README, CLAUDE.md and .env.example now point at llm.toml for LLM settings
(OPENROUTER_API_KEY stays in env) and describe the legacy env override
window, the discovered picker and the startup cross-check. ADR-0003 records
the move and amends ADR-0002's deployment-config home; CONTEXT.md's
Deployment config / Hard default entries follow. Source comments and the
"set LLAMA_SWAP_URL" error strings name the llm.toml keys.
Also fixes README's AGENTIC_MAX_ITERATIONS default (10 -> 12, what the
code uses).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Nothing in the client stopped it: a message carrying images became image_url
parts and went out. Photos stayed local only because callers routed around it.
messages_to_openai, which every OpenRouter request passes through, now refuses
any image (IMAGES_STAY_LOCAL), and describe_image always errors; hybrid mode
describes photos with the local model. A caller bug now fails the turn instead
of uploading a photo.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Long insight/persona chats on llama-swap failed with "exceeds the
available context size" even after raising Context in the app:
- llama-server fixes its window at launch (-c) and ignores num_ctx, but
the history budget trusted the requested num_ctx, so raising it kept
MORE history than the server could hold. LlmClient::context_window()
now reports the served window (llama-swap: /upstream/<model>/props,
cached 10 min) and the budget uses min(requested, served).
- apply_context_budget could only drop tool-call pairs, so plain Q&A
chats never shrank. It now also drops whole earlier exchanges (keeping
user/assistant alternation) and, last, the current turn's tool results,
never the in-progress question.
- The budget only ran once, before the tool loop. ContextFitter now
fits every LLM call (on a copy; the stored transcript is untouched)
and, on a context-overflow error, shrinks the budget from the server's
reported token counts and retries once.
GET /insights/models reports each llama-swap model's context_window when
it's known without loading the model (cached, or currently running), for
the app's Context picker.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
LlmClient methods no longer take an OpenTelemetry Context. llm.request is
now a tracing span (otel.kind=client, llm.provider/operation/model) that
parents on the caller's ambient span; traceparent injection reads the
span's OTel context via OpenTelemetrySpanExt.
The src/ai feature spans (handlers, insight generation, chat, persona
chat, daily summaries, TTS) move to #[instrument] / info_span!, and every
LLM-bearing tokio::spawn carries its parent span with .instrument(): the
streaming chat turn (chat_turn_stream_inner), async chat / persona turns,
background insight jobs and TTS jobs. Sub-spans in the agentic loops
(ai.chat.loop, ai.agentic.loop) instrument each LLM / tool await.
DAO calls keep their explicit Context (derived from the tracing span)
until the DAO migration (#141).
tests/otel_llm_spans.rs pins nesting + traceparent for the blocking and
streaming paths across a spawn on a multi-thread runtime.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Addresses the P2 refactoring follow-ups parked on issue #130:
- otel::with_traceparent(builder, &SpanContext) — one shared W3C
traceparent injector. Replaces the 4-line inline header loop in
llamacpp (3x) and openrouter (3x) and removes ollama's private
inject_traceparent method (4 call sites now use the shared helper).
Takes &SpanContext (not &dyn Span) because the Span trait is not dyn
compatible in opentelemetry 0.31.
- otel::record_http_failure(span, status, message) and
otel::set_http_status(span, status) — collapse the repeated
'set http.status_code attribute + set_status(Status::error)'
boilerplate across all three clients (~10 sites).
Net -45 lines. The remaining if-let-Some(span.as_mut()) guards (token
attributes, parse-failure status) stay inline rather than growing a
SpanHandle wrapper type for a pattern that's already readable.
Add OpenTelemetry tracing to every LLM client call so an insight/chat
turn is a single trace from HTTP entry to each downstream inference.
- LlmClient trait: thread a &opentelemetry::Context through all 5
methods (generate, chat_with_tools, chat_with_tools_stream,
generate_embeddings, describe_image), matching the DAO convention of
explicit context threading rather than ambient Context::current().
- otel::start_llm_request_span: single shared helper that every
provider (openrouter / llamacpp / ollama) funnels through, so the
span shape is identical: name llm.request, SpanKind::Client,
llm.provider / llm.operation / llm.model.
- Per-method spans in all three clients with the spec's operation enum
(chat | chat_stream | embeddings | describe_image | generate); stream
spans are owned and moved into the async_stream generator so they end
when the response drains, not when headers arrive.
- llm.tokens.input / llm.tokens.output recorded on non-stream and
stream completion; llm.cost_usd from usage.cost (OpenRouter);
http.status_code + error status on failure.
- W3C traceparent injected on outgoing LLM HTTP requests (llama-swap,
Ollama, OpenRouter) gated on span_context.is_valid() so debug
no-op-tracer builds don't send an all-zero parent.
- Turn-level spans (ai.chat.turn.execute, ai.persona_chat.turn.execute)
now carry llm.backend + llm.model; with_entry streaming paths thread
the turn context (turn_cx) to child describe_image / agentic-loop
calls instead of the ambient context.
clippy clean; 690 lib + 775 integration + 4 doc tests green.
OpenRouter reserves credit against the requested max_tokens, falling
back to the model's full output window when omitted, so expensive
models failed with credit-limit errors even for short replies. Default
the cap to 16384; OPENROUTER_MAX_TOKENS overrides it and 0 disables it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- search_rag reranker now logs wall-clock time around the ollama.generate
call, the candidate count / top-N going in, and the final reordering.
The "final indices" + swap-count line is info level so it's always
visible; detailed before/after previews stay at debug for when you want
to inspect reranker quality.
- New OllamaClient::generate_no_think convenience that sets Ollama's
top-level think:false on the request, plumbed through try_generate via
a new internal generate_with_options. Used only by the reranker today;
avoids the chain-of-thought tax on reasoning models (Qwen3/VL,
DeepSeek-R1 distills, GPT-OSS) when the task has nothing to reason
about. Server-side no-op on non-reasoning models.
- OpenRouter chat_with_tools "missing choices[0]" error now includes the
actual response body — extracts structured {error: {code, message}}
when OpenRouter surfaces it (common for upstream-provider issues like
rate limits and content moderation), otherwise falls back to a
truncated raw-JSON view.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Few-shot injection on /insights/generate/agentic: compresses prior
training_messages into trajectory blocks (tool calls + result summaries)
and injects into the system prompt. Hardcoded default ids with optional
request override.
- New fewshot_source_ids column on photo_insights (+ migration) to track
which exemplars influenced a given row, for downstream training-set
filtering. Chat amend rows stamp None with a lineage note.
- Ollama client now remembers which server (primary/fallback) most
recently succeeded and tries it first on the next call, via a shared
Arc<AtomicBool>. Avoids re-404ing the primary on every agent iteration
when the chosen model only lives on the fallback.
- Demote noisy logs: daily_summary "Summary match" lines to debug;
inner chat_with_tools non-2xx body log from error to warn (outer
layer owns the terminal-error signal).
- Drift-guard tests for summarize_tool_result covering the success /
empty / error / unknown shape for every tool.
- Tidy: three pre-existing clippy warnings cleaned up.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add LlmClient::chat_with_tools_stream and SSE endpoint
POST /insights/chat/stream that emits text deltas, tool_call /
tool_result pairs, truncated notice, and a terminal done frame as the
agentic loop runs.
- Ollama: parses NDJSON from /api/chat stream, accumulates content
deltas, emits Done with tool_calls from the final chunk.
- OpenRouter: parses OpenAI-compatible SSE, reassembles tool_call
argument deltas by index, asks for stream_options.include_usage.
- InsightChatService spawns the loop on a tokio task, feeds events
through an mpsc channel, persists training_messages at the end.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
OpenAI-compatible client for OpenRouter. Translates canonical wire shapes at
the boundary: tool-call arguments stringify on send / parse on receive
(accepting both string and native-object forms); images rewritten from the
base64 images field into content-parts with image_url entries; role=tool
messages inherit tool_call_id from the preceding assistant's tool calls.
/models parsed into ModelCapabilities via supported_parameters (tool use)
and architecture.input_modalities (vision). 15-minute capabilities cache.
Bearer auth; HTTP-Referer / X-Title attribution headers optional.
Not wired into request routing yet — first consumer arrives with hybrid
backend mode. 11 unit tests cover the translation helpers.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Cameron CordesandClaude Opus 5.5
CameronandClaude Opus 4.7