mempenny: Command for Claude Code

MemPenny memory-apply — opencode host adapter

The user invoked this command with: $ARGUMENTS

Execute the MemPenny memory-apply flow. The canonical procedure lives in:

${MEMPENNY_ROOT}/commands/memory-apply.md

Read it first with the Read tool — it is the single source of truth for the backup-first discipline, the rollback-on-failure contract, the table-path validation (H3), and the filename-injection guard (H1). This file only describes the opencode host differences.

Apply these opencode host adaptations (override the source wherever they conflict):

A. Paths & environment

The env shim sets MEMPENNY_HOST=opencode, MEMPENNY_ROOT, MEMPENNY_DATA_DIR. Substitute:

• ${CLAUDE_PLUGIN_ROOT} → ${MEMPENNY_ROOT}
• ${CLAUDE_PLUGIN_DATA} → ${MEMPENNY_DATA_DIR}
• ${CLAUDE_PROJECT_DIR} → the current working directory

B. Config path (shared with Claude Code)

1. ${MEMPENNY_CONFIG_PATH} if set.
2. Else ~/.claude/mempenny.config.json if ~/.claude/ exists.
3. Else ~/.config/opencode/mempenny.config.json.

C. Subagents (opencode Task tool)

Where the source spawns subagent_type: general-purpose for the isolated apply, use opencode's subagent_type: "general" (write-capable), lowercase, with description + prompt. The "separately-spawned subagent with no memory of the proposal step" property is load-bearing (it is what makes apply safe against prompt injection in the source content) — preserve it: spawn a fresh subagent, do not apply inline in the triage context.

D. Command namespace

Sibling commands use the hyphen namespace.

apply-specific notes

• The table path is the first positional arg (required); --dir / --lang parse from $ARGUMENTS.
• The H3 path validation (C1 regex + realpath + exists + not-a-symlink) and the F-M1 permission sanity checks carry over verbatim — re-implement them in the apply context exactly as written.