workflow-chaining

workflow-chaining - Generic Declarative Workflow Runner

Role

Neutral orchestration engine: reads the YAML, announces each phase, gates on user approval, dispatches specialists via the Skill tool (NL description-match as fallback), writes checkpoints. No domain knowledge - that lives in the .workflow.yaml and the specialist skills. Invoked by odoo-intake after a user approves a multi-step workflow plan at the soft-plan-gate.

Hard rules

1. Prefer the Skill tool for phase dispatch. NL description-match is the fallback when no matching skill name is known at plan time.
2. NEVER spawn a sub-agent directly (no direct agent launch, no context: fork - fan-out is the only exception, ≤3 concurrent workers with the mandatory worker-brief preamble).
3. No recursion. Fork workers are leaf agents and carry: "Do NOT invoke spawner skills via the Skill tool. Do NOT spawn sub-agents. You MAY use the Skill tool for read-only leaf skills (e.g. odoo-feature-check, odoo-override-finding). Only Read/Grep/Glob/Write/Bash."
4. No execution before gate. Emit a gate before each phase; wait for approval.
5. Resume from checkpoint. If resume: true and <output_dir>/<slug>-state.json exists (output_dir is the full .odoo-ai/... path from the YAML - see Phase 0 step 2 for how it resolves to a real location), load it and skip done phases.
6. SSOT for schema → workflows/_schema.md. This body describes behavior, not schema.
7. on_complete EMITs, never dispatches. Matched transitions go to Continuation Contract next[] for the run-harness - this skill never fires a spawner itself.

Phase 0 - Load and validate

1. Read the referenced workflows/<name>.workflow.yaml file.
2. Resolve the run's state root once, via the resolve-capture-substitute protocol in ${CLAUDE_PLUGIN_ROOT}/snippets/state-root-resolution.md: output_dir: from the YAML is a relative .odoo-ai/<name> literal that this skill resolves against the Tier-2 ISOLATE dir (scripts/lib/resolve_project_dir.sh isolate) at RUNTIME - the YAML literal itself and the generator's output_dir must-start-with-.odoo-ai/ assertion are intentionally UNCHANGED; only the runtime resolution moved off a bare project-relative path. Resolve series + profile per ${CLAUDE_PLUGIN_ROOT}/snippets/project-facts-resolution.md.
3. Orphan sweep (do this every invocation, before the resume check below). Run find <resolved ISOLATE output_dir>/ -mindepth 1 -maxdepth 1 -type d -mmin +43200 -exec rm -rf {} + (any sibling <slug>-*/ or <slug>-state.json untouched for over 30 days is presumed consumed; a still-active run keeps writing its own state file/phase artifacts, so its mtime never ages past the threshold). It sweeps whatever output_dir THIS workflow declares, so no per-workflow sweep instruction exists. Full rule + bound: ${CLAUDE_PLUGIN_ROOT}/snippets/visual-evidence-lifecycle-contract.md Clause 3. Enforcer: whoever invokes this runner next, unconditionally, every run - not a separate cleanup agent or cron.
4. If resume: true, check for <output_dir>/<slug>-state.json under the resolved ISOLATE dir (the captured absolute path substitutes for the output_dir prefix from the YAML); if found, load it and determine the last completed phase.
5. Collect any missing inputs[] from the user (one question per missing input).
6. Emit the soft-plan-gate header before any phase runs: