- Python 100%
|
|
||
|---|---|---|
| bohemian_sidecar | ||
| docs | ||
| examples | ||
| tests | ||
| .gitignore | ||
| AGENTS.md | ||
| pyproject.toml | ||
| README.md | ||
bohemian-sidecar
bohemian-sidecar is a tiny MCP/CLI sidecar for agent context hygiene.
It asks the question an agent should ask before touching production, user-facing channels, or multi-repo state while running with a long context:
Is this the real life?
- Czy sprawdziłem realny stan?
Is this just fantasy?
- Czy jadę na narracji z kontekstu?
Caught in a landslide?
- Czy robię za szeroki fix, bo operator nacisnął emocjonalnie?
No escape from reality.
- Sprawdź user-facing proof, nie tylko service health.
The sidecar does not block tools, mutate production, restart services, or
replace operator judgment. It produces a structured green/yellow/red risk
assessment plus a pre-mutation checklist.
Install
cd /Users/pd/Developer/bohemian-sidecar
python3 -m venv .venv
. .venv/bin/activate
pip install -e '.[dev]'
CLI
Assess a risky long-context production change:
bohemian-sidecar assess \
--context-fill 88 \
--mutating \
--live-prod \
--user-facing \
--multiple-repos \
--recent-misdiagnosis \
--operator-pressure
Run a mutation preflight:
bohemian-sidecar preflight \
--context-fill 88 \
--mutating \
--live-prod \
--user-facing \
--touch "openclaw heartbeat routing config" \
--not-touch "Synapse, Element, Honcho DB, sidecar SQLite" \
--rollback "restore timestamped config backup and restart only the touched service" \
--quiet-smoke "manual heartbeat should not send HEARTBEAT_OK to Signal" \
--user-facing-proof "openclaw system heartbeat last does not show status=sent for HEARTBEAT_OK"
MCP server
Run over stdio:
bohemian-sidecar serve
MCP tools:
assess_context_riskmutation_preflight
Example client configuration shape:
{
"mcpServers": {
"bohemian-sidecar": {
"command": "/Users/pd/Developer/bohemian-sidecar/.venv/bin/bohemian-sidecar",
"args": ["serve"]
}
}
}
Risk bands
The default thresholds are intentionally conservative and not presented as a scientific model-quality curve:
green: normal work. Prefer small reversible steps.yellow: assumptions likely matter. Slow down and verify.red: think broad, mutate narrow. Before mutation, nameDotykam,Nie dotykam,Rollback, andQuiet-path smoke.
Long context is treated as useful for synthesis and risky for broad mutation.
Related Work
See docs/related-work.md for nearby MCP/context tools and research on long-context degradation, reflection, and deliberative guardrails.