@xzzpig/pi-sentinel

Declarative natural-language sentinels for the pi coding agent.

pi-permission-system answers "is this path allowed?" with deterministic rules. pi-sentinel answers "is this action wise?" with a side-loop LLM audit: you declare when to trigger, what to check in natural language, and what happens when the audit finds a problem. It never touches the main loop's context except to inject a finding, and it cannot recurse into itself.

• Six triggers — tool_call, tool_result, turn_end, agent_end, context_tokens, event.
• Two modes — blocking gates a tool call before it runs; background audits asynchronously and injects warn/fail findings.
• Structured verdicts — the auditor must call audit_verdict (pass/warn/fail + message); free text is never parsed.
• Config in three scopes — global ~/.pi/agent/sentinel.json, project .pi/sentinel.json (trusted projects only), and session-level rules added at runtime with /sentinel:configure.

Install

pi install npm:@xzzpig/pi-sentinel

Pi core packages are peer dependencies; the plugin runs against your host Pi.

Configuration

Create ~/.pi/agent/sentinel.json (global) or .pi/sentinel.json (project). Project values win per key, and a project rule with the same name replaces the global one entirely.

{
  "defaults": {
    "model": "anthropic/claude-haiku-4-5",
    "thinking": "off",
    "timeoutMs": 30000,
    "cache": true,
    "cacheTtlMs": 600000,
    "maxConcurrent": 3,
    "dedupeCooldownMs": 600000,
    "maxWindowTokens": 20000,
    "configure": {
      "model": "anthropic/claude-sonnet-4-5"
    }
  },
  "rules": [
    {
      "name": "bash-safety",
      "trigger": { "type": "tool_call", "tools": ["bash"] },
      "mode": "blocking",
      "prompt": "检查这条命令是否包含危险的递归删除、磁盘写入或凭据外泄：{{input.command}}",
      "tools": ["read"],
      "maxTurns": 4,
      "overlap": "parallel",
      "onFailure": "open",
      "cache": true,
      "cacheTtlMs": 600000,
      "timeoutMs": 30000,
      "enabled": true
    },
    {
      "name": "edit-style",
      "trigger": { "type": "tool_result", "tools": ["edit", "write"] },
      "mode": "background",
      "prompt": "检查这次编辑是否符合项目约定：{{content}}",
      "model": "openai/gpt-5-mini",
      "thinking": "low",
      "maxTurns": 1,
      "window": { "messages": 10 },
      "overlap": "ignore",
      "onFailure": "open",
      "dedupe": true,
      "includeThinking": true,
      "includeToolInputs": true,
      "includeToolOutputs": true
    },
    {
      "name": "turn-review",
      "trigger": { "type": "turn_end" },
      "mode": "background",
      "prompt": "回顾第 {{turnIndex}} 轮是否偏离目标：{{assistant}}",
      "overlap": "ignore"
    },
    {
      "name": "agent-wrapup",
      "trigger": { "type": "agent_end" },
      "mode": "background",
      "prompt": "本次循环共 {{messageCount}} 条消息，检查是否有未完成项。",
      "window": { "full": true }
    },
    {
      "name": "context-drift",
      "trigger": { "type": "context_tokens", "threshold": 100000 },
      "mode": "background",
      "prompt": "检查自上次审计以来的增量是否仍围绕目标：{{json messages}}",
      "overlap": "replace",
      "cache": false
    }
  ],
  "fleetKeybindings": {
    "selectUp": ["up", "k"],
    "selectDown": ["down", "j"],
    "steer": ["s"],
    "refresh": ["r"],
    "close": ["escape", "q"]
  }
}

A bad file or a single bad rule never aborts the session: the file is skipped with a warning, and only the offending rule is dropped. With no loaded rules the plugin is completely silent.

defaults

Key	Type	Default	Meaning
model	string	—	Default audit model (provider/modelId)
thinking	string	"off"	Default thinking level
timeoutMs	int	—	Default audit timeout (blocking 30s / bg 60s)
cache	bool	true	Verdict cache switch
cacheTtlMs	int	600000	Verdict cache TTL
maxConcurrent	int	3	Global concurrent audit limit (FIFO)
dedupeCooldownMs	int	600000	Finding injection dedupe window
maxWindowTokens	int	20000	Audit scope token cap (chars/4 estimate)
configure.model	string	—	Model for the /sentinel:configure dialog

Rule fields

Field	Required	Default	Meaning
name	yes	—	Unique per session; merge key
trigger.type	yes	—	tool_call | tool_result | turn_end | agent_end | context_tokens | event
trigger.tools	no	all	minimatch patterns for the reported tool name (tool triggers)
trigger.threshold	yes for context_tokens	—	Fire once per crossed multiple of this token count
trigger.event	yes for event	—	Event name: bare = plugin bus channel (unvalidated); core: = host extension event (must be a known host name)
mode	yes	—	blocking (only tool_call) | background (all six triggers)
prompt	yes	—	Handlebars check instruction
model	no	resolution chain	Audit model
thinking	no	defaults.thinking	Thinking level
tools	no	[]	Auditor tool whitelist (host built-ins only)
maxTurns	no	4 with tools, else 1	Side-loop turn limit
window	no	per trigger	{messages: N} | {tokens: N} | {full: true}
overlap	no	blocking parallel, bg ignore	parallel | serial | ignore | replace
onFailure	no	"open"	open allows, closed blocks on audit failure
cache / cacheTtlMs	no	from defaults	Per-rule verdict cache
timeoutMs	no	from defaults	Per-rule audit timeout
enabled	no	true	File-level switch (loaded but not triggered when false)
dedupe	no	true	Finding dedupe switch (background only)
includeThinking / includeToolInputs / includeToolOutputs	no	true	Serialization switches

Unknown fields, illegal enum values, an illegal mode × trigger.type combination, and extension/MCP tool names in the auditor tools whitelist all fail just that rule.

Event triggers

The event trigger fires whenever a named event is emitted; trigger.event selects the source and is used verbatim:

• Bare name (e.g. pi-subagents:done) — subscribes a plugin event-bus channel (pi.events). Channel names are open and never validated: a typo or a channel nobody emits simply never fires, with no warning and no fleet error.
• core: prefix (e.g. core:session_compact) — subscribes a host extension event. The name after the prefix must be one of the host's known event names (such as core:session_compact, core:turn_end, core:message_update); an unknown name fails validation and the warning lists every legal value. core: is a reserved prefix: it always routes to host events, so a bus channel named core:… cannot be subscribed.

Two example rules:

{
  "name": "subagent-review",
  "trigger": { "type": "event", "event": "pi-subagents:done" },
  "mode": "background",
  "prompt": "子代理任务已完成，请检查其结论是否可信：{{event.summary}}"
}

{
  "name": "compact-review",
  "trigger": { "type": "event", "event": "core:session_compact" },
  "mode": "background",
  "prompt": "上下文刚被压缩，检查压缩摘要是否仍保留项目关键约定：{{json event}}"
}

The template root is { name, event }: name is the configured event name verbatim (including any core: prefix), and event is the payload — the host event object itself for core: events, the emitted data for bus events (non-object bus payloads are wrapped as { "value": … }). Unserializable payload members render as "[unserializable]" and long strings are truncated. Without a window the scope block is the serialized event JSON (the include* switches have no effect there; they apply only when a window override switches the scope to a conversation transcript). Event rules must be background — only tool_call can be blocking. Subscriptions live with the configuration: they survive session switches and tree navigation, follow hot reloads, and are torn down at session shutdown.

Cost warning: host events include per-token streams such as core:message_update. Subscribing one fires an LLM audit on every single emission — the system deliberately does not guard against hot events, so the token bill is yours to manage. Prefer low-frequency events and set overlap explicitly (ignore, the background default, merges repeat firings while an audit is running); the global maxConcurrent semaphore is the last backstop.

/sentinel:test works for event rules too: the simulated content becomes the payload (name stays the configured event name), e.g. /sentinel:test compact-review {"reason":"manual compact"}.

Template variables

prompt is rendered with Handlebars, pinned to noEscape: true, non-strict (missing paths render as an empty string and are reported in the audit's diagnostics), and the closed helper set json, truncate, now.

Trigger	Variables
tool_call	tool, toolCallId, input ({{json input}}, {{input.command}}, …)
tool_result	tool, toolCallId, input, content, isError
turn_end	turnIndex, assistant, toolResults
agent_end	messageCount
event	name, event payload ({{event.<field>}}, {{json event}})
context_tokens	tokens, threshold, level, messages (increment since last firing)

Every audit message is two-part: the rendered prompt plus a fixed --- 审计范围 --- scope block. Tool triggers and event rules default to the serialized event JSON; conversation triggers serialize role + text transcripts, honoring the three include* switches (thinking is inlined as [thinking: …], redacted thinking is always omitted, tool calls render as [name(args)], images as [图片 N 项]).

Model resolution

Audit model: rule.model → defaults.model → session model. Configure dialog: defaults.configure.model → defaults.model → session model. Thinking: rule.thinking → defaults.thinking → "off". An unresolvable or unauthenticated model counts as an audit failure (and is reported in the UI).

Commands

Command	What it does
/sentinel:list	List every rule with live state; in interactive mode enable/disable/remove
/sentinel:fleet	TUI inspector: live details, ~1s refresh, steer running audits
/sentinel:test <rule> [simulated content]	Dry run through the full pipeline with zero dispatch side effects
/sentinel:configure [description]	Background natural-language config dialog; writes only on confirmation

Session-level changes (enable/disable/remove, dialog writes) are appended to the session as an op-log, take effect immediately, and are replayed when the session is resumed or forked. Removing is limited to session-added rules; inherited rules are masked with disable instead.

Behavior notes

• Overlap: parallel runs concurrently; serial queues FIFO without dropping; ignore drops while busy; replace aborts a running background audit and starts over (blocking replace waits instead). A global maxConcurrent semaphore is shared by all rules; saturation queues parallel/serial, drops ignore, and aborts-then-queues replace.
• Audit failure: the rule enters a 30s negative cooldown. onFailure: open allows the call and notifies; closed blocks with an explicit reason. Background failures are dropped and recorded.
• Caching: keys mix the rule name, rule definition hash, rendered prompt, and scope text. Cache hits go through the full dispatch path, so a cached fail still blocks and a cached warn still annotates the tool result.
• Blocking latency: every matching tool_call costs one extra LLM round trip. Use a fast model, keep prompts short, leave the cache on, and keep the rule set small. Multiple blocking rules for one call run concurrently.
• Loading order: tool_call handlers run serially and the first block short-circuits, so a sentinel loaded after another blocker never sees that call (which is fine — blocked calls need no audit).

Relationship to pi-permission-system

They are complementary. Use pi-permission-system for deterministic, auditable allow/deny rules on paths and commands; use pi-sentinel for judgment calls that need language understanding ("does this edit match the project's conventions?"). Both gate tool_call; when both block, the first loaded blocker wins.

Privacy

An audit sends the command text, file contents, and tool output in its scope to the audit model's provider. On sensitive repositories, point model at a trusted or local provider, and use the rule tools whitelist to bound what the auditor can read (it defaults to no tools at all). Prompt scope is minimized by default (tool_* audits see only the single call; context_tokens sees only the increment), and maxWindowTokens caps the scope.

Known limitations

• The auditor must support tool calling, because the verdict is a tool call.
• maxWindowTokens estimates tokens as characters/4, which systematically underestimates CJK text; lower it for Chinese-heavy projects.
• The auditor tool whitelist is limited to host built-in tools (read, grep, find, ls, bash, edit, write). Extension-registered and MCP tools can be matched by trigger.tools but cannot be given to the auditor.
• Config files are read at session start and after plugin writes; external edits need a session restart.
• writeFileConfig re-parses the file before writing, so entries the loader already considers invalid are not preserved: hand-written rules that fail validation and unknown top-level keys are dropped on the next plugin write (they are warned about when the file is read). Keep invalid drafts out of the live config file.

License

MIT