feat(app): add native sessions and durable local memory

[NarwhalChen]

Summary

• make Codex persistent threads and Claude native session IDs the authoritative owners of provider history, while Convera keeps a durable public transcript and crash-recoverable turn journal
• replace the experimental Letta/remote path with a local-only Off / Local memory engine; Local provides versioned blocks, archive search, correction, explicit forget, status, WAL recovery, and asynchronous subscription-native curation
• add a repeatable Off/Local evaluation runner that emits standalone JSON and HTML reports for deterministic memory semantics and paired real-Codex recall trials
• keep the subconscious curator text-only and tool-free, isolate memory/session state by conversation, actor, provider, and revision, and preserve source-agent provenance across multi-agent turns
• rotate or rebase native sessions only when provider, prompt/sandbox fingerprint, branch, reset, deletion, or memory version requires it instead of resending the entire transcript every turn
• apply one main-process host policy to Codex and Claude tools, bind web-bridge streams to their owning client, abort work on disconnect, and bound shutdown even when curation hangs
• harden local Electron automation discovery, loopback proxy handling, and executable launch behavior; fail early on unsupported Node 26 instead of reporting a package success without an application artifact

Product behavior

• Memory defaults to Off; existing local memory remains stored but is neither injected nor curated while paused
• Local memory uses the selected native subscription provider for a hidden curation pass after completed turns; it does not require Letta Cloud, a Letta subscription, or a separate backend
• memory context is injected only when a native session is created or the memory version changes
• destructive forgetting remains approval-gated and survives crash/restart through durable intent replay
• agents in the same channel share the public conversation while retaining actor-specific provider sessions, sandbox workspaces, and memory provenance
• pnpm --filter @convera/app memory:eval runs the deterministic benchmark; pnpm --filter @convera/app memory:eval:real adds paired real-Codex trials and writes local ignored artifacts under .automation/artifacts/memory-eval

Verification

• pnpm --filter @convera/app test: 64 files passed; 411 tests passed, with 1 gated real-provider test skipped by default
• real Codex persistent-session integration passed: a nonce created before runtime disposal was recovered after restart from the same native thread
• real Codex Off/Local A/B recall passed across three independent paired trials: Off recalled 0/3 and Local recalled 3/3; both modes satisfied their response contract in 3/3 trials
• the latest real run measured a mean Local input-token overhead of 162.67 tokens per turn and a mean latency delta of +271.12 ms; three trials are a functional smoke sample, not enough to establish a stable latency direction
• deterministic evaluation passed cross-restart recall, correction, user scope, workspace isolation, conversation isolation, and explicit forget; overall task accuracy was Off 33.3% versus Local 100%
• an actual hidden Electron automation smoke launched one Convera renderer, found 14 visible elements before and after an atomic Escape action, captured a non-blank welcome-screen screenshot, and closed cleanly
• automation TypeScript check, Prettier, git diff --check, and ESLint passed; ESLint reported 0 errors and 5 existing React hook warnings
• Electron Forge completed under temporary Node 20 and prepared Chromedriver 132.0.6834.210 for Electron 34; the default Node 26 path now stops before packaging with an actionable compatibility error

Boundaries

• no real Claude request was made
• the benchmark report and screenshot are intentionally local ignored artifacts rather than committed generated outputs
• opaque external MCP calls and unsandboxable host shell execution remain denied; computer control remains approval-gated
• RobotJS removal and default Cua MCP integration remain isolated in PR Replace builtin computer control with managed Cua MCP #212
• this PR is intentionally left open for review and does not merge itself into main

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

2 Skipped Deployments

Project	Deployment	Actions	Updated (UTC)
foxyfox-docs	Skipped		Jul 31, 2026 7:12am
foxyfox-website	Skipped		Jul 31, 2026 7:12am

[vercel]

@NarwhalChen must be a member of the codefox-repo team on Vercel to deploy.
- Click here to add @NarwhalChen to the team.
- If you initiated this build, request access.

Learn more about collaboration on Vercel and other options here.