v1.0.0

Noah Code v1.0.0

See model output as it arrives, choose your permission mode during setup, and connect with a
Codex / ChatGPT account. Noah also gains an HTTP service, editor integration through ACP,
persistent child tasks, optional child worktrees, and browser tools.

Faster feedback and easier setup

• Completion providers now stream incremental generation into the terminal and service event
  stream. Previews are provisional and may contain CodeAct code; finalized messages remain the
  authoritative response. The Responses API transport still buffers its output.
• Guided provider setup offers searchable model discovery, bundled catalogs where appropriate,
  and manual entry when discovery is unavailable. Custom OpenAI-compatible endpoints use only
  their explicitly configured credentials.
• First interactive setup asks users to choose Normal, Auto, or YOLO, with a one-line
  explanation of each. Normal is the default. The choice is saved in trusted user configuration;
  --permissions normal|auto|yolo overrides it for one launch. Existing --auto and --yolo
  flags remain available. Auto still blocks interpreter commands; YOLO skips Noah permission
  checks and approval prompts without disabling the execution sandbox.
• Codex / ChatGPT account is available in setup and /model, or through
  noah providers login codex. Browser sign-in and --device-code are supported, with a
  separate private Codex credential directory. This requires the official Codex CLI 0.153.4
  or newer and uses experimental app-server fields. Noah retains control of its own tools and
  permissions. Account usage follows Codex limits and is not reported as API-dollar spend.
  Protocol behavior is covered by automated tests; live account sign-in was not exercised.

Services, editors, and independent child tasks

• noah serve exposes persisted sessions, prompts, cancellation, history, approvals, recovery,
  and replayable SSE events through a shared service boundary. It defaults to loopback, requires
  a bearer token, and rejects browser-origin requests. See the service guide.
• noah acp connects ACP v1 editors over stdio, supporting new and loaded sessions, prompts,
  cancellation, permission requests, and finalized conversation and tool updates. Capabilities
  are advertised only where implemented.
• Child tasks save their conversations and status, support follow-ups and independent
  cancellation, and can be inspected with /tasks while the parent is busy. Background writers
  require isolated Git worktrees starting from committed HEAD. Changes remain available
  for review and explicit integration. Children stop when Noah exits; persistence does not
  automatically restart them. See persistent tasks.
• noah browser setup configures a pinned Playwright MCP server with an isolated browser
  profile. Browser actions follow Noah's MCP permissions. Lifecycle hooks now observe session,
  turn, and worktree events, alongside existing tool hooks, skills, and MCP extensions.

More reliable edits and verification

• Sandbox execution preserves multiline string literals instead of inserting indentation into
  their contents. Bounded tool help exposes signatures, return shapes, and examples to agents.
• Workspace tools compile Python edits before writing and reject an entire patch batch on a
  syntax error. Shell writes remain outside this preflight.
• When an execution changes the workspace, observed failed, stale, or incomplete checks prevent
  the agent from declaring completion until it reruns them or reports a blocker. Unittest
  commands are recognized, and generated caches no longer stale their own successful results.
• Empty responses caused by output-token exhaustion receive bounded continuation within the
  existing iteration and budget limits, with original usage and response evidence preserved.

Reproducible comparison

The new evaluation harness runs Noah and OpenCode against identical fixture tasks, models,
prompts, budgets, and frozen acceptance tests. In the final trial, both agents passed both
fixtures and completed successfully. Earlier failures, the harness permission correction, and
known and unknown charges remain documented. These two small tasks validate targeted repairs;
they do not establish a general coding-quality ranking. See the results
and reproduction guide.

Upgrade

noah update

See configuration for provider sign-in, permission choices, and transport
limitations. Browser tools require Noah's MCP extra and a compatible local Node/browser setup.

Full changelog: v0.8.0...v1.0.0