[LinuxDevil]

What changed

New apps/agent-forge/server/author/draft/ module — the zod contract and pure spec builders the Author's propose tools (AA08/AA09) will sit on.

schema.ts

• authorDraftSchema: derived from the SDK's agentSpecV2Schema minus version, evals, flow (DRAFT_OMITTED_KEYS), strict. description is required and capped at 500 chars (stricter than the spec); instructions is a plain string 1–50,000 chars (never { file }); name and model stay required exactly where the spec requires them. Never re-declared by hand, so the schema cannot drift from the SDK.
• authorPatchSchema: every draft key optional (omitted = keep); the spec-optional keys also accept null (reset to default) — NULLABLE_PATCH_KEYS is computed from the spec schema's optional fields minus description. Strict.
• FIELD_DESCRIPTIONS: one sentence per draftable field; a coverage test pins it to exactly the schema keys.

toSpec.ts

• draftToSpec(draft): { version: 2, ...draft } plus a chat-trigger backstop, validated with agentSpecV2Schema.
• applyAuthorPatch(base, patch): undefined → keep, null → delete key, any other value → replace (lists replace wholesale, [] clears); same chat-trigger backstop and validation.
• draftInvalidFailure(toolName, issues): the DRAFT_INVALID tool failure (first 10 issues as path: message).
• deriveAgentId(name, taken): name → slug (lowercase, non-[a-z0-9] runs → -, 48-char cap, agent fallback) with -2, -3… collision suffixes; every output passes isValidAgentId.
• Issues carry tools[2].builtin-style paths ([n] for array indexes).
• Chat-trigger rule: appended only when no type: 'chat' entry exists, with the first unused t<N> id.

No wiring yet — AA08 is the first caller (test usage satisfies fallow until then).

Test summary

• server/author/draft/__tests__/schema.test.ts + toSpec.test.ts: 125 tests — required/omitted/unknown keys, null rules, full patch-semantics table, t1 taken → t2, guardrails[0].name issue path, deriveAgentId cases (empty, unicode, 60-char, collisions).
• typecheck:server, typecheck, test:server (69 files / 944 tests), test (85 files / 942 tests) — all pass.
• Root tsc --noEmit, lint, fallow (no issues), docs:ratchet, api:check, docs:llms:check, size-check — all pass.
• test:coverage: 11 pre-existing/environmental failures unrelated to this change — 6× templates.test.ts README startsWith checks fail under a CRLF checkout (core.autocrlf=true), 5× create-lousho tests need the package's dist/ build (not produced in this worktree). All other 7,152 tests pass; coverage report generated via --coverage.reportOnFailure so fallow could consume it.

Acceptance criteria

• authorDraftSchema rejects version, evals, flow and unknown keys; requires name, description, model, instructions.
• applyAuthorPatch(base, { tools: [] }) clears tools; { reasoning: null } removes reasoning; {} returns base unchanged (plus chat trigger).
• { description: null } fails patch validation.
• deriveAgentId('Support Triage!', new Set(['support-triage'])) → support-triage-2; every output passes isValidAgentId.
• Every top-level key of both schemas has a description.

Generated with Devin