Reference
Guardrails & Detector Catalog
Highbeam runs 32 core detectors (DET-01..DET-32), 8 habit coaches, and 6 local Blind-Spot Sensors in fast local JavaScript. Zero extra LLM judge tokens are ever spent.
Canonical Detector Catalog (DET-01 to DET-32)
Four rules carry a Ring-0 Security tag (DET-01, DET-02, DET-18, DET-19). Ring-0 guards protect secrets, shell safety, repo settings, and subagent inputs, and block unsafe calls with exit 2. If the hook itself hits an internal fault it fails open by default (see Hooks Lifecycle).
| ID | Detector Name | Default / Severity | What It Detects & How to Remediate |
|---|---|---|---|
DET-01 |
Option Argument Skipping & Destructive Git/Shell Guard Ring-0 | guard · CRITICAL |
Reads git and shell flags safely to stop flag-smuggling tricks. Blocks risky commands like git reset --hard, git push --force, broad rm -rf, and branch deletes unless you ask for them. |
DET-02 |
Zero-Token Secret & Sensitive File Guard Ring-0 | guard · CRITICAL |
Blocks reads or writes to .env*, SSH private keys, cloud keys, ~/.gemini/oauth_creds.json, and token files before secrets enter chat context. |
DET-03 |
Lockfile & Migration Scope Drift Guard | guard · WARN |
Stops hand edits to generated lockfiles (package-lock.json, pnpm-lock.yaml, Cargo.lock) or past DB migrations unless you ask for them. |
DET-04 |
3-Strike Consecutive Tool Failure Breaker | guard · HIGH |
Trips when 3 tool calls in a row fail with the same unfixed error (consecutiveFailureLimit: 3). Check the root cause before retrying. |
DET-05 |
Destructive Code-Strip, Test-Weakening & Flip-Flop Guard | guard · CRITICAL |
Blocks edits that delete test asserts in bulk, add .skip or only to hide broken tests, or flip-flop between A and B code states. |
DET-06 |
Unmodified File Re-Read & Blind-Edit Guard | advisory · MEDIUM |
Flags repeat full reads of files unchanged on disk since earlier turns. Also warns before editing a file never read in the session. |
DET-07 |
Exit-Safe CLI Compressor & Output Bomb Guard | guard · HIGH |
Catches shell outputs over 4,096 chars (or 25k tokens), strips ANSI noise, and saves full logs in SQLite FTS5 (with centered snippets and a built-in table polyfill on Node 22 builds lacking SQLITE_ENABLE_FTS5). Keeps exact exit codes and error lines intact. |
DET-08 |
KV-Cache Gravity Cross-Model Switch Guard | guard · HIGH |
Checks single-turn C* and multi-turn C_max payback on PreModelSwitch. Uses a $0.02 floor (minSwitchPenaltyUSD), lets unknown non-Anthropic models pass, and defaults holds to permissionDecision: "ask" ("deny" in headless runs). |
DET-09 |
Warm Prefix-Cache Mutation Fence | guard · HIGH |
Holds writes to AGENTS.md, CLAUDE.md, and Turn-0 rules—plus mid-TTL effort, thinking, speed, and mcp_tools changes—above computeDet09BreakEvenTokens (det09MinCachedTokens). Queues edits in TurnZeroGate until TTL ends or PostCompact. |
DET-10 |
Idle 5m Implicit-Cache Watchdog | advisory · WARN |
Watches idle breaks nearing the 300s cache limit. Shows an amber warning at 45s and saves the State Baton at 10s. |
DET-11 |
Shadow File Context Bloat & Confabulation Guard | advisory · MEDIUM |
Spots stale backup copies, scratch duplicates, or guessed file names before edits drift away from the real source file. |
DET-12 |
Haiku 100k Cliff Breaker & Zombie Subprocess Leak Guard | guard · CRITICAL |
Stops Claude Haiku sessions before hitting the 100,000-token 5x price jump (98k warn, 100k stop) and flags orphan background processes. |
DET-13 |
Eager MCP Override & Startup Bloat Auditor | advisory · MEDIUM |
Checks Turn-0 eager MCP tool schemas that pad every turn and suggests lazy-loading servers you rarely use. |
DET-14 |
Premature Stop Gate & Test Verification Guard | guard · HIGH |
Stops the agent from finishing on Stop when edited source files still have failing tests or no test run yet. |
DET-15 |
Batched Stop-Phase Formatter & Placeholder Slop Guard | advisory · MEDIUM |
Saves formatter runs for the end of a task and flags unfinished TODO, FIXME, or stub code in diffs. |
DET-16 |
Hybrid Quota Fan-Out Throttle & Subagent Recursion Cap | advisory · HIGH |
Caps parallel subagents to 1 worker at 80% quota, asks at 90%, and limits workers to 40,000 tokens or 12 turns to avoid 168-hour lockouts. |
DET-17 |
Blind Edit Mid-Batch & Progressive Checkpoint Deferral | advisory · MEDIUM |
Counts dirty files without a test check (default: 4 files) and nudges you to run a quick check before editing more. |
DET-18 |
Untrusted Workspace Settings & Hook Injection Guard Ring-0 | guard · CRITICAL |
Blocks agent edits to .claude/settings.json, hooks.json, or IDE run settings that could widen permissions or add shell hooks. |
DET-19 |
Subagent IPI Boundary & Unicode Tag Sanitizer Ring-0 | guard · HIGH |
Strips hidden Unicode tag chars (U+E0000..U+E007F), zero-width joiners, and prompt-injection lines from subagent and web outputs. |
DET-20 |
Multi-Agent Scope Collision & Hard Quota Brake | guard · CRITICAL |
Enforces file locks in .agents/locks/active_scopes.json across parallel agents and warns on Vision Overage HTTP 400 image bugs. |
DET-21 |
Scratchpad Loop Guard (Compatibility Alias → DET-30) |
guard · CRITICAL |
Saved alias that routes straight to DET-30 thinking-loop detection. |
DET-22 |
Diff-Thrashing & Rapid Code Mutation Sensor | guard · HIGH |
Spots rapid rewrites of the exact same lines over a few turns and nudges you to re-check requirements. |
DET-23 |
Empty Catch-Block & Swallowed Error Sensor | advisory · HIGH |
Flags new empty catch {} blocks that hide runtime errors without logging or handling them. |
DET-24 |
Context Degradation & High-Turn Prompt Bloat Sensor | advisory · MEDIUM |
Warns when turn count passes trust_decay_turn_threshold (10 turns) while middle-of-prompt clutter climbs. |
DET-25 |
Cascade 70k Progressive Checkpoint Sensor | advisory · INFO |
Prompts a short progress note near 70,000 tokens so long chats never hit 140k without a saved checkpoint. |
DET-26 |
Monolithic Prompt Decomposer Sensor | advisory · INFO |
Spots huge multi-goal prompts and suggests splitting separate tasks into clean turns or worktrees. |
DET-27 |
Rule File Bloat & Prefix Diet Auditor | advisory · WARN |
Flags AGENTS.md or CLAUDE.md files over 8,000 tokens and suggests running /evolve to split out skills. |
DET-28 |
Subagent Phantom & Orphan Process Sensor | advisory · HIGH |
Checks that files claimed by subagents exist on disk and spots stuck language-server or test PIDs (clean up with highbeam doctor --reap-orphans). |
DET-29 |
Reflector Queue Starvation & Async Feedback Gate | guard · MEDIUM |
Keeps staged learning queues from growing unchecked and surfaces saved lessons once the cache is cold. |
DET-30 |
Scratchpad Loop & Reasoning Hallucination Guard | guard · CRITICAL |
Spots stuck thinking loops and reasoning spikes with no real tool progress, setting loopTriggered: true in Cost Radar. |
DET-31 |
Cross-Worktree Git Merge Conflict Sensor | advisory · HIGH |
Runs safe in-memory git merge-tree checks across active worktrees to warn before two branches clash on the same lines. |
DET-32 |
IOEDR / REMORY Semantic Drift & Contradiction Guard | guard · HIGH |
Checks new plans and edits against locked user rules and the State Baton to catch silent rule drift. |
Semantic Habit Detectors (semanticDetectors)
Along with numbered IDs, .attention-horizon.json includes eight readable habit slugs you can set via highbeam config set <slug> <default|off|warn|block>:
formatter-bypass: Suggests running prettier or eslint once at the end instead of burning turns after every small edit.drip-feed: Spots rapid 1-line follow-up prompts that pay extra cache-write fees when steps could be grouped.vague-prompt-rubric(aliases:prompt-decomposer,scope-brake): Scores broad prompts that miss target files or test goals.effort-overkill: Flags max or high effort frontier use on simple typos, renames, or formatting.post-commit-handoff: Nudges you to compact or start fresh after finishing a keygit commit.vision-screenshot-coach: Suggests reading DOM or CSS text instead of sending repeat full screenshots (and warns onDET-20overage HTTP 400 bugs).micro-skills: Suggests turning repeat workflow fixes (aftermicroSkillConfirmSessions: 2) into reusable skills.relay-breaker(alias:spec-relay): Flags Round 2+ multi-model spec debates and suggests switching to/relay.
Antigravity 6 Local Blind-Spot Sensors
During Antigravity's PreInvocation hook, src/core/blindspot_sensors.mjs runs six fast local disk checks:
DIRTY_SNAP: Spots when a file read earlier in chat was changed on disk by you, a formatter, or git.SCHEMA_TAX: Measures Turn-0 MCP and tool schema token size against your window limit.UNSEEN_CALLERS: Scans repo imports when an exported function changes so callers outsideACTIVE_SETdo not break.LEGACY_POISON: Warns when reads pull code fromdeprecated_paths(**/legacy/**,**/deprecated/**,**/v1/**).QUOTA_AND_STREAM_SHIELD: Guards against low 5-hour quota (<10%) and single-file writes over600lines.LOOP_AND_GIT_GUARD: Pairs 3-strike command loop checks with dirty-tree reminders.