[currentbits]

Adds the privacy gateway from #295. It's a deterministic filter (not a prompt) on every prompt Solenta hands a provider CLI, with a separate level for each provider. It's off by default, so nothing changes until you opt in under Settings › Advanced › Privacy gateway.

Closes #295

Changes

• Levels per provider, with an "every provider" fallback:
  • off
  • secrets: API keys and tokens (the guardrail rules), bearer tokens, whole PEM private keys, and this thread's secret-card values (Secret placeholders: orchestrator substitutes real values, agents never see them #289)
  • strict: everything in secrets, plus emails, card numbers (Luhn-checked), US SSNs, and your own regexes
• Where it applies: the final outbound prompt of every run, so it covers hand-off context, memory bootstrap, PDF text, folded queued messages, and turns started by worker forks and thread_send. It also covers steers, Ask, and /btw. Each match becomes [redacted:<rule>].
• Logged per thread: each turn with redactions gets an event row, e.g. Privacy gateway (strict, grok) redacted 2 items before sending: github-token ×1, email ×1. The row has rules and counts, never values. The transcript keeps the raw prompt you typed.
• The Export a thread as a self-contained, redacted bug report #394 bug-report export now uses the same redactor. Its rules and output are unchanged.

Verification

• New electron/test/privacy-gateway.test.js:
  • unit tests for levels, rules, Luhn, secret cards, and settings validation
  • a store round-trip
  • a runner test through a fake grok CLI, which checks that the -p prompt is redacted while the transcript keeps the raw text and gets the event row
• New renderer test in test/settingsModal.test.tsx for the editor's saves. The settings modal file passes 96/0.
• Full electron suite with a CI-like env (inherited CODER_* removed): 4965 pass / 0 fail / 0 cancelled.
• npm run typecheck and npx vite build are clean.
• Not tested: clicking through it in the packaged app.

Notes for reviewers

• Coverage limit (by design): the gateway only sees what Solenta puts together. Files and command output the agent reads itself go straight from the CLI to its provider. The settings note says this. Closing that gap would need an egress proxy in front of the CLI.
• Not covered: Build-workflow runs (workflowEngine) build their own prompts and aren't filtered yet.
• False positives at strict: a 13–19 digit number starting with 3–6 that passes Luhn still gets redacted as a card number. Millisecond timestamps start with 1, so they're safe.
• The redacted copy goes only to the CLI. Hiding secrets in the rendered transcript (raised in the issue comments) is a separate follow-up.
• About 650 changed lines, which is over the app's 400-line PR cap. Most of it is the new module, its tests, and the settings editor.

---

claude-opus-5-5 via Claude Code

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-19T09:33:16.684111Z	41297a8	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 41297a8e02

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Synchronize the pattern draft after settings hydrate

If the Settings modal mounts before the asynchronous settings request completes (or while a retry is pending), this initializer permanently captures an empty pattern list even after settings.privacyGateway.patterns arrives. The textarea then displays blank, and merely focusing and blurring it compares that blank draft with the loaded patterns and saves patterns: [], silently deleting the user's custom redaction rules. Synchronize patternsText when the settings prop hydrates or changes.

Useful? React with 👍 / 👎.