Commit Graph
8 Commits
Author SHA1 Message Date
amosblomqvist fb66606555 Wave 6 chunk B: the embeddings surfaces — picker, re-embed gate, gray-outs, cost pill
- Settings ▸ Graph grows the embedding-model picker (graph-row-embedding-model):
  GET /api/models/embeddings' live-federated BYOK catalog (re-fetched on mount
  and every settings save), rows priced from EMBEDDING_PRICES ("$x.xx/1M in",
  honestly "—" when unlisted), ids shown AND stored as full colon-safe slugs
  (openrouter:liquid/…:free is never split), plus an Unset row writing null.
  Empty catalog = nothing keyed: the trigger BECOMES the deep link to Keys
  ("Add a provider key first…") — and deep-links fired while the dialog is
  open now actually switch sections (initialSection was read once at mount;
  the body follows it now — pre-existing gap, first exercised by this link).
- Re-embed gate (owner design): picking a model that differs from the stored
  one holds the write behind a confirm strip carrying the honest node count
  from GET /api/graph; zero propositions skip the strip (nothing to re-embed)
  and land directly. Unset never confirms — the port just stops resolving.
  The copy says the re-embed runs with the next background pass (the staleness
  machinery — a settings PUT alone never triggers it; no server endpoints
  were added).
- Gray-out, the searchStatus precedent: with no model configured, the ladder
  rungs Embedding/Clustering/Labeling dim + disable with a "Needs an
  embedding model" hint (master switch stays lit), and the stats-bar
  GraphContextControl swaps its slider for a dimmed deep-link button to the
  Graph section. Loading never grays; every settings save re-reads (signal).
- Cost chrome: the pipeline popover grows the Embeddings stage row
  (costs.stages.embedding, wave 6's folded stage — "— · 0 in / 0 out" until
  runs exist, same as its siblings).
- Master's ride-along: the core-barrel ChatDeleteMode comment now says two
  type-only things (it always exported both).

Gates: tsc clean, 912/912 (8 new: price lines, model-dependency predicate,
unset/set SSR dimming; 2 ladder fixtures updated to the new no-model
semantics), lint 0, build ok. Live-verified on :3001: unset deep-link →
graph section, menu pricing + colon-safety (37 options), confirm strip with
"20 propositions", Cancel keeps the record, Unset writes immediately,
in-dialog empty-catalog deep-link lands Keys, Embeddings row in the popover.
2026-09-23 12:50:59 +03:00
amosblomqvist 22266973f3 Fix wave: dropdown keyboard/focus, appearance resets, inline key editors, content-sized slider
- MenuShell: after placement, focus lands in the search when the menu has
  one, else the body (tabIndex -1) — the lists' mount-time focus ran during
  the hidden measuring pass and silently failed, so no dropdown ever opened
  focused. The body owns the arrow machine for searchless menus:
  ArrowDown/Up cycle options with wrap (first/last entry from the body,
  native-select convention), Home/End, Enter activates, preventDefault so
  arrows never scroll the page.
- Appearance: the CSS-var applier only ever setProperty'd — a later write
  omitting a var left the stale inline value forever (empty background
  never reset, serif toggle-back looked dead). appliedNamesRef tracks what
  was set; staleVarNames (pure, exported) returns what to removeProperty.
- CommitInput: controlled — after a commit the field snaps to the stored
  record, so the clamp on a -5 prose size is visible (12), Esc reverts.
- Keys: Add-provider dropdown closes on pick; the paste row is inline
  (label | input | Save | Cancel); existing-row editors go inline with an
  explicit Cancel beside Save/Delete — folding is visible, not Esc-only.
- Graph context: readout is content-sized (tabular-nums) — no fixed-width
  dead gap when "off"; the dictation strip fills whatever space remains.

Gates: tsc clean, 866/866 tests (new staleVarNames suite), lint 0, build ok.
Live-verified via Playwright: dropdown focus+arrow cycles+wrap+Enter pick,
background set→empty reset, serif round-trip, visible clamp to 12, provider
pick closes menu, inline editor + Cancel + Esc fold, slider width 270→277px.
2026-09-23 00:57:28 +03:00
amosblomqvist cbecb8df03 ui: rename learner-facing 'knowledge context' copy to 'graph context' (components + render tests) 2026-09-18 18:03:25 +03:00
amosblomqvist 4022e3f0d0 fix: the knowledge-context slider is a discrete 6-stop scale — off, then powers of four from 64
The continuous 0-4096 dial was precision theater (a 128-token budget is
~2 propositions; nobody wants that granularity). Decided stops (owner,
2026-09-15): 0 / 64 / 256 / 1024 / 4096 / 16384 — the native range walks
positions 0-5 and maps through GRAPH_CONTEXT_TOKENS_STOPS in contracts.
Max rises to 16384 (route range follows); the server still accepts any
integer in range so a stale client cannot brick settings; non-stop
legacy values snap to the nearest stop for the thumb, readout stays
honest.
2026-09-15 08:09:56 +03:00
amosblomqvist f6b0c3fb86 fix: the knowledge-context readout shows the full number (1024 tok, no k-rounding) 2026-09-15 08:02:07 +03:00
amosblomqvist b1416089c6 fix: the knowledge-context readout aligns left (it read right-aligned off the slider) 2026-09-15 08:01:33 +03:00
amosblomqvist 4c7ad48c73 fix: the knowledge-context slider lives in the stats bar's left slot — the session-data spot — with a "tok" readout
The original vision, restored (owner, 2026-09-15): pre-chat the session-
data side of the under-composer stats bar is free real estate ("—" only),
so the budget slider mounts exactly where session data will sit once a
session exists — via a new StatsBar leftSlot — vertically centered with
the model + thinking pills by the row's own geometry. Short 96px slider
instead of a full-width one; readout abbreviates "1k tok" so >1k budgets
stay on one line ("4.1k tokens" wrapped). Home and /chat/new both.
2026-09-15 08:00:04 +03:00
amosblomqvist 1d45d77303 feat: graph context injection — the session's prior knowledge, steered in at session start (vision-context-injection.md)
One shot per chat, fired at the first turn's onFinish: the session's
topic (its only typed message) is embedded as an EmbeddingGemma retrieval
query; the semantically closest graph propositions are selected under the
session's token budget (pure budget, no similarity threshold — the
decided doctrine) and delivered as a marker-only steer through the wake
pipeline, queuing behind pending questions like every sub-agent result.

- contracts: ChatSettings.graphContextTokens (default 1024, max 4096,
  step 128, 0 = off); GraphContextData + graphContextText; the
  graph-context JobKind + buildGraphContextSteer (id steer_ctx_<chatId>)
- core/graph-context.ts: brute-force cosine top-K, greedy chars/4 fill
  under budget, most-recent UnderstandingLevel fold, count cap 64,
  parse-don't-cast at the marker seam
- server: embedQuery (role-asymmetric query format, same warm singleton
  — vector-space parity by construction); the runner (startGraphContext-
  Injection one-shot gates; runGraphContextJob validates ledger parity);
  wake branches (chatIdOf/jobIdOf, finalizeJob done, recoverStaleJobs
  re-deliver or contextRespawn); abortJob kills context jobs SILENTLY
  (nothing was promised in the lesson); turn-service trigger seam
- adapter: data-graph-context flattens to the canonical block (malformed
  parts have no model face)
- UI: the budget slider (home dock + /chat/new draft, the session-data
  strip); the quiet SystemLine "~537 · 10 propositions"; dock kind
  "Knowledge context"; model-picker PUT now spreads settings (a model
  pick can no longer clear the sibling override)

Live-smoked end to end on the real embedder + real GLM: seeded ledger
(10 propositions), session on "modal logic: possible worlds and the
necessity box" — job in the dock, queued behind the prober's question,
delivered on the answer, SystemLine in the transcript, wake turn reads
the block and continues the probe. Gates: tsc, lint, 459/459, build.
2026-09-15 07:54:50 +03:00