Signadot Plan

Signadot: Authoring and Running Plans

Not for explaining plans conceptually. If the user asks "what are plans," "how do plans work," or wants a tutorial or overview, do not use this skill. Point them at Signadot documentation instead. This skill is for producing, executing, or iterating on a concrete plan.

A Signadot Plan is an immutable, compiled DAG of action invocations. Each step invokes one action from the org's catalog and wires inputs from plan params, literal values, or upstream step outputs. Schema, action availability, and action contracts are discoverable at runtime; fetch them before authoring.

Reference Map

Load these one-hop references only when the workflow reaches that topic:

• references/schema-and-actions.md: discovering the plan schema, action catalog, action IDs, disabled actions, and action-body contracts.
• references/spec-authoring.md: composing YAML specs, refs vs values, drill-in rules, conditions, extraInputs, extraOutputs, secrets, and common validator failures.
• references/routing-and-cluster.md: routingContext, routing-key injection, custom routing headers, query-param fallback, and spec.cluster.
• references/running-and-debugging.md: create/run commands, execution JSON, logs, output artifacts, exit codes, and distinguishing runner failures from script failures.
• references/tagging.md: when to tag, retagging safety, and selectionHint.
• references/worked-examples.md: small plan snippets for baseline traffic, sandbox-routed traffic, conditional steps, and outputs.

Core Workflow

1. Clarify the goal and side effects. Identify what the plan should prove or automate, what target it will run against, and whether creating, running, or tagging is expected. Ask only when the target, cluster/routing context, or side-effect risk is ambiguous.
2. Discover the schema and catalog. Fetch signadot plan schema and scan enabled actions before drafting. Read schema-and-actions.md.
3. Choose actions by contract, not by name alone. For every selected action, fetch the action body and ID. The step spec uses action.actionID, not the action name.
4. Resolve routing and cluster shape. If the plan targets a sandbox, route group, routing key, or .svc URL that should hit isolated code, read routing-and-cluster.md before drafting.
5. Draft the YAML spec. Compose params, steps, output, and optional cluster/runner directly. Under each step's action, include only actionID; the server snapshots the rest of the action contract at create time. Read spec-authoring.md.
6. Self-check before create. Verify refs use path expressions only, every arg is declared by the action or as an extraInput, drill-in sources have a schema, conditions read resolved args, secrets are not in literals, and every traffic-issuing step has the right routing context.
7. Create, then inspect validation errors precisely. Use signadot plan create -f <spec.yaml> -o json. The server validates at create time and fails on the first issue; fix the exact issue rather than rewriting broadly.
8. Run and iterate when needed. Run by plan ID or tag with -o json, fetch failed step logs or output artifacts explicitly, then re-run with corrected params or a fresh plan spec. Read running-and-debugging.md.
9. Tag only when the plan is reusable. Default to leaving one-off plans untagged. If a tag is needed, inspect existing tag state before applying and add a useful selectionHint when creating the plan. Read tagging.md.
10. Close with the durable identifiers and evidence. Report what was created or run, the target, params used, outputs inspected, and any remaining operator action.

Tool Integrations And CLI Use

If an integrated Signadot tool (MCP server, plugin, IDE extension, or similar) exposes control-plane reads — clusters, sandboxes, route groups, workloads, endpoints — prefer its output over recalled values. Otherwise use the CLI for the same reads.

Use the Signadot CLI for plan-specific operations unless an integration clearly covers the same command:

signadot plan schema
signadot plan action list -o json
signadot plan action get <name> -o json
signadot plan create -f <spec.yaml> -o json
signadot plan run <plan-id> --param k=v -o json
signadot plan x logs <exec-id> <step-id>
signadot plan x get-output <exec-id> <name>
signadot plan get <plan-id> -o json
signadot plan tag get <name> -o json
signadot plan tag apply <name> --plan <plan-id>

Side-Effect Policy

Essential Rules

• Schema wins. If this skill conflicts with signadot plan schema, follow the schema.
• Actions are org-scoped. Filter to status.enabled == true; disabled actions cannot be referenced by a created plan.
• Plans are immutable. Updating an action later does not affect existing plans. Author a fresh spec and create a new plan to pick up action updates.
• Step order is dependency-driven. Array order is not execution order; refs imply the DAG.
• Refs are paths, not expressions. Use params.<name> or steps.<id>.outputs.<name> with optional drill-in; use a composing action for operators, conditionals, string concatenation, or transformations.
• Drill-in requires schema. The source param/output must declare a schema and the runtime value must be JSON.
• Routing has two halves. routingContext sets env vars such as SIGNADOT_ROUTING_KEY; outbound requests still need the routing key injected in accepted headers or query params.
• spec.cluster is not routingContext. Cluster decides where the runner runs; routing context decides how a step directs traffic.
• Tagging is optional. Tags are stable pointers for reusable plans, not a default naming mechanism for every draft.
• Executions are at-least-once. A step can run more than once across retries or re-runs. Action code should be idempotent; do not author plans that assume exactly-once execution.

Quick Reference

Reads use -o json; the CLI already pretty-prints JSON, so pipe through jq only when filtering. Writes (plan create) use a YAML file.

Final Report

Include the parts that exist for the task:

• Plan ID, execution ID, and tag name if one was created or updated.
• Cluster, sandbox, route group, routing key, or explicit note that the run was baseline/cluster-agnostic.
• Params used, with secret values redacted and secret refs named only when safe.
• Action names and IDs selected, especially if a close alternative was used.
• Outputs inspected and the final pass/fail signal.
• Failed step errors, log locations, or artifact commands when unresolved.
• Side effects left behind: untagged plans, retagged names, generated YAML files, or any operator follow-up.