shell-command（外壳命令）

Shell Non-Interactive Strategy (Global)

Context: OpenCode's shell environment is strictly non-interactive. It lacks a TTY/PTY, meaning any command that waits for user input, confirmation, or launches a UI (editor/pager) will hang indefinitely and timeout.

Goal: Achieve parity with Claude Code's shell capabilities through internalized knowledge of non-interactive flags and environment variables.

Cognitive & Behavioral Standards

To match the high-agency, autonomous capabilities of advanced models (like Claude Sonnet), this strategy enforces strict cognitive patterns. These are not just shell tips; they are behavioral requirements for success in this environment.

Goal: Eliminate "human-in-the-loop" dependency during task execution.

Key Behaviors:

1. Process Continuity (Turn-Taking):

- Rule: Never stop after a tool output to "wait for instructions" unless the task is complete. - Why: The environment is non-interactive. You must drive the workflow. - Mechanism: Commands expecting input MUST use timeouts or explicit "yes" pipes.

1. Explicit Action Framing (Positive Constraints):

- Rule: Follow "GOOD" (positive) instructions, ignore "BAD" (negative) assumptions. - Why: Models follow explicit directives ("Use -y") better than prohibitions ("Don't prompt"). - Mechanism: Always preemptively supply non-interactive flags.

1. Environment Rigor (Context Awareness):

- Rule: Assume a headless CI environment where any prompt = failure. - Why: There is no TTY. "Asking the user" via a shell prompt causes a hang. - Mechanism: Strictly avoid editors, pagers, and interactive modes.

Core Mandates

1. Assume CI=true: Act as if running in a headless CI/CD pipeline.
2. No Editors/Pagers: vim, nano, less, more, man are BANNED.
3. Force & Yes: Always preemptively supply "yes" or "force" flags.
4. Use Tools: Prefer Read/Write/Edit tools over shell manipulation (sed, echo, cat).
5. No Interactive Modes: Never use -i or -p flags that require user input.

Environment Variables (Auto-Set)

These environment variables help prevent interactive prompts:

Command Reference

Package Managers

Git Operations

System & Files

Docker

Python/Node REPLs

Banned Commands (Will Always Hang)

These commands will hang indefinitely - never use them:

• Editors: vim, vi, nano, emacs, pico, ed
• Pagers: less, more, most, pg
• Manual pages: man
• Interactive git: git add -p, git rebase -i, git commit (without -m)
• REPLs: python, node, irb, ghci (without script/command)
• Interactive shells: bash -i, zsh -i

Handling Prompts

When a command doesn't have a non-interactive flag:

The "Yes" Pipe

yes | ./install_script.sh

Heredoc Input

./configure.sh <<EOF
option1
option2
EOF

Echo Pipe

echo "password" | sudo -S command

Timeout Wrapper (last resort)

timeout 30 ./potentially_hanging_script.sh || echo "Timed out"

Alternatively, use a subshell to start the hanging command and kill it after completion (e.g., for starting a dev server).

Best Practices

1. Always test commands mentally for interactive prompts before running
2. Check man pages (via web search) for -y, --yes, --non-interactive, -f, --force flags
3. Use --help to discover non-interactive options: cmd --help | grep -i "non-interactive\|force\|yes"
4. Prefer OpenCode tools over shell commands for file operations
5. Set timeout for any command that might unexpectedly prompt

Advanced Instruction Patterns (Cognitive Optimization)

The Problem: Implicit Constraints

Large Language Models (LLMs) often struggle with:

1. Negative constraints: Inverting or ignoring "don't do X" instructions.
2. Turn termination: Stopping after tool execution instead of auto-continuing.
3. Context weighting: Failing to prioritize authoritative instructions over general knowledge.

Strategy 1: Explicit Action Framing (BAD vs GOOD)

This plugin uses the BAD vs GOOD pattern to enforce positive constraints. Instead of saying "Don't use interactive flags", we provide a concrete "Good" alternative.

Why it works:

• "BAD: npm init" → Model identifies the failure pattern.
• "GOOD: npm init -y" → Model receives a specific, executable instruction.
• Result: Reduces hallucination of interactive commands by providing a verified substitute.

Strategy 2: Process Continuity

In non-interactive environments, the agent must drive the process forward.

The Rule: Never stop after a tool execution unless the task is complete.

Pattern:

1. Execute command (e.g., git status)
2. Analyze output
3. Explicitly state next step: "Status is clean. Next: I will run tests."
4. Execute next step immediately

Strategy 3: Context Hierarchy

When instructions conflict (e.g., generic docs vs this specific strategy), establish precedence:

1. Cite the Authority: "Per shell_strategy.md..."
2. Follow the Specifics: Rules in this file override general model training or other documentation.

Strategy 4: Applying These Patterns Beyond Shell

The cognitive strategies used here (Explicit Action Framing) apply to all coding tasks:

Instead of:

Do not use logging.getLogger()
Don't create CLI code here

Use:

ALWAYS USE: config.logging_config.get_logger()
USE THIS REPO FOR: API backend only

By framing instructions as "Actionable Positive Constraints", you reduce hallucination and improve compliance across all models.