swannysec Anti-Laziness Guard
Stop hook that detects and blocks work-skipping rationalizations at stop time. Fires automatically on every Stop event when the workflow-toolkit plugin is enabled.
How It Works
Three hooks run in parallel every time the agent tries to stop. Any one blocking prevents the stop.
Layer 1: Command Hook (always runs, <1s)
Deterministic regex matching against the agent's last_assistant_message. Detects Tier 1-2 laziness phrases with >75% confidence.
Blocks on phrases like:
- "skip the remaining review stages"
- "in the interest of time"
- "context constraints" / "running low on context"
- "accelerate through the remaining"
- "consolidate the remaining"
- "not worth the complexity"
- "to save time" / "for brevity"
See references/phrase-taxonomy.md for the complete taxonomy with sources.
Layer 2: Prompt Hook (always runs, ~2-3s)
Haiku evaluates the agent's final message for a single criterion: internal contradiction — the message claims completion but contains contradicting evidence of incompleteness in the same message.
Biased toward allowing: when in doubt, the evaluator allows the stop. False positives (blocking a legitimate stop) cause more damage than false negatives (missing a lazy stop).
Layer 3: Agent Hook (conditional, 10-30s)
Context-aware deep verification that reads task lists, plan files, and the session transcript. Auto-activates when structured work exists:
- Plan files in
.claude/plans/*.mdor~/.claude/plans/*.md - Task files in
.claude/tasks/**/*.jsonor~/.claude/tasks/**/*.json - Manual flag file:
.claude/anti-laziness-deep-check
When active, a subagent with Read/Grep/Glob access:
- Checks
~/.claude/tasks/and.claude/tasks/for incomplete tasks - Reads plan files in
.claude/plans/and cross-references against the agent's claims - Searches the transcript for TodoWrite entries with pending/in_progress status
- Checks for offering instead of doing — agent offers to do assigned work instead of completing it (with carve-outs for completed-then-offering, backlog items, credential questions, and CLAUDE.md-directed confirmations)
- Checks for silent omission — agent summarizes completed work without mentioning assigned items it didn't do
Context budget: The subagent reads selectively — only user messages from the transcript (not assistant messages, tool results, or progress events). This keeps evaluation within Haiku's 200k token context window even for long sessions.
Infinite Loop Prevention
The command hook checks stop_hook_active — if the agent was already blocked by a stop hook and is trying to stop again, it is allowed through. This prevents the hook from blocking indefinitely.
Background-Work Gate
All three layers allow the stop when the orchestrator is waiting on a still-running subagent it launched. Claude Code re-invokes the agent when background work completes, so blocking a "waiting" stop is both wrong — the awaited work is not the main agent's to do yet — and the direct cause of wasteful in-turn sleep/wait polling loops (the agent waits inside the turn because the guard won't let it end).
- Layer 1 (command): checks the sibling
subagents/directory derived fromtranscript_pathfor anagent-*.jsonlwritten in the last 3 minutes whose last line is not a completed assistant turn. - Layer 3 (agent): runs the same structural check before resolving the in-scope file set.
- Layer 2 (prompt): allows when the final message indicates the agent is awaiting delegated results (it only receives the message text, so it relies on that signal).
Detection covers subagents (the common fan-out case). A backgrounded bash shell that signals completion out-of-band is not directly detected by Layer 1's structural check; Layer 2's prose signal is the backstop there.
Guidance for agents: if you have fanned work out to subagents or launched a background job that will signal its own completion, stop and let the harness re-invoke you — do not run a sleep/wait loop to keep the turn alive. Reserve intentional in-turn waiting for work that genuinely cannot be resumed after a stop. Light administrative steps (pure transcription, reading context for the next step, fetching an unrelated piece of data) remain fine while you wait.
Configuration
Enable/disable the entire hook: Enable or disable the workflow-toolkit plugin in Claude Code settings.
Layer 3 auto-activation: Layer 3 activates automatically when plan or task files exist. No configuration needed for structured work sessions.
Manual activation (for sessions without plans/tasks):
# Before a critical session
touch .claude/anti-laziness-deep-check
# After the session
rm .claude/anti-laziness-deep-check
Disable just the command hook: Not possible without modifying hooks.json. If you experience false positives, please report them so we can refine the phrase taxonomy.
Tier Summary
| Tier | Confidence | Action | Detection Method |
|---|---|---|---|
| 1 | >90% | BLOCK | Regex (command hook) |
| 2 | 75-90% | BLOCK | Regex with context scoping (command hook) |
| 3 | 50-75% | Documented only | Not enforced — see references/phrase-taxonomy.md |
| 4 | Varies | BLOCK | Haiku evaluation — internal contradiction only (prompt hook) |
| 5 | Varies | BLOCK | Context-aware agent verification — offering, omission, task/plan checks (agent hook) |
Coverage model: Layers 1+2 are intentionally conservative — near-zero false positives, may miss subtle patterns like pure rationalization without internal contradiction. Layer 3 auto-activates during structured work sessions (plan or task files present) and provides deeper, context-aware coverage including patterns that require transcript access.
Background
This hook was designed based on research across 10+ verified GitHub issues (anthropics/claude-code #26691, #21604, #23368, #16506, #1632, #24129, #6159, #8738, #20270, #1113), the Columbia DAPLab "Vibe Coding Needs Policy Enforcement" research, the Cognition/Anthropic "context anxiety" discovery, and analysis of real session logs showing agents using euphemistic language ("accelerate," "consolidate," "streamline") to disguise work-skipping as productivity optimization.