Agently TriggerFlow Execution State

This skill covers runtime-instance persistence for TriggerFlow executions. It focuses on execution.save(), execution.load(), resume-after-restore, waiting-interrupt recovery, ready-result recovery, file or string state loading, and runtime-resource reinjection. It does not cover flow-definition export/import or Mermaid.

Prerequisite: Agently >= 4.0.8.5.

Scope

Use this skill for:

• execution.save()
• execution.load()
• saving to or loading from dict, JSON string, YAML string, JSON file, or YAML file
• restoring waiting executions
• restoring executions whose final result is already ready
• using continue_with(...) after restore
• understanding what execution state contains and what it does not contain
• reinjecting runtime resources after restore

Do not use this skill for:

• get_flow_config(), get_json_flow(), get_yaml_flow(), or Mermaid
• blueprint copy and definition export/import
• choosing between runtime_data, flow_data, and runtime-resource placement
• provider setup, model schema design, or non-TriggerFlow persistence topics

Workflow

1. Start with references/state-surfaces.md to distinguish execution state from flow config and from general state-placement design.
2. If the task is about save/load roundtrip, read references/save-load-roundtrip.md.
3. If the task is about waiting flows, pending interrupts, or resume-after-restore, read references/waiting-and-resume.md.
4. If the task is about restoring a suspended business workflow or conversation after restart, read references/business-restore-recipes.md.
5. If the task is about runtime resources after restore, read references/runtime-resource-reinjection.md.
6. If behavior still looks wrong, use references/troubleshooting.md.

Core Mental Model

Execution state is the saved runtime surface of one execution instance.

It captures:

• execution status
• runtime data
• flow data snapshot
• pending interrupts
• last signal
• final result readiness and value
• resource key names

It does not carry executable resource objects themselves.

So the standard restore pattern is:

1. rebuild or reuse the correct flow definition
2. create a fresh execution
3. load the saved execution state
4. reinject any required runtime resources
5. continue or read the result

In real systems, a suspended workflow usually needs more than execution.load(...) alone.

Typical business restore sequence:

1. restore or rebuild the correct flow definition
2. create a fresh execution on that definition
3. load the saved execution state
4. reinject runtime resources
5. if the execution is waiting, inspect interrupts and continue it
6. if the result is already ready, read the result directly

Selection Rules

• resume a waiting execution later -> execution.save() then execution.load() on a fresh execution
• restore a completed execution whose result was already ready -> execution.load() then get_result() / async_get_result()
• persist to transportable artifact -> save to JSON or YAML file
• restore from in-memory payload -> load from dict or JSON/YAML string
• restored execution needs runtime-only tools or services -> pass runtime_resources=... to load(...) or reinject before continuing
• choosing where state or resources should live before persistence -> use agently-triggerflow-state-and-resources
• definition itself must be exported or imported -> use agently-triggerflow-config, not this skill

Important Boundaries

• execution state is for one running or completed execution, not the reusable flow definition
• runtime resources are represented only by resource_keys, not by serialized callable objects or clients
• waiting interrupts can be restored and resumed, but they still need the proper execution definition and resources in place
• flow config and execution state are complementary, not interchangeable

References

• references/source-map.md
• references/state-surfaces.md
• references/save-load-roundtrip.md
• references/waiting-and-resume.md
• references/business-restore-recipes.md
• references/runtime-resource-reinjection.md
• references/troubleshooting.md