Cross-LLM Delivery
jhesham v0.4.2
Test-gated build delivery
Save this plugin to Following and return to review observed changes. Follow up to 10 plugins free.
Description
From the marketplace listing
Codex leads delivery of test-gated slices and the Codex CLI (provider: codex) implements each one headlessly. The lead writes contracts and committed acceptance tests; the engine independently verifies each Git candidate, checks its writable allowlist and integrates only accepted commits. For developers who already use Codex and want contract-driven, multi-slice implementation with independent acceptance; debugging and tiny edits are better done directly. Requirements: Python 3.11+, Git and an installed, authenticated Codex CLI on your own account; no account access or models are provided. Use a standalone skill for local Codex CLI or IDE workflows. Plugin discovery depends on your host version and provider; executor plugin installation and IDE plugin discovery remain unverified. See the project's support matrix for recorded evidence. ChatGPT web is not a supported execution surface. Runs locally with your user privileges; slices use Git worktrees, which are not a security sandbox. The selected provider may receive prompts and repository content under your account terms, and usage is billed to your plan or provider account. Read the bundled PRIVACY.md and SECURITY.md for data handling, retention and support. Project details: https://github.com/jhesham/cross-llm-delivery. Independent project; not affiliated with or endorsed by OpenAI. Product names are trademarks of their owners.
Language: English · Automatically detected from descriptions.
Competitor profile
From the developer’s listing and package; capabilities are not independently tested. Package observed Nov 2, 2026.
Declared capabilities
- Test-gated slice delivery
- Independent acceptance
- Verified Git integration
Declared limitations
Specific limitations have not been established. Missing information does not mean unrestricted capabilities.
Cost & access
Pricing and account requirements have not been established from the available package declarations. Unknown pricing does not mean free access.
Starter prompt
A starter prompt has not been provided in the catalog.
Files & skills
File archives
Skill instructions
cross-llm-codex4.28 KB
--- name: cross-llm-codex description: >- Use when a large build has been decomposed into a multi-slice plan (contracts, acceptance tests, dependency DAG) so the codex headless executor implements each slice while Codex orchestrates and judges. --- <!-- GENERATED from cross-llm-delivery (provider: codex, v0.4.2) - do not edit here; edit the monorepo source. --> # Cross-LLM Delivery -- Codex host, codex executor **Codex is the lead.** You decompose the build into testable slices, fix the contracts, author the acceptance tests, judge every result, and integrate verified work. The **codex headless executor** does the bulk implementation typing inside isolated git worktrees. The vendored `cld` engine and driver under `scripts/` run in place -- no package install and no source checkout are needed. An exact model ID is required; this provider has no default model. Pass `--executor codex:<model-id>@<effort>` explicitly. For requested fast mode, use `codex:gpt-6-luna@max+fast`; do not silently drop the tier. For executor choices, run `python scripts/list_models.py --json` and follow the picker section of `references/delivery-core.md`. Present exact returned IDs and supported efforts; discovery does not authorize dispatch. ## When to use (and when not to) - **Use it** on a large build expressed as a plan of independently testable slices with a dependency DAG, where executor tokens dwarf lead tokens. - **Do not use it** for small one-file fixes; the per-dispatch overhead only pays off on large builds. Small work stays with Codex directly. ## Ground rules - Respect the user's existing project instructions (AGENTS.md and friends). Never install, overwrite, or replace them. - Preserve the user's exact executor/model/effort/tier selection. Present the shortlist only when no selection exists; a listing is not dispatch permission. Describe discovery only for the recorded host version and surface. - Never claim the provider's cost is free. Check that billed dispatches are covered by the user's existing authorization; ask only if that is unclear. - For gate-4 repairs, follow the user's authorization and the host's active approval policy. Never silently dispatch another paid attempt. - `--step` and validation make paid network calls through the executor CLI. Check the active host permissions before dispatch. If network is disabled, use the host's supported authorization flow; do not bypass its restrictions. CLD blocks dispatch when `CODEX_SANDBOX_NETWORK_DISABLED=1` is set. ## Drive the build Run these commands from the installed skill directory, using absolute plan and target-repo paths. The driver is vendored in this bundle. Prefer `--json` for machine-checkable gates and act on the JSON `gate_code`. On Windows a shell wrapper (`pwsh -Command`) can report exit 1 for any non-zero driver exit, so the JSON gate, not the shell's exit status, is authoritative. ```bash python scripts/run_delivery.py <plan.md> --repo <dir> --dry-run --json # preview layers python scripts/run_delivery.py <plan.md> --repo <dir> --step --workers 1 --executor <exact-spec> --validation-policy <deny|unmetered|allow> --budget-attempts <N> --json python scripts/run_delivery.py --status --repo <dir> --json # cheap digest python scripts/run_delivery.py <plan.md> --repo <dir> --integrate --integration-tests <selector> --json ``` Choose the validation policy and cumulative attempt limit under the user's authorization. Validation and retries count as dispatches. Prefer one worker for a bounded sitting; the CLI default is four. Exit codes: 0 ok, work remains; 2 failure/defer; 3 integrated and verified; 4 lead repair required; 5 invalid plan/state or missing prerequisite; 6 accepted work awaits integration. ## References - `references/delivery-core.md` -- the shared, host-neutral gate (exit-code) contract and authorization rules; identical in every host bundle. - `references/codex-workflow.md` -- the full Codex-led workflow: plan authoring, batch-stepping, the model picker, gates, and the repair loop. - `references/provider-setup.md` -- install and verify the codex CLI. - `references/provider.md` -- codex executor specifics: locked invocation, auth, cost, and platform notes. - [README.md](README.md) -- data handling, privacy, security reporting and a no-inference reviewer preview of this bundle.
Referenced files: 68 — file references are metadata, not package contents.
Package details
Publisher declarations from the archived package. These are separate from our research and the live service's terms.
- Package license
- MIT
- Package author
- jhesham
Declared capabilities
- Test-gated slice delivery
- Independent acceptance
- Verified Git integration
Package observed Nov 2, 2026.
Technical details
- First seen
- Oct 31, 2026 · 06:00 UTC
- Last seen
- Nov 3, 2026 · 18:00 UTC
- Collection status
- Collected
plugins_6ac4ce7367e88191bdf6378dc7d46962
Download plugin data (JSON)