Workspace Init

Debug Prefix Protocol

• When called, print this first: [Skills: workspace-init] Follow: <rules / context constraints that really apply now>; Current step: <one-line note>.
• If this skill clearly depends on one rule, add one more line like [Rules: <rule>] ..., then continue with the real work.
• The debug prefix should stay stable, short, and easy to grep.

Flow

• Run the init script first to initialize all runtime files in one shot:

  bash .opencode/skills/workspace-init/init.sh
  

  This handles git init, trace file creation, and all memory file creation from templates. It never overwrites existing files. Cross-platform: works on macOS (bash/zsh), Linux (bash/zsh), and Windows (Git Bash).
• Read possible project-note files in the root directory, such as README.md, task notes, and existing .opencode/trace.md, to align project background and context.
• Infer the user interaction language from the current task, and reply in that language by default, unless the task contract requires another delivery language.
• Separate interface mode from execution strategy: run / tui are interface modes, while autonomous / interactive are execution strategies. If there is no stronger signal, the default execution strategy is autonomous; run + interactive is not a supported default path.
• If .opencode/trace.md or any .opencode/memory/*.md already exists, read its current contents before any later write, append, or overwrite. Do not treat runtime files like empty templates that may be rebuilt blindly.
• After .opencode/memory/initial-prompt.md is initialized, the main agent must write the original user prompt, earliest task goal, explicit metrics, and non-goals for this round into it at once as an append-only baseline. Later it may only append clarifications, not overwrite the original prompt block.
• After .opencode/memory/context-snapshot.md is initialized, the main agent must refresh the first snapshot before workspace-init closes: at minimum it should write task anchors, current phase, explicit constraints, the most obvious gaps, and the default next-dispatch focus. This is not an optional summary. It is the canonical entry point for later multi-agent compaction.
• Remind the main agent: local git is for detail-level checkpoints, while memory / trace are for abstract judgments and reviews. They should be maintained in parallel, not used as substitutes for each other.
• Remind the main agent: after the workspace-init baseline is established, it should create the first local git commit soon, then keep committing again at plan switches, key milestones, major rollback points, and the final stable pre-delivery state.
• If the current task will clearly enter a multi-agent / multi-stage / harness flow, also remind the main agent to read .opencode/protocols/ and .opencode/evals/, especially dispatch, compaction, memory-sync, and runtime-events, so it does not initialize memory only and miss runtime protocols or eval contracts.
• If the current task may involve Python, do one round of environment detection before asking the user anything:
  • whether Python is actually needed for this task
  • whether a reusable workspace-root .venv already exists
  • whether the project already declares stronger environment constraints, such as .python-version, pyproject.toml, environment.yml, requirements*.txt, or uv.lock
  • whether multiple equally reasonable environment candidates still remain after local detection
• The default Python environment priority is fixed as: reuse existing .venv > follow project constraints > create a workspace-root .venv.
• If the current execution strategy is autonomous and the environment detection found no real conflict, follow that default priority directly instead of asking the user about interpreter preferences.
• If the current execution strategy is interactive, still do the same local detection first; only when multiple equally reasonable environment candidates remain and they would materially change the later path may this fork be escalated back to the main agent for a short user confirmation.
• Remind the main agent that interactive does not mean asking on every step. It only means exposing real forks to the user after local detection has already narrowed the space.
• Make clear the current known goal, available inputs, key constraints, and the most obvious gaps.

Output Format

• Read materials
• Initialized files
• Git initialization status
• Initial prompt baseline status
• Context-snapshot status
• Current language
• Execution-strategy status
• Python-needed judgment
• Python-environment detection
• Python-environment default recommendation
• Known goals and gaps