oa-state-workspace

name: oa-state-workspace description: "Explains oa agent workspace layout, --direct flag, and how agents write output. Use when understanding where agent files go or how to access agent output. Activates for: workspace, agent output, --direct, /tmp/oa-agent, result.md, project_root." user-invocable: false

Critical Rules

• ALWAYS use --direct in every oa run call — because without it, output is written to volatile /tmp/oa-agent-<uuid>/ which is lost on reboot (L-010).
• NEVER reference agent output with relative paths — because workspace location depends on the --direct flag and relative paths break cross-agent references.

Workspace Locations

Workspace Structure

/tmp/oa-agent-<uuid>/          # default workspace (no --direct)
├── CLAUDE.md                  # auto-generated task instructions
└── output/
    └── result.md              # agent writes primary output here

# With --direct, workspace = project root:
/mnt/c/Users/Freek/Project/
├── CLAUDE.md                  # injected task instructions
└── output/
    └── result.md

How Agents Write Output

1. Agent reads its task from the injected CLAUDE.md in the workspace root.
2. Agent writes primary output to ./output/result.md (relative to its workspace).
3. Agent creates .done in the workspace root when fully complete.
4. Orchestrator collects output with oa collect <agent-name>.

Patterns

Pattern 1: Collect agent output

# Check completion
oa status

# Collect output (printed to stdout, saved to ~/.oa/collected/<name>.md)
oa collect worker-1

Pattern 2: Pass output between agents using absolute paths

# researcher-a wrote to /mnt/c/Project/output/result.md
# Pass it to combiner via absolute path in the prompt:
oa run "Read /mnt/c/Project/output/result.md and summarize" --name combiner --model claude/sonnet --direct

project_root inheritance

When --direct is used, the agent workspace root is the orchestrator project root. All agents with --direct share the same project root, enabling file sharing via absolute paths.

Anti-Patterns

• Bad: Referencing ./output/result.md from another agent — resolves differently per workspace
• Good: Use absolute path /mnt/c/Project/output/worker-name/result.md
• Bad: Omitting --direct — output lost after system reboot
• Good: Include --direct in every oa run call

References

• Related: oa-orchestration-spawn, oa-orchestration-pipeline