TheWorkshop Open Source Edition is a skill for Codex and Claude Code that runs non-coding work in a structured, auditable way.
It turns ambiguous requests into a living execution workflow:
Project -> Workstreams -> Jobs
with explicit gates, orchestration, monitoring, lessons learned, and spend visibility.
When this repo says "project OS", it means:
- a repeatable workflow system for the agent
- not an operating system
- implemented as a skill the agent runs inside Codex/Claude Code
The diagram below shows TheWorkshop's control plane (planning, gating, orchestration) and execution plane (work execution, outputs, monitoring).
- A Codex/Claude Code skill, not a standalone app
- A structured runtime for non-coding projects
- Agreement-gated before execution starts
- Truth-gated and reward-gated before completion claims
- Parallel-orchestration aware (sub-agents when independent jobs exist)
- Dashboard-first monitoring with token/spend telemetry
- A replacement for human strategic ownership
- A generic code framework or web app product
- A system that marks work complete on artifact presence alone
Project: top-level outcome and success definitionWorkstream: coherent thread in support of project goalJob(Work Item): smallest executable/verifiable unitWave(optional): timeboxed grouping across workstreams
Completion promises are explicit:
<promise>{ID}-DONE</promise>
A job can only transition to done when all gates pass:
- Agreement gate (scope accepted before execution)
- Dependency/freshness gate (inputs are current)
- TruthGate (verification of correctness)
- Reward gate (meets
reward_target)
Execution quality defaults:
job_start.pyauto-applies ranked lessons into# Relevant Lessons Learned(override:--no-apply-lessons).plan_check.pywarns on weak placeholder content forplannedjobs and hard-fails weak content forin_progress/donejobs.
# One command from the repo root
git clone https://github.com/CongressionalInsights/theworkshop.git
mkdir -p "$CODEX_HOME/skills"
cp -R theworkshop "$CODEX_HOME/skills/theworkshop"Typical destination:
$CODEX_HOME/skills/theworkshop- usually
~/.codex/skills/theworkshop
To update later:
cd "$CODEX_HOME/skills/theworkshop" && git pull origin main# create project
python3 scripts/project_new.py --name "Workshop Demo"
# add workstream + job
python3 scripts/workstream_add.py --project /path/to/project --title "Research"
python3 scripts/job_add.py --project /path/to/project --workstream WS-YYYYMMDD-001 --title "Draft options memo"
python3 scripts/job_add.py --project /path/to/project --workstream WS-YYYYMMDD-001 --title "Attribution sweep" --job-profile investigation_attribution
python3 scripts/job_add.py --project /path/to/project --workstream WS-YYYYMMDD-001 --title "Entity resolution" --job-profile identity_resolution
python3 scripts/discuss.py --project /path/to/project --work-item-id WI-YYYYMMDD-001 --decision "Use concise format" --required --no-interactive
# validate and orchestrate
python3 scripts/plan_check.py --project /path/to/project
python3 scripts/schema_validate.py --project /path/to/project
python3 scripts/optimize_plan.py --project /path/to/project
python3 scripts/orchestrate_plan.py --project /path/to/project
python3 scripts/dispatch_orchestration.py --project /path/to/project --dry-run
python3 scripts/council_plan.py --project /path/to/project --dry-run
# execute one job
python3 scripts/job_start.py --project /path/to/project --work-item-id WI-YYYYMMDD-001
python3 scripts/job_start.py --project /path/to/project --work-item-id WI-YYYYMMDD-001 --lessons-limit 5 --lessons-include-global
python3 scripts/verify_work.py --project /path/to/project --work-item-id WI-YYYYMMDD-001
python3 scripts/job_complete.py --project /path/to/project --work-item-id WI-YYYYMMDD-001 --cascade
# optional utility lanes
python3 scripts/health.py --project /path/to/project --repair
python3 scripts/quick.py --project /path/to/project --title "One-off patch" --command "echo done"
python3 scripts/dashboard_server.py --project /path/to/project --openExpected core outputs:
outputs/dashboard.htmloutputs/dashboard.jsonoutputs/dashboard.mdoutputs/<date>-task-tracker.csvlogs/execution.jsonlartifacts/truth-report.jsonnotes/context/<WS-or-WI>-CONTEXT.mdoutputs/uat/<run-id>-UAT.mdoutputs/uat/<run-id>-UAT.jsonoutputs/health.jsonquick/<id>-<slug>/plan.mdquick/<id>-<slug>/summary.md
- Dashboard auto-opens best-effort at execution start (unless disabled)
- Auto-refresh supports stale detection and pause/resume
- Optional local live transport:
python3 scripts/dashboard_server.py --project /path/to/project- serves
dashboard.htmloverhttp://127.0.0.1:* - publishes
/eventsSSE updates so the page can switch from poll mode to live mode
- serves
- Cost display is billing-aware:
subscription_auth: billed cost shown as$0marginal, API-equivalent shown secondarilymetered_api: billed cost from exact telemetry when availableunknown: estimate-first fallback
Use work-item scoped image generation:
python3 scripts/imagegen_job.py --project /path/to/project --work-item-id WI-YYYYMMDD-002
python3 scripts/imagegen_job.py --project /path/to/project --work-item-id WI-YYYYMMDD-002 --credential-provider env
python3 scripts/imagegen_job.py --project /path/to/project --work-item-id WI-YYYYMMDD-002 --credential-provider keychain --approve ttl:1hSet one provider before first run:
export THEWORKSHOP_IMAGEGEN_API_KEY=...Compatibility for existing local setups:
export OPENAI_API_KEY=...Optional legacy keychain flow:
export THEWORKSHOP_IMAGEGEN_CREDENTIAL_SOURCE=keychain
export THEWORKSHOP_KEYCHAIN_SERVICE=OPENAI_KEYThe apple-keychain skill remains optional and cross-platform fallback behavior is env-first.
python3 scripts/doctor.py
cd scripts && for t in *_test.py; do python3 "$t"; done./scripts/install_skill.sh --forceUse --link for a symlinked dev install.
- Contribution guidelines: CONTRIBUTING.md
- Support boundaries: SUPPORT.md
- Security reporting: SECURITY.md
- Stable
v0.1.0baseline for Project -> Workstreams -> Jobs control plane - TruthGate + stale invalidation + orchestration artifacts
- Billing-aware spend in dashboard
- More robust synthetic scenario suite for document-quality outcomes
- Additional dashboard drilldowns for truth/reward failure analysis
- GitHub mirror ergonomics and dry-run diagnostics
- Optional docs site for deeper operators manual
- Broader template library for non-coding domains
- Extended export/report bundles for stakeholder handoff
theworkshop/
README.md
SKILL.md
CHANGELOG.md
scripts/
references/
examples/
docs/assets/
.github/
MIT. See LICENSE.
