nanobot-ts/memory-bank/activeContext.md

Active Context

Current Focus

Docs directory created with 4 files (PRD.md, Architecture.md, API.md, Discoveries.md). All source files previously written and verified — typecheck and lint are both clean.

Session State (as of this writing)

• All source files complete and passing tsc --noEmit (0 errors) and oxlint (0 errors, 0 warnings)
• package.json scripts added: start, dev, typecheck
• Ready for runtime / integration testing

Key Fixes Applied This Session

• Zod v4 .default(): nested object schemas need factory functions returning full output types (e.g. .default(() => ({ field: value, ... })))
• Zod v4 z.record(): requires two args: z.record(z.string(), z.unknown())
• AI SDK v6: LanguageModelV2 → LanguageModel; maxTokens → maxOutputTokens; maxSteps → stopWhen: stepCountIs(1); usage fields: inputTokens / outputTokens (not promptTokens / completionTokens)
• ollama-ai-provider v1.2.0: returns LanguageModelV1 — cast with as unknown as LanguageModel
• js-tiktoken: get_encoding → getEncoding
• web.ts: Document global not available in Bun/Node — return Record<string, unknown> from makePseudoDocument, cast at call site
• loop.ts: syntax error ContextBuilder(workspace: ...) → new ContextBuilder(opts.workspace)
• lint: all ${err} in template literals → ${String(err)}; String(args['key'] ?? '') → strArg(args, 'key') helper; unused onProgress param → _onProgress; WebSocket onerror err type is Event → use err.type

Work Queue (next steps)

1. Create workspace helper module (src/cli/utils.ts) with ensureWorkspace() and syncTemplates()
2. Create onboard command (src/cli/onboard.ts) with path argument and directory-not-empty guard
3. Agent/gateway commands check workspace exists (throw if not found)
4. Added required provider field to agent config (values: anthropic, openai, google, openrouter, ollama)
5. Provider resolution uses explicit provider from config (no model prefix parsing)
6. Typecheck and lint pass (0 errors)
7. Test onboard and agent commands work correctly
8. Test with a real Mattermost config (optional — user can do this)

Key Decisions Made

• Mattermost channel uses raw WebSocket + fetch (no mattermostdriver, no SSL hack)
• No MCP support (use shell tools / CLI instead)
• No reasoning/thinking token handling (can add later)
• Config is fresh Zod schema (no migration from Python config needed)
• ollama-ai-provider package (not @ai-sdk/ollama which 404s on npm)
• strArg(args, key, fallback?) helper exported from agent/tools/base.ts for safe unknown→string extraction
• Agent config requires explicit provider field (no more model prefix like "anthropic/claude-...")
• Model names are now just the raw model ID (e.g., "claude-sonnet-4-5" not "anthropic/claude-sonnet-4-5")