| project | cli-bridge-multi-demo | ||||
|---|---|---|---|---|---|
| brief_id | 2026-05-17-CLIM | ||||
| scope_locked | true | ||||
| mode | gated | ||||
| authorized_costs |
|
||||
| roles |
|
||||
| budgets |
|
Run the multi-agent coordinator (v0.4) where every role uses the same cli-bridge adapter. The whole run delegates to a single local CLI agent (Claude Code by default). The phase machine + deterministic detectors run against the CLI's stdout for each role.
This is the single-adapter shape of the bridge in v0.5. Per-role adapter mixing (planner on a cheap API, executor on a CLI) lands in v0.5.1; until then, pick one adapter for the whole brief.
A two-sub-task plan:
- List the files in
examples/. - Describe what each example brief covers in one short paragraph.
Both sub-tasks pass review. Phase 4 wraps with a one-paragraph summary.
- Read-only filesystem. No writes.
- Reviewer verifies executor claims against the files, not just the executor's word.
- The bridged CLI handles tool calls in its own way (its own approval prompts, its own MCP servers); OpenWar does not relitigate them.
- A CLI agent installed and on PATH (default:
claude).
Run from this repo's root:
npx @pythonluvr/openwar run examples/cli-bridge-multi-agent-brief.md \
--adapter cli-bridge \
--cli-binary claudeThe cost-tier preview before Phase 0 will show the run as tier=free
because cli-bridge defaults to free (local subscription). The framework
doc gets prepended to every role invocation so the CLI is OpenWar-aware
even without changes to its own config.
When per-role adapter mixing ships in v0.5.1 the brief format extends to let planner / reviewer use a cheap API while executor stays on the CLI. The shape is still being designed; do not depend on the syntax landing exactly as drafted in the framework doc.