Pavan's Workflow

A reusable multi-model, multi-agent engineering workflow for Oh My Pi (omp): fresh specialized workers, durable file-backed state, independent primary/backup role models, scoped Graphify navigation, Coder-only Ponytail, and optional product-design roles — with the important boundaries checked in code, not only in prompts.

3.7: fast-coder, automatic failover, and red proof for fix rounds. First attempts use deterministic routing for Fast Coder (workflow-coder-fast), gate timeouts are no longer blamed on the Coder, primary model/provider failures automatically fail over to configured backup agents without pausing for Human authorization, and fix rounds require reproducible red proof for fixed findings. Details: CHANGELOG.md.

How the workflow runs

flowchart LR
    H[Human supervisor] <--> M[Main Orchestrator]
    M --> C[Fresh Coder + Ponytail]
    C --> G{{close check: gates + guard + blast radius}}
    G -- close_quick --> M
    G -- review --> R[Fresh Reviewer]
    R --> M
    M --> T[Fresh Tester]
    T --> M
    M -. system uncertainty .-> A[Fresh Architect + Grilling]
    M -. visual advice .-> DA[Design Advisor]
    M -. visual implementation .-> D[Designer + UI skill]
    M -. optional pre-release .-> S[Security Reviewer]

Loading

All routing goes through Main. Workers never route another worker, write workflow state, commit, or push — and the guard blocks those actions inside the worker before they run.

What is enforced, and by what

Rule	Enforced by
Read-only roles change nothing; Coder/Designer stay in target_files; Tester in test paths; no worker edits workflow files or changes git state	workflow-guard extension: blocks the tool call inside the worker before it runs; workflow_guard.py verifies what the worker edited (changes by Main, the Human, or parallel sessions are never the worker's)
Backup workers start on recorded primary model/provider failure	before_subagent_spawn hook blocks unauthorized spawns
quick closes only with green command gates, manual gates checked, a clean Coder/Designer guard verdict and no open violation, no blast-radius hit, and risk not high	workflow_close.py check
Objective Gates are re-run by Main, not trusted from worker reports; the whole-project suite ((close-only)) runs once in the close check, not again in the Coder, and a failure arrives with its failing lines and full log	workflow_gates.py
The Tester runs on every step unless the close check decides close_quick or the card names an R14 skip reason (human_opt_out, presentation_only, docs_only, mechanical_rename)	workflow_close.py check (tester.required) · dashboard routing
Reviewer is independent of the Coder model (primary and backup)	doctor warning (workflow_model_diversity.py)
One worker at a time, no recursive spawning, no automatic model fallback	.omp/config.yml task/retry policy

The full rule set (R1–R22) with each rule's enforcement is in TEAM_CONTRACT.md; rules marked "Main" or "prompt" still rely on the model.

Install

Requirements: omp, git, Python 3.9+. Graphify is installed on first run when missing (optional; the workflow falls back to source tools).

Into an existing repository (recommended)

cd /path/to/your/project
bash <(curl -fsSL https://raw.githubusercontent.com/Pavan-Gopa/Pavans-Workflow/main/install.sh) .

Your README.md, CHANGELOG.md, VERSION, and an existing .omp/config.yml are kept (workflow roles are merged into the config). The installer refuses to overwrite files it would otherwise replace and lists them. Commit the installed files before the first step so the blast-radius check starts clean.

New project

mkdir my-project && cd my-project && git init
bash <(curl -fsSL https://raw.githubusercontent.com/Pavan-Gopa/Pavans-Workflow/main/install.sh) .

Cloning this repository as your project also works (bash install.sh .), but then point origin at your own repository so project work is never pushed here.

Update

Close OMP for the project, then from its root:

bash <(curl -fsSL https://raw.githubusercontent.com/Pavan-Gopa/Pavans-Workflow/main/install.sh) --update

Inside OMP: /workflow-update check (plan) or /workflow-update. From a shell: bash AI_Workflow_Kit/script/workflow_update.sh check|apply. Add --ref v3.8.0 to pin a release, --refresh-graphify to rebuild the graph.

Updates install the newest vX.Y.Z release from AI_Workflow_Kit/framework.manifest, remove framework files the release deleted (unless you modified them), keep project state, model selections, custom .omp files, .graphifyignore rules, and the Graphify index, and back up every framework file they touch under <git-common-dir>/pavans-workflow/update-backups/. If the framework step fails it is rolled back automatically; the state migration keeps its own backups, and a failing doctor is reported (not rolled back). Restart OMP afterwards.

Coming from 3.4.x: run the curl command above once. It removes the obsolete experiments/ payloads, restores root files an old updater overwrote (from its backups), and tells you if a leftover root VERSION/CHANGELOG.md came from the workflow. The old in-project workflow_update.sh apply also works: it hands over to the 3.5 manager mid-run and completes the migration.

Start

bash AI_Workflow_Kit/script/omp_workflow.sh

Main model control

Goal	Action
Change the persistent Main model	Alt+M → Roles → DEFAULT, choose model and effort
Temporarily use the backup	quick-switch (Alt+Q)
Inspect the backup mapping	workflow_orchestrator_backup in Alt+M → Roles

workflow_orchestrator is a hidden managed alias to @default.

Role model pairs

Configure through Alt+M → Roles. On a primary worker model/provider failure Main fails over to the role's backup automatically (the Fast Coder's backup is optional; without a distinct one it escalates to the Coder); the Human is asked only when no distinct backup is configured or the backup fails. Main's own backup is a manual live switch (@workflow_orchestrator_backup, Alt+Q).

Role	Primary	Backup (automatic failover)
Main	DEFAULT via @workflow_orchestrator	@workflow_orchestrator_backup
Coder	@workflow_coder	@workflow_coder_backup
Reviewer	@workflow_reviewer	@workflow_reviewer_backup
Tester	@workflow_tester	@workflow_tester_backup
Architect	@workflow_architect	@workflow_architect_backup
Security	@workflow_security	@workflow_security_backup
Design Advisor	@workflow_design_advisor	@workflow_design_advisor_backup
Designer	@workflow_designer	@workflow_designer_backup

Keep the Reviewer on a different model than both Coder entries, and spread backups across providers; the doctor warns otherwise.

Role skills

Each worker role loads the skills named on the autoloadSkills line of its agent files (.omp/agents/workflow-<role>[-backup].md): ponytail for the Coders, ui-designer for the Designer and Design Advisor, grilling for the Architect. Those are the Workflow's own and stay. A project can add more skills to a role; the additions survive updates:

python3 AI_Workflow_Kit/script/workflow_skills.py show --target . --json
python3 AI_Workflow_Kit/script/workflow_skills.py set --target . \
  --agent workflow-coder --agent workflow-coder-backup --skills=graphify,my-skill
python3 AI_Workflow_Kit/script/workflow_skills.py set --target . --agent workflow-coder --skills=   # remove all additions

• set makes the list the agent's complete additions; name both the primary and the backup so a failover loads the same skills. It rejects unknown agents, malformed names, skills the Workflow already loads for that agent, and ponytail for the Designer and Design Advisor (the doctor requires them to stay free of it). Either every file changes or none does.
• Additions live in .omp/workflow-skills.json ({"version": 1, "added": {"workflow-coder": ["graphify"]}}). It belongs to the project: commit it, updates never replace it, and workers cannot edit it (.omp/ is guard-protected).
• Update copies the new agent files, then writes the additions back after the Workflow's defaults, so new defaults arrive and your skills stay. workflow_skills.py apply does the same by hand. A skill you added that a later release makes a default drops out of the overlay.
• The installed-files record and workflow_framework.py verify ignore the additions: only an edit of the agent file itself is reported as a local modification. Additions are appended to the existing line and removed as its exact inverse, whatever its YAML spelling (autoloadSkills : [ 'x' ], "autoloadSkills": [...], CRLF newlines, an empty [], or no line at all), so removing every addition restores the shipped file byte for byte.
• If an agent file loses an addition the overlay records, verify and the doctor fail it, and workflow_framework.py update --check lists it as [REPAIR]; the next update writes it back.
• The doctor requires ponytail on both Coders and ui-designer (without ponytail) on both design roles, as parsed members of the list (not comments), and that every addition in the overlay is on its agent's line.

Useful controls

Action	Control
Live workflow dashboard	Alt+W
Model leaderboard across all projects (inside the dashboard)	l
Full Agent Hub	Alt+A
Quick-focus the running worker / return	Tab on an empty composer / Tab or Esc
Reconcile and continue	/workflow status
Explain routing	/workflow why
Close decision for the current step	python3 AI_Workflow_Kit/script/workflow_close.py check
Guard verdicts / record the Human's decision	python3 AI_Workflow_Kit/script/workflow_guard.py status --all / resolve --id … --note …
Guard mode for this repository	python3 AI_Workflow_Kit/script/workflow_guard.py mode [enforce|report|off] (or WF_GUARD_MODE)
Designer advice / edits	/workflow designer advise <surface> / redesign <surface>
Manual OMP Stats	o in Alt+W or /workflow-stats
Diagnostics	bash AI_Workflow_Kit/script/workflow_doctor.sh

Verify an installed project

cat AI_Workflow_Kit/VERSION
bash AI_Workflow_Kit/script/workflow_doctor.sh

Expected version:

3.8.0

Repository map

.omp/                          agents, commands, extensions (dashboard, guard, quick focus, Main attribution), libs, selftests
AI_Workflow_Kit/framework.manifest   what the framework installs (single source for install/update/doctor/CI)
AI_Workflow_Kit/templates/     pristine project-state templates (rendered into projects when missing)
AI_Workflow_Kit/docs/AI/       rules (TEAM_CONTRACT.md), Main procedure, role contracts
AI_Workflow_Kit/script/        installer core, updater, doctor, gates, guard, close check, metrics, Graphify tools
AI_Workflow_Kit/vendor/        pinned dependency versions (Graphify, Ponytail, tested OMP)
ci/                            repository checks, end-to-end install/update test, tsc config (not installed)
ponytail*/ grilling/ ui-designer/   skills

Developing the framework

python3 ci/repo_checks.py                 # template purity, versions, manifest coverage, rule IDs
bash ci/e2e_install_update.sh             # install, update, 3.4.2 migration with stub omp
for t in AI_Workflow_Kit/script/*.selftest.py; do python3 "$t"; done
for t in AI_Workflow_Kit/script/*.selftest.sh; do bash "$t"; done
for t in .omp/tests/*.selftest.ts; do node --experimental-strip-types "$t"; done
bun ci/omp_runtime_guard.ts               # guard inside a real OMP session (needs ci/typecheck deps)

Never commit rendered project state: edit AI_Workflow_Kit/templates/ instead. A release is VERSION + AI_Workflow_Kit/VERSION + dependencies.lock + a CHANGELOG entry, then a vX.Y.Z tag; the release workflow checks they agree.

MIT. See LICENSE.