Skip to content

fix(docker): pin container-side workspace and config dirs in compose#77446

Merged
vincentkoc merged 1 commit intoopenclaw:mainfrom
lonexreb:fix/77436-docker-workspace-host-leak
May 4, 2026
Merged

fix(docker): pin container-side workspace and config dirs in compose#77446
vincentkoc merged 1 commit intoopenclaw:mainfrom
lonexreb:fix/77436-docker-workspace-host-leak

Conversation

@lonexreb
Copy link
Copy Markdown
Contributor

@lonexreb lonexreb commented May 4, 2026

Summary

Fixes #77436 — fresh `./scripts/docker/setup.sh` Docker setups on macOS started failing after 04.29 with the first agent reply dying on:

```
EACCES: permission denied, mkdir '/Users'
```

The reporter confirmed the same flow worked on 2026.04.29 and reproduces consistently from 2026.05.02 onward.

Root cause

`scripts/docker/setup.sh` defaults `OPENCLAW_WORKSPACE_DIR` and `OPENCLAW_CONFIG_DIR` from the host `$HOME` (so on macOS this is `/Users//.openclaw/...`), then writes both into repo `.env` so Compose can use them as bind-mount sources for `/home/node/.openclaw` and `/home/node/.openclaw/workspace`.

Both services then loaded the same `.env` via `env_file:` without re-overriding those two variables in the container's `environment:` block. Inside Linux Docker the env vars carried host-style paths, and once a runtime/wizard resolution path persisted that into `agents.defaults.workspace`, the next `ensureAgentWorkspace` on the reply path called `fs.mkdir` against `/Users/...` — which fails because Linux Docker has no `/Users` and the container user cannot create top-level dirs anyway.

The clawsweeper review on the issue traced the chain end-to-end (`scripts/docker/setup.sh:241`, `docker-compose.yml:5`, `src/wizard/setup.ts:567`, `src/auto-reply/reply/get-reply.ts:237`) and recommended fixing the boundary between host bind-mount source and container runtime path.

Fix

Override `OPENCLAW_CONFIG_DIR` and `OPENCLAW_WORKSPACE_DIR` in the `environment:` block on both `openclaw-gateway` and `openclaw-cli` services so they always resolve to the canonical container paths:

```yaml
OPENCLAW_CONFIG_DIR: /home/node/.openclaw
OPENCLAW_WORKSPACE_DIR: /home/node/.openclaw/workspace
```

Compose's `environment:` takes precedence over `env_file:`, so:

  • The bind-mount source on `volumes:` still uses the host-side value substituted from `.env` (correct — that's what we want for the Docker mount).
  • The in-container env always sees the container-side path (correct — that's what runtime code resolves against).

This separates host-mount-source from container-runtime-path the way the existing `HOME=/home/node` override already does for `HOME`. It's defensive against any current or future code path that resolves a workspace from `OPENCLAW_WORKSPACE_DIR`.

Verification

  • `pnpm test src/docker-setup.e2e.test.ts` — 27 passed (1 new sync test locks the contract on both services).
  • `pnpm exec oxfmt --check --threads=1 docker-compose.yml src/docker-setup.e2e.test.ts CHANGELOG.md` — clean.

I did not run the live Docker/Ollama Cloud flow from the reporter (no Docker on this dev machine), matching the read-only constraint clawsweeper noted. The unit-level evidence locks the wire contract: both services emit container-side paths even when `.env` carries host-side ones.

Closes

Fixes #77436

@openclaw-barnacle openclaw-barnacle Bot added docker Docker and sandbox tooling size: XS labels May 4, 2026
@clawsweeper
Copy link
Copy Markdown
Contributor

clawsweeper Bot commented May 4, 2026 •
edited
Loading

Codex review: needs maintainer review before merge.

Summary
The branch adds container-local OPENCLAW_CONFIG_DIR and OPENCLAW_WORKSPACE_DIR overrides to both Compose services, a Docker setup sync test, and a changelog fix entry.

Reproducibility: yes. by source inspection for the env leak: setup writes host-side paths to .env, Compose imports them into both containers on current main, and the reply path creates the configured workspace. I did not live-reproduce the Docker flow or independently prove the exact /Users persistence step beyond the linked report and PR discussion.

Next step before merge
No repair job is needed; the open PR is a narrow implementation candidate that needs normal maintainer review and checks rather than an automated replacement.

Security
Cleared: The diff pins existing local container path variables, adds a regression assertion, and updates the changelog without adding dependencies, secrets, permissions, downloads, or new execution paths.

Review details

Best possible solution:

Land the focused Compose boundary fix after normal PR checks so #77436 closes through the merge.

Do we have a high-confidence way to reproduce the issue?

Yes by source inspection for the env leak: setup writes host-side paths to .env, Compose imports them into both containers on current main, and the reply path creates the configured workspace. I did not live-reproduce the Docker flow or independently prove the exact /Users persistence step beyond the linked report and PR discussion.

Is this the best way to solve the issue?

Yes; service-level container path overrides are the narrowest maintainable fix because they keep host paths available for bind-mount interpolation while preventing those values from becoming the container runtime env.

What I checked:

  • Current main imports host .env into both services: Both openclaw-gateway and openclaw-cli load .env; current main sets HOME but does not override OPENCLAW_CONFIG_DIR or OPENCLAW_WORKSPACE_DIR in the container environment. (docker-compose.yml:5, 103cdd9d96f8)
  • Setup persists host-side mount paths: scripts/docker/setup.sh defaults config/workspace paths from host $HOME, exports them, and writes them into repo .env, which Compose also uses for the bind-mount sources. (scripts/docker/setup.sh:241, 103cdd9d96f8)
  • Reply path creates the resolved workspace: The reply path resolves the agent workspace and calls ensureAgentWorkspace; that helper resolves the path and immediately runs recursive fs.mkdir, matching the reported failure mode when a host-style /Users/... workspace leaks into runtime config. (src/auto-reply/reply/get-reply.ts:237, 103cdd9d96f8)
  • Compose precedence supports the fix: Docker's official Compose environment precedence page says explicit environment values take precedence over env_file values, so service-level container paths should override .env inside the container without changing Compose interpolation for bind mounts.
  • PR diff matches the intended boundary fix: The patch adds the two container-side path overrides under both service environment blocks, adds a sync test requiring both services to contain those values, and adds a user-facing changelog entry. (docker-compose.yml:8, 18347cf45100)
  • PR discussion records targeted verification: The PR body and follow-up comment report pnpm test src/docker-setup.e2e.test.ts passing 27 tests after rebase, including the new sync test; I did not rerun tests in this read-only review. (src/docker-setup.e2e.test.ts:623, 18347cf45100)

Likely related people:

  • vincentkoc: Recent Docker setup/docs/test stabilization and image work touched the same setup, Compose, and Docker documentation surface. (role: recent Docker setup maintainer; confidence: medium; commits: a7c5a0425988, 601d1ccd2499, 3f3f66a5f75b; files: scripts/docker/setup.sh, docker-compose.yml, docs/install/docker.md)
  • Peter Steinberger: Historical Docker onboarding and mount-validation hardening commits are on the same setup flow and Compose boundary. (role: historical Docker setup owner; confidence: medium; commits: 35976da7a078, dfef943f0a18, 7255c20ddcc3; files: scripts/docker/setup.sh, docker-compose.yml, docs/install/docker.md)
  • jamtujest: Authored the opt-in Docker sandbox support that touched the same Compose, setup script, docs, and Docker setup test boundary. (role: adjacent Docker Compose contributor; confidence: medium; commits: cb491dfde5cb; files: docker-compose.yml, scripts/docker/setup.sh, docs/install/docker.md)

Remaining risk / open question:

  • I did not run the live Docker/Ollama flow or local tests in this read-only pass; validation relies on source inspection, the PR discussion, and the added contract test.

Codex review notes: model gpt-5.5, reasoning high; reviewed against 103cdd9d96f8.

@lonexreb lonexreb force-pushed the fix/77436-docker-workspace-host-leak branch from 52b41b2 to 18347cf Compare May 4, 2026 17:58
@lonexreb
Copy link
Copy Markdown
Contributor Author

lonexreb commented May 4, 2026

Rebased onto current origin/main to address the merge-conflict status. No code change; clawsweeper review confirmed no correctness findings, just the dirty-merge state. Targeted tests still green: 27 passing in src/docker-setup.e2e.test.ts (including the new sync test that locks the contract on both services).

@lonexreb @vincentkoc
fix(docker): pin container-side workspace and config dirs in compose
5db7774 
scripts/docker/setup.sh writes the host-side OPENCLAW_WORKSPACE_DIR and
OPENCLAW_CONFIG_DIR (e.g. /Users/<you>/.openclaw/...) into .env so Compose
can use them as bind-mount sources. Both gateway and CLI services then
imported the same .env via env_file, so inside the Linux container the env
vars carried host-style paths. After 04.29 this regressed: the first agent
reply died with 'EACCES: permission denied, mkdir /Users' when the wizard
or runtime resolved a workspace from those env values.

Override OPENCLAW_CONFIG_DIR and OPENCLAW_WORKSPACE_DIR in the environment:
block on both services so the container always sees /home/node/.openclaw
and /home/node/.openclaw/workspace, regardless of what .env carries for
Compose substitution. Compose's environment: takes precedence over
env_file:, so the bind-mount source still uses the host path while the
in-container env stays correct.

Add a sync test alongside the existing docker-compose contract tests that
locks the override on both services so this contract cannot regress
silently.

Fixes openclaw#77436.
@vincentkoc vincentkoc force-pushed the fix/77436-docker-workspace-host-leak branch from 18347cf to 5db7774 Compare May 4, 2026 22:57
@vincentkoc vincentkoc self-assigned this May 4, 2026
@vincentkoc vincentkoc merged commit 34f805a into openclaw:main May 4, 2026
93 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

@vincentkoc vincentkoc

Labels

docker Docker and sandbox tooling size: XS

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

[Bug]: Docker-setup.sh, does not work after 04.29 version

2 participants

@lonexreb @vincentkoc