v1.0.7

[1.0.7] — 2026-09-28

Jev judge layer — calibrated second opinion over any backend

• New --judge jev flag (run/update) layers TypeSafe's Jev — a System One decision model that returns typed judgments with calibrated probabilities, not generated text — on top of the selected --backend (claude, openai-compatible, or gemini). The engine still generates every extraction; the judge re-judges it. --backend jev is not accepted and errors with a pointer to --judge (ASTRIA_LLM_JUDGE=jev selects it via env).
• Trivial-file gate (on by default, bounded batch sizes) — before a file's first extraction, batched keep/drop judgments (≈1 billed request per 50 files, files >64 KB presumed rich) skip files the judge finds empty or trivial, so they never cost an engine call. Gated files keep their structural extraction; the run summary reports them ("N files gated by Jev").
• Per-file verification — one request per file re-chooses node types and relations from the schema allowlists (replacing the lossy relates_to/concept clamps) and gets a keep/drop existence verdict per edge. Spurious edges are dropped (ASTRIA_LLM_JEV_MIN_EDGE_PROBABILITY, default 0.40); kept edges carry the judge's keep probability as a calibrated confidence_score in the edges table — semantic edges previously left it null.
• Suggested-question ranking — on runs that rebuilt the graph, the report's suggested questions are re-ordered by judge keep-scores so the most useful one leads. Best-effort: any judge failure keeps the generated order.
• Judge calls count toward ASTRIA_LLM_BUDGET like every other response, and the judge configuration (model, thresholds, gate settings, prompt text) fingerprints into the semantic extraction cache — changing it invalidates cached extractions. --judge without --backend errors: the judge wraps an engine, it cannot generate extractions.
• Configuration: ASTRIA_LLM_JUDGE_API_KEY (or TYPESAFE_API_KEY) and ASTRIA_LLM_JUDGE_MODEL (default jev-latest) are vendor-generic; behavior knobs keep the honest ASTRIA_LLM_JEV_* names (_VERIFY, _MIN_EDGE_PROBABILITY, _GATE, _GATE_MAX_BYTES, _GATE_DROP_THRESHOLD, _GATE_BATCH).

HTML visualization rewritten around drill-down

• astria export --format html now ships a self-contained canvas viewer (no vis-network, no network access required) that opens as community bubbles — one per community, sized by membership, with edge-weighted links between bubbles. Click a bubble to expand it into member nodes, click a member to focus its 1-hop neighborhood, and search to jump straight to any symbol; "All nodes" expands everything with level-of-detail labels.
• The exported layout stays fully precomputed (physics-free), and the viewer draws only what is on screen, so large graphs open and zoom instantly even in sandboxed HTML previewers.
• Viewer source lives in packages/viewer (TypeScript, npm run build); the minified bundle is embedded at crates/astria-napi/src/assets/viewer.js. Community bubbles use themed labels from the communities table when --label-communities produced them.
• Relation-aware focus — the exported edge payload now carries the edge kind (calls, imports, …): the focus panel lists a selected node's neighbors with their relation, and the highlighted 1-hop edges gain direction arrowheads (direction shown where it matters, not on the hairball).
• Community search — search matches community names as well as symbols and files; picking a community expands and centers its bubble.
• Accessibility floor — the canvas exposes a role="img" label with node/community/edge counts plus a visually hidden summary of the controls, so screen readers get a usable description of the export.
• Quiet, throttled git hooks — astria hook install now writes v4 hooks that invoke update . --quiet --if-stale 10: hook-driven rebuilds print nothing (no progress lines, no token benchmark), and skip entirely when the graph was published less than 10 minutes ago, so a burst of commits rebuilds once instead of per commit. astria update gained matching --quiet / --if-stale <minutes> flags; hooks retry plain update . against any CLI version that predates the flags, and still never break a commit.

Skills and MCP updated for the new features

• The shipped skills (packages/astria-cli/skills/skill*.md, full + per-assistant variants) now teach agents the new capabilities: the interactive bubble-viewer export (export --format html, --mode standard|large, the tree view, and --neo4j-push/--redis-push), the full update flag set (--no-dedup, --embed, --label-communities, --deep, --quiet, --if-stale), and the git hooks (astria hook install|uninstall|status, hook-guard) with their automatic post-commit refresh. Existing installs refresh by re-running astria install.
• The MCP server's client instructions now point agents at the hooks (astria hook install) for automatic post-edit freshness, and the skill's MCP tool list is corrected to include health.

Retrieval ranking tightened for prose corpora

• Chunk labels are a truncated first line of the chunk's own body; scoring no longer amplifies that prefix at label weight for chunk nodes, so a later session whose opening line re-mentions a topic cannot outrank the chunk whose body actually answers the question.
• The IDF pre-pass now counts document bodies as well as labels, so terms that are common in bodies but rare in first lines ("group", "friends" in transcripts) stop acting as near-max discriminators, and rare proper nouns carry the ranking.
• Measured on the full LoCoMo set (1,977 questions, structural, no embeddings): recall@1 63.5% → 66.1%, recall@3 79.3% → 80.7%, MRR 0.717 → 0.736, with recall@5/10 at 85.0%. The 35-question code self-check (quality harness) holds recall@5 at 82.9% with MRR 0.636 → 0.659 (a different series from the paired-runner self numbers below — different harness, pinned graphs).

Qualified-name retrieval and the first blind answer-correctness run

• Question terms now score against each node's scope-qualified id (BaseCommand.get_usage reaches src_click_core_basecommand::get_usage through the id even though every same-name symbol shares one bare label), and id tokens join the IDF pre-pass so ubiquitous scope words ("src", "core") cannot act as rare discriminators.
• The seed reservation honors qualified names too: an explicitly named qualified symbol reserves its node a traversal seed instead of losing the slot to a label-tie stranger. Click's additional validation went from 0% to 2/2 exact definitions surfaced, additional ripgrep from 25% to 3/4, and the paired-runner self set's MRR from 0.618 to 0.687 at unchanged file recall; LoCoMo is unchanged.
• Blind answer-correctness judging finally ran (TypeSafe System One judge, scripts/bench/quality/blind-judge.mjs): both tools answered the same 35 rubric-grounded questions, graded without tool identity — astria 100% PASS, Graphify 77.1% PASS / 2.9% PARTIAL / 20% FAIL. First generated-answer-correctness measurement in the project (single judge, single run, 35 self-corpus questions — not a statistical claim). The same pairs re-graded by the independent promptfoo/OpenRouter judge (gpt-4o-mini) agreed on the ordering at 77.1% vs 65.7% pass.
• The paired runner's budgets are configurable (budgets array). A four-point budget-response curve (250/500/1000/2000) shows astria's 250-token answers outscoring Graphify's 2,000-token answers (MRR 0.680 vs 0.531, recall@5 74% vs 69%) while Graphify exceeds each of the two smallest budgets on 48/50 raw responses and astria stays inside budget on all 200 (structural only, one observation per condition).
• Two reserved golden tracks exist, authored from pinned source and unused during development: click.doc-intent-v1.jsonl (8 doc-intent cases) and click.reserved-v1.jsonl (12 cases, doc- and code-intent, line-exact definitions). First use must be an evaluation run; afterwards they count as exercised.
• Held-out evidence grew: scripts/bench/paired/*.heldout-v2.jsonl adds 14 separately authored, line-exact grounded cases (5 Click, 5 Express, 4 ripgrep); current runtime retrieves 5/5, 5/5, 2/4 files and 11/13 v2 definitions in the top five.

Fixed

• Tree export hardening — the symbol-tree hover inspector builds its panel with DOM textContent instead of innerHTML, and the embedded JSON escapes </script/<!-- breakout sequences: the tree viewer now upholds the bubble viewer's labels-as-text safety property, with matching tests.
• Docs drift — reference pages corrected against the code: query-log env semantics (ASTRIA_QUERY_LOG is a literal path; ASTRIA_QUERY_LOG_ENABLE selects the default), the --json 20-neighbor cap, --detail high as an EXTRACTED/DECLARED class filter, --label-communities/--deep/update --embed flags, missing env-var rows (ASTRIA_LLM_BUDGET, ASTRIA_LLM_COMMUNITY_MAX, NEO4J_*), the real scip_* relations replacing the never-emitted method/inherits/forks, schema table columns, memory ingestion timing, and same_type_as label-based grouping. The docs-sync guard no longer counts #[cfg(test)] fixtures as relation emitters (a clamp-test relation: "forks" had been satisfying the check for a documented relation that does not exist).
• Docs drift guard, both directions — the docs-sync check now also fails when ARCHITECTURE.md documents a relation that no production code emits or references (with an explicit (external only) escape), when an ASTRIA_* variable is read but undocumented or documented but never read, when a registered CLI command/flag or MCP tool is missing from its reference page, and when the new generated SQLite schema block is stale. The schema block is generated from db.rs into the architecture page by scripts/generate-schema-docs.mjs (column lists can no longer drift — the first generated block surfaced five previously undocumented columns); ARCHITECTURE.md's schema section now links there instead of restating columns, and the website's relation list defers to the graph-model reference. The distributed skill (skills/astria/SKILL.md) gained an explicit scope note pointing at the full surface.

What's Changed

• ci: Node 24 action upgrades + runner label pinning by @nodesify-technology in #89

Full Changelog: v1.0.6...v1.0.7