Files

T

Bohdan Triapitsyn 28f0736d69 feat: small-model utility calls on existing OpenCode providers (#2049 )

Adds a server-side "small model" capability: direct, cheap LLM calls that
reuse the user's existing OpenCode provider logins — the mechanism OpenCode
uses internally for titles and summaries but does not expose through the
SDK or plugins. Zero new dependencies; plain fetch with per-provider wire
formats, credentials never leave the server.

Core (packages/web/server/lib/small-model):
- Resolution mirrors OpenCode's session scoping: explicit settings override
  → small_model from the OpenCode config → family scan within the session's
  provider → the session's own model. The global provider scan only serves
  callers without a session context, and background callers forbid it
  entirely (restrictToPreferredProvider), so conversation content never
  reaches a provider the user didn't pick — explicit choices excepted.
- Per-provider auth replicating OpenCode's plugin loaders: GitHub Copilot
  (device token as bearer, no exchange), ChatGPT plan via the codex
  Responses API (single-flight OAuth refresh written back to auth.json),
  Anthropic messages, Google generateContent, generic OpenAI-compatible.
- OpenCode's free models (opencode/big-pickle, *-free) are never called
  directly; unauthenticated providers are skipped by design.
- Prompt clamping to the model's catalog context limit; thinking disabled
  where a wire switch exists (Z.AI/GLM, MiniMax-M3, Gemini Flash); robust
  content parsing with a clear error when a thinking model spends its whol
  budget on reasoning.
- Settings → Sessions gains a Small Model group: use-default checkbox plus
  an override picker limited to authenticated providers, persisted with
  web/desktop/VS Code sanitization parity.

Consumers:
- Session assist: a server-side watcher on the global SSE hub generates a
  short recap and one suggested follow-up after a session idles quietly fo
  a minute, stored on session metadata (openchamber.assist). Freshness is
  keyed to the last assistant message id, so new activity invalidates the
  payload everywhere with no extra writes. The chat shows the recap under
  the last message after five quiet minutes and the suggestion as a
  dismissible chip above the composer (tap fills the input, never sends).
  Gated by a new Chat setting (default on) that is a hard generation
  switch. Language is anchored to the conversation itself, with a
  script-mismatch guard against model/backend language hallucination.
- TTS: a third input mode, summarized — long replies are condensed to
  spoken prose before playback on any TTS engine.
- Git: commit-message and PR generation moved off the active chat session
  onto the small model fed with real diffs and the commit list (bodies
  included), with a session-transport fallback for free-model-only setups.
- Notes: Add to notes distills long selections into 1-3 dense sentences
  preserving exact identifiers, with verbatim fallback on failure.

Fixes along the way:
- The global event watcher now starts unconditionally; it was gated behind
  the desktop-notify env, leaving the server-side event hub dead in
  packaged apps.
- OpenCode re-emits message.updated for old user messages after idle; the
  watcher no longer mistakes those for new activity.
- Session metadata merges from a fresh read right before the PATCH, so
  writes made during the generation window (suggestion dismissals, review
  links) are preserved; the assist runtime stops during graceful shutdown.

2026-07-05 23:19:10 +03:00

3.3 KiB

Raw Blame History

Session Assist

Server-side watcher that generates a short recap of the agent's last reply and one suggested user follow-up with the small model (lib/small-model), storing both on the session's metadata under metadata.openchamber.assist.

Flow

createSessionAssistRuntime is a consumer of the server's global SSE fan-out (index.js → onPayload), riding the same upstream connection as notifications. Purely event-driven — dormant sessions never generate anything, there is no backfill and no session scanning.
session.status: idle arms a 60-second per-session timer; any busy/ retry status or a user message.updated clears it (the "1 minute of quiet" rule).
On fire: fetch the session (skip sub-agent sessions with parentID), take the LAST exchange only — the final assistant reply plus the user message it answered (assistant parentID → user id) — and call generateSmallModelText with the session's own provider/model taken from the last assistant message — so the utility call spends the same subscription as the conversation. restrictToPreferredProvider forbids the resolver's global fallback: conversation content never goes to a provider the user didn't pick for the session, unless the small model was chosen explicitly (settings override or opencode config). A resolver 404 is silently skipped.
The {recap, suggestion} JSON is clamped and PATCHed onto the session metadata together with forMessageID (the last assistant message id) and generatedAt. Before writing, the session tail is re-checked (a stale result is dropped) and the metadata is merged from a fresh session read so concurrent metadata writes made during generation are preserved.

Settings gate

sessionAssistEnabled in OpenChamber settings (Settings → Chat, default on) is a hard generation switch checked at fire time: when off, no small-model calls run and nothing is written. Existing payloads keep rendering and can still be dismissed — the switch is about generation, not visibility.

Freshness contract (no clearing writes)

Clients do not need the payload to be deleted: they render it only while assist.forMessageID still equals the session's last assistant message id (and the session is idle). Any new message invalidates the payload everywhere instantly and offline; the next idle cycle overwrites it.

UI consumers (packages/ui)

lib/sessionAssistMetadata.ts — payload parsing.
hooks/useSessionAssist.ts — freshness gating + the 5-minute quiet window for the recap (single timeout to the boundary, no polling).
components/chat/SessionRecapSpacer.tsx — renders the recap inside the fixed-height reserved gap under the last message (height never changes).
components/chat/SessionSuggestionChip.tsx — one tappable suggestion chip near the composer (desktop chips row + above the mobile pill); hidden as soon as the composer has any content. Tap fills the input, never sends.

Limitations

The watcher lives in the web server, so VS Code (extension-only, no web server) does not generate assists; it still renders payloads produced by a web/desktop instance of the same OpenCode server via session.updated.
Metadata payloads ride every session.updated event — keep the clamps (RECAP_CHAR_LIMIT, SUGGESTION_CHAR_LIMIT) small.

3.3 KiB Raw Blame History