FantomMcpServer 1.1.0

Retrieval quality

• Contextual retrieval. Every code chunk is embedded with a one-line
  LLM-written context (EMBED_TEXT_VERSION 4) that says where it sits and
  what it is for, so a search for a concept finds the function that
  implements it, not only the one that names it. Contexts are cached per
  project and reused across re-embeds; 20 nodes per request, one retry.
• Model switch to qwen/qwen3-embedding-8b (4096d) for code, docs and
  Axon vectors, pinned to one upstream provider. The 4B model's only host
  answered "model busy" under load; the 8B lane fans out across the
  configured sidecars and is checked for vector compatibility (cosine ≥ 0.99
  against a reference) before it may serve.
• Version groups. A project belongs to a product/version line
  (haxall 4.0.6, skyspark 3.1, a range such as skyspark 3.1.1-3.1.12);
  searchFantomCode, semanticCodeSearch, askCodebase and the dashboard
  take a versionGroup filter. The sidebar and the 3D graph show the tree.

askCodebase (RLM loop)

• Jev routing gates (TypeSafe, advisory, one typed judgment each,
  ~1 s, fractions of a cent). Gate 1 decides whether a question needs the
  investigation loop at all, given the retrieval hits and the identifiers
  from the question that appear in them. Gate 2 scores the research plan's
  coverage and appends a search for each identifier the plan missed —
  search_symbols for a symbol, search_files for an indexed file name,
  search_code for a file the code generates (funcs.xeto, doc.md).
  Gate 3 judges whether the gathered evidence answers the question before
  the model drafts, and nudges one more search when a named symbol is
  absent. Every gate is null-safe: no key, a timeout or an uncertain answer
  leaves the loop as it was. Enabled from the LLM Providers page
  (llmProviders.typesafe, key TYPESAFE_API_KEY).
• The plan parses. Thinking models spent the whole max_tokens on
  chain-of-thought and returned an empty plan on every deep ask; reasoning
  is now off for the plan and every loop round, in each endpoint's dialect.
• Provider-direct RLM models. The RLM dropdown on the OpenRouter page
  also offers Gemini 3.8 / 3.7 / 3.5 Flash and Claude Sonnet 5.5 / Opus 5.5 /
  Haiku 4.5 on the provider's own API with the key from the LLM Providers
  page — no sidecar, no OpenRouter — shown only while that key is present.
  The loop and the final write-up both run there. Per-provider request
  shaping (Gemini's thought_signature on replayed tool calls and
  reasoning_effort; Anthropic's rejected temperature) is handled.
  Measured on the same question: 78 s on DeepSeek V4 Flash via OpenRouter →
  48 s on Gemini 3.8 Flash direct, longer answer, same citations; about
  1.4 ¢ per deep ask.
• Loop cost cut. No cross-encoder rerank and no per-hit graph metrics
  inside the loop (the model never read them; a search fell from 18–34 s to
  1–9 s); tool calls are bounded by the deadline instead of overrunning it;
  the forced-final draft round is skipped once evidence is gathered
  (synthesis writes the answer; the draft was empty every time).
• Synthesis tier and routing: LLM work goes only to sidecars that run the
  rlm-sandbox role.

Embedding pipeline

• Live re-embed into the active table (no shadow) that resumes: a
  restart skips every project already complete and clears a half-written
  one. Five projects run at once; the dashboard pins them to the top of the
  table and names all of them.
• Scheduler: an upstream 429 re-queues a chunk without counting against the
  lane; a lane parked by one bad probe is re-probed; a gate refusal counts
  as a lane failure and falls back to the next lane.
• build-missing counts against the graph's node count (Prisma's ran ~30 %
  high, so a six-project fill walked all 318).
• The docs table is read-probed on open — a dropped table left a stale
  catalogue entry and a re-embed wrote 0 of 45,123 items.

Stability

• LadybugDB SIGSEGV fixed. Two callers missing the connection cache at
  once opened two native databases on one file; the second replaced the
  first, which was freed under a running auto-checkpoint
  (BufferManager::pin). Opens are now single-flight per project, the
  engine's auto-checkpoint is off (we checkpoint on close and at shutdown),
  and a connection is never closed under a live query.
• Every project graph store is pre-flighted at boot; a corrupt file is
  quarantined (never deleted) and the project rebuilt.
• Sidecar registry: a register frame matches by hostname so a re-pointed
  entry is reused; an HTTP-fallback heartbeat never creates an entry.
• scripts/start-dev-launchd.sh starts the dev server as a launchd job so
  it holds the macOS Local Network permission a terminal-spawned process
  inherits as denied.

Dashboard

• Vector Viewer remembers every filter per browser and restores it on
  refresh (?projectId= / ?vg= in the URL still win); "Reset filters".
• Project table: default sort by id (queue order), in-flight projects pinned
  on top with an "embedding" chip; "Last run" column; "Embed missing"
  refuses while a job is running.
• LLM Providers: Claude Opus 5.5 / Sonnet 5.5 / Haiku 4.5; Gemini 3.6 Flash
  and 3.5 Flash Lite. An env-only API key is persisted to .env when its
  provider is saved.