Sensors Catalog

Sensors are shell scripts wired into Claude Code’s hook system. They run automatically on specific events — agents cannot bypass them. Each sensor has an event that triggers it and an optional matcher that filters which tools it applies to.

Note

Sensors marked as required are auto-installed for all compatible agents and cannot be unchecked in the interactive picker. They are the enforcement layer of the Bonsai instruction stack.

All Sensors

Name	Description	Event	Matcher	Compatible Agents	Required
agent-review	Outputs a review checklist after a dispatched agent completes work	PostToolUse	Agent	tech-lead	No
api-security-check	Detects security anti-patterns in code changes — SQL injection, hardcoded secrets, eval, CORS wildcards	PreToolUse	Edit	Write	backend, fullstack, security
compact-recovery	Re-injects minimal context after /compact (Quick Triggers + Work State only)	SessionStart	compact	all	Yes (all)
context-guard	Injects context-aware behavioral constraints and detects session wrap-up trigger words before each prompt	UserPromptSubmit		all	Yes (all)
dispatch-guard	Validates code agent dispatches — requires worktree isolation, plan reference, and plan existence before execution	PreToolUse	Agent	tech-lead	No
iac-safety-guard	Blocks dangerous infrastructure commands — terraform destroy, kubectl delete namespace, unsafe docker ops	PreToolUse	Bash	devops	No
routine-check	Checks routine dashboard at session start and flags overdue maintenance routines	SessionStart		all	Auto-managed
scope-guard-commands	Blocks agent from running application execution commands (tests, builds, servers)	PreToolUse	Bash	security, tech-lead	No
scope-guard-files	Blocks agent from editing files outside its workspace	PreToolUse	Edit	Write	all
session-context	Injects core identity, memory, protocols, and project status at session start	SessionStart	startup	resume	clear
status-bar	Persistent status line showing context usage, session health, and git state after every response	Stop		all	Yes (all)
subagent-stop-review	Outputs a structured review checklist when a dispatched agent finishes work	SubagentStop		tech-lead	No
test-integrity-guard	Detects test quality issues — removed assertions, .skip/.only, empty test bodies, debug artifacts	PreToolUse	Edit	Write	backend, frontend, fullstack

Sensors by Function

Guards

Guards block actions that violate boundaries or safety rules. They intercept tool calls before they execute and prevent dangerous operations.

• scope-guard-files (required) — Prevents agents from editing files outside their workspace directory. This is the primary workspace isolation mechanism.
• scope-guard-commands — Prevents the Tech Lead and Security agents from running application execution commands (tests, builds, servers) — these agents orchestrate, not execute.
• dispatch-guard — Validates that code agent dispatches use worktree isolation, reference a plan, and that the plan file exists. Tech Lead only.
• iac-safety-guard — Blocks destructive infrastructure commands like terraform destroy and kubectl delete namespace. DevOps only.

Context Injection

Context injection sensors add information at key moments — session start, before prompts, or after responses.

• session-context (required) — Injects the agent’s core identity, working memory, protocols, and current project status at session start. This is how agents “wake up” with context.
• context-guard (required) — Injects behavioral constraints before each user prompt and detects session wrap-up trigger words (e.g., “that’s all”, “wrap up”).
• status-bar (required) — Displays a persistent status line after every response showing context usage percentage, session health, and current git state.

Review

Review sensors output checklists after agents complete work, prompting the Tech Lead to verify quality.

• agent-review — Outputs a review checklist after a dispatched code agent completes work via the Agent tool (PostToolUse).
• subagent-stop-review — Outputs a structured review checklist when a dispatched agent finishes work (SubagentStop event).

Code Quality

Code quality sensors detect anti-patterns in code changes before they are written.

• api-security-check — Scans code edits for security anti-patterns: SQL injection, hardcoded secrets, eval(), CORS wildcards, and similar issues.
• test-integrity-guard — Catches test quality issues: removed assertions, .skip/.only left in tests, empty test bodies, and debug artifacts.

Routine Monitoring

• routine-check — Parses the routine dashboard at session start and flags any overdue maintenance routines. This sensor is auto-managed: it is installed when any routines are present and removed when the last routine is removed. You do not install or remove it manually.

Hook Events

Sensors listen to Claude Code hook events. The available events are:

Event	When it fires
SessionStart	Once at the beginning of a new session
UserPromptSubmit	Before each user message is processed
PreToolUse	Before a tool call executes (use matcher to filter by tool)
PostToolUse	After a tool call completes (use matcher to filter by tool)
Stop	After each agent response
SubagentStop	When a dispatched subagent finishes

Default Sensors by Agent Type

Agent Type	Default Sensors
Tech Lead	session-context, scope-guard-files, scope-guard-commands, dispatch-guard, subagent-stop-review
Backend	session-context, scope-guard-files
Frontend	session-context, scope-guard-files
Full-Stack	session-context, scope-guard-files
DevOps	session-context, scope-guard-files, iac-safety-guard
Security	session-context, scope-guard-files, scope-guard-commands, api-security-check

Tip

The 4 required sensors (context-guard, scope-guard-files, session-context, status-bar) are always installed for all agents. The table above shows the additional default sensors per agent type.