v0.1.46 — Token Economy Index, persisted workspace, AGP as the built-in governance layer

Released 2026-05-29. Multi-platform signed/notarized DMG + EXE + AppImage. Auto-update from v0.1.45 happens on next launch.

This is the economics + governance release.

For ten releases since v0.1.38 the SMART-Code surface has been the headline. v0.1.46 turns the camera around: instead of what the assistant can do, this release is about what every turn costs you — and what runs alongside every agent in your Fleet to keep it safe. Two themes:

1. Token Economy Index (Phase 0 + Phase 1) — buildRepoContext now returns tiered context, sendCloud emits Anthropic prompt-cache control markers at static / semi-static boundaries, and the routing drawer gets a per-turn token-economy panel.
2. AGP (Agent Governance Plane) by TrustModel, built-in — Edge sidecar Docker runner, standalone AGP CLI, Fleet-level governance scaffold, and a verified /install API contract against the QA control plane. Every agent in your Fleet view will get a TrustScore next release.

Plus a long-overdue piece of polish: open editor tabs and SMART-Code chat history now persist across project switches and across app restarts. CodeNow now reopens where you left it.

What changed since v0.1.38

The wiki skipped seven releases. Here is the arc in one screen, before the v0.1.46 deep-dive.

The throughline: v0.1.39–v0.1.42 expanded what the assistant knows; v0.1.44–v0.1.46 wire that knowledge into a cost-aware, governed runtime.

v0.1.46 — Token Economy Index

The single most expensive thing CodeNow does is mint Anthropic / OpenAI tokens on your behalf. Until this release, every SMART-Code turn re-sent the full repo context as a fresh, uncached system prompt. Long sessions paid full input price for context that hadn’t changed in 40 minutes.

v0.1.46 ships Phase 1 of the Token Economy Index — the structural fix — and Phase 0 of the observability layer that proves it.

Phase 1 — Tiered context + Anthropic prompt caching

buildRepoContext used to return one flat string. It now returns:

{
  static:     '...persona + project ID + CLAUDE.md...',
  semiStatic: '...knowledge bundle + dir listing...',
  volatile:   '...active file path + last 15 ledger entries...',
}

sendCloud assembles the system field as Anthropic content blocks with cache_control: { type: 'ephemeral' } markers at the static / semi-static boundaries:

The boundary at the end of semi-static means Anthropic computes the cache key over static + semi-static. If only volatile changed (you typed a follow-up), the prefix hits cache and Anthropic bills cache-read price (~10% of input) instead of full input price.

Targets (not yet measured in prod):

The Phase 1 wiring is in sendCloud (site/api/codenow-code/chat.js) and buildRepoContext (renderer-side services/repoContext.js).

Phase 0 — Per-turn usage panel in the routing drawer

The routing drawer that v0.1.38 introduced now has a second panel: Token economy (this turn).

Server emits a done SSE frame at the end of every chat stream. v0.1.46 extends that frame to optionally include the usage object Anthropic / OpenAI return:

event: done
data: {
  "usage": {
    "input_tokens": 18421,
    "output_tokens": 612,
    "cache_creation_input_tokens": 18033,
    "cache_read_input_tokens": 0
  }
}

The renderer parks this on routingInfo.usage. The drawer renders:

Important honesty constraint: the server may not always emit usage (older serverless deploy, OpenAI route with tools where the SDK strips it). The panel is conditionally rendered. If no usage object arrives, the panel hides — we do not fabricate 0 / 0 / —. No faked numbers.

Strategy reference — $ / resolved-task

Internal note for anyone reading the code: the published USP for SMART-Code is $ per resolved task, not $ per token. Token Economy Index is the instrument; $ / resolved-task is the axis. Phase 2 (next release) wires the resolved-task counter — until then, the drawer reports tokens; the pricing page reports tasks; the two are not yet joined.

v0.1.46 — Persistence (open tabs + chat history)

A long-standing rough edge: switch project, lose context. v0.1.46 closes it.

Open editor tabs + active file

settings.projectStates[projectId] now stores:

{
  openTabs: ['src/auth.ts', 'src/auth.test.ts', 'README.md'],
  activeFile: 'src/auth.ts',
  lastOpenedAt: 1716998400000,
}

• Written on every tab open / close / focus change (debounced 500ms).
• Read by switchProject(projectId) — restored tabs open in the same order, active file gets focus.
• Survives app restart (it’s in ~/.codenow/settings.json).

If a file in the saved set no longer exists, the restore quietly drops it (no error toast — the project moved on).

SMART-Code chat history

Per-project, per-mode, capped at 100 messages:

<projectRoot>/.codenow/sessions/smart-code-cloud.json
<projectRoot>/.codenow/sessions/smart-code-local.json

Restored on project open. Cleared by the existing Clear chat action in the SMART-Code header (it now also unlinks the file on disk).

Why per-project on disk and not in ~/.codenow/settings.json: chat transcripts can hit hundreds of KB; settings.json is hot-read on every app start. Keeping chat under <project>/.codenow/ also means a project move-or-share carries its own history.

v0.1.46 — AGP by TrustModel, built-in

The strategic shift in this release: TrustModel’s AGP (Agent Governance Plane) is now CodeNow’s built-in agent-governance layer. Not an integration. Not an add-on. Built-in.

Three pieces shipped:

1. AGP client + Fleet governance scaffold (TRUS-1042)

A new agp/ module under the renderer wraps the TrustModel /install API and exposes a typed client:

The Fleet view (introduced v0.1.40) now reserves a column for Governance — empty in this release, populated by the TrustScore badge UI shipping in v0.1.47.

2. Edge sidecar — Docker runner + IPC layer (TRUS-1044 / TRUS-987)

Every agent in the Fleet runs inside a TrustModel AGP Edge sidecar:

The runner is opt-in per agent in this release — flipping the governance: 'agp' flag in an agent’s config wires the sidecar at next start. Default is governance: 'none' to keep existing Fleets unchanged.

3. Standalone AGP CLI

codenow agp is a standalone subcommand that launches an Edge sidecar without a parent CodeNow process — for CI agents, ServiceNow Build Agent harnesses, and any externally hosted agent that wants TrustScore coverage.

codenow agp run \
  --agent-id <id> \
  --policy <policy-id> \
  --endpoint https://api-trustmodel.pdxqa.com

Same Docker image, same IPC contract, no CodeNow desktop dependency.

4. /install API contract — verified on QA

Wire-level contract is verified live against https://api-trustmodel.pdxqa.com:

Production endpoint (api.trustmodel.ai) accepts the same payloads; the CodeNow client switches via AGP_ENDPOINT env override.

Result

Every agent that runs in Fleet view from this release forward is eligible for a TrustScore. The score badge ships in v0.1.47. The plumbing — sidecar, IPC, install, telemetry — is in v0.1.46.

This is the first time an IDE has shipped agent-governance as a built-in primitive rather than a Marketplace add-on.

v0.1.39 → v0.1.45 changelog

Terse — for the reader who skipped the wiki for seven releases.

v0.1.39 — Speculative dual-execution + Codebase Map

• Speculative dual-exec: when the router picks GPT-5, fire Haiku in parallel. Show Haiku’s draft instantly; swap to GPT only if the answers disagree substantially. Disabled when tools are attached or permissionMode === 'full'. Env kill-switch: SPECULATIVE_DUAL_EXEC_DISABLED.
• Codebase Map dashboard: a new tab under the project view rendering the v0.1.38 semantic index as a clustered force-graph. Click a cluster to filter SMART-Code to that subtree.
• Routing drawer gains a speculative_swap row when Haiku draft was discarded in favor of GPT.

v0.1.40 — Agent Fleet + Agent Foundry rename

• Agent Fleet: cross-project view of every agent across every CodeNow workspace on this machine. One scrollable list, filter by project / state / kit. The view that v0.1.46’s AGP governance column plugs into.
• Agent Foundry: the build surface (formerly “Build tab”) renamed. Self-serve agent creation — pick a kit, name it, push to a project — no terminal required.
• Build-tab self-serve uses the same kit picker as the existing five local-process kits (claude, codex, gemini, aider, cursor).

v0.1.41 — SIGHUP fix + Fleet UX

• Permanent fix for the job-control / SIGHUP warning that v0.1.19 partially addressed. Long-running PTY children (npm run dev, vitest —watch) no longer write bash: warning: shell level (N) too high to stderr on focus change. Root cause was the wrapper script trapping SIGHUP and re-emitting; v0.1.41 detaches the PTY with setsid on Linux/macOS and the equivalent CREATE_NEW_PROCESS_GROUP on Windows.
• Fleet view: keyboard nav (j/k), bulk-state filter, last-touched sort.

v0.1.42 — Memory ledger + Skillify + Promotion modal

• EPIC-A: append-only memory ledger. SMART-Code emits <codenow-memory>...</codenow-memory> sigils that the renderer harvests post-stream and writes to <project>/.codenow/memory/ledger.jsonl. Used by buildRepoContext (v0.1.44).
• EPIC-B: Skillify — ~/.codenow/skills/ wiki + retriever. Skills are markdown files with frontmatter; the retriever does BM25 over titles + descriptions and surfaces the top-K into the system prompt.
• EPIC-C: Promotion modal. When a chat thread looks like it should become a reusable skill, SMART-Code suggests promotion — one click writes the markdown file.

v0.1.43 — Re-tag

• v0.1.42 was originally tagged v0.1.42.1 for an auto-updater fix. The 4-segment semver broke the channel-detection regex in electron-updater (it expects \d+\.\d+\.\d+). Re-cut as v0.1.43 with identical binary contents.
• No code change vs. v0.1.42.

v0.1.44 — SMART-Code cross-agent context + scanner re-tune

• buildRepoContext absorbs the v0.1.42 knowledge bundle (Skillify retriever output) and memory ledger (last 15 entries). This is the data structure v0.1.46 then tiers + caches.
• Auto-detect scanner: three new framework detectors (SvelteKit, Astro, RedwoodJS). HealthAI false-positive (it was matching on the literal string “health” in package descriptions) fixed by requiring a package.json dep match.

v0.1.45 — Cursor-missing fix + FLT-9 + browser polish

• Cursor-missing bug: when a Claude+Terminal tab’s PTY exited (cursor binary upgraded out from under the agent), the tab silently froze. Now detects PTY exit, surfaces a one-line banner, and auto-rebuilds the tab.
• FLT-9: built-in agents (the five local-process kits + Foundry-created ones) now appear in Fleet view alongside externally registered ones.
• Browser default flipped to codenow.pro (was google.com).
• file:// scheme support in the embedded browser — local previews open without a dev server.
• FLT-7: per-framework scan toggle in project settings. Disable detectors you don’t care about; the scanner skips them on next walk.

Upgrade notes

User-facing behavior that changes in v0.1.46:

Auto-update from v0.1.45 happens on next launch. No migration steps.

What’s next — v0.1.47

Three items, scoped:

1. Token Economics Phase 2 — proactive retrieval. Instead of stuffing the full knowledge bundle into semi-static every turn, run a fast retrieval pass first and only inject the top-K relevant skills. Cuts semi-static size; raises cache-hit rate further; cuts first-turn cost. Target: another 2× input cost reduction on long sessions on top of the Phase 1 cache hit.

2. Cost Cockpit UI. A dedicated dashboard (under the Codebase Map’s neighbor tab) rolling up the per-turn usage panel into per-session, per-day, per-project cost. Joins to the resolved-task counter so $ / resolved-task becomes a number you can see, not just a tagline. Depends on Phase 0 emitting usage consistently across providers — that work is queued separately.

3. TrustScore badges in Fleet view. The UI half of v0.1.46’s AGP plumbing. Every Fleet row gets a 0–100 TrustScore badge with a click-through to the governance policy + recent violations. Edge sidecar is already emitting the telemetry — this release just adds the renderer.

Phase 0 telemetry from v0.1.46 — the actual cache-hit %, the actual input cost reduction — will be reported as numbers in the v0.1.47 release notes. No targets posing as results.