meetmind 0.1.0

Turn meeting transcripts into a temporal context graph for coding agents.

Meeting → Context Graph

Turn meeting transcripts into a temporal context graph that a coding agent reads to write and update code. Each meeting updates the same graph.json, and git history records how decisions evolved.

Meetily (capture + transcribe, free)  ─►  context_graph CLI  ─►  graph.json (git)  ─►  coding agent

Why a graph, not just notes

Every node has a status (active / superseded / deprecated) and a supersedes list. When a later meeting reverses a decision, the old node is marked superseded and a new active node links to it via a SUPERSEDES edge. So the graph always shows current truth and what changed.

Setup

pip install -r requirements.txt
export GROQ_API_KEY=...      # free key from console.groq.com  (PowerShell: $env:GROQ_API_KEY="...")

Usage

# --- recommended: pull straight from Meetily's local DB (recording -> transcript) ---
python -m context_graph meetily-list                 # see recent Meetily meetings
python -m context_graph ingest --meetily             # ingest the LATEST Meetily meeting
python -m context_graph ingest --meetily-id <id>     # ingest a specific Meetily meeting

# --- batch: catch up on many meetings at once (oldest first, one commit each) ---
python -m context_graph ingest --new                 # every Meetily meeting not yet in the graph
python -m context_graph ingest --since 2026-07-01    # meetings created on/after a date
python -m context_graph ingest --since 2026-07-01 --new   # both filters combined
python -m context_graph ingest --all                 # every Meetily meeting

# --- other input sources ---
python -m context_graph ingest transcripts/m1.txt --meeting-id 2026-07-12-kickoff
python -m context_graph ingest --paste "we'll use REST and Postgres" --meeting-id m1
python -m context_graph ingest --audio meeting.mp3 --meeting-id m1   # Groq Whisper fallback

python -m context_graph ingest ... --dry-run   # preview changes, no write/commit
python -m context_graph show                    # print current active nodes
python -m context_graph delta                   # git diff of last meeting's changes

# render the graph from the terminal (no Slack needed)
python -m context_graph viz --out graph.png                 # PNG
python -m context_graph viz --format mermaid --out g.mmd    # Mermaid text
python -m context_graph viz --format dot --active-only      # DOT, hide superseded

# check graph integrity (dangling edges, supersedes cycles, duplicate concepts)
python -m context_graph validate

# export the ACTIVE-only spec for the coding agent (grouped by feature)
python -m context_graph export --format md --out spec.md
python -m context_graph export --format json --out spec.json

# scope to a separate project graph -> context-graphs/<name>.json
python -m context_graph ingest samples/m1-kickoff.txt --meeting-id m1 --project payments
python -m context_graph export --project payments

# multi-project meeting: split ONE transcript across per-project graphs automatically
python -m context_graph ingest big-meeting.txt --meeting-id m1 --route

Multi-project meetings (routing)

When one meeting discusses several projects, --route makes the extractor label each node with a project and fans the nodes out into separate context-graphs/<project>.json files — each with its own temporal history, supersession, and export spec. Known projects are discovered from the existing graph files, so labels stay consistent meeting to meeting; a genuinely new project name auto-creates its graph. Cross-cutting items go to general.

Cross-project dependencies are preserved: an edge between nodes in different projects (e.g. Billing DEPENDS_ON an Auth decision) is stored in the source project's graph as a namespaced reference auth:decision-api-graphql, surfaced in that project's export spec (so the coding agent sees the dependency) and exempt from validation's dangling-edge check.

In the Slack bot, set CG_ROUTE_PROJECTS=1 and the bot routes each uploaded meeting by content, replying with a per-project breakdown + one graph image per project touched (instead of assuming the channel = the project).

Entity resolution (how meetings merge without duplicating)

When a later meeting revises a decision, the pipeline decides NEW vs DUPLICATE vs CHANGE (supersede) using, in order:

1. Explicit supersession — the extractor is shown the current active nodes and emits supersedes: [id] when a statement replaces one. Most reliable.
2. Embedding similarity (optional) — if sentence-transformers is installed (or GEMINI_API_KEY is set) nodes are matched semantically, so req-user-login and req-authentication-flow can still merge. pip install sentence-transformers to enable; without it the system falls back to string/concept matching (no error).
3. String/concept matching — id-prefix + token overlap. Always available.

Use export to hand the coding agent a clean current-state spec (no history), and validate in CI to catch a corrupt graph before an agent acts on it.

Meetily integration (recording → transcript)

Meetily records system+mic audio locally and transcribes it for free — no meeting bot needed. It stores everything in a SQLite database, which this CLI reads read-only:

OS	Database path
Windows	%APPDATA%\com.meetily.ai\meeting_minutes.sqlite
macOS	~/Library/Application Support/com.meetily.ai/meeting_minutes.sqlite
Linux	~/.local/share/com.meetily.ai/meeting_minutes.sqlite

The path is auto-detected; override with --meetily-db PATH if needed. With --meetily the meeting-id is auto-derived from the meeting's date + title (e.g. 2026-07-12-design-review), so you don't pass --meeting-id. Speaker source is labeled You: (mic) / Others: (system audio).

Meetings need not be live. A previously-recorded call works two ways: (1) import the recording into Meetily (its audio-import feature transcribes it into the same DB, then --meetily picks it up), or (2) point the CLI straight at the file with --audio old-meeting.mp3 (Groq Whisper transcribes it, no Meetily needed).

Other inputs: pasted minutes (--paste), a plain transcript file, or the built-in Groq Whisper fallback for a raw audio file (--audio).

Slack bot (audio in → graph image out)

Let anyone on the team drop a meeting recording into a channel and get the context graph back as a rendered image — no CLI needed.

User uploads meeting.wav ─► bot transcribes (Groq) ─► ingest ─► posts changelog
                                                              └─► uploads graph PNG

Mentions: @bot graph (post the image), @bot show (list active nodes), @bot delta (git diff since last meeting).

Run it (Socket Mode — no public URL needed)

pip install -r requirements.txt          # includes slack-bolt, graphviz, requests
export SLACK_BOT_TOKEN=xoxb-...           # bot token
export SLACK_APP_TOKEN=xapp-...           # app-level token, scope connections:write
export GROQ_API_KEY=...
export CG_GRAPH_PATH=graph.json           # optional
python -m context_graph.slackbot

Slack app setup: enable Socket Mode; subscribe to events message.channels and app_mention; bot scopes app_mentions:read, channels:history, files:read, files:write, chat:write (add groups:history for private channels). Invite the bot to the channel.

Graph rendering

The image is rendered by context_graph/viz.py:

• Graphviz (preferred, fully offline) — install the system binary (choco install graphviz on Windows, apt install graphviz, brew install graphviz); the graphviz pip package wraps it.
• mermaid.ink (automatic fallback if dot isn't found) — renders via a free web service. ⚠️ This sends the graph to an external host; install Graphviz to keep confidential meetings local.

Encoding: node color = status (green active / grey superseded / amber open question), shape = type (Decision ◆, Requirement ▢, Constraint ⬡, …), and SUPERSEDES edges are red dashed. Check viz.available_renderer() to see which backend is active.

Contract for the coding agent

Read graph.json and treat it as follows:

• Current requirements = every node with "status": "active". Obey these.
• Ignore nodes with "status": "superseded" or "deprecated" as current truth — but use them to know what to remove or replace.
• OpenQuestion nodes are NOT decided — do not assume an answer; flag them.
• What changed since last meeting = git diff HEAD~1 HEAD -- graph.json (or python -m context_graph delta). A node flipping to superseded plus a new active node linked by a SUPERSEDES edge = "replace the old implementation with the new one."
• Edges tell you structure: IMPLEMENTS, DEPENDS_ON, BLOCKS, SUPERSEDES, DECIDED_IN, RELATES_TO.

Node shape

{
  "id": "decision-api-style",
  "type": "Decision",
  "title": "API uses GraphQL",
  "content": "Switch the public API from REST to GraphQL.",
  "status": "active",
  "source_meeting": "2026-07-19-review",
  "created_at": "2026-07-19T10:00:00+00:00",
  "supersedes": ["decision-api-rest"],
  "confidence": 0.9,
  "owner": "Priya"
}

Scope (v1)

Included: transcript/paste/audio input, Groq extraction, temporal merge with supersession, JSON graph + git history, CLI.

Deferred: MCP server, Neo4j/Graphiti engine, meeting bot, embedding-based dedup, GUI. See the plan file for rationale.