Weave Multi-Agent Orchestration Plugin

In the SDD workflow, Analysis is distinct from Review gates:

• Analysis (Thread, step 8): A mechanical consistency check that finds gaps, conflicts, and coverage holes. It produces a findings report and does NOT block workflow progression. However, the report is an input for the implementation step.
• Code Review (Weft, step 10): A quality judgment performed after implementation. It assesses correctness, readability, and testing. It CAN block the workflow with a [REJECT] status.
• Security Review (Warp, step 11): A security audit performed after implementation. It CAN block the workflow with a [REJECT] status.

To maintain a clean execution flow, organize tasks into the following standard phases:

The Spec-Driven Development (SDD) process generates two distinct sets of artifacts to separate specification from execution.

SDD Specification Artifacts (.specify/)

These belong to the feature and contain research, requirements, and plans.

• .specify/memory/constitution.md: Project governance.
• .specify/features/{slug}/spec.md: Feature specification (FR-001, SC-001 format).
• .specify/features/{slug}/plan.md: SDD implementation plan.
• .specify/features/{slug}/tasks.md: Granular task list.
• .specify/features/{slug}/analysis.md: Consistency findings.
• .specify/features/{slug}/checklists/requirements.md: FR coverage checklist.
• .specify/working-spec.json: Shared state tracking current phase and paths.

Weave Execution Artifacts (.weave/)

These belong to the execution engine.

• .weave/plans/{slug}.md: The Tapestry-executable version of the SDD plan.

Weave agents are categorized into three tiers based on their reasoning requirements. Choosing the correct tier is critical for workflow quality:

• Top Tier (Loom, Pattern, Warp): Requires the strongest reasoning. These agents make high-level decisions (orchestration, planning, security) that impact the entire workflow. Bad judgment here causes cascading failures.
• Mid Tier (Tapestry, Weft, Shuttle): Requires good coding ability and comprehension. These agents follow plans and analyze code. They must be reliable and not sloppy.
• Economy Tier (Thread, Spindle): Requires speed over depth. These agents perform high-volume, low-complexity tasks like codebase searching and web research.

The project uses npm dist-tags to separate stable releases from development/preview versions. All prerelease types (e.g., alpha, beta, rc, preview) are published under the next tag.

| Dist-tag | Used for | Install command |
|----------|----------|-----------------|
| `latest` | Stable releases (`0.7.0`) | `npm install @opencode_weave/weave` |
| `next` | Preview/prerelease versions (`0.7.0-beta.1`) | `npm install @opencode_weave/weave@next` |

The evaluation website should perform a single fetch to the canonical suite JSONL and normalize the rows client-side. The expected normalized object shape for the UI is:

{
  suiteId,
  timestamp,
  modelKey,
  provider,
  model,
  score,
  passedCases,
  totalCases,
  durationMs,
  caseResults: [{ caseId, description, passed, score }]
}

Recommended UI Components:

• Scenario-by-model matrix: Rows represent cases, columns represent models. Primary content shows the latest pass/fail and score; secondary content shows a compact history of the last 6 runs.

When an agent needs a model, Weave resolves the selection using the following priority order:

1. Config override: The value set at agents.<name>.model in your configuration (highest priority).
2. UI-selected model: Only applies to Loom, Tapestry, and Shuttle.
3. Category model: If the agent belongs to a category with a configured model.
4. Built-in fallback chain: Attempts to match default models against your available models.
5. System default: The global fallback model.
6. Hardcoded fallback: github-copilot/claude-opus-4.8 (lowest priority).