Sentinel config
View or update Sentinel configurationFrom its SKILL.md
npx -y skills add AndurilCode/sentinel --skill sentinel-configAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
4.7 KB, as published. Nobody here has run it
Sentinel Config
View or update the Sentinel configuration in the current repository.
What to do
-
Check that
.claude/sentinel/config.yamlexists. If not, tell the user to run/sentinel-initfirst and stop. -
Read
.claude/sentinel/config.yaml. -
If the user asked to view the config, show the current values and explain what each one does. Stop here.
-
If the user asked to change something, use the Edit tool to update the specific value(s) in the config file. Preserve all comments. Do not rewrite the entire file.
Configuration reference
| Key | Default | Description |
|---|---|---|
backend | ollama | LLM backend: ollama, claude, or copilot |
model | gemma3:4b | Default model for evaluation (backend-specific) |
backends.ollama.url | http://localhost:11434 | Ollama endpoint |
backends.ollama.model | (top-level model) | Default Ollama model |
backends.claude.model | haiku | Default Claude model |
backends.copilot.model | gpt-5-mini | Default Copilot model |
ollama_url | http://localhost:11434 | Ollama endpoint (legacy; prefer backends.ollama.url) |
timeout_ms | 5000 | Per-rule evaluation timeout |
confidence_threshold | 0.7 | Minimum confidence to count as violation |
max_parallel | 4 | Concurrent LLM calls |
ollama_concurrency | 1 | Max concurrent Ollama HTTP calls (GPU-bound) |
think | false | Enable /think mode (slower, more accurate) |
fail_open | true | Skip rule on error vs block |
content_max_chars | 800 | File content truncation in prompts |
log_file | null | JSONL telemetry path |
rules_dir | rules | Rules directory (relative to config dir) |
tool_map | (see reference.md) | Tool name → trigger type mapping |
mcp_prefix | mcp__ | Prefix for detecting MCP tool names |
mcp_separator | __ | Separator for parsing MCP server/tool from tool name |
context.enabled | true | Enable session context accumulator |
context.backend | (top-level backend) | LLM backend for accumulator |
context.model | gemma3:4b | Model for accumulator |
context.min_events | 3 | Minimum events before accumulator update |
context.lock_timeout_s | 30 | Max wait for GPU lock (Ollama only) |
context.summary_max_words | 150 | Token budget for rolling summary |
scribe.enabled | true | Enable the Scribe convention learning pipeline |
scribe.backend | (top-level backend) | LLM backend for scribe |
scribe.model | (top-level model) | Default model for all scribe steps |
scribe.extraction_model | (scribe.model) | Override for extraction (reflect + learn) |
scribe.synthesis_model | (scribe.model) | Override for validation + synthesis |
scribe.guidance | null | Priority guidance text for extraction |
scribe.think | false | Enable /think mode for validation+synthesis |
scribe.extraction_timeout_ms | 15000 | Timeout for extraction LLM calls |
scribe.extraction_num_predict | 1000 | Max output tokens for extraction |
scribe.synthesis_timeout_ms | 15000 | Timeout for validation+synthesis LLM calls |
scribe.synthesis_num_predict | 1000 | Max output tokens for validation+synthesis |
scribe.temperature | 0.1 | LLM temperature for all scribe calls |
scribe.transcript_budget_chars | 4000 | Max compacted transcript size before truncation |
scribe.thresholds.extraction_confidence | 0.7 | Minimum confidence to store an observation |
scribe.thresholds.draft_confidence | 0.8 | Minimum confidence for learn mode drafts |
scribe.sources.documentation | true | Enable doc scanning via /sentinel-learn |
scribe.doc_globs | [CLAUDE.md, AGENTS.md, ...] | File patterns for /sentinel-learn |
scribe.notification.max_age_days | 7 | Max age for draft notifications |
Backend and model recommendations
| Backend | Model | Notes |
|---|---|---|
ollama (8 GB RAM) | gemma3:4b | Dense, ~3 GB at Q4, local |
ollama (16 GB RAM) | gemma3:12b | Dense, best local accuracy for block rules |
claude | haiku | Fast cloud model, good for most rules |
copilot | gpt-5-mini | Fast cloud model via Copilot CLI |
Backends can be mixed: use Ollama globally for speed, override specific block rules with backend: claude for higher accuracy.
Next steps
After viewing or updating config, suggest these to the user:
/sentinel-stats— see how the current configuration affects rule performance/sentinel-rule— create a new rule if the rules directory is empty/sentinel-learn— scan documentation for conventions to enforce automatically