main
4
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Algis DumbrisandClaude Opus 4.6
|
121d05f875 |
feat(docker): auto-stage host OAuth credentials into agent containers
The docker harness now detects and stages host CLI auth files (~/.gemini/oauth_creds.json, ~/.claude/.credentials.json) into a writable agent-home directory mounted at /home/agent. This lets containerized agents reuse the host's Gemini Pro / Claude Pro OAuth sessions without manual secret management or API keys. Only auth files are copied — not the host's settings.json or MCP configs (which contain stale localhost URLs that would hang Gemini CLI inside containers). The staged dir is writable so CLIs can create projects.json, history, etc. alongside the auth files. Also sets GEMINI_DEFAULT_AUTH_TYPE=oauth-personal and GEMINI_CLI_NO_RELAUNCH=true when OAuth creds are detected, writes Claude's hasCompletedOnboarding flag, and simplifies the doc-gardener example to use the harness-level credential staging instead of manual HOME directory seeding. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|
Algis DumbrisandClaude Opus 4.6
|
ddbb1bd329 |
fix(docker): tmpfs /tmp mounted rw,exec,size=128m
Docker Desktop's default noexec on tmpfs broke the "download a CLI to /tmp, chmod +x, run it" workflow — exactly what the doc-gardener inspector needs to verify docs.mcpproxy.app against the real mcpproxy binary. Previously the agent spent ~10 minutes in a self- debug loop discovering the noexec, falling back to /home/agent, running into externally-managed Python, missing python3-venv, etc. With /tmp exec, the inspector's own install pipeline works on the first try: curl | tar | chmod | run. First real run produced a 72-claim drift report (21 matched / 1 drifted / 50 missing) against mcpproxy v0.24.4 in ~8 minutes, no REVISE loop. The 64m → 128m bump gives breathing room for curl'd tarballs that need a temp extraction directory alongside the final binary. Inspector prompt updated to tell the agent about the /tmp install path explicitly and forbid the previous /home/agent detours. Also updated the coordinator brief template to match. Note the image itself is UNCHANGED — we deliberately do NOT bake mcpproxy (or any other domain-specific tool) into synapbus-agent. The image stays a blank Linux shell with Node + Python + core tools, and each example's prompt teaches its agent how to install whatever it needs. This keeps the gardener universal: swap in any other docs domain and the inspector figures out what to install on demand. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|
Algis DumbrisandClaude Opus 4.6
|
f1e2b1fa38 |
feat(doc-gardener): MCP-native + docker-isolated multi-agent demo
Replace the legacy cmd/docgardener orchestration (~2400 LOC of Go
spawning subprocess workers via local_command + admin socket) with
three Docker-isolated agents that all reach SynapBus through MCP:
doc-coordinator — Gemini Pro, triages goal, calls create_goal +
propose_task_tree + send_message via MCP
docs-inspector — Gemini Flash, fetches docs, installs mcpproxy,
shells out to verify, reports findings via MCP
docs-critic — Gemini Flash, independent reviewer with its
own MCP API key + config_hash, audits the
inspector's evidence and DMs the owner
Every agent runs inside synapbus-agent:latest with --cap-drop=ALL,
--security-opt=no-new-privileges, --read-only root + tmpfs /tmp,
--pids-limit, memory + CPU quotas. The container reaches the
SynapBus MCP server on the host at host.docker.internal:18089
because the docker harness rewrites .gemini/settings.json URLs
from 127.0.0.1 automatically.
Wrapper baked into the image at /usr/local/bin/synapbus-agent-wrapper.sh
so configs don't need to mount or template a per-example wrapper.
The harness's default no longer overrides docker CMD — the image's
baked entry script is used unless docker.command is set explicitly.
start.sh changes:
- Preflight: docker daemon, GEMINI_API_KEY (or ~/.gemini/oauth_creds.json)
- Builds synapbus-agent image lazily on first run
- Mints one MCP API key per agent via `agent revoke-key`
- Templates each config with __PORT__, __*_APIKEY__, __MODEL__,
__GEMINI_API_KEY__, __EXTRA_MOUNTS__
- With OAuth fallback: copies host ~/.gemini → data/agent-home/.gemini
once and bind-mounts the whole agent-home rw at /home/agent so
in-container gemini has a writable HOME without polluting the host
- SYNAPBUS_KEEP_WORKDIR=1 preserves per-run docker workdirs for
debugging
- Sets harness_name=docker explicitly so the resolver picks the
right backend even with empty local_command
stop.sh: best-effort cleanup of lingering synapbus-* containers so a
killed parent doesn't leave bind-mount holders that block the next
start.sh from re-mounting the same paths.
run_task.sh: snapshot-baseline pattern (only watches replies newer
than the max msg id at send time), 600s deadline, treats any reply
from doc-coordinator that isn't DELEGATED:/REVISING: as terminal,
plus FINAL:/CANNOT: from any sender.
cmd/docgardener slimmed from 7 files / 2580 LOC to 3 files / ~370 LOC.
The remaining binary only renders the HTML report (queries goals +
goal_tasks + traces + harness_runs from the SynapBus DB read-only).
agent.go, channels.go, flow.go, gemini_tree.go all deleted.
Verified end-to-end against gemini-2.5-pro coordinator + gemini-2.5-flash
workers (with OAuth fallback mount):
./run_task.sh "what does this demo do?"
→ coordinator TRIVIAL: replies directly via MCP send_message
./run_task.sh "Verify the CLI commands on docs.mcpproxy.app/cli/command-reference"
→ coordinator calls create_goal (slug verify-mcpproxy-cli-...),
propose_task_tree (3-node tree: coordinator/plan,
doc-gardener/scan, doc-gardener/audit) and send_message to
docs-inspector
→ inspector container runs ~10 minutes inside the sandbox:
installs mcpproxy from real release URL (linux-arm64), curls
the docs page, falls back from BeautifulSoup → grep when
python3-venv is missing, debugs its own f-string syntax, writes
extract_flags.py, runs `mcpproxy --help` for ground truth
→ real multi-agent iteration loop: critic REVISE: → inspector
retry → critic REVISE: with new feedback
The agents discovered real environment quirks (tmpfs noexec on /tmp,
externally-managed Python, missing python3-venv) and worked around
them inside the sandbox without touching the host.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
|
||
|
Algis DumbrisandClaude Opus 4.6
|
560d9d4125 |
feat(harness): docker isolation backend + canonical synapbus-agent image
New `internal/harness/docker` package: per-run ephemeral container
backend that runs each agent in `docker run --rm`, bind-mounts the
materialized workdir at /workspace, and captures stdout/stderr/exit
code/result.json the same way the subprocess backend does.
Inspired by scion's pkg/runtime/docker.go: shell out to the docker
CLI (zero new Go deps, zero CGO), per-task ephemeral containers with
no warm pool, host-side scratch dir bind-mounted in.
Default security posture (overridable per agent):
--rm
--cap-drop=ALL
--security-opt=no-new-privileges
--read-only with tmpfs /tmp
--pids-limit=512
--user=<host uid:gid>
--network=bridge (configurable; --network=none for air-gap)
--memory / --cpus from agent config
--add-host host.docker.internal:host-gateway on Linux
The backend reuses subprocess.AgentConfig for gemini_md/claude_md/
mcp_servers/skills materialization so existing example configs work
unchanged. Per-agent docker tunables go under a new `docker` block
in harness_config_json: image, memory, cpus, network, extra_mounts,
cap_add, read_only_root, user, entrypoint, command, extra_args.
MCP host rewrite: `.gemini/settings.json` URLs of the form
http://127.0.0.1:<port>/mcp are rewritten to
http://host.docker.internal:<port>/mcp at materialization time so the
in-container Gemini CLI can reach the SynapBus MCP server on the host
without code changes in the example wrappers.
Wired into the reactor and Registry resolver:
- Registry.Resolve picks "docker" when harness_config_json contains a
`"docker"` block, taking precedence over local_command so explicit
isolation never silently downgrades.
- reactor.agentBackendKind() returns backendDocker for the same case.
- evaluateTrigger's harness-backend gate accepts backendDocker
alongside subprocess + webhook.
- main.go registers docker.Harness with the harness registry, passing
the SynapBus listen port so the URL rewrite uses the correct host
port.
Smoke tests in docker_test.go (skipped when no docker daemon):
- TestExecute_Hello: env injection + bind-mount writeback + message.json
+ result.json + stdout capture using alpine:3.20
- TestExecute_NoImage: rejects agents missing docker.image
- TestExecute_TimeoutCancel: wall-clock budget kills the container
New canonical agent image at image-build/synapbus-agent/:
- Debian bookworm-slim base
- Node 22 + @google/gemini-cli + @anthropic-ai/claude-code
- jq, sqlite3, curl, git, python3, tini (PID 1 for signal forwarding)
- Non-root agent user uid/gid 1000
- ENTRYPOINT tini, CMD /workspace/wrapper.sh
No SynapBus binary inside the image — agents reach the host MCP server
over the network at host.docker.internal:<port>.
Pre-existing reactor test failures (TestReactorNoK8sImage,
TestReactorDepthExceeded, TestReactorBudgetExhausted,
TestReactorCooldownSkipped, TestReactorSequentialExecution) verified
to exist on
|