Skip to content

Changelog

Detailed notes for every release from v0.1.0-alpha.2 on are in docs/release_notes/ and on the documentation site; the entries below summarise them.

v0.3.4

Fixes from the v0.3.3 retest. Full notes: docs/release_notes/v0.3.4.md.

  • Local API: path-exists answers only inside your configured folders; the editor and zsum command settings accept only known programs when set from the app; saved API keys are never echoed back; prompt variant filenames are checked; hidden vault files return 404.
  • Fixed: Ctrl-K on Home, a card for every mode on Home, a small corner logo by default, a clearer zsum Collections empty state, real backoff on provider rate limits, AKMS_GLOBAL_VAULT in first-run setup.
  • Releases: SHA256SUMS.txt no longer lists itself.

v0.3.3

The rest of the v0.3.1 QA pass. Full notes: docs/release_notes/v0.3.3.md.

  • No dead ends: mode setup messages keep the mode switcher and side menu, with a Copy snippet action for Failure Memory and Records; related papers, batch run history, View in Graph and zsum Collections lead somewhere.
  • The Picker lists every batch (app-created and starter batches were missing) and picks up changes without a restart.
  • Fixed: vault preview math and nested frontmatter, prompt Preview (always empty), graph review message, starter zsum config, hidden vault entries.
  • Clearer messages: Discover provider failures, provider test advice, Zotero not connected, Open in Editor result, MechDSL status.
  • API: Discover input validation and provider_errors, picker routes in OpenAPI, graph refresh stream is POST, UTC graph status, zsum test errors, learn export stub removed.
  • Polish: app title and icon, Linux package description, Logo switch for the backdrop logo, dark-mode contrast, Ctrl-K fixes.

v0.3.2

Security and reliability fixes from the v0.3.1 QA pass. Full notes: docs/release_notes/v0.3.2.md.

  • Local API requires a per-launch token and a loopback Host, closing access by other local programs and by web pages through DNS rebinding.
  • Hardened inputs: prompt and workflow ids can't escape the library, node file read/open is limited to the node and vault folders, saved API keys are masked, command settings must name a real program, and error responses carry no tracebacks.
  • No orphaned sidecars: the sidecar exits when the app dies for any reason; Windows kills the whole process tree; a second launch focuses the first window.
  • Windows sidecar serves again: in v0.3.1 it crashed right after starting (uvicorn rebuilt the listening socket as a Unix socket).
  • Docs: install links point at the installer releases; checksum steps per OS.

v0.3.1

Linux packages that run on current distributions. Full notes: docs/release_notes/v0.3.1.md.

  • Linux builds need glibc 2.35, not 2.43. v0.3.0's AppImage and .deb were built on the newest runner and did not start on Ubuntu 24.04 or Debian 13. They are now built on Ubuntu 22.04, and the release build fails if a shipped binary needs a newer glibc or the AppImage lacks its bundled WebKitGTK.
  • Docs: Linux requirements, AppImage vs .deb, and running from source in a browser without WebKitGTK.

v0.3.0

The first public-source release. Full notes: docs/release_notes/v0.3.0.md. It also carries everything prepared as v0.2.1, which was never tagged.

  • Records mode: read-only browse of explicitly configured akms-records stores, their schemas and their record graph.
  • No source checkouts. AKMS, MechDSL and zotero-summarizer are installed from PyPI; the picker, nodes, batches and graph work without an AKMS checkout.

  • The packaged app supports all of AKMS 0.4. The sidecar build installs the dependency set uv.lock records and bundles every optional mode: the AKMS tabs and Learn, the MechDSL emit bridge, Failure Memory and the Records browser. Earlier builds pinned AKMS below 0.4 and left Failure Memory and Records out of the frozen binary.

  • No vendored source trees. The bundled agent assets and the zsum prompts are read from the installed akms and zotero-summarizer packages, and the old vendored thirdparty/ snapshots are deleted from the repository. LOGIC_LOOM_SIDECAR_CORE_ONLY=1 builds a zsum-only sidecar.
  • Every version field agrees (scripts/check_versions.py; the API package and tauri.conf.json had drifted), and the release workflow refuses a tag that does not match them or has no changelog section.
  • Security upgrades in the lockfile. The new uv audit gate reported 30 advisories in nine transitive packages; all are cleared by upgrading anyio, cryptography, filelock (with litellm, which capped it), fsspec, multidict, pydantic-settings, pypdf, starlette and urllib3. No AKMS, MechDSL or zotero-summarizer pin moved.
  • Documentation is a website at https://cemm2.github.io/Logic-Loom/, built from docs/ and published on each release tag. It replaces the HTML design guides that used to live in docs/ and the docs pages in the CEmM2/Logic-Loom-releases repository. Release notes live in docs/release_notes/ and are part of the site.
  • Repository hygiene for a public tree: LICENSE (MIT), SECURITY.md, CONTRIBUTING.md, CODE_OF_CONDUCT.md, SUPPORT.md, a pull-request template, an always-on CI workflow, a security workflow (dependency review, lockfile audits, CodeQL), Dependabot, read-only workflow tokens, and scripts/check_public_tree.py, which fails on private working folders, private repository references, real home-directory paths and secrets.
  • Third-party notices. THIRD_PARTY_NOTICES.md, generated by scripts/gen_third_party_notices.py, lists every bundled Python, JavaScript and Rust package with their licenses and license texts, and ships inside the installers with LICENSE. The release refuses a tag while it is out of date with the lockfiles.

v0.2.1

Patch release; no Logic-Loom feature changes. Refreshes the vendored AKMS toolchain (7e07ace → f32d81f) and moves the akms / akms-failure-memory git pins (72494ac → f32d81f), so the bundled agent assets and the failure-memory package are current for the first time since v0.1.0-alpha.7.

Newly bundled agent assets, now sourced from AKMS's published root paths rather than its internal .claude/ tree: skills/akms, skills/deepwiki-nodes, skills/akms-spec-check, hooks/ (vault read-only + frozen-schema guards), and docs/akms/spec/invariants.md. Existing assets refreshed — which picks up an upstream fix to akms-spec-reviewer / akms-spec-check, both of which had been reading a design_docs path that does not exist.

akms-failure-memory 0.1.0 → 0.3.0: preflight no longer enforces artifact identity on every operation, and the provider fingerprint no longer tracks AKMS source bytes.

Known issue unchanged from v0.2.0: Backend E2E (windows-latest) exits 1 after the suite passes, from non-daemon ThreadPoolExecutor workers at interpreter shutdown (lesson L001). No test fails; product behaviour unaffected.

Full notes: docs/release_notes/v0.2.1.md

v0.2.0

Logic-Loom graduates out of alpha. Adds the Failure Memory workspace — a read-only surface for inspecting .failure-memory lesson registries, plus an off-by-default operator mode for recording and compiling them — and fixes AKMS node-extraction batch runs hanging in the packaged app. Known issue: backend E2E on windows-latest exits 1 after all tests pass (tracked in this repo's own failure-memory ledger). Full detail in docs/release_notes/v0.2.0.md.

v0.1.0-alpha.4 — 2026-07-01

Learn — host LLM provider chooser (AKMS ▸ Settings)

Unblocks the LLM-gated Learn modes (adaptive_path, llm_expanded) with a provider chooser in AKMS ▸ Settings, built on akms-learn's public provider registry (no upstream change):

  • Pick the API route (LiteLLM, keyed by anthropic_api_key / openai_api_key / google_api_key) or an agent-CLI route — claude -p, codex exec, or a user-defined custom command ({prompt} arg-substituted, else piped to stdin). nlm grounded generation stays; copilot / antigravity ship as empty user-fillable command slots.
  • CLI agents activate only on an explicit choice (never on mere PATH presence), so the no-provider gate is unchanged. Agent stdout is normalized (a single enclosing markdown code-fence is unwrapped) before it becomes the generated section body.
  • Saved keys / CLIs go live in-process via a Reload backend action — no app restart (POST /api/learn/providers/reload). Choosing a not-yet-ready provider surfaces a warning naming exactly what to fill or install.
  • New read-only discovery endpoint GET /api/learn/providers.

MechDSL — in-app Taichi install (verify backend)

The MechDSL Overview gains an Install taichi 1.7.4 action: the packaged sidecar ships the emit bridge Taichi-free, and the pinned wheel is side-installed at runtime into an ABI-namespaced dir added to the sidecar's sys.path, so verify is unblocked without re-freezing. emit/transpile keep working with no Taichi.

v0.1.0-alpha.3 — Research Assistant (Learn), LLM providers, MechDSL bridge

Third alpha. Adds the Learn workflow to the Research Assistant, activates LLM-backed learning modes through a host-provided provider, and ships the MechDSL executable bridge — LaTeX → FEM-solver-code generation, bundled Taichi-free. Also pulls a large zotero-summarizer upgrade (see below).

AKMS backend re-vendor (2026-06-17 — alpha.3 re-cut)

The vendored AKMS stack was advanced to upstream main (017996ec) and the alpha.3 binaries rebuilt from it. No UI changes — the AKMS generation backend is overhauled under the hood:

  • akms-nodes-gen dropped its headless-browser dependency. It no longer hard-depends on notebooklm-py[browser]/Playwright; node generation is now LLM-provider-pluggable through LiteLLM (an optional [llm] extra) and CLI-grounded (the nlm CLI / an Anthropic MCP-server URL). This removes the heavy browser stack from the sidecar and fixes the long-standing uv sync --extra akms install hang — the full akms extra (Picker / Nodes / Graph) now installs cleanly.
  • akms core gained CLI/agent backends — claude -p, codex exec, and a local OpenAI-compatible backend, selectable via --backend (pulls litellm + openai-agents + claude-agent-sdk + opentelemetry-* into the bundled sidecar in place of the browser stack).
  • akms-learn baked nbformat/jinja2 into its base dependencies (the [notebook]/[html] extras are now no-ops) and auto-selects an LLM provider.
  • Added scripts/install_nlm.sh to provision the NotebookLM (nlm) CLI.
  • MechDSL vendored packages (mechdsl-core + algo2code) bumped to 0.2.0.

Research Assistant — Learn tab (AKMS-Learn)

A new Learn tab compiles AKMS knowledge into learning artifacts via the vendored akms-learn engine. The UI is fully capability-driven — modes and exporters come from the backend (/api/learn/profiles, /exporters) and unavailable/planned modes render disabled rather than failing on submit.

  • Backend /api/learn/* surface (status, profiles, exporters, compile, artifacts) with per-mode request validation and a capability gate.
  • Pass B: mode-specific forms, preview panels (notebook / assessment / provenance-JSON / generated-warnings), artifact history with open-export-folder, and a domain-pack status card (source-pack table + capability badges, payload-driven).
  • "Send to Learn" handoff from the Graph view.
  • Structured generation warnings (generated-content / skipped-prereq / missing-dep) and a Learning Source Review Bundle export.
  • Compile artifacts cache under learning/<id> with a global-vault write guard; LF-canonical artifact text for cross-platform reproducibility.

LLM provider activation

Learning modes that need an LLM (e.g. llm_expanded) now activate when the host supplies a provider — no separate key wiring in the engine.

  • The host-resolved LLM provider is injected into akms-learn LearningRequests at compile time; capability gates and /profiles reflect real host availability so a provider-less host degrades gracefully instead of erroring.

MechDSL executable bridge (Tier 3)

New MechDSL mode (/api/mechdsl/* + Overview / Compile / Artifacts UI, behind MechdslGate): compile % mechanics LaTeX problems into FEM solver code and transpile algorithm boxes, via compmech-reference-pack + mechdsl-core.

  • Bundled Taichi-free. The emit/transpile/capabilities surface needs no Taichi runtime (it lives in mechdsl-core's optional [verify] extra, not installed); only run_verify=True would, and it is gated off. Keeps the sidecar ~170 MB lighter — /api/mechdsl/status reports available: true with taichi_required_for: ["verify"].
  • akms-learn is now bundled in the sidecar (compmech's dependency), so the Learn modes above ship in the packaged app.

Vendored dependencies — MechDSL (new)

mechdsl-core + algo2code are vendored under thirdparty/MechDSL/ as a partial source mirror (refresh with scripts/sync_mechdsl_vendor.sh).

Vendored dependencies — zotero-summarizer

The bundled zotero-summarizer has been updated with a large upstream merge (thirdparty/zotero-summarizer via git subtree pull --squash). All items below are available in the zsum CLI bundled with Logic-Loom; Pipeline UI integration for starred items is in progress.

New: NotebookLM pipeline (zsum nlm) ★

Full integration of NotebookLM as the LLM backend for card and digest generation. Activate with extraction.mode: notebooklm in config.yaml.

  • zsum nlm upload|cards|digests|all|status|cleanup command group
  • Parent collections use direct-member PDFs + child digest .md files as notebook sources (max 30 sources per notebook, configurable)
  • Automatic fallback to page-range extraction for NLM-failed papers, with extraction_method: fallback_pages provenance threaded through cards, cache, publish, and the Obsidian plugin (amber indicator + tooltip)
  • notebooklm.parse_locally option pipes the NLM answer through a local LLM for JSON normalization when NotebookLM's raw response is unparseable
  • Cache manifests: cache/nlm_notebooks.json, cache/nlm_digests.json, cache/nlm_redigests.json

New: Section-aware PDF extraction ★

extraction.mode: sections uses TOC boundaries and fuzzy heading matching (via rapidfuzz) to extract the conceptually relevant portion of each PDF rather than fixed page ranges. Falls back to pages mode when fewer than two target sections are found. Three modes now available:

Mode Behaviour
pages (default) First N + last M pages of each PDF
sections TOC/heading-detected sections with fuzzy matching
notebooklm Delegates entirely to NLM; page-range fallback for failures

New: Enrichment commands ★

Three standalone analysis commands that augment an existing vault without a full pipeline rerun:

  • zsum citations — citation counts from OpenAlex + Semantic Scholar; cached with configurable staleness; surfaces in paper frontmatter and per-collection ## Most Cited Papers section
  • zsum reviews — heuristic review/survey/meta-analysis/tutorial classifier; renders in collection notes and sets paper_type frontmatter
  • zsum restructure — single-LLM-call advisory on collection hierarchy; proposes renames, splits, merges; cached to cache/restructure_report.md

New: LLM providers

  • codex_exec — Codex CLI backend (codex exec) with --output-schema for schema-constrained digest generation
  • gemini_cli — Gemini CLI (gemini -p) backend
  • notebooklm — see NotebookLM pipeline above

New: Setup command ★

  • zsum setup-model — one-command Ollama configuration: checks daemon, pulls a model by size tier (small/medium/large/xlarge defined in an external model_tiers.yaml), writes config.yaml, verifies the model responds. Flags: --size, --model, --list, --dry-run.

Vendored dependencies — AKMS

New: AKMS Learn ★ (specced, API + UI in progress)

akms-learn is vendored alongside akms core. It is a separate learner- facing projection layer over the AKMS graph: given a graph slice and a learning request, it compiles a Learning Source Packet (LSP) — a reproducible YAML artifact with reading order, per-node section extraction, pitfalls, assessment items, code links, and a validation block. Twelve generation modes are defined (deterministic outline through adaptive paths); modes 1 and 12 are implemented in v0.1. Logic-Loom API routes (/api/learn/*) and the Learn tab are in progress.

Pass B — backend (Plan 4 Phase 4). The /api/learn/* surface now describes its own capabilities so the UI can be driven entirely from backend status. All additions are consumer-only (the backend reads akms_learn's catalog and descriptor data; it never reimplements compiler logic) and additive (Pass-A response shapes are unchanged):

  • GET /api/learn/profiles and /exporters now return the full catalogue, each entry {id, label, status} with availability (available / unavailable / planned) derived from capabilities_with_status() and optional-dependency state — never a hard-coded list.
  • POST /api/learn/compile validates the requested generation mode up front: unknown mode → 422 unknown_mode, unavailable mode → 412 capability_unavailable, missing topic/goal → 422 compile_precondition_error. Structured envelopes only — never a 500.
  • Compile responses carry per-family artifact metadata (notebook cell-count/executed, HTML static/size, assessment public/answer-key presence) and structured warning categories (generated-content, skipped-prerequisite, missing-optional-dependency, source-quality).
  • GET /api/learn/status now includes a domain_packs[] array (domain/source pack ids, capabilities, and per-source-pack role + status), sourced from domain-pack descriptor YAML — with no companion-package import and no hard-coded domain concepts.
  • GET /api/learn/status also includes a generation_modes[] array — the filtered, user-selectable modes the compile gate accepts (exporters, execution-backend adapters, and infra capabilities such as learning_source_packet excluded). The mode selector derives from this, so the default selection can no longer POST an infra capability and be rejected 422 unknown_mode.
  • Compile/export remain synchronous (measured sub-100 ms); background-run support is deferred with a documented re-evaluation trigger.

Backend test infrastructure: pytest now uses --import-mode=importlib, and a session fixture materialises the bundled AKMS Claude assets when the vendored source is present, so the full pytest services/logic_loom_api/tests suite runs green on a fresh checkout.

Release pipeline

  • New scripts/release.sh — single-script entrypoint that picks up CI-built binaries from the latest successful Release workflow run (or a pinned --run-id), normalizes filenames into the canonical Logic-Loom-<ver>-<platform>-<arch>.<ext> scheme, stages them under the public release repo, regenerates SHA256SUMS.txt + per-file .sha256.txt, copies release notes into place, commits + pushes the release repo, and creates (or, with --replace, recreates) the GitHub Release with all assets attached. Auto-detects prerelease from -alpha/-beta/-rc suffixes. See README.md → "Releases".
  • Added --download-artifacts yes|no flag to skip re-downloading CI artifacts when iterating on a failed release attempt. The 500 MB download is cached under .tmp/release-<tag>-<runid>/ and reused on the next run.
  • README.md and CHANGELOG.md are now shipped with every release: the script stages versioned copies (Logic-Loom-<ver>-README.md, Logic-Loom-<ver>-CHANGELOG.md) alongside the binaries, includes them in the checksum manifest, and attaches them to the GitHub Release.

CI

  • release.yml no longer relies on bash globstar, which the macos-latest runner's system bash 3.2 doesn't support. The artifact-staging glob now uses brace expansion + nullglob.

0.1.0-alpha.2 — 2026-05-19

Second multi-platform alpha. Full release notes: docs/release_notes/v0.1.0-alpha.2.md.

Three-mode shell

Logic-Loom is now a three-mode desktop app: zsum, akms (gated), and assistant. Each mode has its own sub-routes (/zsum/*, /akms/*, /assistant/*); legacy deep links (/picker, /nodes, /graph, /pipeline, /vault) are kept as redirects.

Assistant mode

Brand-new subsystem for paper discovery + source batching + workflow runs, living under services/logic_loom_api/src/logic_loom_api/assistant/ and apps/desktop/src/routes/assistant/.

  • Source batches — SourceBatch model with nine §8.3 actions (create, edit, duplicate, archive, add/remove/reorder sources, send to AKMS) persisted to <assistant_workspace>/batches.json
  • Paper discovery — Multi-provider clients (OpenAlex, Semantic Scholar) with normalised result merging, on-disk cache observability (<assistant_workspace>/discovery_cache/), and Zotero catalog matching
  • Related papers — Five exploration modes (citations, references, similar, co-citation, co-reference) with ranking presets and a cache lifecycle
  • Workflow runs — Five bundled Assistant workflows + prompt-slot resolution + dispatcher supporting both claude and notebooklm executors; reproducibility artefacts under <assistant_workspace>/assistant/outputs/<batch_id>/<workflow_id>/<run_id>/
  • Outputs — Filtered run-output browser with four-tab detail view + rerun/comparison flows; /api/assistant/outputs is a filtered view over the shared run registry
  • HTTP surface — Five route modules under /api/assistant/*: batches, discover, workflows, outputs, index. All mounted via the degrade-gracefully try/except pattern in main.py
  • Workspace bootstrap — bootstrap_assistant_workspace runs in the FastAPI lifespan alongside the AKMS bootstrap; both wrapped in try/except so a workspace failure never blocks the sidecar coming up

Workspace search (Cmd-K)

A SQLite FTS5-backed research-memory index plus an end-to-end Cmd-K modal.

  • Seven indexable artefact kinds — zotero, abstract, zsum_card, zsum_digest, assistant_output, akms_node, run_metadata (closed set in assistant/research_index/schema.py)
  • HTTP endpoints:
  • POST /api/assistant/index/search — full-text search
  • POST /api/assistant/index/rebuild — walk every artefact and re-ingest
  • GET /api/assistant/index/status — per-kind counts + last_rebuild_at
  • On-disk state — <assistant_workspace>/index/fts.sqlite (records + FTS5 virtual table) plus index.meta.json so last_rebuild_at survives sidecar restarts
  • Frontend — Cmd-K / Ctrl-K modal in AppShell with seven kind-filter chips, ranked result list, deep-link routing to the owning Assistant route, and an empty-state Rebuild index affordance
  • Auto-updates — A shared IndexLifecycle singleton (per Settings instance) is wired into SourceBatchService and RunnerService so the index stays current across the four hot paths:
  • Batch create / edit fires on_change for every zotero source
  • Batch remove_sources fires on_delete for evicted sources + on_change for the remaining state
  • Batch delete fires on_delete for every zotero source on the deleted batch
  • Workflow runs fire on_change for run_metadata at start and for assistant_output on subprocess completion (via a new on_done callback on RunRegistry.attach_workflow_subprocess)
  • 500 ms debounce — Rapid bulk writes collapse into a single FTS5 flush
  • Adapter robustness — The assistant_output + run_metadata rebuild walkers prefer the global RunRegistry (canonical source of truth via log_path.parent) and fall back to the legacy filesystem walk so workspace layout changes don't break the rebuild path

Effective-prompts filename unified

The shared OutputStore previously wrote effective-prompts.json (dash) while the Assistant runner wrote effective_prompts.json (underscore). Both writers now use the underscore form. The HTTP route stays at /effective-prompts for REST convention.

Packaging and release workflow

  • Added PyInstaller-based sidecar entrypoint at packaging/pyinstaller/logic_loom_api_entry.py
  • Added platform-specific sidecar build scripts:
  • scripts/build_sidecar_macos.sh
  • scripts/build_sidecar_linux.sh
  • scripts/build_sidecar_windows.ps1
  • Updated Tauri bundling to include apps/desktop/src-tauri/binaries/* as packaged resources
  • Updated the Rust sidecar launcher to use uv run ... in debug builds and the bundled logic-loom-api executable in packaged builds
  • Added scripts/logic_loom_build.mjs so pnpm logic-loom:build now auto-builds the sidecar for the current operating system before invoking the Tauri release build
  • Added release-build artifact ignores for .venv-sidecar-build/, .pyinstaller-build/, .pyinstaller-dist/, logic-loom-api.spec, and apps/desktop/src-tauri/binaries/
  • Verified locally on macOS:
  • PyInstaller sidecar build succeeded
  • direct sidecar launch + /api/app/health check succeeded
  • packaged app and .dmg build succeeded

Vendoring

AKMS and zotero-summarizer are now vendored under thirdparty/ via git subtree --squash. Path deps in pyproject.toml [tool.uv.sources] point at the vendored copies for both editable dev installs and reproducible packaged builds. The Nodes_Vault subtree lives in the vendored AKMS but is explicitly excluded from the wheel (Phase 12 scripts/check_resources_no_vault.py guard).

Documentation

  • New services/logic_loom_api/src/logic_loom_api/assistant/README.md — one-page module map, service dependency graph, on-disk shape, and known follow-ups for the Assistant subsystem
  • New CLAUDE.md covers the three-mode shell, the full mounted-routes list (including the try/except-guarded routers), and the Assistant subsystem gotcha

0.2.0 — 2026-05-10

Graph visualization

Logic-Loom v0.2 adds a read-only AKMS graph explorer backed by the AKMS core graph compiler and a Logic-Loom-owned cache.

Backend:

  • New /api/graph/* namespace:
  • GET /api/graph/status
  • POST /api/graph/refresh
  • GET /api/graph/nodes
  • GET /api/graph/edges
  • GET /api/graph/ego
  • GET /api/graph/cluster
  • New graph adapter compiles via akms.graph.build_graph.build_graph() and loads cached graphs via load_graph()
  • Graph cache is owned by Logic-Loom (cache_path/graph/graph.json) rather than written into the AKMS repo
  • Graph nodes are enriched with file_path metadata for inspector/editor integration

Frontend:

  • New Graph tab guarded by AKMS graph readiness
  • React Flow canvas with dagre layout
  • Filtered graph, ego-graph, and cluster views
  • Inspector with metadata, connected-edge summary, Open in Nodes, and Open in Editor actions
  • Nodes review page can now jump directly into the Graph tab for the selected node

Configuration:

  • Added akms_global_vault setting with fallback resolution:
  • explicit config
  • <akms_repo>/Packages/Nodes_Vault
  • AKMS_GLOBAL_VAULT
  • ~/.claude/akms/nodes
  • GET /api/app/config now exposes graph cache and effective AKMS graph-vault paths
  • GET /api/app/akms-status now distinguishes picker availability from graph availability

AKMS is now optional

Logic-Loom can run standalone with just zotero-summarizer. The akms-nodes-gen integration moved behind [project.optional-dependencies]. The optional extra now installs both akms and akms-nodes-gen.

Install with uv sync --extra akms to enable the Picker, Nodes, and Graph tabs. Without it, those tabs show a clear availability placeholder instead of failing.

  • akms_repo removed from first-run required fields
  • New GET /api/app/akms-status endpoint reports package + repo availability
  • Frontend <AkmsGate> wrapper on Picker, Nodes, and Graph routes

NotebookLM integration

New /api/nlm/* route module with four task types, all running as background subprocesses through the shared RunRegistry with SSE log streaming:

  • POST /api/nlm/notebook — Create NotebookLM notebook and upload PDFs (wraps prepare_notebook.py)
  • POST /api/nlm/generate — Generate nodes via Claude API with optional NLM MCP grounding (wraps generate_nodes_pipeline.py)
  • POST /api/nlm/nlm-batch — Generate nodes directly via NotebookLM, no Claude tokens (wraps akms_nodes_gen.nlm_batch)
  • POST /api/nlm/validate — Validate nodes against source notebooks (wraps validate_nodes.py)

Picker UI changes:

  • Generate (Claude) and Generate (NLM) buttons in BatchDetailPane
  • Validate Nodes button in BatchDetailPane
  • Buttons unlock progressively after notebook creation
  • NlmBatchDialog collects plan path, prompt file, and options before launch
  • NlmRunDialog streams subprocess output in real time

Config path fix

Fixed getBatch() not reading notebook_id from the AKMS assignment response, which caused Generate/Validate buttons to stay disabled after notebook creation.

Settings

  • Added nlm_mcp_url setting for NotebookLM MCP server URL
  • Added akms_global_vault setting for AKMS graph source resolution

0.1.0 — 2026-05-09

Initial release: zsum pipeline runner, vault browser, AKMS batch picker, settings page, first-run setup dialog.