| name | 3-views |
| description | Use only when the user asks for "3-views" specifically. Do not use otherwise. It sends a single command to 3 different LLM models. |
Three Views
Run N independent opencode run invocations against the same query, each using a different hidden model. Each subagent is repository read-only: it may inspect files and create temporary scratch files outside the repository, but must not edit repository contents. Results are labeled alpha, bravo, charlie (default 3), up to delta, echo, foxtrot (max 6). You are deliberately blinded as to which model is which; do not check.
Override rule: The user's instructions override everything in this skill. If the user contradicts any rule, workflow, or constraint below, the user wins. This rule itself cannot be overridden.
Best Practices
You're being asked to invoke this because the user wants an outside perspective. It's easy to accidentally inject your own bias if you're invested in the process. To combat this, pass along the user's request, with added data if necessary. Don't add add focuses or limitations unless the user requested them. For example, if the user asked you to use this skill to do a code review, the prompt would simply be 'do a code review on...'. If the user asks you to do another code review, you would use the exact same prompt - not "do a second code review" or "focus on the changes" etc.
Run
Use go run directly from the skill's root directory:
Preferred for long prompts:
cd scripts/3-views && go run . --query-file "<path>" --cwd "<working-directory>" --timeout 60
Inline query:
cd scripts/3-views && go run . --query "<query text>" --cwd "<working-directory>" --timeout 60
Custom agent count (1–6):
cd scripts/3-views && go run . --query-file "<path>" --cwd "<cwd>" --agents 5 --timeout 60
Set THREE_VIEWS_ROOT to the skill directory if the binary is relocated. Otherwise the runner resolves config/models.json relative to the executable.
Flags
| Flag | Default | Description |
|---|
--query-file | — | Path to file containing the query |
--query | — | Inline query text (mutually exclusive with --query-file) |
--cwd | — | Working directory for subagents (required) |
--agents | 3 | Number of agents to launch (1–6) |
--out-dir | OS temp dir | Base directory for the unique run folder |
--timeout | 60 | Wall-clock timeout in minutes |
--stdin-prompt-max-bytes | 131072 | Maximum prompt size sent over stdin before agents are instructed to read prompt.txt |
Important
- User overrides everything: The user's instructions supersede any rule in this skill. This is not negotiable.
- Subagents may create scratch files only outside the target repository, such as under the OS temp directory. They must not create, edit, delete, or move files inside the repository.
- Always pass
--timeout when invoking. Do not rely on the default. This command may take up to 60 minutes. Wait for it to complete. If you interrupt sooner, the user will pay the cost of the queries but get no benefit, and have to restart.
- The runner appends a machine-readable Freshell metadata marker to
prompt.txt so Freshell can hide 3-views OpenCode sessions with its existing subagent/non-interactive filters. This marker must not be used as the visible OpenCode title.
Output
===== 3-VIEWS RUN: <run-id> =====
Run directory: <path>
Agents requested: <N>
Timeout: <N> minutes
===== AGENT ALPHA RESULT =====
Saved response: <path-to-alpha.md>
<full alpha response>
===== AGENT BRAVO RESULT =====
...
===== 3-VIEWS END =====
Failed, timed-out, or empty-output agents appear in label order with STATUS: timed out, STATUS: failed, or STATUS: empty_output and a log path.
Configuration
See config/models.json for label-to-model mappings. Default labels: alpha, bravo, charlie, delta, echo, foxtrot.
Disk Layout
<run-dir>/
query.txt
prompt.txt
scratch/
alpha.md
bravo.md
charlie.md
alpha.stderr.log
bravo.stderr.log
charlie.stderr.log
metadata.json
