lemming-cli 0.4.1__tar.gz → 0.4.3__tar.gz

lemming_cli-0.4.3/.dockerignore

@@ -0,0 +1,30 @@
+# The image builds its own environments, so host-built ones are not just
+# wasted bytes in the context — they are the wrong platform's binaries
+# landing on top of what `uv sync` and `npm install` are about to create.
+.venv
+node_modules
+
+# History and build products the image never reads.
+.git
+.github
+dist
+test-results
+docs/screenshots
+
+# Agent worktrees are whole checkouts of this repo, each with its own
+# dependencies installed, and they dwarf the source they sit beside.
+.claude
+
+# Local state and caches.
+.pytest_cache
+.ruff_cache
+__pycache__
+*.pyc
+.DS_Store
+
+# Roadmaps and their locks belong to whoever is running lemming, not to
+# the image; a stale one baked in would be read as that project's state.
+tasks.yml
+*.lock
+!uv.lock
+!package-lock.json

{lemming_cli-0.4.1 → lemming_cli-0.4.3}/.github/workflows/ci.yml

@@ -19,7 +19,7 @@ jobs:
           python-version: '3.14'
 
       - name: Install Python dependencies
-        run: uv sync
+        run: uv sync --locked
 
       - name: Run Python linting
         run: uv run ruff check .

{lemming_cli-0.4.1 → lemming_cli-0.4.3}/Dockerfile

@@ -11,7 +11,8 @@ RUN curl -fsSL https://deb.nodesource.com/setup_24.x | bash - \
     && rm -rf /var/lib/apt/lists/*
 
 # Install AI coding CLIs (antigravity ships a native binary, not an npm package)
-RUN npm install -g @openai/codex @anthropic-ai/claude-code \
+RUN npm install -g --allow-scripts=opencode-ai \
+      @openai/codex @anthropic-ai/claude-code opencode-ai \
     && curl -fsSL https://antigravity.google/cli/install.sh | bash -s -- --dir /usr/local/bin
 
 # Copy cloudflared and uv binaries directly from their official images
@@ -22,8 +23,14 @@ WORKDIR /opt/lemming
 COPY . .
 RUN uv sync --no-dev --frozen
 
-# Install frontend dependencies and Playwright browsers
-RUN npm install && npx playwright install --with-deps
+# Frontend dependencies are always needed: postinstall vendors mancha into
+# the web assets. The browsers are not, and they are more than half the
+# image, so anything that will never drive one builds without them.
+ARG INSTALL_BROWSERS=true
+RUN npm install \
+    && if [ "$INSTALL_BROWSERS" = "true" ]; then \
+         npx playwright install --with-deps; \
+       fi
 
 # Pre-create the lemming home directory so users can bind-mount a .env file into it
 RUN mkdir -p /root/.local/lemming

lemming_cli-0.4.1/README.md → lemming_cli-0.4.3/PKG-INFO

@@ -1,3 +1,25 @@
+Metadata-Version: 2.5
+Name: lemming-cli
+Version: 0.4.3
+Summary: An autonomous, iterative task runner for AI coding agents.
+Project-URL: Homepage, https://github.com/owahltinez/lemming
+Project-URL: Repository, https://github.com/owahltinez/lemming
+Author: Lemming Contributors
+License: MIT
+License-File: LICENSE
+Keywords: agents,ai,antigravity,automation,claude,opencode
+Requires-Python: >=3.12
+Requires-Dist: beautifulsoup4>=4.13.3
+Requires-Dist: click>=8.0.0
+Requires-Dist: fastapi>=0.124.4
+Requires-Dist: markdownify>=0.14.1
+Requires-Dist: pydantic>=2.12.5
+Requires-Dist: pyyaml>=6.0.3
+Requires-Dist: readability-cli>=0.7.0
+Requires-Dist: requests>=2.32.3
+Requires-Dist: uvicorn>=0.33.0
+Description-Content-Type: text/markdown
+
 # Lemming 🐹
 
 **The transparent, tool-agnostic orchestrator for autonomous AI coding agents.**
@@ -16,8 +38,8 @@ human-readable `tasks.yml` file.
   update is recorded in a human-readable `tasks.yml` file. You can step in,
   adjust the roadmap, or swap agents at any time.
 - **Tool Agnostic**: Lemming doesn't care which agent you use. It works
-  out-of-the-box with `agy`, `aider`, `claude`, `codex`, or even your own custom
-  scripts.
+  out-of-the-box with `agy`, `opencode`, `claude`, `codex`, or even your own
+  custom scripts.
 - **Resilient Execution**: With built-in heartbeat monitoring, automatic
   retries, and progress tracking, Lemming handles process crashes and rate
   limits gracefully.
@@ -71,7 +93,7 @@ Start the autonomous loop.
 
 ```bash
 # Run using the project's configured agent (defaults to the first
-# supported runner found on PATH: agy, aider, claude, or codex)
+# supported runner found on PATH: agy, opencode, claude, or codex)
 lemming run
 
 # Flags passed after -- are sent directly to the underlying runner
@@ -118,22 +140,25 @@ removed unless it failed — in which case it is kept, and its path printed, so
 the log can be read. Kept directories live in `~/.local/lemming/exec-*` and
 are retired automatically a week later, so a recent failure is always still
 there to inspect. Stdout carries the agent's message alone and everything
-else goes to stderr, so the output can be consumed directly.
+else goes to stderr, so the output can be consumed directly. Stderr reports
+the ephemeral tasks file and task ID at startup; pass those to `lemming
+--tasks-file <path> status <id>` or `logs <id>` to inspect an active run. Use
+the global verbose flag (`lemming -v exec ...`) to stream runner activity.
 
 Note that the agent runs unattended (`--yolo` by default), so it does not
 inherit the permission prompts of whatever launched it.
 
 ### Teaching Your Agent to Use It
 
-Lemming ships an Agent Skill so that an agent already running — Claude Code,
-Gemini CLI, Cursor — can reach any other agent CLI without knowing how each
-one spells its flags. Install it once:
+Lemming ships an Agent Skill so that an agent already running — Codex, Claude
+Code, Gemini CLI, Cursor — can reach any other agent CLI without knowing how
+each one spells its flags. Install it once:
 
 ```bash
 # Cross-tool location (~/.agents/skills), and report any tool-specific ones
 lemming skill install
 
-# Cover every agent tool detected on this machine
+# Cover detected tools that require their own skill directory too
 lemming skill install --all
 
 # This repository only
@@ -268,7 +293,7 @@ lemming hooks list
 lemming config list
 
 # Persist configuration to tasks.yml
-lemming config set runner aider
+lemming config set runner opencode
 ```
 
 ### Evaluating Prompt Changes
@@ -373,8 +398,9 @@ These come before the subcommand and apply to all of them.
     per-task `--runner`/`--model` overrides anything passed here.
 - **`exec [<description>]`**: Run one task, or one set of reviews, outside any
   roadmap. Prints the agent's closing message to stdout and everything else to
-  stderr. Exits non-zero if the task did not complete. See
-  [One-off tasks](#one-off-tasks-without-a-roadmap).
+  stderr, including its tasks file and task ID at startup. Use `lemming -v
+  exec ...` to stream runner activity. Exits non-zero if the task did not
+  complete. See [One-off tasks](#one-off-tasks-without-a-roadmap).
   - `-f/--file`: Read the description from a file, or `-` for stdin. Unlike
     `add`, there is no length cap.
   - `--review <names>`: Reviews to run after the task, comma-separated or
@@ -449,7 +475,10 @@ Lemming uses **fuzzy matching** to automatically inject the correct "YOLO"
 
 - **Antigravity (`agy`)**: Adds `--dangerously-skip-permissions` and exposes
   the project workspace with `--add-dir`
-- **Aider**: Adds `--yes --quiet`
+- **OpenCode**: Runs non-interactively via `opencode run`, adds `--format json`,
+  and, in YOLO mode, adds `--auto`. For its Google provider, Lemming makes an
+  existing `GOOGLE_API_KEY` or `GEMINI_API_KEY` available under OpenCode's
+  native `GOOGLE_GENERATIVE_AI_API_KEY` name.
 - **Claude**: Adds `--dangerously-skip-permissions`
 - **Codex**: Runs non-interactively via `codex exec`, adds `--json`, and, in
   YOLO mode, adds `--dangerously-bypass-approvals-and-sandbox`
@@ -511,10 +540,16 @@ pushing a `v*` tag triggers the `publish.yml` GitHub Actions workflow, which
 builds the package with `uv build` and uploads it.
 
 ```bash
-# 1. Bump the version in pyproject.toml, commit, and push
-# 2. Tag the release and push the tag
-git tag v0.1.1
-git push origin v0.1.1
+# 1. Bump pyproject.toml and uv.lock together, commit, and push
+uv version --bump patch --no-sync
+version="$(uv version --short)"
+git add pyproject.toml uv.lock
+git commit -m "Bump version to $version"
+git push origin main
+
+# 2. After CI passes, tag the release and push the tag
+git tag "v$version"
+git push origin "v$version"
 ```
 
 ---

lemming_cli-0.4.1/PKG-INFO → lemming_cli-0.4.3/README.md

@@ -1,25 +1,3 @@
-Metadata-Version: 2.5
-Name: lemming-cli
-Version: 0.4.1
-Summary: An autonomous, iterative task runner for AI coding agents.
-Project-URL: Homepage, https://github.com/owahltinez/lemming
-Project-URL: Repository, https://github.com/owahltinez/lemming
-Author: Lemming Contributors
-License: MIT
-License-File: LICENSE
-Keywords: agents,ai,aider,antigravity,automation,claude
-Requires-Python: >=3.12
-Requires-Dist: beautifulsoup4>=4.13.3
-Requires-Dist: click>=8.0.0
-Requires-Dist: fastapi>=0.124.4
-Requires-Dist: markdownify>=0.14.1
-Requires-Dist: pydantic>=2.12.5
-Requires-Dist: pyyaml>=6.0.3
-Requires-Dist: readability-cli>=0.7.0
-Requires-Dist: requests>=2.32.3
-Requires-Dist: uvicorn>=0.33.0
-Description-Content-Type: text/markdown
-
 # Lemming 🐹
 
 **The transparent, tool-agnostic orchestrator for autonomous AI coding agents.**
@@ -38,8 +16,8 @@ human-readable `tasks.yml` file.
   update is recorded in a human-readable `tasks.yml` file. You can step in,
   adjust the roadmap, or swap agents at any time.
 - **Tool Agnostic**: Lemming doesn't care which agent you use. It works
-  out-of-the-box with `agy`, `aider`, `claude`, `codex`, or even your own custom
-  scripts.
+  out-of-the-box with `agy`, `opencode`, `claude`, `codex`, or even your own
+  custom scripts.
 - **Resilient Execution**: With built-in heartbeat monitoring, automatic
   retries, and progress tracking, Lemming handles process crashes and rate
   limits gracefully.
@@ -93,7 +71,7 @@ Start the autonomous loop.
 
 ```bash
 # Run using the project's configured agent (defaults to the first
-# supported runner found on PATH: agy, aider, claude, or codex)
+# supported runner found on PATH: agy, opencode, claude, or codex)
 lemming run
 
 # Flags passed after -- are sent directly to the underlying runner
@@ -140,22 +118,25 @@ removed unless it failed — in which case it is kept, and its path printed, so
 the log can be read. Kept directories live in `~/.local/lemming/exec-*` and
 are retired automatically a week later, so a recent failure is always still
 there to inspect. Stdout carries the agent's message alone and everything
-else goes to stderr, so the output can be consumed directly.
+else goes to stderr, so the output can be consumed directly. Stderr reports
+the ephemeral tasks file and task ID at startup; pass those to `lemming
+--tasks-file <path> status <id>` or `logs <id>` to inspect an active run. Use
+the global verbose flag (`lemming -v exec ...`) to stream runner activity.
 
 Note that the agent runs unattended (`--yolo` by default), so it does not
 inherit the permission prompts of whatever launched it.
 
 ### Teaching Your Agent to Use It
 
-Lemming ships an Agent Skill so that an agent already running — Claude Code,
-Gemini CLI, Cursor — can reach any other agent CLI without knowing how each
-one spells its flags. Install it once:
+Lemming ships an Agent Skill so that an agent already running — Codex, Claude
+Code, Gemini CLI, Cursor — can reach any other agent CLI without knowing how
+each one spells its flags. Install it once:
 
 ```bash
 # Cross-tool location (~/.agents/skills), and report any tool-specific ones
 lemming skill install
 
-# Cover every agent tool detected on this machine
+# Cover detected tools that require their own skill directory too
 lemming skill install --all
 
 # This repository only
@@ -290,7 +271,7 @@ lemming hooks list
 lemming config list
 
 # Persist configuration to tasks.yml
-lemming config set runner aider
+lemming config set runner opencode
 ```
 
 ### Evaluating Prompt Changes
@@ -395,8 +376,9 @@ These come before the subcommand and apply to all of them.
     per-task `--runner`/`--model` overrides anything passed here.
 - **`exec [<description>]`**: Run one task, or one set of reviews, outside any
   roadmap. Prints the agent's closing message to stdout and everything else to
-  stderr. Exits non-zero if the task did not complete. See
-  [One-off tasks](#one-off-tasks-without-a-roadmap).
+  stderr, including its tasks file and task ID at startup. Use `lemming -v
+  exec ...` to stream runner activity. Exits non-zero if the task did not
+  complete. See [One-off tasks](#one-off-tasks-without-a-roadmap).
   - `-f/--file`: Read the description from a file, or `-` for stdin. Unlike
     `add`, there is no length cap.
   - `--review <names>`: Reviews to run after the task, comma-separated or
@@ -471,7 +453,10 @@ Lemming uses **fuzzy matching** to automatically inject the correct "YOLO"
 
 - **Antigravity (`agy`)**: Adds `--dangerously-skip-permissions` and exposes
   the project workspace with `--add-dir`
-- **Aider**: Adds `--yes --quiet`
+- **OpenCode**: Runs non-interactively via `opencode run`, adds `--format json`,
+  and, in YOLO mode, adds `--auto`. For its Google provider, Lemming makes an
+  existing `GOOGLE_API_KEY` or `GEMINI_API_KEY` available under OpenCode's
+  native `GOOGLE_GENERATIVE_AI_API_KEY` name.
 - **Claude**: Adds `--dangerously-skip-permissions`
 - **Codex**: Runs non-interactively via `codex exec`, adds `--json`, and, in
   YOLO mode, adds `--dangerously-bypass-approvals-and-sandbox`
@@ -533,10 +518,16 @@ pushing a `v*` tag triggers the `publish.yml` GitHub Actions workflow, which
 builds the package with `uv build` and uploads it.
 
 ```bash
-# 1. Bump the version in pyproject.toml, commit, and push
-# 2. Tag the release and push the tag
-git tag v0.1.1
-git push origin v0.1.1
+# 1. Bump pyproject.toml and uv.lock together, commit, and push
+uv version --bump patch --no-sync
+version="$(uv version --short)"
+git add pyproject.toml uv.lock
+git commit -m "Bump version to $version"
+git push origin main
+
+# 2. After CI passes, tag the release and push the tag
+git tag "v$version"
+git push origin "v$version"
 ```
 
 ---

{lemming_cli-0.4.1 → lemming_cli-0.4.3}/SKILL.md

@@ -1,13 +1,13 @@
 ---
 name: lemming
-description: One interface over the AI coding agent CLIs (agy, claude, codex). Delegate a single task to another agent to spare a quota or play to its strengths; run a readability, testing, or UX review over changed files, a branch, or a path; or drive a multi-task project through a roadmap that survives context limits. Use whenever work should run under a different agent than the current one, when reviewing a diff before opening a pull request, or when a job is too large for one agent run.
+description: One interface over the AI coding agent CLIs (agy, claude, codex, opencode). Delegate a single task to another agent to spare a quota or play to its strengths; run a readability, testing, or UX review over changed files, a branch, or a path; or drive a multi-task project through a roadmap that survives context limits. Use whenever work should run under a different agent than the current one, when reviewing a diff before opening a pull request, or when a job is too large for one agent run.
 license: MIT
 ---
 
 # lemming
 
 One interface over several agent CLIs. You do not need to know how `agy`,
-`claude`, or `codex` spell their flags — only `lemming`.
+`claude`, `codex`, or `opencode` spell their flags — only `lemming`.
 
 Use `exec` for one unit of work, and a roadmap for anything larger than a
 single agent run can hold.
@@ -15,16 +15,32 @@ single agent run can hold.
 ## Delegate one task
 
 ```sh
-lemming exec "Fix the flaky heartbeat test in src/runner_test.py" --runner codex
+lemming -v exec "Fix the flaky heartbeat test in src/runner_test.py" --runner codex
 
 # Pipe a longer handoff rather than fighting shell quoting; no length limit
-cat handoff.md | lemming exec -f - --runner agy
+cat handoff.md | lemming -v exec -f - --runner agy
 ```
 
 The agent's closing message comes back on **stdout**, everything else on
 stderr, and the exit code is 0 only if the task completed. Read that message
 instead of hunting through logs.
 
+When you launch `exec` as a subprocess, use the global `-v` flag (before
+`exec`) and continue monitoring its output until it exits. Verbose mode streams
+the runner's activity on stderr. At startup, stderr also reports the exact
+`Tasks file` and `Task ID`; use them from another process to inspect the
+agent's durable findings or full log without guessing which `exec-*` directory
+belongs to the run:
+
+```sh
+lemming --tasks-file <reported-path> status <reported-task-id>
+lemming --tasks-file <reported-path> logs <reported-task-id>
+```
+
+The task runner is required to record its approach first and concise findings
+as it works. These entries are checkpoints rather than a continuous trace, so
+use verbose output or the log to tell whether it is actively making progress.
+
 The task starts with an empty context and sees nothing of your conversation,
 so give it everything it needs.
 
@@ -94,9 +110,10 @@ turns one off for the project.
   atomicity to rely on.
 - **Concurrent runs in one checkout interleave their edits.** Give each
   writing run its own worktree and address it with `-C`.
-- **Failures keep their state.** The directory is printed on stderr and holds
-  the run's log directly, so `cat <that dir>/*-runner.log` is the whole
-  recovery. It is retired automatically after a week.
+- **Every run reports its state.** The tasks file and task ID are printed on
+  stderr at startup. Successful state is removed on exit unless `--keep` is
+  used; failures keep their directory and print it again for recovery. Kept
+  failure state is retired automatically after a week.
 - **`--tasks-file` and `-C` are group options.** They go before the
   subcommand — `lemming -C ../other exec ...`, never `exec -C ../other`.
 

{lemming_cli-0.4.1 → lemming_cli-0.4.3}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "lemming-cli"
-version = "0.4.1"
+version = "0.4.3"
 description = "An autonomous, iterative task runner for AI coding agents."
 readme = "README.md"
 authors = [
@@ -12,7 +12,7 @@ authors = [
 ]
 license = { text = "MIT" }
 requires-python = ">=3.12"
-keywords = ["ai", "agents", "automation", "antigravity", "claude", "aider"]
+keywords = ["ai", "agents", "automation", "antigravity", "claude", "opencode"]
 dependencies = [
     "beautifulsoup4>=4.13.3",
     "click>=8.0.0",

{lemming_cli-0.4.1 → lemming_cli-0.4.3}/src/lemming/api/config_test.py

@@ -20,7 +20,7 @@ def test_run_loop(client, test_tasks):
 def test_get_runners(client):
     response = client.get("/api/runners")
     assert response.status_code == 200
-    assert response.json() == ["agy", "aider", "claude", "codex"]
+    assert response.json() == ["agy", "opencode", "claude", "codex"]
 
 
 def test_project_goal_isolation(client, test_tasks):

{lemming_cli-0.4.1 → lemming_cli-0.4.3}/src/lemming/cli/config.py

@@ -50,7 +50,7 @@ def config_set(ctx: click.Context, key: str, value: str):
     discard the pinned model.
 
     Examples:
-      lemming config set runner aider
+      lemming config set runner opencode
       lemming config set model gemini-3.6-flash-high
       lemming config set model default
       lemming config set retries 5

{lemming_cli-0.4.1 → lemming_cli-0.4.3}/src/lemming/cli/exec_cli.py

@@ -324,6 +324,12 @@ def exec_command(
         else _create_review_task(exec_dir, reviews, config)
     )
 
+    # A supervising process needs an exact, race-free handle to the ephemeral
+    # roadmap while the run is active. Keep it on stderr so stdout remains the
+    # agent's closing-message interface.
+    click.echo(f"Tasks file: {tasks_file}", err=True)
+    click.echo(f"Task ID: {task_id}", err=True)
+
     shutdown.clear_drain()
     shutdown.install_handlers()
 

{lemming_cli-0.4.1 → lemming_cli-0.4.3}/src/lemming/cli/exec_cli_test.py

@@ -80,6 +80,24 @@ def test_exec_keeps_progress_chatter_off_stdout(workspace):
     assert "Attempt" in result.stderr
 
 
+def test_exec_announces_monitoring_coordinates_on_stderr(workspace):
+    """A supervisor can inspect an active exec without guessing its state."""
+    seen = {}
+
+    def finish(cmd, tasks_file, task_id, *args, **kwargs):
+        seen["tasks_file"] = tasks_file
+        seen["task_id"] = task_id
+        return _finish("Done.")(cmd, tasks_file, task_id, *args, **kwargs)
+
+    with mock.patch("lemming.runner.run_with_heartbeat", finish):
+        result = CliRunner().invoke(cli, ["exec", "Fix the flaky test"])
+
+    assert result.exit_code == 0, result.stderr
+    assert result.stdout.strip() == "Done."
+    assert f"Tasks file: {seen['tasks_file']}" in result.stderr
+    assert f"Task ID: {seen['task_id']}" in result.stderr
+
+
 def test_exec_removes_its_state_directory_on_success(workspace):
     """A one-shot leaves nothing behind when it worked."""
     with mock.patch("lemming.runner.run_with_heartbeat", _finish("Done.")):

{lemming_cli-0.4.1 → lemming_cli-0.4.3}/src/lemming/evals/cli.py

@@ -178,6 +178,10 @@ def run(
             )
         click.echo(f"Building eval image '{image}'...")
         container.build_image(context, image=image, docker=docker)
+        # The image is what the trials need; the cache that produced it is
+        # dead weight, and it outgrows the image several times over across
+        # repeated runs.
+        container.prune_build_cache(docker=docker)
 
     if run_dir is None:
         run_dir = pathlib.Path(tempfile.mkdtemp(prefix="lemming-evals-"))

{lemming_cli-0.4.1 → lemming_cli-0.4.3}/src/lemming/evals/container.py

@@ -19,6 +19,7 @@ DEFAULT_FORWARD_ENV = (
     "CLAUDE_CODE_OAUTH_TOKEN",
     "GEMINI_API_KEY",
     "GOOGLE_API_KEY",
+    "GOOGLE_GENERATIVE_AI_API_KEY",
 )
 
 # Paths inside the container; the harness mounts host dirs onto these.
@@ -41,7 +42,41 @@ def build_image(
         image: Tag for the built image.
         docker: Docker-compatible CLI binary to invoke.
     """
-    subprocess.run([docker, "build", "--tag", image, str(context)], check=True)
+    # No scenario drives a browser, and they are more than half the image.
+    subprocess.run(
+        [
+            docker,
+            "build",
+            "--build-arg",
+            "INSTALL_BROWSERS=false",
+            "--tag",
+            image,
+            str(context),
+        ],
+        check=True,
+    )
+
+
+def prune_build_cache(docker: str = "docker") -> None:
+    """Discards the build cache left behind by an eval image build.
+
+    Cache accumulates faster than images do — a run rebuilding this image a
+    few times leaves tens of gigabytes behind — and nothing reuses it once
+    the trials have finished.
+
+    Args:
+        docker: Docker-compatible CLI binary to invoke.
+    """
+    try:
+        subprocess.run(
+            [docker, "builder", "prune", "--force", "--all"],
+            check=False,
+            capture_output=True,
+        )
+    except OSError:
+        # This runs after the results are in; failing to tidy up must never
+        # be the reason a completed eval run reports an error.
+        pass
 
 
 def trial_command(

{lemming_cli-0.4.1 → lemming_cli-0.4.3}/src/lemming/evals/container_test.py

@@ -84,3 +84,34 @@ class TestRunTrial(unittest.TestCase):
 
 if __name__ == "__main__":
     unittest.main()
+
+
+class BuildImageTest(unittest.TestCase):
+    """The build must not carry what the eval trials never use."""
+
+    @unittest.mock.patch("subprocess.run")
+    def test_builds_without_browsers(
+        self, run: unittest.mock.MagicMock
+    ) -> None:
+        """Browsers are over half the image and no trial drives one."""
+        container.build_image(pathlib.Path("/repo"))
+
+        argv = run.call_args[0][0]
+        self.assertIn("--build-arg", argv)
+        self.assertIn("INSTALL_BROWSERS=false", argv)
+
+    @unittest.mock.patch("subprocess.run")
+    def test_prunes_the_build_cache(self, run: unittest.mock.MagicMock) -> None:
+        """Repeated eval runs accumulate cache faster than images."""
+        container.prune_build_cache()
+
+        argv = run.call_args[0][0]
+        self.assertEqual(argv[:3], ["docker", "builder", "prune"])
+        self.assertIn("--force", argv)
+
+    @unittest.mock.patch("subprocess.run", side_effect=OSError("no docker"))
+    def test_pruning_never_fails_a_run(
+        self, run: unittest.mock.MagicMock
+    ) -> None:
+        """Cleanup runs after the results are in; it must not lose them."""
+        container.prune_build_cache()

{lemming_cli-0.4.1 → lemming_cli-0.4.3}/src/lemming/models.py

@@ -8,7 +8,7 @@ import time
 import pydantic
 
 # Runner CLIs the orchestrator knows how to drive, in default-preference order.
-KNOWN_RUNNERS = ("agy", "aider", "claude", "codex")
+KNOWN_RUNNERS = ("agy", "opencode", "claude", "codex")
 
 
 @functools.cache