Wiki · Research
Durable/local product track template
Shelf
Research

Durable/local product track template

Use with the shared spine for Letta Code, Aider, goose, and Google ADK. This track admits Implemented, Executed, Measured, Documented, and Opinion claims. Protocol text consumed by a product may be labelled Normative only for what that protocol requires, never as evidence that the product conforms.

P1. Default product path

Trace one default user request from ingress through model/runtime decisions, external actions, observations, durable writes, caller-facing output, and termination. Put optional backends, extensions, delegation, and hosted-only behaviour in separate paths.

Required questions:

  • Which component owns the loop and terminal state?
  • Where do commands run, and which host surfaces cross that boundary?
  • Which artifact remains when the happy path, process, or external action fails?
  • Which test executes the default path rather than only constructing its components?

P2. Continuity-layer matrix

Complete all five shared-spine state rows. For each applicable layer additionally record:

LayerUser or agent can write?Retrieval/selection ruleScope and isolationForget/delete pathSurvives
Conversation history{{turn / process / machine / account}}
Recovery state
Documentary intent
Repository state
Agent-managed long-term memory

Do not infer long-term memory from a conversation store, embeddings from a search API, learning from persistence, or identity isolation from a generated id. Trace the write and read paths.

P3. Repository change and validation

Where the product edits a repository, trace context selection → proposed edit → filesystem write → validation → diff/commit/revert. Record Git/worktree semantics separately from process, filesystem, credential, and network isolation. If repository change is not a product concern, mark this section not applicable with the reason.

P4. Extension and evolution boundary

Trace one extension, plugin, tool protocol, sub-agent, or public-language binding from discovery to invocation and failure. Separate stable public contract, capability negotiation, compatibility policy, and current implementation. For multi-language subjects, show whether tests establish semantic parity or only similar surface names.

Product counter-evidence prompt

Search for the shortest path that bypasses the proposed durability, isolation, validation, or extension boundary. A documented feature with no implementation/test evidence and an optional path mistaken for the default are first-class counter-findings.