socket.
• The agent runtime stays host-agnostic (no DB dependency) — see companion runtime PR.

Phase 1 — API core (chat slice)

• ChatSession + ChatFeedback Prisma models (FK cascade to Agent) + migration.
• ChatSyncService reconciles from S3 (IFileGateway.list → data/sessions/*.jsonl);
  env-gated setInterval (CHAT_SYNC_INTERVAL_SEC) + manual POST /chats/sync.
  Counts are monotonic (seed/raise, never lower — compaction shrinks the file).
• Endpoints (Owner/Admin): GET /chats, /chats/:id, /chats/:id/messages, POST /chats/sync.
• Shared TranscriptReaderService (agent/file): surfaces compaction summary events
  as a collapsed marker instead of a silent gap, and filters synthetic loop-control
  noise (continuation prompts + the duplicated pre-cutoff partial). Bridle's transcript
  endpoint now delegates to it (byte-identical).

Phase 2 — Admin UI (chat slice)

• Chats list: search / channel / archived / internal filters, server pagination, Sync.
• Chat detail: meta header, tail-first transcript with scroll-up "load older",
  "show tool events" debug toggle, collapsible summary marker, assistant markdown.
• Sidebar "Chats" nav + a "Chats" tab on the agent detail page (scoped by agentId).
• Regenerated admin + app OpenAPI clients (ChatsService).

Phase 3 — Realtime

• IChatGateway.recordActivity: atomic increment (updateMany, race-free), dedup by
  eventId, create fallback with a retry on the concurrent-create unique violation.
  Realtime owns the monotonic counts.
• bridleAgentWs.handler subscribes to session_activity → recordActivity, taking
  agentId from the authenticated socket (never the payload).
• Fixed a DI cycle (Bridle → Chat → File → Bridle) with forwardRef.

Phase 4 — Insights cron-batch

ChatInsightService: LLM summary + structured insights (topics / sentiment / resolved /
language) written onto ChatSession, driven by the same env-gated setInterval
(CHAT_INSIGHT_INTERVAL_SEC). Eligibility gate so we never spend tokens on forgotten/empty
chats: userMessageCount ≥ 3, settled (cooldown ≥ 1h), new activity since summaryAt,
channel ≠ internal, per-run batch cap. Manual POST /chats/:id/summarize kept; a rejected
credential (401/403) aborts the batch instead of hammering every session.

Phase 5 — Feedback + Export (web/admin)

ChatFeedback upsert, unique per (session, message, author); POST/DELETE/GET /chats/:id/feedback (source='admin', authorId = jwt.sub) — admin 👍/👎 in the bubble.
formatChatExport → json / markdown / csv (CSV-escaped); GET /chats/:id/export uses
@res() to bypass the {success,data} interceptor and stream the file with download
headers. Telegram inline-keyboard feedback (→ session_feedback over the agent socket)
deferred as a follow-up.

Phase 6 — App "my history" (app/slices/chat)

End-user-facing Nuxt-layer slice: list + read-only transcript with the summary gist,
insights, feedback, and export; nav link + /chats route guard. MyChatController at
/me/chats is JWT-only (any authenticated user), scoped server-side to the caller's OWN
bridle sessions — every id ownership-checked (404, not 403, so we never confirm someone
else's chat), tool events never exposed (capped types even in json export). Identity mirrors
the bridle client/HTTP handlers: owners/admins resolve to the shared admin channel (the
runtime recognizes them via adminIds containing the literal admin), everyone else their
JWT sub; feedback is authored by the real sub, not the channel. Self-service POST /me/chats/sync (syncForExternalUser) reconciles only the caller's own sessions — a manual
fallback when realtime indexing hasn't caught up.

Hardening (beyond the plan)

fix(llm) — anthropicAuthHeaders: sk-ant-oat… subscription OAuth tokens (what the agents
run on) go as Authorization: Bearer + anthropic-beta oauth,claude-code, not x-api-key
(a hard 401). Applied to chat summaries, the credential health check, and the scenario
generator — so all three work on the exact credential the bots use.
fix(bridle) — HTTP chat (POST /:agentId/message[/sync]) verifies the Bearer JWT the app
already sends and uses its sub/admin as clientId, instead of minting a throwaway
sync- per request; the runtime keys access-approval and session history on that id, so
a stable identity stops the "send the owner your code" prompt firing on every message.
fix(app) — default layout renders a single with a toggled wrapper class instead of
duplicating it across v-if/v-else on isFlush; the branch swap was remounting the page and
re-firing every useAsyncData request when navigating in/out of /agents/:id.

On the "source of truth"

Compaction windows long chats down (the JSONL is a rolling window, not a full ledger).
Product decision: history = gist, not a verbatim log — compaction is untouched and
the summary is surfaced instead.

Verification

• tsc --noEmit clean. Unit specs: chat gateway (reconcile monotonic counts;
  recordActivity create/increment/dedup/create-race), transcript reader (hygiene,
  summary, pagination), reconcile.
• Live e2e vs local Postgres: POST /chats/sync indexes S3 sessions; emitting
  session_activity updates the row live (count/dedup/freshness) with no Sync.

Deploy notes

• Run the migration (migrate:prod) — additive (two new tables, FKs, indexes).
• Set CHAT_SYNC_INTERVAL_SEC to enable periodic reconcile (else chats index only on
  manual Sync).
• Realtime updates require agents on runtime ≥ 0.27.0 (companion PR); without it,
  reconcile still populates the list.