v0.1.38 — SMART-Code grows up: write, see why, search by meaning

Released 2026-05-25. Multi-platform signed/notarized DMG + EXE + AppImage. Auto-update from v0.1.37 happens automatically on next launch.

This is the SMART-Code maturity release. In v0.1.36 the assistant could only read your code. In v0.1.37 we fixed the regressions that made it look broken. In v0.1.38 it grows up — it can edit files (with your approval), it can search your codebase by meaning not just keywords, and the routing layer that picks between Claude and GPT-5 is now fully transparent.

Five items shipped:

1. Write tools + Approve/Reject diff modal — edit_file, create_file, delete_file
2. Confirm / Safe / Full permission modes — finally wired
3. GPT-5 with tools — DEEP_INTENT prompts no longer fall back to Anthropic when tools are attached
4. Codebase semantic index — Voyage voyage-code-3 + Supabase pgvector for cloud; Ollama nomic-embed-text + per-project SQLite for local
5. Routing transparency drawer — click the model pill on any SMART-Code answer to see which signals matched and why it routed where it did

What’s NOT in this release (deferred to v0.1.39 with a focused cycle): the speculative dual-execution half of item #5 — running Haiku in parallel with GPT and swapping in the better answer. The implementing agent stalled mid-task; we chose to ship the cleaner transparency drawer first rather than rush the parallel-streaming work.

1. Write tools — edit_file, create_file, delete_file

SMART-Code can now change your files. Before this release, every “fix this” ended in copy-paste from chat. Now the assistant proposes the change, you see a clean diff, and you click Approve or Reject.

Three new tools wired into the agent loop:

The safety list — what SMART-Code is NEVER allowed to write or delete:

• .env, .env.* (any dotenv)
• .ssh/, id_rsa*, id_ed25519*, *.pem, *.key
• .aws/, .docker/, .netrc, .gnupg/
• node_modules/, .git/, package-lock.json, yarn.lock, pnpm-lock.yaml
• Anything outside your project root

Hard cap: 1MB per file write. Reject larger.

The Approve / Reject modal:

Every write tool call surfaces a modal with the diff. For edit_file you see the file’s current content with the old_string highlighted red and the result with new_string highlighted green. For create_file you see the full new contents. For delete_file you see the file contents and a delete warning.

Click Approve → the IPC codenow-code:edit-file (or :create-file / :delete-file) runs the operation atomically (temp + rename) and the assistant gets a normal tool_result success. Click Reject → the assistant gets tool_result with is_error: true, content: "User rejected the edit. Try a different approach." and re-plans.

2. Permission modes — finally functional

The Confirm / Safe / Full pills in the topbar weren’t cosmetic — they already drove per-project AI-CLI autonomy (claude / codex / gemini / aider auto-approve). v0.1.38 extends them to also drive SMART-Code’s permission mode. One pill click, one mental model:

Persisted to ~/.codenow/settings.json key smartCodePermissionMode. Default: confirm. Setting survives restart.

3. GPT-5 with tools — no more Anthropic fallback

In v0.1.36 the smart router would route DEEP_INTENT prompts to GPT-5 — unless tools were attached, in which case it silently fell back to Anthropic Haiku 4.5. The whole point of routing is undermined when tool turns can’t reach the right brain.

v0.1.38 fixes this. streamOpenAI in site/api/codenow-code/chat.js now:

• Translates Anthropic tool schemas to OpenAI function-calling schemas (input_schema → parameters).
• Parses tool_calls deltas from OpenAI’s SSE stream (buffered per index).
• Emits the same normalized event: tool_use events as streamAnthropic — the renderer agent loop stays provider-agnostic.
• Translates tool_result continuation messages from Anthropic format to OpenAI’s role: 'tool' format with tool_call_id.

Net result: a DEEP_INTENT prompt like “refactor my auth across these 4 files” now routes to GPT-5 with tools, instead of falling back to Haiku.

Honest gap: the smoke rig drives the renderer only — there’s no fetch-mocking seam against the Vercel serverless handler, so the full server-side OpenAI tool-call SSE roundtrip isn’t asserted end-to-end. Server unit coverage belongs in a separate Node test rig (tools/smoke/server-chat.mjs) — queued. Per the no-mocks rule we did not fabricate renderer-side coverage that would claim more than we have.

4. Codebase semantic index — RAG-grade retrieval

Single biggest quality jump in this release.

Before: SMART-Code’s grep_code ran a regex across hundreds of files when you asked “where does login work?” Cursor and Windsurf give you the right 3 files in 200ms because they pre-indexed embeddings. v0.1.38 closes that gap.

How it works:

Two backends, mode-aware:

Lifecycle:

• On project open, CodeNow walks the project respecting .gitignore + the same safety list as the write tools. Chunks each file into ~500-token windows (200 token overlap). Caps: 10MB total content, 1000 files per project. Batches up to 128 chunks per Voyage request.
• Incremental re-index via chokidar file watcher (2s debounce per change).
• Per-user, per-project isolation — Supabase RLS scopes every row to the authenticated user; the codenow_code_index_search RPC also enforces user_id explicitly so even an accidental missing filter at the API layer can’t leak across users.

New tool: semantic_search(query, topK?=8)

Wired into the agent loop. When SMART-Code wants to find relevant code, it can now call semantic_search("authentication and login flow") and get back the top-K most semantically related chunks — not a regex grep. grep_code stays for exact-string searches (file path, function name, error message); semantic_search is for natural-language queries.

For local-mode users: run ollama pull nomic-embed-text once. CodeNow detects Ollama on http://localhost:11434 and uses it. No internet round-trips.

Cost (cloud users): ~$0.04-0.12 to index a 500-file repo once; cents per session for incremental re-index. At 1,000 users: ~$120/month for indexing. Pre-work is well inside the $9.99/seat economics.

5. Routing transparency drawer

The model-tag pill on every SMART-Code answer (e.g. SMART-Code · GPT) is now clickable. It opens a right-side drawer that shows the routing decision in full:

• Header: Routing decision · <provider>/<model>
• Reason line: Routed to <provider> because: <decidedReason> matched (or Manually pinned by user if explicit, or Default: speed wins for short turns if no signal matched)
• Signal table — six rows, one per signal:

Each row shows whether it matched (✓ / ✗) and the evidence — the matched substring with ~20 chars of context (e.g. "...let's refactor the auth..."), or total: 9482 chars for HUGE_CONTEXT, or 7 user turns for LONG_THREAD. The winning row is highlighted.

ESC, backdrop click, or X button to close.

Why this matters: other tools route models silently. You ask a question, an answer appears, and you have no idea why this lab vs that lab. CodeNow’s whole positioning is we show you. This drawer makes the routing decision auditable on every single turn.

The matching server-side change: smartRoute(messages) now returns { provider, trace, decidedReason } instead of a bare string, and the first SSE event of every chat stream is event: routing with the trace + reason + a isExplicit: true|false flag for user-pinned cases.

What’s coming in v0.1.39

Three items, focused:

1. Speculative dual-execution (the deferred half of v0.1.38 item #5) — when the router picks GPT-5 (slow path), fire Haiku in parallel; show Haiku’s draft instantly, swap to GPT only when GPT’s answer disagrees substantially. Perceived speed jump vs Windsurf without owning their inference stack. Cost-control gates: disabled when permissionMode === 'full', disabled when tools are attached, env kill-switch SPECULATIVE_DUAL_EXEC_DISABLED.
2. ServiceNow Build Agent SDK kit — embed as a first-class harness in the kit picker.
3. Salesforce Agentforce DX kit — same.

Both #2 and #3 need OAuth-against-tenant + deploy-first-then-test flows, which is structurally heavier than the existing five local-process kits. Scoped here so v0.1.39 has a coherent enterprise-harness theme.