Everything else here is local: skills on your machine, a vault on your disk.
This puts a vault behind a URL so an agent elsewhere can search and write it.
One container, one vault, one API key. The vault stays plain markdown on a
mounted volume, so Obsidian still opens it. The container does no LLM work --
search and packing delegate to the existing graphrag/context_pack modules, and
the caller's agent does the thinking.
- obsidian_wiki/server.py: four vault functions, five REST routes and four MCP
tools as transport over them. Auth is one bearer key; the process refuses to
start without WIKI_API_KEY unless WIKI_ALLOW_ANONYMOUS=1. Every path is
resolved inside the vault or refused.
- Dockerfile, docker-compose.yml, .dockerignore, and a ghcr push on v* tags.
- New [server] extra; the core wheel stays dependency-free.
- docs/deployment.md, plus the WIKI_* vars in configuration.md and .env.example.
Requires mcp>=2.0: 2.0 renamed FastMCP to MCPServer, and session_manager only
exists once streamable_http_app() has been called -- a mounted sub-app's own
lifespan is ignored, so it is wired into the FastAPI lifespan by hand.
The brain-capture extension only went one way: browser -> vault. This adds
the return path — read the form on the current page, answer what the vault
supports, and leave the rest blank — and folds both directions into a single
two-tab extension renamed from "Brain Vault Capture" to "Brain".
Retrieval never calls a model. `obsidian-wiki context-pack --json` ranks the
vault deterministically and the model does one narrow job: map retrieved facts
onto fields. One `claude -p` / `codex exec` invocation per fill, billed to the
subscription the user already has. Nothing is ever submitted.
A native messaging host bridges the two, staged into ~/.obsidian-wiki/ rather
than run from the checkout. Chrome's children do not inherit your shell, and
all three consequences bite:
- macOS TCC: ~/Documents is mode 700 with an ACL, and a native host inherits
Chrome's TCC context, so a checkout there cannot be exec'd at all. Chrome
reports this as "Native host has exited", identical to a crash.
- PATH: Chrome passes a minimal PATH, hiding claude/codex in ~/.local/bin.
- USER/LOGNAME: Claude Code reads OAuth credentials from the macOS login
Keychain and needs these to find them, else "Not logged in" despite a valid
session.
The host bookends every run in its log so those failures stay distinguishable,
re-validates every returned fill against the real form schema (an injected
value cannot reach a field the page never advertised), and forwards the pack's
instruction_policy so vault excerpts stay data rather than instruction.
Form-schema extraction resolves labels via aria-labelledby before aria-label:
Google Forms puts aria-label="Your answer" on every paragraph field while the
real question sits behind aria-labelledby, so checking aria-label first
silently mislabelled exactly the textarea fields. Values are written through
the native setter with input/change dispatched, so React-controlled forms see
a genuine edit.
Docs: docs/browser-extension.md, linked from both READMEs and the docs index.
Also backfills the untranslated tagline from #159 into README_TW.md.
Two strands of work.
Session brain — a topic graph over raw agent session history, so you can
find the session where something happened instead of remembering its
UUID. Reads every Claude transcript plus the pruned sessions that survive
only as prompts in history.jsonl, clusters them with local TF-IDF and
community detection (no embeddings, no API calls, no new dependencies),
and writes a sidecar at ~/.claude/session-brain/ — the vault is never
touched. Ships five modules, the sessions-* CLI commands, an interactive
graph.html, and the /session-brain and /wiki-sessions skills. This is
retrieval, deliberately distinct from wiki-history-ingest, which distils
sessions into permanent vault pages.
Docs — README.md was 598 lines and put trust-check flags and a 17-row
agent matrix in front of anyone evaluating the project. Cut it to a 129-
line landing page (problem statement, 60-second install, four blocks of
runnable commands, graph screenshot, why-not-a-notes-folder) and move the
reference material into docs/: installation, agents, skills, cli,
configuration, architecture, session-brain, contributing. README_TW.md
mirrors the new structure; docs/ stays English-only for now.
The restructure also closed real drift:
- Ten skills existed but appeared in no reference — session-brain,
session-search, wiki-dedup, wiki-digest, wiki-import, wiki-context-pack,
wiki-stage-commit, wiki-switch, vault-skill-factory, and
obsidian-layout-adjustment.
- SETUP.md listed 12 of 39 skills and still claimed "no scripts, no
dependencies" despite the pip package. It is now a signpost into docs/.
- Eight config variables were undocumented anywhere, including
WIKI_STAGED_WRITES, which wiki-setup prompts for and three skills read.
Added to .env.example and docs/configuration.md.
- sessions-name and a number of CLI flags were missing from the docs.
Two details worth noting. README links are absolute GitHub URLs because
pyproject sets readme = "README.md" as the PyPI long_description, and PyPI
resolves relative hrefs against pypi.org rather than the repo — docs/
links would have 404'd on the package page. And three doc-drift guards
asserted their facts against README.md; since those facts moved rather
than disappeared, the assertions were repointed at the page that now owns
each one instead of padding content back into the README.
Also gitignore /vault/ (a local knowledge vault, not a fixture) and the
packed browser extension.
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Arnav