redbarron23/obs-agent

**Turn missing or suspect Azure/GCP labels into a reviewed change plan.** Import an inventory,
inspect ownership evidence, accept or override suggestions, and export the before/after changes.
Label review runs locally without an LLM or cloud credentials. Optional cost chat uses synthetic data.
Quick start
Requirements: Python **3.10+** and uv.
Open the local URL printed by Streamlit. **Label review** is the landing page. Enter your reviewer
name or ID, inspect a suggestion, save a decision, and download **Reviewed change plan**.
Start with the bundled synthetic inventory; no API keys or Ollama setup are needed.
For disposable demo decisions, run OBS_AGENT_DEMO=1 uv run streamlit run app.py.
See the five-minute walkthrough.
Use your own inventory
1. Open **Import / reopen inventory** and download the CSV and vocabulary examples.
2. Upload your sanitized Azure subscription/GCP project snapshot. If its headers differ from the
CSV schema, map them on the page. Supply your own team/environment values and aliases in the
optional vocabulary JSON.
3. Select **Save and open review**, then review suggestions and export a plan.
4. Return to **Import / reopen inventory** to resume a saved snapshot after a browser or app restart.
5. Next cycle, save the new snapshot and open **Compare snapshots** to review only what changed.
Snapshots, vocabularies, and approvals persist locally in review-workspaces/ (configurable with
OBS_AGENT_REVIEW_DIR). Keep this private directory outside version control and back it up.
Imports are available in the review UI; the REST API and optional chat use the synthetic estate.
The public demo disables private imports. See Import inventory.
What you get
- Missing/suspect team and env suggestions with evidence, confidence, and alternatives.
- A SQLite decision history with reviewer identity, timestamps, inventory fingerprint, and inference version.
- Reviewed-only JSON change plans and CSV, shell, or Terraform exports. A cost center accompanies
a team change only when the inventory establishes the mapping.
- Explicit opt-in to include unreviewed high-confidence suggestions in convenience exports.
An example exported plan shows the full schema.
For example, an accepted suggestion produces this change:
The complete plan includes the approval evidence and snapshot/version identifiers. Changed inputs
or inference versions require a fresh review. obs-agent exports plans; it never applies cloud changes.
Reviewer identifiers are self-reported locally; authenticate reviewers at your deployment boundary.
Run with Docker
Open http://localhost:8501 for review and http://localhost:8000/docs for the API.
Both bind to loopback. The named volume retains decisions and imported reviews. Ollama is needed
only for optional cost chat; Compose reaches the host automatically. See Interfaces.
How it works
Label inference uses observed labels and ownership signals; reviewers resolve uncertain suggestions.
Plans bind approvals to their inventory and inference version. Optional cost chat is a separate