@agent-forge/plugin-alignment
Requirement alignment capability plugin for Agent Forge: a three-step alignment funnel (objective → detail → design) with block-level bidirectional feedback confirmed on a local HTML page, feeding the plan layer with a user-approved baseline.
- Design contract (authoritative):
docs/design/需求对齐插件设计.md - Position: an upstream quality gate for
plan— not a replacement. The goal/plan/todo layers are untouched; the completed baseline is consumed through the shared session-entry protocol.
Tools
| Tool | Purpose |
|---|---|
alignment_set | Create / update the active step's full block set. Unchanged blocks inherit user approval by content fingerprint; comments are append-only (user text survives resends verbatim). |
alignment_round | Open the local confirmation page (127.0.0.1, random token path) and return its URL plus progress facts. Non-blocking: feedback arrives later as an alignment.feedback session message. |
alignment_advance | inspect (status/gate/baseline), retreat (roll back to an earlier step, resetting downstream confirmations), cancel. |
Command: /alignment · /alignment why · /alignment log [N].
Persistence protocol
Structurally identical to the goal/plan/todo/workflow plugins: append-only
alignment.state full snapshots plus one alignment.transition audit entry
per conversion, announced as alignment.changed.
Notes
- Requires a host session that exposes
SessionAPI.sendMessage(feedback injection); without it onlyalignment_roundfails — the rest of the surface stays usable. - On completion the engine writes a read-only baseline snapshot to
<workspace>/alignments/<id>.md(image-generation workspace-product precedent). - Deterministic tests only (fake CapabilityAPI + real
node:httpon an OS-assigned port); no real model calls.