Updating to the current development API#

The current development branch simplifies public module boundaries. Removed Python paths have no compatibility aliases. Model, experience and session data formats remain separate from their implementation modules.

Earlier API	Current API
`tensorcode.runtime` evidence records	`tensorcode.tools.cognition`; `Evidence` is also available beside `Investigator`
Direct cognitive-session and memory construction	`investigator.new_cognitive_session(policy=..., memory=..., max_records=...)`
Cognitive-session restoration with a supplied model	`investigator.load_cognitive_session(path)`
`runtime.PlanExecutor(...)`	`planner.new_executor(actions=..., replan=..., max_steps=...)`
`runtime.ActionLoop(...)`	`tools.actions.action_loop(chooser=..., actions=..., max_steps=...)`
Action callback records	`tensorcode.tools.actions`
Structured plan and outcome records	`tensorcode.tools.planner`
`tensorcode.tracing`	Root `trace`, `Trace`, `InputRef` and `OutputRef`
`training.ToolTrainer(model, ...)`	`training.Trainer.from_tool(model, ...)`
`training.Trainer(operations, ...)`	`training.Trainer.from_ops(operations, ...)`
`training.load(...)`	`training.load_experience(...)`
`runtime.SelectionPolicy`	`policy={...}` passed to `new_cognitive_session`
`runtime.CognitiveState`	Read-only `session.state` / `bot.cognitive_state`
`runtime.CognitiveSession`, `EpisodicMemory`, `LearnedEpisodicMemory`	`investigator.new_cognitive_session(memory=...)`; memory is tool-owned
`runtime.JsonMemory`, `MemoryRecord`, `MemorySearch`, message-sequence codecs	No replacement; storage is an application concern
`tracing.Session`	`tensorcode.Trace` (returned by `tensorcode.trace()`)
`Hypothesis(id, text, origin, provenance)` positional	`Hypothesis(id, text, origin, model_provenance=...)`; provenance is a required keyword
`Trace.capture(...)` / `capture_async(...)`	Internal; use operations inside `trace()` or `Trainer.capture(...)`
Free training checkpoint functions	`trainer.save_checkpoint(...)` and `trainer.load_checkpoint(...)`

Investigator and Planner keep their existing ranking-history new_session() behavior. Use the explicit cognitive-session factory for revisable evidence and episodic retrieval. InvestigationSession remains available from tools.investigator for annotations; normal construction belongs to the tool. Chatbot keeps new_session() for independent conversations sharing its weights, and ChatSession remains available from tools.chatbot.

The generic DecisionPipeline wrapper is removed: call your encoder and decision operation directly, then apply any explicit application policy. General-purpose storage remains an application concern; cognitive session factories own their memory implementation. There is no replacement public storage-engine namespace.

Both trainer factories preserve their previous learning-rate defaults: .001 for a tool objective and .01 for operation graphs, using SGD unless an optimizer is supplied. The tool factory applies the model's training-mode policy. The ops factory leaves module modes unchanged. Graph supervision still comes from trace() and explicit supervise(...); it is not inferred by capture().

Saved data#

Canonical tool and operation identities, model parameter layouts, and unchanged experience/session schemas are preserved. Loading does not import Python types named by an artifact. Application dataclasses still require the same explicit codec names bound to trusted current classes on save and load.

Complete directory training checkpoints retain their existing JSON/safetensors envelope, including optimizer state, module modes, steps, caller progress and supported random state. The older standalone tensorcode.checkpoint file contains only model/optimizer state; loading it cannot recover metadata that was never saved. New trainer saves produce complete directory checkpoints.

Use training for collect/replay/resume examples and cognition for evidence revision and session restoration. Historical evaluation reports retain the source revision and API used for their measurements.