BigBoss: 399 passing tests, 0 skips across 4 CI jobs

BigBoss is a local approval queue and SQLite decision store for AI coding agents. I built it with the team (the AI agents I direct) using Python's standard library, Server-Sent Events and a static browser interface (thisisntjon/bigboss-approval-plane, inspected at commit 5615b72 on September 23, 2026). Tagged release v0.1.0 records 399 passing tests and 0 skips across 4 CI jobs (GitHub Actions run 35043376854).

Method: adapters submit a proposed agent action with its workspace and run context. Policy routing classifies the action (src/bigboss/policy.py), SQLite stores the request and operator decision (src/bigboss/store.py), and the calling adapter matches execution to the approved action_hash (src/bigboss/contracts.py).

BigBoss case study, agent security write-up, local quickstart, about me, evidence index.

Architecture of the inspected BigBoss approval plane at commit 5615b72: action hashing, three-way policy routing, SQLite persistence and adapter-level enforcement.

What was built

Policy classifier with 3 action routes: classifies each proposed action into 1 of 3 routes: automatically allowed for structured read-only actions, pending review for file edits and shell commands, or blocked for selected command patterns.

Source: src/bigboss/policy.py and tests/test_policy.py Status: Public source and unit tests at commit 5615b72.
Action-bound approval hash: computes 1 action_hash from 3 inputs (the proposed action, workspace and policy version) so an operator decision applies only to that exact request.

Source: src/bigboss/contracts.py Status: Public source at commit 5615b72.
SQLite store for 4 record types: persists requests, operator decisions, enrolled devices and events in local SQLite tables.

Source: src/bigboss/store.py and tests/test_store.py Status: Public source and store tests at commit 5615b72.
4 integration surfaces: provides local HTTP endpoints with Server-Sent Events, a local MCP server, command hooks and a JSON-RPC app-server bridge.

Source: README.md, src/bigboss/server.py and src/bigboss/codex_app_bridge.py Status: Public source at commit 5615b72.

Exact results

399 passing tests, 0 skips across 4 CI jobs: tagged release v0.1.0 (commit 10acc53) passed 399 tests with 0 skips in each of 4 CI matrix jobs across 2 operating systems (Ubuntu and Windows) and 2 Python versions (Python 3.12 and Python 3.13).

Source: Release v0.1.0, GitHub Actions run 35043376854 and .github/workflows/ci.yml Status: Public release and CI workflow run on GitHub.
Commit 5615b72 passes all 4 CI matrix jobs: commit 5615b72 (September 19, 2026) restricts pairing-page access to loopback addresses (127.0.0.1 and ::1) and browser navigation headers, passing all 4 CI matrix jobs.

Source: Commit 5615b72, src/bigboss/server.py and GitHub Actions run 35467029298 Status: Public commit and CI workflow run on GitHub.

Verify in one minute

Open release v0.1.0 and GitHub Actions run 35043376854 to inspect the 399 passing tests across 4 CI jobs, or clone thisisntjon/bigboss-approval-plane at commit 5615b72 and run the test suite and local demo:

git clone https://github.com/thisisntjon/bigboss-approval-plane
cd bigboss-approval-plane
git checkout 5615b729126cb829e3e5eedc209b71016eb231e2
pytest -q
python -m bigboss.cli demo

Scope of these checks

Adapter enforcement boundary: the app-server bridge pauses execution until a decision returns; HTTP and MCP clients must enforce the returned decision in the calling wrapper.

Source: README.md#security-model and src/bigboss/codex_app_bridge.py Status: Documented adapter contract; BigBoss does not sandbox the host OS.
Independent reproduction: CI runs were executed on my repository runners (run 35043376854, run 35467029298).

Source: BigBoss case study: Run the checks yourself Status: My repository CI runs; no external third-party test report recorded yet.