1.6.0

This release is about the cook backlog: which of your never-compiled notes
an agent should read first, and which should never have been on the list.
Everything here is additive under the 1.0 workspace contract — one new
optional config file (.knowlery/config.json), one new optional store
(.knowlery/triage.json), three new CLI subcommands, and optional keys on
health --json / stale --json. A vault that uses none of it behaves as
before.

Jev ranks the uncooked backlog (#79, #81)

• Optional Jev cook triage. Jev is TypeSafe AI's
  decision model. With a TypeSafe API key, Knowlery asks it two questions per
  uncooked note — is this worth compiling? and which existing compiled
  page is about the same subject? — and turns the answers into a reading
  order for /cook plus a suggested update target per note. The experiments
  in #77 showed Jev is a usable ranker (its confident-yes band was clean
  on real notes) but not a filter, so nothing is ever hidden: scored notes
  above 0.2 come first by score, unscored notes next (newest first), the
  ≤ 0.2 band last — still listed, still read. Without a key nothing changes.
• Configure the key once. The key goes in the plugin's settings (new
  group Jev cook triage (optional): what it unlocks, the privacy boundary,
  a Get a key button, a password field with Clear). It is stored in the
  plugin's data.json under the vault config folder — exclude that file
  from sync/git if you share the vault. The Node shells read the same file:
  knowlery stale, the MCP stale tool and knowlery triage resolve the
  key as TYPESAFE_API_KEY → .obsidian/plugins/knowlery/data.json, so an
  agent in Codex, Claude Code, Cursor or any MCP client gets the ranking
  with no second setup step. Vaults with a non-default config folder use the
  environment variable.
• Scores persist in .knowlery/triage.json (schemaVersion 1): path →
  worth, target, mtimeMs, scoredAt, plus the model id. A note is
  scored once until it changes; entries for notes that are no longer
  uncooked are pruned on write. The file holds paths and numbers only —
  never the key, never note text. All four staleness transports (MCP
  stale, knowlery stale, obsidian knowlery:stale, and the offline
  node .knowlery/bin/query.mjs --stale) read it and print the same order;
  uncookedNotes[].jev = { worth, target } is an additive optional field in
  stale --json and the MCP result.
• Three ways to produce scores.
  • Dashboard → Knowledge health: Rank N with Jev scores the whole
    backlog with overall progress, then shows the scored list (green ≥ 0.8,
    faint ≤ 0.2) with → update <page> / → new page, skipped/failed
    counts, a cost estimate, Copy cook prompt and Re-rank. Without a
    key the screen shows a one-line hint with an Open settings link above
    the unchanged list.
  • knowlery triage [--json] (new) scores the whole backlog from a
    terminal in passes of 40, printing progress, the confidence bands
    (> 0.8 / 0.2–0.8 / ≤ 0.2), token count and cost. Idempotent and
    resumable — the store is the cache. Refuses without a key and names
    both places to put one.
  • knowlery stale / MCP stale top up at most 40 unscored notes (newest
    first) before rendering, so an agent's critical path never waits on a
    long run; failures degrade to the mechanical report with diagnostics on
    stderr, keeping --json clean.
  • Optional Score new notes automatically toggle (default off):
    about 30 s after notes settle, the plugin scores up to 40 unscored notes
    in the background and writes the store — for people whose workbench is
    an agent rather than Obsidian.
• Cost and privacy. About $0.0001 per note; a 381-note backlog cost
  ≈ $0.03 in ~42 s, a 4,000-note vault is on the order of $0.35. Jev
  receives each note's path, title, frontmatter and first 15 non-empty
  lines, plus the titles and descriptions of up to five shortlisted compiled
  pages — never full bodies. Retrieval (runQuery, query.mjs) never calls
  Jev. The model is pinned (jev-1.13.0).
• Cook skill. Incremental Mode gains a Jev-ranked uncooked notes
  paragraph: the jev field is a reading order, not a filter; open the
  suggested target before writing; still glance at the ≤ 0.2 band; never
  ask the user for the key or call Jev yourself; when most of a long list is
  unscored, suggest knowlery triage rather than run it (it spends the
  user's budget). The knowlery-cli skill documents knowlery triage.
• Cross-shell cache key. Node reports fractional mtimes on APFS and
  Obsidian reports the same instant rounded to whole milliseconds; the store
  normalises to Obsidian's rounding and compares with a sub-millisecond
  tolerance, so the plugin and the CLI/MCP recognise each other's scores
  instead of re-scoring (and paying for) the same notes. The fs scanner now
  reports whole-millisecond mtimeMs for the same reason (value precision
  only).

What counts as uncooked (#78, #80, #82)

• Reference directories: queryable but never uncooked. Imported
  knowledge — a team KB pulled into the vault, say — must stay searchable
  but is not your material to cook; on one work vault it put ~2,850 notes
  on the uncooked list and buried the real work. referenceDirs in the new
  .knowlery/config.json gives named directories the treatment Library/
  bundles already had: scanned at user tier, excluded from the uncooked
  report, in both the CLI scanner and the Obsidian live snapshot. Managed
  with knowlery reference list [--json] | add <dir> | remove <dir> or
  the new Reference directories section in Dashboard → Config (folder
  picker, per-directory note count, remove). Absolute paths, .., hidden
  directories and the reserved top-level directories are rejected.
• Scan exclusions that are safe for vaults that are not repos. On one
  vault 871 of 4,108 "uncooked notes" were READMEs under node_modules/.
  The rules, shared by every transport, now layer as: hidden directories
  (silent, as before) → node_modules/, __pycache__/ always → dist/ build/ target/ outputs/ coverage/ vendor/ only beside a project
  manifest (package.json, Cargo.toml, pyproject.toml, go.mod,
  pom.xml, Makefile, …) and never at the vault root — a notes folder
  called build/ stays indexed, and so does a top-level outputs/ of
  deliverables in a vault that happens to carry a root package.json →
  ignoreDirs, whatever you name, in the same config file (a directory
  may be in at most one list). Managed with knowlery ignore list [--json] | add <dir> | remove <dir> or the Ignored directories
  section in Config. The live snapshot re-judges siblings when a manifest
  appears or disappears and prunes/re-reads pages the moment a list changes.
• Every skip is visible. Because an excluded folder silently vanishes
  from retrieval, every non-hidden skip is reported with its reason:
  knowlery health prints a Skipped directories (not indexed) section
  (--json gains an additive skipped key) and warns about configured
  reference/ignored directories that no longer exist and about a malformed
  .knowlery/config.json; Dashboard → Diagnostics gets an Indexing
  section with the same list and a persistent warning row while the config
  cannot be read; the Config tab shows that warning inline above both
  lists. A malformed config degrades to "no reference or ignored
  directories" identically in both transports, and says why.
• Generated agent files are not notes. The vault-root AGENTS.md that
  Knowlery regenerates from KNOWLEDGE.md and your rules joined the system
  files (KNOWLEDGE.md, SCHEMA.md, log.md): it is never a page, an
  uncooked note, or a query candidate — it used to rank first in
  fresh-vault queries and get Jev-scored as uncooked. .claude/CLAUDE.md
  and .agents/ live under hidden directories and were never scanned.
• /cook --all now states that it excludes Library/, reference
  directories, ignored directories and the skipped directories; the
  knowlery-cli skill documents knowlery reference and knowlery ignore.

Fixes

• Plugin reload refreshes query.mjs at the same version. Swapping a
  beta main.js in place used to leave the generated
  .knowlery/bin/query.mjs stale until the version number changed; it is
  now re-synced by content diff on every load, with the same newer-shell
  downgrade guard as the full version sync.
• Knowledge health: the home card no longer disappears while the live
  snapshot is warming after a reload or while .knowlery/config.json is
  malformed; the drill-in re-renders after a Rank run completes.
• knowlery triage no longer prints "Scores saved" when every note already
  had a current score and nothing was sent.
• tests/cli/ignore-commands.test.ts no longer collides on
  case-insensitive filesystems.

Contract

• POSITIONAL_LIMITS: reference: 2, ignore: 2, triage: 0 (added
  1.6.0). health --json gains optional skipped; stale --json and the
  MCP stale result gain optional uncookedNotes[].jev. No MCP write tool
  for reference or ignored directories — agents observe the effect through
  stale and health. Eval drift guard: the cook-skill edits in #78, #80
  and #81 were prose-only for the existing fixtures and waived explicitly in
  each PR; the agreed follow-up is to label a sample of the ≤ 0.2 band from
  a real vault before hardening the skill's stance.