Forge
What this is for. Forge is the human-in-the-loop path for changing how agents work. The system proposes (soul, skill, or tool); you Approve & apply or Reject. Nothing of identity rewrites itself unless autonomy explicitly allows self-update — the default safe path is a proposal on this page.
When to use it
Section titled “When to use it”- An agent wants to change IDENTITY / soul and must not edit the file itself.
- Feedback on a skill or tool has piled up and you want Scan Now to turn it into proposals.
- You need to see current vs proposed value, reasoning, and confidence before applying.
- You want a log of collected feedback (useful / friction) without applying anything yet.
Typical workflow
Section titled “Typical workflow”- Open Forge in the sidebar (AI section) — route
/forge. - Stay on Proposals (or switch to Feedback). Filter All / Pending / Testing / Approved / Rejected / Applied.
- Expand a card — especially under Soul proposals. Read current vs proposed, then Approve & apply or Reject.
- The status should move to Applied (or Rejected). Open the agent’s Workspace tab to confirm IDENTITY matches the applied text.
Features
Section titled “Features”Forge is the human-in-the-loop path for changing deep agent identity/soul, skills, and tools. Agents (or the system) propose; you review and apply — or reject.
Why Forge exists
Section titled “Why Forge exists”| Path | When |
|---|---|
| Direct workspace edit | You edit IDENTITY/SOUL files yourself |
| Agent self-update | Only if autonomy identitySelfUpdate allows it |
| Forge proposal | Default safe path for autonomous improvement |
When identity self-update is disabled in config/autonomy, agents must use forge proposals instead of rewriting IDENTITY.md directly.
Page chrome
Section titled “Page chrome”Subtitle: Feedback-driven improvement proposals for tools and skills.
| Control | Meaning |
|---|---|
| Scan Now | Analyse collected feedback and generate proposals |
| Stats Total / Pending / Applied | Counts across soul + tool/skill proposals |
| Tab Proposals | Review queue |
| Tab Feedback | Raw useful/friction items recorded when tools and skills are used |
| Filter All + status chips | Restrict the list |
Empty proposals: No proposals yet. Click “Scan Now” to analyze feedback and generate improvement proposals.
Empty feedback: No feedback collected yet. Feedback is recorded automatically when tools and skills are used.
Proposal statuses
Section titled “Proposal statuses”Pending · Testing · Approved · Rejected · Applied
Targets
Section titled “Targets”| Target | Meaning |
|---|---|
| Soul | Identity / workspace prose — listed under Soul proposals |
| Skill | Skill text or behaviour |
| Tool | Tool description, schema, or code |
Scopes on a proposal: description · schema · prompt · behavior · code.
Soul proposal card
Section titled “Soul proposal card”Soul proposals are listed separately from tool/skill cards. They are human-approved — applying writes the workspace file.
| Control | Meaning |
|---|---|
| Current value | What is on disk now |
| Proposed value | What would replace it |
| Reasoning | Why the change |
| Based on N feedback items | Evidence count |
| Approve & apply | Accept into workspace |
| Reject | Discard |
Tool/skill cards add Show details / Hide details, confidence, and Approve & Apply / Reject.