Skip to main content

Run data

You are in the deep end

This page is the stage-handoff contract: goal, slots, and notes. Start at the request lifecycle if you want the sequence.

The run graph has three channels for stage handoff inside one correlation id:

ChannelLives onWho consumes it
goalStart body (job payload / chat utterance). Never updated.Every LLM stage; base of every HTTP body
slotsworking.slots[stage_id] — structured JSON after each stageNext HTTP merge (schema keys only); branch/gate; prefetch pack; child kind=agent projection
notesAppend-only strings on workingLLM stages (_user_blob(goal, notes)); optional payload["notes"] on HTTP

GraphState in agent-runtime/app/graph/workflow.py is result, goal, notes, slots. HTTP assembly is in agent-runtime/app/graph/payload.py:

payload = dict(goal)
# merge keys from prior slots and CALL JSON args when declared on input_schema.properties
# when properties exist, drop undeclared keys (including chat utterance)

Fail closed if input_schema.required keys are missing. Do not dump all slots or all notes onto every HTTP call.

query_formulation sets payload["query"] from the LLM. classify / synthesis append prose to notes and, when JSON parses, write slots[stage_id]. LLM roles with output_schema validate through with_structured_output. How to write that schema: schemas.

Example: stage 2 gets stage 1’s JSON​

Route purchase_refund (D5 proof): job payload has doc_id, account_id only.

  1. extract_fields (classify) returns { merchant, amount, date } → stored in slots["extract_fields"] and appended to notes.
  2. match_purchase HTTP body = goal ∪ { merchant, amount, date } because those keys are required on match_purchase’s input_schema.
  3. If classify omits a required field, the run fails before match_purchase invokes.

Route card_freeze today: identity/limit tools return prose; structured slot merge is not the proof path unless the workflow adds schema-bound hops.

Prefetch​

On deterministic_prefetch, empty-invoke prefetch stages POST corpus gateways and write working.slots.prefetch. Downstream LLM/HTTP stages consume packed text. See retrieve.

Still catalogue-only​

  • conversation / long_term — Shared Memory (not the run pin). See memory.
  • Stage allowlists on Pattern 3 — metadata only; AR does not enforce inner tool picks yet.

Catalogue vs Runtime matrix: autonomy modes. Verification checklist: dataflow-plan.md.

Last updated on