agentsclimarketplace
Sign in

Agent behavior constraints

Skill josix/agent-flow/skills/agent-behavior-constraints

This skill should be used when handling agent model selection, tool access permissions, behavioral guardrails, MCP tool preferences, or any question about what agents can/cannot do.From its SKILL.md

Install
npx -y skills add josix/agent-flow --skill agent-behavior-constraints

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 7 stars7 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

6.0 KB, as published. Nobody here has run it

Agent Behavior Constraints

Define behavioral rules governing model selection, tool access, and operational guardrails.

Overview

This skill consolidates four core constraint domains:

  1. Model Routing - Which AI model powers each agent
  2. Tool Access - What tools each agent can use
  3. Behavioral Guardrails - Non-negotiable rules for all agents
  4. MCP Tool Preferences - Domain-specific tool selection

Apply these constraints when spawning agents, checking permissions, or reviewing behavior.

Model Routing

AgentModelRationale
Senku (Planner)OpusStrategic planning needs deep reasoning
Riko (Explorer)OpusComplex exploration needs thorough analysis
Loid (Executor)SonnetBalanced speed and capability for implementation
Lawliet (Reviewer)SonnetFast iteration for review feedback loops
Alphonse (Verifier)SonnetQuick verification command execution
Speedwagon (Authoring)SonnetFast content authoring for explainer modules

Decision Rule:

  • Opus for strategic/planning tasks requiring deep reasoning
  • Sonnet for execution/verification tasks requiring speed

See Model Selection Guide for detailed criteria.

Tool Access Matrix

Riko (Explorer):       [Read] [Grep] [Glob] [Bash]* [WebSearch] [WebFetch]
Senku (Planner):       [Read] [Grep] [Glob] [TodoWrite]
Loid (Executor):       [Read] [Write] [Edit] [Bash] [Grep] [Glob]
Lawliet (Reviewer):    [Read] [Grep] [Glob] [Bash]
Alphonse (Verifier):   [Read] [Bash] [Grep]
Speedwagon (Authoring):[Read] [Grep] [Glob] [Write]† [Edit]† [Bash]‡

Key Restrictions:

  • Only Loid can modify files (Write, Edit) — except Speedwagon's scoped authoring exception
  • Only Riko can access web (WebSearch, WebFetch)
  • Only Senku can manage tasks (TodoWrite)

Footnotes:

    • Riko's Bash access is limited to AST analysis tools only (ast-grep, tree-sitter, language parsers)
  • † Speedwagon Write/Edit scoped to explain-out/ and .claude/explain-briefs/ only
  • ‡ Speedwagon Bash limited to bash scripts/compile-explain.sh only

See Tool Access Details for per-agent breakdowns.

External CLI dispatch — Codex co-review (Phase 4)

The orchestrator may invoke the OpenAI Codex CLI as an external Bash dispatch during Phase 4 (Review) when codex.available: true in orchestration state. This is the only sanctioned non-persona tool call from the orchestrator.

  • Codex receives: task description, Lawliet's full reply, git diff under review. See docs/guides/using-codex-review.md for the data boundary.
  • Codex runs in -s read-only --ignore-user-config sandbox.
  • Verdict-merge rules: commands/orchestrate.md Phase 4 (canonical).
  • Per-run opt-out: AGENT_FLOW_NO_CODEX=1 env var.

Personas (Riko/Senku/Loid/Lawliet/Alphonse) must NOT invoke codex directly.

Behavioral Guardrails

Universal Non-Negotiables

  1. Never speculate about unread code - Read files before making assertions
  2. Never suppress type errors - Fix root causes, not symptoms
  3. Prefer existing patterns - Follow the codebase's established style
  4. Avoid irreversible actions - Do not delete or force-push without confirmation
  5. Read before deciding - Gather context when uncertain
  6. Ask one targeted question - Only if truly blocked and cannot find answer in code

Agent-Specific Rules

AgentKey Constraints
RikoRead-only; summarize findings concisely
SenkuCreate actionable plans; estimate complexity
LoidRun tests after changes; follow the plan exactly
LawlietCite specific code; distinguish blockers from suggestions
AlphonseRun all verification commands; report exact output
SpeedwagonWrite only to explain-out/ and .claude/explain-briefs/; Bash only for compile-explain.sh

MCP Tool Preferences

Prefer MCP tools over shell commands for domain operations.

DomainPreferredFallback
GitHubgh CLI or MCPAPI calls
ObsidianMCP toolsFile operations
PlaywrightMCP tools-
DatabaseMCP toolsDirect SQL

See MCP Tool Guide for domain-specific guidance.

Quick Reference

Tool Access Check

ToolRikoSenkuLoidLawlietAlphonseSpeedwagon
ReadYesYesYesYesYesYes
GrepYesYesYesYesYesYes
GlobYesYesYesYes-Yes
Write--Yes--Scoped†
Edit--Yes--Scoped†
BashYes*-YesYesYesScoped‡
WebSearchYes-----
TodoWrite-Yes----

*Riko: Bash restricted to AST analysis tools only (ast-grep, tree-sitter, language parsers)

Violation Protocol

  1. Stop - Halt forbidden operation
  2. Document - Record what was blocked
  3. Delegate - Hand off to appropriate agent
  4. Continue - Proceed with permitted operations

Resources

Related Skills

Keep looking

Skills are one crate of 628,853. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.