A fork of jcode fusing in the best orchestration, safety, and swarm ideas from OpenAI's Codex Harness and xAI's Grok Build.
What this is · What shipped · Build docs · Status · Credits
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.
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.
The full build history is preserved, in order:
DESIGN.md— the original design doc: why jcode was chosen as the base, what was ported from where, and whyclaude-code-build/SESSION_1_MEMORY.md— session 1 narrativePROGRESS.md— phase-by-phase build logclaude-code-build/SESSION_2_MEMORY.md— session 2 narrativeclaude-code-build/PROJECT_FINAL_SUMMARY.md— the short version of all of the above
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.
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.