Skip to content

Repository files navigation

jcode-fusion

A fork of jcode fusing in the best orchestration, safety, and swarm ideas from OpenAI's Codex Harness and xAI's Grok Build.

License: MIT Base Status

What this is · What shipped · Build docs · Status · Credits


What this is

jcode was chosen as the base harness over Codex Harness and Grok Build for one reason: multi-provider OAuth / multi-account auth. Neither competitor has it — both are single-vendor tools tied to their own subscription or API billing. jcode's 40+ provider profiles with built-in OAuth login is the entire reason a "bring your own subscriptions, no API billing" harness is possible at all — so it's kept completely untouched here, not one feature among several.

On top of that base, this fork ports nine specific features from Codex Harness and Grok Build across five phases — reimplemented in idiomatic jcode-style Rust, not copy-pasted (jcode is MIT, both donor projects are Apache-2.0).

Full rationale for every decision — including what was rejected and why — lives in DESIGN.md.

What shipped

Nine features, ported from two donor harnesses across five phases — each reimplemented in idiomatic jcode-style Rust, never copy-pasted (jcode is MIT, both donors are Apache-2.0):

Phase Feature Source What it does
0 — Foundation Mission Engine Codex Harness + Grok Build Budget-aware state machine (token/wall-clock accounting, auto-continuation) + a rule that an agent can never grade its own "done" — a separate verifier must confirm completion.
0 — Foundation Provable-safe rewind Grok Build Rewind checkpoints that refuse to guess rather than reconstruct unsafely after compaction.
1 — Safety Guardian Codex Harness LLM-as-judge auto-adjudicating sandbox-escalation requests — deny-only, fails closed.
1 — Safety Layered OS sandboxing Codex Harness Whole-process macOS Seatbelt sandboxing (re-execs the entire binary, not just subprocess calls).
1 — Safety Execpolicy-as-Starlark Codex Harness Starlark-based command classification layered on top of jcode's existing risk classifier, never downgrading it.
2 — Swarm rework Worktree-per-subagent isolation Codex Harness + Grok Build (convergent) Every spawned agent gets its own git worktree; conflicts resolve via plain git merge instead of mid-edit negotiation. A real apply action merges a worker's commits back into the coordinator's tree.
3 — Ecosystem Full Agent Client Protocol Grok Build Session lifecycle, session-scoped MCP servers, bidirectional client callbacks over ACP.
3 — Ecosystem Orchestration-as-script Grok Build workflow tool: saved, replayable, parameterized multi-agent task graphs.
4 — Memory Two-phase consolidation Codex Harness Leased/claimed background extraction jobs per session (with retry backoff), then a locked-write consolidator rendering a global MEMORY.md.

Every phase is complete per the original DESIGN.md scope. The full per-feature rationale — including convergent cases like worktree isolation, where both donors independently arrived at the same idea — is in DESIGN.md §6.

Quality process: two full-repo AI code scans surfaced 54 findings, every one triaged — fixed, confirmed pre-existing, or deliberately deferred, never rubber-stamped. Real bugs caught before shipping: a cross-session ACP callback forgery, a lock-contention leak, a permanent-data-loss regression in memory consolidation, and others. The recurring lesson: review the fix, not just the original bug — several "fixed" issues had a second, subtler bug hiding inside the fix itself, caught only by a follow-up adversarial pass.

Build docs

The full build history is preserved, in order:

  1. DESIGN.md — the original design doc: why jcode was chosen as the base, what was ported from where, and why
  2. claude-code-build/SESSION_1_MEMORY.md — session 1 narrative
  3. PROGRESS.md — phase-by-phase build log
  4. claude-code-build/SESSION_2_MEMORY.md — session 2 narrative
  5. claude-code-build/PROJECT_FINAL_SUMMARY.md — the short version of all of the above

Current status

Paused, not abandoned. All five phases are complete, tested, and reusable as-is. The one open item — a persistent 429 on the Antigravity provider — was investigated exhaustively (see PROJECT_FINAL_SUMMARY.md for the full seven-hypothesis writeup) and traced to a real, already-acknowledged Google bug affecting Jio-managed Google AI Pro subscriptions, tracked upstream at Google Issue Tracker #525093265 — not a jcode-fusion (or jcode) code problem.

Picking this back up: git clone this repo and read the build docs above in order.

Credits

  • 1jehuang/jcode — the base harness this fork builds on. MIT license.
  • openai/codex (Codex Harness) — source of the Guardian, layered OS sandboxing, execpolicy-as-Starlark, two-phase memory consolidation, and (jointly with Grok Build) Mission Engine and worktree-per-subagent isolation. Apache-2.0.
  • xai-org/grok-build — source of provable-safe rewind, full ACP support, orchestration-as-script, and (jointly with Codex Harness) Mission Engine and worktree-per-subagent isolation. Apache-2.0.

This is an independently maintained fork, not an upstream contribution — jcode's CONTRIBUTING.md discourages large external PRs, so this project is planned and resourced as a hard fork from day one.

About

A fork of jcode merging the best orchestration/safety/swarm features from OpenAI's Codex Harness and xAI's Grok Build. Full design doc in DESIGN.md.

Resources

Contributing

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages