[LinuxDevil]

What changed

proposeAgent({ draft }) — the Author's create-proposal tool. It validates the draft via AA07's draftToSpec, checks every catalog-referenced value (model + fallback models, builtin tools, guardrail names, trigger types, subagents.<key>.ref and handoffs[i].agent against workspace agent ids) in a single failure that lists up to 10 problems with the first problem's code, computes the card's readiness rows via the requirements dep ([] on dep error), runs the injected reviewDraft (passReview until the real review lands), records the card under the 'create' identity in per-turn proposal state, and returns the ProposalOutput. proposalId is the call's toolCallId; nothing is written to disk.

New modules under server/author/proposal/:

• turnState.ts — ProposalIdentity, identityKey, ProposalTurnState, proposalTurnState (lazily stored in turn.state['proposals']).
• review.ts — ReviewInput/ReviewVerdict/ReviewDraft contract + passReview default.
• catalogChecks.ts — CatalogFacts, loadCatalogFacts (memoized per turn in turn.state['catalogFacts']), checkCatalogs.

Also:

• shared/authorWire.ts — ProposalKind, ProposalOutput; SpecChange re-exported alongside Requirement.
• tools/index.ts — AuthorToolDeps gains reviewDraft and requirements; proposeAgentTool registered in AUTHOR_TOOL_FACTORIES (now 14 tools).
• app.ts — createApp wires reviewDraft: passReview and requirements to computeRequirements with requirementDeps resolved per call against the session's workspace root (workspace seam preserved).

How it was tested

• npm run typecheck:server, npm run typecheck — clean.
• npm run test:server — 1020 tests pass (76 files), incl. new proposeAgent.test.ts (8) and catalogChecks.test.ts (14).
• npm test -- --run — 946 client tests pass.
• Root: npx tsc --noEmit, npm run lint, npm run lint --workspace apps/agent-forge (0 errors), npm run test:coverage (7152 pass; coverage report written), npm run fallow (0 above threshold), npm run docs:ratchet, npm run api:check, npm run docs:llms:check, npm run size-check — all green.
• Note: the first test:coverage run showed 11 failures unrelated to this diff — 5 in packages/create-lousho (its dist/ was unbuilt) and 6 src/spec/v2/templates.test.ts README assertions caused by CRLF checkout line endings on Windows. Both were environmental; after building the sibling package and normalizing the checkouts to LF, the full suite passes with 0 failures. No content changes were committed for either.

Acceptance criteria

• A valid draft returns ok: true with proposalId === ctx.toolCallId, kind: 'create', enabled: false, a chat trigger present, and nothing written to .lousho/agents/.
• A draft with an unknown model and an unknown tool returns one UNKNOWN_MODEL failure listing both problems.
• A review returning pass: false makes the tool return that failure and leaves proposed empty.
• After a successful call proposalTurnState(turn).proposed.get('create') is the output.

Generated with Devin