The platform-ops agent-compose project: agents, deployed on agent-compose itself, that watch the agent-compose platform they run on. The aim is to catch platform problems early and give agent-compose development concrete, evidence-backed direction.
| Agent | Role | Triggers | Skills | Writes to |
|---|---|---|---|---|
source-sync |
Refreshes the agent-compose source checkout | nightly-sync: 00:00 Asia/Shanghai |
— | the ac-src volume; publishes external.source-synced |
platform-analyst |
Analyzes platform health and usage | nightly-report: wakes on external.source-synced |
health-report |
its final answer (Run history) and the DingTalk group |
The nightly chain is: source-sync runs a fixed shell command (clone/fetch through the host-side proxy) → publishes external.source-synced with {ok, commit, syncedAt} → platform-analyst's event trigger starts the report. The sync itself is not an agent run, so no model is used to pull code.
On the nightly-report trigger, or when asked by hand, platform-analyst does the following:
- Collects one day's Runs, Sandboxes, and Scheduler Runs across all projects (
tool/collect.py). The result is a compact digest with every failure's details and failures grouped by error signature. - Classifies each failure as one of: platform bug, LLM upstream, image/runtime mismatch, user script/config, or user cancel.
- For suspected platform bugs, uses the checkout refreshed by
source-sync(the event carriesokand the commit), locates the code at the deployed version's tag, then checksorigin/mainto see whether the bug is already fixed upstream. When the sync failed, it says so in the report instead of guessing code locations. - Summarizes usage per agent: activity, success rate, provider/model/driver/image, and scheduler frequency.
- Answers with the report. The text is the run's final answer, so it stays in Run history and is visible in the agent-compose UI. The
nightly-reporttrigger also sends it to the DingTalk group throughtool/notify.py.
The details are in skills/health-report/SKILL.md.
| Path | What |
|---|---|
agent-compose.yml |
Project platform-ops: agents, their skills, triggers, and the ac-src volume. The name sorts after devboard-ai on purpose, because the agent-compose UI opens the alphabetically first project by default |
skills/ |
Skills, organized by job, so agents can share them |
tool/collect.py |
Health-report data collector (Python stdlib only) |
tool/notify.py |
Sends a report to the DingTalk robot (Python stdlib only): tables become lists and long reports are split |
image/Dockerfile |
The agents' image: the guest image (pinned by digest) plus agent-compose-ro |
tool/acli |
Wrapper around agent-compose-ro that points it at the daemon the sandbox runs on |
tool/acsrc |
Read-only git access to the agent-compose checkout ($ACOPS_SRC, default /src/agent-compose) |
cli/ |
Source of agent-compose-ro: chaitin-cli's agent-compose command alone (~12 MB), with dry-run fixed on, so every write command is refused before a request is sent. The chaitin-cli commit is pinned in cli/go.mod |
scripts/ |
Build, sync, and deploy helpers |
bin/, src/ |
Not in git: the built CLI (copied into the image, not deployed with the project) and, optionally, a host-side checkout written by scripts/sync-src.sh for manual analysis |
Every agent runs in ac-ops-agent:local, which carries the CLI at /usr/local/bin/agent-compose-ro. Inside a sandbox, the project is mounted read-only at /workspace, and is shared by all agents. The ac-src project volume is mounted at /src: writable for source-sync, read-only for platform-analyst. Agents use /tmp for scratch files.
A report is the run's final answer. The daemon keeps it in the Run history, and the UI shows it. The next report finds the previous one there, through acli run ls and acli inspect run. Bind mounts outside SANDBOX_ROOT are rejected on devboard (with DOCKER_HOST_SANDBOX_ROOT set); a local project volume lives under the daemon's DATA_ROOT, so it is the supported place for state that outlives a run.
To add an agent:
- Define it in
agent-compose.yml. - Put any new skill under
skills/.
Devboard cannot reach GitHub, and the daemon reads file sources from its own filesystem. So the project is copied to the host under the daemon's data mount: /data/agent-compose/projects/platform-ops on the host, which is /data/projects/platform-ops inside the daemon.
# 0. Once: create the host directory. /data/agent-compose is root-owned.
ssh devboard 'sudo mkdir -p /data/agent-compose/projects/platform-ops && sudo chown "$USER" /data/agent-compose/projects/platform-ops'
# 1. Once: store the HTTP proxy used by the source sync (see below).
scripts/set-notify-secret.sh http-proxy
# 2. Build bin/agent-compose-ro (linux/amd64).
scripts/build-cli.sh
# 3. Copy the project and validate, then build the image on the host and deploy.
scripts/deploy.sh --check
scripts/deploy.shThe agents no longer read src/ on the host; source-sync fills the ac-src volume at run time. scripts/sync-src.sh is still useful for a host-side checkout when analyzing by hand, but it is not part of the deploy.
scripts/deploy.sh builds the image with scripts/build-image.sh, which streams image/Dockerfile and bin/agent-compose-ro to the host's Docker and tags the result ac-ops-agent:local. The daemon uses that local image without pulling, so no registry is involved. To update only the CLI, run scripts/build-cli.sh and scripts/build-image.sh; new runs pick up the new image.
At 00:00 Asia/Shanghai, source-sync runs a fixed scheduler.shell command that clones or fetches https://github.com/chaitin/agent-compose.git into /src/agent-compose (the ac-src volume) and publishes external.source-synced with {ok, commit, syncedAt}. platform-analyst's external.source-synced trigger then starts the report.
The sandbox has no direct GitHub access, so the sync command reads the proxy URL from a host-side secret:
scripts/set-notify-secret.sh http-proxy # paste the HTTP proxy URL; input is hiddenThis writes secrets/http-proxy under the project directory on the host. scripts/deploy.sh excludes secrets/ from rsync, so redeploying keeps it. The URL never appears in this repository. If the sync fails, the event carries ok: false and the error; the report still runs and records that the source was unavailable.
The nightly report is sent to the group whose robot webhook lives on the host, never in git:
scripts/set-notify-secret.sh # paste the webhook URL; input is hiddenThis writes /data/agent-compose/projects/platform-ops/secrets/dingtalk-webhook (mode 0600, owned by the SSH user), which the sandbox sees read-only at /workspace/secrets/dingtalk-webhook. scripts/deploy.sh excludes secrets/ from rsync, so redeploying keeps it.
If the robot uses 加签, store the signing secret as well; tool/notify.py picks it up automatically:
scripts/set-notify-secret.sh dingtalk-secretIf it uses a keyword, DINGTALK_KEYWORD in agent-compose.yml holds it (每日报告 today); tool/notify.py puts it in every message, including each part of a split report. If it uses an IP allowlist, add the devboard host's egress address. The send is triggered only by the nightly-report prompt ("并发送到钉钉"); a manual run sends only when its prompt asks for it. To inspect the payload without sending:
DINGTALK_KEYWORD=每日报告 python3 tool/notify.py --file /tmp/report.md --dry-runTo move chaitin-cli forward, cd cli && go get github.com/chaitin/chaitin-cli@<commit>. Its vYYMM.x.y tags are not usable Go versions, so it is always pinned by commit. To move the guest image forward, change the digest in image/Dockerfile.
The source refresh is automatic from now on. Run scripts/sync-src.sh only when you want a host-side checkout under src/ for manual analysis.
-
Ask
platform-analystfor a report on a specific day:agent-compose run platform-analyst --prompt "生成 2026-10-14 的 agent-compose 健康报告" -
Collect only, without an agent (for example, from a machine with a tunnel to the daemon):
GOOS=darwin GOARCH=arm64 OUT=/tmp/agent-compose-ro scripts/build-cli.sh ACOPS_CLI=/tmp/agent-compose-ro AGENT_COMPOSE_URL=http://127.0.0.1:17410 \ python3 tool/collect.py --date 2026-10-14 # digest under /tmp/platform-ops/
- Read-only queries. All queries go through
agent-compose-ro. It is built with dry-run fixed on, so the following are refused before anything is sent: run, stop, rm, resume, exec, trigger, invoke, and auth login/logout. No flag turns this off. - Two outbound writes, both explicit.
source-syncfetches GitHub through the host-side proxy secret only during its sync command;tool/notify.pyPOSTs the report to the DingTalk robot only when the task asks for it. The proxy URL, webhook and signing secret are host-side files undersecrets/(excluded from git and from deploys); no script prints them, and reports never contain them. The proxy is not set globally:platform-analystnever gets it. - Minimal mounts. The project mount is read-only. The only writable mount is the
ac-srcvolume, writable forsource-syncand read-only forplatform-analyst. The daemon's data directory anddata.dbare not mounted, because they hold LLM tokens and webhook secrets. - Untrusted content. Other agents' prompts, outputs, and logs may carry text from external webhooks. The agents treat them as evidence only and never follow instructions in them.
- Privacy. Reports give aggregates and short IDs, and do not quote prompts or conversations.
collect.pyredacts common credential patterns before anything reaches the digest.