Skip to content

flue-agent-scaffold.md

Latest commit

 

History

History
82 lines (66 loc) · 2.89 KB

flue-agent-scaffold.md

File metadata and controls

82 lines (66 loc) · 2.89 KB
Raw

Flue Agent Scaffold Issue Template

Goal

Create or review the smallest useful Flue agent scaffold for one bounded workbench workflow.

Context

  • Workbench lane: docs/flue-agent-harness-lane.md
  • Skill: skills/workbench-flue-agent-harness/SKILL.md
  • Upstream entry: https://flueframework.com/start.md
  • Default local dev port: 3583

SDD Stage Card

SDD_STAGE: Technical Design
OWNER: <assigned agent>
STATUS: READY_FOR_REVIEW
REVIEWER: Workbench Supervisor
EVIDENCE: docs/flue-agent-harness-lane.md, skills/workbench-flue-agent-harness/SKILL.md
HANDOFF_SUMMARY: Package one proven workflow as a Flue starter agent without changing Multica runtime state.
SCOPED_EVIDENCE: docs/flue-agent-harness-lane.md, skills/workbench-flue-agent-harness/SKILL.md, target project files only
ANTI_OVER_READ: Do not inspect unrelated issue history, full agent roster, raw run transcripts, or private runtime config.

Required Contract

FLUE_AGENT_CONTRACT:
  purpose: "<what the agent should do>"
  project_directory: "<absolute or repo-relative path>"
  workspace_layout: "root | .flue"
  agent_file: "./agents/<name>.ts | ./.flue/agents/<name>.ts"
  deploy_target: "node | cloudflare | github-actions | gitlab-ci | other-node-baseline"
  model_id: "anthropic/claude-sonnet-4-6 | anthropic/claude-opus-4-7 | openai/gpt-5.5 | openrouter/moonshotai/kimi-k2.6 | <models.json id>"
  sandbox_mode: "virtual | local | cloudflare-r2 | container | external"
  trigger: "webhook | cli | ci"
  secrets_policy: "none | env-only | host-command-env"
  validation_command: "<exact command>"
  public_artifact_policy: "summaries-only"

Implementation Steps

  1. Read docs/flue-agent-harness-lane.md.
  2. Fetch https://flueframework.com/start.md and the Flue README if fresh upstream behavior matters.
  3. Verify the contract fields.
  4. Infer layout from the target directory:
    • new or empty: agents/ and roles/;
    • existing non-empty: .flue/agents/ and .flue/roles/.
  5. Create the smallest useful starter agent.
  6. Pass the model ID explicitly to init({ model: "<model_id>" }).
  7. Run the chosen validation command.
  8. Report only compact evidence and residual risk.

Verification Commands

Use the command that matches the target:

flue dev --target node
flue build --target node
flue dev --target cloudflare
flue build --target cloudflare
flue run <agent-name> --target node --id smoke --payload '{}'

Do not use flue run --target cloudflare.

PASS Criteria

  • FLUE_AGENT_CONTRACT is complete.
  • Agent file path matches the inferred layout.
  • Model ID is exact and explicitly passed to init.
  • Deploy target and verification command match.
  • No secrets, tokens, raw payloads, OAuth material, or raw run transcripts are committed or pasted.
  • Starter agent is intentionally narrow and does not mutate live Multica state.
  • Closeout includes FLUE_AGENT_REPORT and VERDICT: PASS | FLAG | BLOCK.