Thanks to visit codestin.com
Credit goes to github.com

Skip to content
 
 

Repository files navigation

elmakus codex_workflow fork

Version 1.1.17-private.5, based directly on upstream prerelease v1.1.17 commit 414a5d301ff17ca6e655330474c8346863d0d5d0.

This public fork keeps upstream's lifecycle/runtime foundation while applying owner-specific orchestration, model, update-channel, and safety choices.

Private behavior

  • Heavy is the only workflow route. Leaf-state questions and small bounded tasks work directly without subagents and without reading heavy_route.md.
  • Deployment state has a hard user-visible output gate in AGENTS.md: if work can continue safely without user input, Main remains silent until the final response. Intermediate findings, plan changes, worker results, repository state, and next-step narration are not reasons to speak.
  • Heavy uses progressive disclosure: standing orchestration stays in heavy_route.md; detailed work-package, Micro Execution, follow-up, and recovery guidance lives in delegation.md and is loaded only when needed.
  • Worker compute is selected by one global runtime profile. plus preserves the historical Luna-heavy allocation. pro-x5 moves ordinary worker roles to GPT-5.6 Sol Low while keeping Senior Executor at Sol Medium. Main is never changed by the profile.
  • Micro Executor is a distinct seventh worker below Default Executor. Its stable fallback is Luna High in plus and Sol Low in pro-x5; when the current Codex runtime/account supports it, the same role may run with GPT-5.3-Codex-Spark at high reasoning. Spark is optional.
  • Investigator may inspect one bounded project evidence gap, Internet sources, or both, while remaining read-only. Main retains causal, architecture, solution, integration, and acceptance decisions.
  • Companion is created once at the first deployment-state entry while Main's context is still small, then reused for bounded context work. Its bootstrap is scoped to the current goal and does not trigger a full agent_docs/ or unrelated-module intake.
  • Main owns deployment updates to project_progress.md, project_diary.md, and latest_session_work.md. Archivist handles other assigned documentation and the read-only closing handoff.
  • Heavy has no workflow-imposed aggregate worker limit. The workflow does not write a fixed max_concurrent_threads_per_session; available concurrency is left to the Codex platform/account.
  • Heavy is orchestrator-first, uses long event-driven waits, treats an empty timeout as a reason to wait again rather than poll, and keeps routine orchestration silent unless the user needs a decision or risk update.
  • Running workers may use send_message to /root only for rare material mid-task BLOCKER, COURSE_CHANGE, or CRITICAL_PARTIAL events. Routine progress and normal completion never use that channel; receipt of a material event, including a wait_agent wake, complements the existing long-wait lifecycle rather than replacing it.
  • No token-accounting skill, deployment counting marker, usage-report table, or reporting obligation is included.
  • Release discovery and downloads are restricted to GitHub Releases published from elmakus/codex_workflow. There is no background/startup auto-update and no release channel from upstream.
  • Installation, updates, and compute-profile changes happen only when the owner asks. Subscription type is never detected or inferred automatically.

Workflow and roles

For substantive work Main enters deployment state, immediately bootstraps the single session Companion, and loads codex_workflow/heavy_route.md. Companion starts only bounded goal-relevant context work; Main can continue independent Heavy intake and orchestration without waiting for it. Main chooses only useful worker capabilities and owns scope, architecture, scheduling, integration, acceptance, and final claims. Before preparing/following up a worker package, using Micro Execution, or recovering a worker, Main loads ~/.codex/codex_workflow/delegation.md.

Role plus pro-x5 Responsibility
Micro Executor Spark/high when available; fallback Luna/high Spark/high when available; fallback Sol/low Tiny deterministic implementation subtasks inside Heavy.
Default Executor Luna/max Sol/low Normal bounded implementation and repair.
Senior Executor Sol/medium Sol/medium Exceptionally difficult bounded production or solution work.
Tester Luna/max Sol/low Independent verification.
Companion Luna/max Sol/low Persistent read-only project context, created once at first deployment entry and reused.
Investigator Luna/max Sol/low Disposable read-only project/Internet evidence investigation.
Archivist Luna/max Sol/low Verified documentation outside Main-owned deployment-state docs and closing handoff.

If a micro task ceases to be tiny and deterministic, Main reclassifies it directly to Default Executor or Senior Executor; there is no Micro reasoning escalation ladder.

The active selection is stored in ~/.codex/codex_workflow/settings.toml. Missing settings on a pre-profile installation mean plus. Profile switches rewrite only the workflow-owned worker model/reasoning fields plus that settings file in one compensating transaction. Updates preserve the selected profile.

Documentation locations

Project documents stay in <project>/agent_docs/: project_overview.md, project_core_tech.md, project_structure.md, project_progress.md, project_diary.md, latest_session_work.md, and assigned module documents. Existing documents and protected project instructions are preserved during migration. The progress and latest-session templates contain project state only, not embedded maintenance instructions.

~/.codex/codex_workflow/operate/ holds workflow command guides and version metadata. The lifecycle launcher is ~/.codex/codex_workflow/runtime/workflow.py.

Updates

The normal update channel is the owner's GitHub Releases in elmakus/codex_workflow.

codex_workflow --check-update performs a read-only check. It accepts only non-draft SemVer releases containing both the expected versioned ZIP and SHA256SUMS.

codex_workflow --update downloads those assets, verifies SHA-256, applies strict URL/path/size checks, safely extracts and validates the package, and delegates migration to the incoming runtime. A verified local --source path remains available internally for recovery and explicit/manual migrations.

The incoming package is the desired state of workflow-owned files. Unrelated Codex settings/workers/skills, user content outside managed regions, project-local instructions, personalization, agent_docs/, and the selected compute profile are preserved. Updates create a timestamped backup and apply through a compensating transaction.

Commands

Prompt Purpose
codex_workflow --install Install into the explicitly selected project.
codex_workflow --check-update Read-only check of owner releases.
codex_workflow --update Install the latest verified owner release into the selected project/runtime.
codex_workflow --profile Show the active global compute profile and worker mapping.
codex_workflow --profile plus Use the historical Luna-heavy worker allocation.
codex_workflow --profile pro-x5 Use Sol Low for ordinary workers and Sol Medium for Senior.
codex_workflow --personal Change project workflow preferences.
codex_workflow --disable / --enable Disable or enable the selected project.
codex_workflow --remove Preview removal, then remove owned files after explicit confirmation; preserve project documents.

For release preparation and provenance see RELEASING.md. For the ownership map see workflow_break_down.md.

About

A swarm orchestration system in Codex, drastically reduce overall token usage.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages