---

name: opencode-runner description: "Run coding tasks via opencode using free cloud models. Use when asked to offload work to opencode.ai or run a free model. Don't use for local models (Ollama, LM Studio), Claude/OpenAI calls, or when Claude should do the work itself." license: MIT effort: medium metadata: version: 1.6.0 author: "Luong NGUYEN luongnv89@gmail.com"

OpenCode Runner

Delegate coding tasks to opencode using free models — zero cost, fully automated.

OpenCode (opencode.ai) is a terminal AI coding assistant that supports multiple providers and models. This skill automates the process of selecting the best available free model, launching the task, and reporting progress back to you.

Prerequisites

• opencode installed: which opencode must succeed; install via npm i -g opencode-ai@latest or brew install opencode if missing
• Internet access: required for cloud model selection and task execution via OpenCode Zen
• Project context: the current working directory should be the project root for file-context tasks

Critical Rules

1. Never do the task yourself. This skill exists solely to delegate work to opencode. If opencode is not installed, fails to run, or no free cloud model is available — report the problem to the user and stop. Do not fall back to editing files directly, writing code yourself, or using any other tool to accomplish the user's coding task. The whole point is that opencode does the work.

2. Only select cloud models. Never select local models (e.g., ollama/*, lmstudio/*, or any model running on localhost). Only select models from the opencode/* provider namespace, which are cloud-hosted on OpenCode Zen. Local models have unpredictable availability, performance, and may not support the tool-use capabilities opencode needs.

3. Always clean up after yourself. opencode spawns background processes (LSP servers, MCP servers, node workers) that persist after the task finishes. Every execution path — success, failure, error, timeout — must end with the cleanup steps in Phase 6. Orphaned opencode processes silently eat CPU and memory, and users won't notice until their machine slows to a crawl.

Repo Sync Before Edits (mandatory)

Before creating/updating/deleting files in an existing repository, sync the current branch with remote:

branch="$(git rev-parse --abbrev-ref HEAD)"
git fetch origin
git pull --rebase origin "$branch"

If the working tree is not clean, stash first, sync, then restore:

git stash push -u -m "pre-sync"
branch="$(git rev-parse --abbrev-ref HEAD)"
git fetch origin && git pull --rebase origin "$branch"
git stash pop

If origin is missing or conflicts occur, stop and ask the user before continuing.

Phase 1: Verify Installation

Check that opencode is installed and at the latest version.

Step 1: Check if installed

which opencode && opencode --version

If opencode is not found, tell the user:

opencode is not installed. Install it with one of these commands:

curl -fsSL https://opencode.ai/install | bash

or

npm i -g opencode-ai@latest

or (macOS)

brew install opencode

Then stop completely with BLOCKED — opencode is not installed (Final Report) — do not proceed to any other phase, do not attempt the task yourself, do not edit any files. Wait for the user to install opencode and re-invoke this skill.

Step 2: Check for updates

1. Run opencode upgrade. It upgrades to the latest version, or confirms the installed one is current.
2. If opencode upgrade exits non-zero, tell the user its error and suggest running opencode upgrade manually. Then continue to item 3.
3. Run opencode --version again.
4. If opencode --version exits non-zero, stop with BLOCKED — opencode does not start after the upgrade. Do not continue to Phase 2.

Phase 2: Discover Free Models & Let User Pick

Query the available models, present the free ones to the user, and let them choose. If the user doesn't choose, fall back to the priority order.

opencode models 2>/dev/null

Free model priority list (default order if user defers)

Priority	Model ID	Notes
1	opencode/deepseek-v4-flash-free	Strong recent free coding model — preferred default
2	opencode/minimax-m2.5-free	MiniMax free tier — good general-purpose
3	opencode/nemotron-3-super-free	NVIDIA Nemotron free tier
4	opencode/big-pickle	Free fallback
5	opencode/gpt-5-nano	Small fallback; free only when opencode models prices it at $0

Model IDs evolve. A model is free-tier eligible when its ID is in the opencode/* namespace and one of these holds: the ID ends in -free, the ID is opencode/big-pickle, or the opencode models output prices it at $0. Match by suffix, not by exact ID. Without a $0 price in the output, leave opencode/gpt-5-nano out.

Selection logic

1. Run opencode models and collect all entries.

2. Filter out all non-opencode/* models — ignore anything from ollama/*, lmstudio/*, nvidia/*, or any other namespace. Only cloud-hosted opencode/* models qualify.

3. Among the remaining list, keep only the free-tier eligible models.

4. Present the free models to the user as numbered options, in priority order, with priority 1 marked as the default. Use the <options> format if possible. Example:

   I found these free cloud models available via opencode. Pick one, or accept the default.

   1. opencode/deepseek-v4-flash-free (default — priority 1)
   2. opencode/minimax-m2.5-free
   3. opencode/nemotron-3-super-free
   4. opencode/big-pickle

5. If the user names a model, use that one. If the user says "default", "you pick", "any", or doesn't specify, use priority 1 (the highest-priority available free model).

6. If no free-tier eligible model remains, stop with BLOCKED — no free cloud model available — do not fall back to local models, paid models, or doing the task yourself.

Privacy note (always show this with the model list): Free models on OpenCode Zen may use collected data for model improvement.

Phase 3: Confirm Before Executing

Before invoking opencode, show the user a one-screen summary and get explicit confirmation. This catches wrong-model or wrong-prompt mistakes before any tokens are burned.

Present this block:

Ready to delegate to opencode

• Model: opencode/deepseek-v4-flash-free (free tier)
• Working directory: /Users/.../current-project
• Context files: path/to/foo.py, path/to/bar.py (or "none")
• Prompt: (quote the prompt verbatim, multi-line OK)
• Estimated duration: unknown — opencode is non-deterministic; cleanup runs even on timeout

Confirm to proceed, or tell me what to change (model, prompt, files).

Then offer the user two <options>: "Proceed" and "Change something". Wait for confirmation. Do not invoke opencode run until the user confirms.

If the user asks to change anything (different model, edit prompt, add/remove context files), loop back: update the field, re-show the summary, and ask again. If the user cancels instead, stop with BLOCKED — the user cancelled before the run.

Phase 4: Execute the Task

Run the coding task with the confirmed model. Always run in the background with output redirected to a log file — this is required for the low-token monitoring strategy in Phase 5.

RUN=/tmp/opencode-run-$(date +%s)
( opencode run -m "[confirmed-model-id]" "[confirmed prompt]" > "$RUN.log" 2>&1; echo $? > "$RUN.exit" ) &
echo $! > "$RUN.pid"
echo "opencode started: run=$RUN pid=$(cat "$RUN.pid")"

Record the run= value this prints and substitute it literally in every later Phase 5 and Phase 6 shell. Each Bash call starts fresh, so rebind RUN to that recorded prefix before reading "$RUN.log", "$RUN.pid" or "$RUN.exit". The .pid file is the ownership record: it holds the id of the wrapper subshell, whose child is opencode run. The wrapper writes opencode's exit code to .exit when opencode finishes, because a later shell cannot wait on it.

Handling multi-line or complex prompts

For tasks that reference files or need detailed context, use the --file flag:

RUN=/tmp/opencode-run-$(date +%s)
( opencode run -m "[confirmed-model-id]" --file path/to/relevant-file.py "[task description]" > "$RUN.log" 2>&1; echo $? > "$RUN.exit" ) &
echo $! > "$RUN.pid"
echo "opencode started: run=$RUN pid=$(cat "$RUN.pid")"

Foreground execution is discouraged — streaming the full opencode output back into your context wastes tokens. The Phase 5 monitor reads only the log tail.

Phase 5: Monitor with Minimum Tokens

opencode output is verbose. Streaming the full log back into your context is expensive — a single long run can easily push past 10k tokens of stream chatter. Use the lightweight polling protocol below instead.

Polling protocol

Run one tiny status command per check. It returns at most ~200 bytes — enough to know status, elapsed time, and the latest activity line — without ingesting the whole log.

RUN=/tmp/opencode-run-1234567890   # paste the run= value Phase 4 printed
status() {
  pid=$(cat "$RUN.pid" 2>/dev/null)
  if [ -n "$pid" ] && kill -0 "$pid" 2>/dev/null; then s=running; else s=done; fi
  ec=$(cat "$RUN.exit" 2>/dev/null)
  bytes=$(wc -c < "$RUN.log" 2>/dev/null || echo 0)
  last=$(tail -n 1 "$RUN.log" 2>/dev/null | tr -d '\r' | cut -c1-160)
  printf 'status=%s pid=%s exit=%s bytes=%s last=%q\n' "$s" "${pid:-none}" "${ec:-none}" "$bytes" "$last"
}
status

That single line is your full progress sample. Do not tail -n 50, do not cat "$RUN.log", do not stream stdout — those defeat the purpose.

Cadence (mandatory)

• Wait 30 seconds before polls 1 and 2. Wait 60–120 seconds before each later poll.
• Hard cap: 6 polls total per run. If poll 6 still shows status=running, ask the user whether to keep waiting or kill the process. Do not poll again until the user answers. If the user says keep waiting, allow up to 6 more polls under the same rules.
• A run is stalled when bytes= is the same in two consecutive polls. A run has timed out when 5 minutes pass with no bytes= growth. On a stall or a timeout, ask the user whether to keep waiting or kill the process. Kill it only when the user approves, then follow On error, stall, or timeout.

What to report between polls

One short line per check, derived from the status() output:

Poll 2 (t+60s): running, 4.2 KB written, last: "Editing src/foo.py …"

Do not paste the raw log. Do not summarize what opencode is "thinking" — you can't tell from a tail line. Stick to: status, elapsed, byte growth, last line.

On completion

When status=done:

1. Read the tail only, not the whole log: tail -n 40 "$RUN.log". That's the summary opencode prints at the end (files changed, tokens used, errors).
2. Classify the run. exit=0 is a success. Any other exit= value is a failure. exit=none means the wrapper was killed before opencode finished: treat it as a failure.
3. If the working directory is a git repository, run git status --porcelain to list the files that changed. This is a read only: never edit, stage, or revert those files.
4. Do not paste the full log. If the user asks to keep it, skip deleting "$RUN.log" in Phase 6 Step 3 and give its path in the final report.
5. If the task failed, suggest retrying with the next free model from the Phase 2 list.
6. Run Phase 6 cleanup — mandatory, even on success.
7. Print the final report (Final Report).

On error, stall, or timeout

1. Report the error or stall to the user (one line, derived from the last poll).
2. Suggest retrying with the next free cloud model.
3. If all free models have been tried, suggest the user run opencode auth list to check provider auth.
4. Never attempt the task yourself as a fallback.
5. Run Phase 6 cleanup — always.
6. Print the final report (Final Report).

Phase 6: Cleanup (mandatory)

Every execution — success, failure, error, or timeout — must end with cleanup. opencode spawns child processes (LSP servers, MCP servers, node workers) that persist after the main process exits. Without cleanup, these orphaned processes accumulate and drain system resources.

Step 1: Kill the opencode process tree

If you launched opencode in the background, use the pidfile for the recorded run prefix as the ownership record:

RUN=/tmp/opencode-run-1234567890   # paste the run= value Phase 4 printed
pid=$(cat "$RUN.pid" 2>/dev/null)
if [ -n "$pid" ]; then
  pkill -TERM -P "$pid" 2>/dev/null   # opencode runs as a child of the wrapper subshell
  kill "$pid" 2>/dev/null
  sleep 2
  pkill -KILL -P "$pid" 2>/dev/null
  kill -9 "$pid" 2>/dev/null
fi

Step 2: Find and kill orphaned opencode processes

1. Run pgrep -fl "opencode run" to list lingering opencode run processes.
2. If it lists any process, run pkill -f "opencode run".

This pattern never matches the user's interactive TUI session (opencode without a subcommand). Leave that session alone.

Step 3: Clean up temp files

rm -f "$RUN.log" "$RUN.pid" "$RUN.exit" 2>/dev/null

If the user asked to keep the log (Phase 5, On completion item 4), drop "$RUN.log" from this command.

Step 4: Confirm cleanup

1. Run pgrep -f "opencode run" >/dev/null && echo cleanup=incomplete || echo cleanup=complete.
2. If it prints cleanup=complete, report: Cleanup complete — no opencode run process remains.
3. If it prints cleanup=incomplete, warn: Warning: some opencode run processes are still running. Run pkill -f "opencode run" manually to clean up.

Final Report

After Phase 6, print one final report in concise text, in this order: Result: (COMPLETE, PARTIAL — <reason>, or BLOCKED — <reason>), Evidence: (only checks that ran), Uncertainty:, then Decision: (the approval needed, or No approval needed.). A stop before opencode run starts skips Phases 4–6 and prints the report directly. Status rules, examples and the reader checks: references/final-report.md.

Expected Output

See references/expected-output.md for the full set of example blocks the user should see at each phase (model picker, confirmation summary, low-token progress polls, cleanup confirmation). On error or timeout, the cleanup confirmation still runs. Final report examples: references/final-report.md.

---

Edge Cases

When an edge case fires, read references/edge-cases.md: opencode missing, a failed upgrade, no free model, every model failing, a stall or timeout, a rejected confirmation, an open TUI session, multi-line prompts, a non-zero exit, or a killed wrapper. It also holds the opencode command reference.

---

Acceptance Criteria

The skill run is considered successful when all of the following are verifiable:

• Installation verified — Phase 1 confirms opencode is on PATH and reports its version before any other action.
• Only cloud opencode/* models considered — Filter step rejects ollama/*, lmstudio/*, nvidia/*, and any other namespace. No local model is ever selected.
• Free models presented to the user — Phase 2 lists every available free opencode/* model with priority 1 as default, plus the privacy note. The user may pick one; if they defer, priority 1 is used.
• Confirmation captured before execution — Phase 3 shows model + cwd + context files + prompt, and waits for explicit user confirmation before invoking opencode run.
• Task delegated to opencode — The coding task is executed via opencode run, not by the skill editing files directly or writing code itself.
• Low-token monitoring used — Phase 5 polls via the single-line status() helper. The raw log is never streamed back; only the last line, byte count, and status are reported per poll. Max 6 polls per run.
• Stall/timeout handled — No byte growth across two consecutive polls, or 5 minutes with no byte growth, triggers a confirmation with the user and (if approved) a kill + cleanup.
• Completion summary delivered — On status=done, the skill reads only tail -n 40 "$RUN.log", classifies the run by its exit= value, and summarizes files changed and token usage.
• Cleanup runs on every exit path — Phase 6 runs whether the task succeeded, failed, errored, stalled, or timed out, and Step 4 prints cleanup=complete before the report claims cleanup.
• Temp files removed — "$RUN.log", "$RUN.pid" and "$RUN.exit" are deleted during cleanup, unless the user asked to keep the log.
• No fallback to self — If opencode is unavailable, the user rejects the confirmation, or all free models fail, the skill stops. It never falls back to editing files or writing code itself.
• Final report delivered — every run, stops included, ends with Result:, Evidence:, Uncertainty: and Decision:, and passes the four reader checks in references/final-report.md: the result is findable in the first line, facts are separated from assumptions, each claim traces to a check that ran, and the next decision is named. Human understanding stays unconfirmed until a user answers those checks.

---

Step Completion Reports

After each phase, emit a compact status block so pass/fail is scannable:

◆ [Step Name] ([step N of M] — [context])
··································································
  [Check 1]:          √ pass
  [Check 2]:          × fail — [reason]
  [Criteria]:         √ N/M met
  ____________________________
  Result:             PASS | FAIL | PARTIAL

Use √ for pass, × for fail, and — to add brief context. Per-phase example blocks (Installation, Model Discovery, Confirmation, Execution, Monitor, Cleanup) are in references/expected-output.md.