JAZ boundary

Type: agentic-system-analyses/types/agentic-system-boundary.md

Boundary and evidence

The target is the JAZ package distributed by this repository, including its shipped console, experimental components and retained-execution tooling. The intended use is to explain how this package runs model-directed Python agents, carries state through iterations and recursive calls, controls execution through hooks, and records or reuses execution content. This is not an analysis of an arbitrary application built with JAZ.

JAZ qualifies as an enclosing runtime, with a whole-system boundary: it owns the model-query/REPL iteration loop and terminal result handling, even when a Python application starts it through a library call. A callable API is not the complete target in isolation. The installed console and module entry point are shipped consumers and remain included. SRC-1, pyproject.toml, src/jaz/__main__.py, src/jaz/invoke.py, src/jaz/_agent.py and src/jaz/console.py establish these entry points and responsibilities.

The package declares its loop directly:

An agent runs as a conversation with a Python REPL: the LLM writes code, sees what it evaluated to, and iterates until it finishes with a native return (or raise):: --- src/jaz/__init__.py

The implementation builds an Agent for an invocation, initializes run-local REPL state and opening messages, repeatedly calls do_one_repl_iteration, appends observations for Continue, and completes on Return or Raise. The sync and async paths are both within scope. SRC-1, src/jaz/invoke.py and src/jaz/_agent.py supply implementation evidence; this is not merely a README classification.

repl_iter_result = self.do_one_repl_iteration( i, depth, can_recurse, messages, state, invoke_id, repl_history=repl_history, ) --- src/jaz/_agent.py

The functional boundary includes input resolution and descriptions, scoped values and recursive invocation, configuration, protocol and prompt rendering, Python execution and permission controls, model adapters and accounting, hooks and their state/effects, console history and settings, trace production and replay, workflow materialization, rollout collection and trace/log utilities. Provider infrastructure and caller-defined tool implementations are external dependencies, not capabilities implemented by JAZ. The shipped SWE tool implementation is included rather than treated as an external caller.

The console is material to the state loop. Its chat wrapper passes earlier session turns into subsequent invocations and records both results and exceptions; its namespace creates fresh main/helper histories. It also persists model defaults and reloads them at console startup. These are JAZ responsibilities, not duties assigned to an excluded application. SRC-1, src/jaz/console.py:

if history.turns: extra["conversation_history"] = history --- src/jaz/console.py

The evidence boundary is Git revision 0803d4971be785e95b80054b02259664d70fa3da, applicable through 2026-10-15. The acquired checkout's HEAD matches that pin and its status was clean when inspected. The source object above is unchanged from acquisition. Evidence is code-grounded: source text establishes implemented paths, while package documentation and inline contracts establish declared intent. No target code, examples or tests were executed. No observed-run or causal-experiment evidence is registered.

Shipped-path coverage

The tracked root contains exactly CHANGELOG.md, LICENSE, README.md, pyproject.toml and src/. The table accounts for each; subdivisions of src/ distinguish material responsibilities. “Included” defines the analytical boundary, not a claim that every line has already been audited. Initially inspected source ranges are recorded below; available but uninspected implementation details are not access gaps and remain available to subsequent jobs under SRC-1.

The retained-content paths are not excluded as mere logging. TrajectoryRecorder writes ATIF JSON, TrajectoryReplay loads saved responses and supplies them at query time, and WorkflowReplay emits standalone Python entry blocks. trace_to_directory is another shipped consumer of recorded ATIF files. SRC-1, src/jaz/hooks/builtin/atif_trace.py, src/jaz/hooks/builtin/replay.py, src/jaz/hooks/builtin/workflow_replay.py, src/jaz/utils/trace_to_directory.py.

self.output_path.write_text(json.dumps(data, indent=self.indent)) --- src/jaz/hooks/builtin/atif_trace.py

return [ SupplyLLMResponse( content=r.content, prompt_tokens=r.prompt_tokens, completion_tokens=completion_tokens, cost_usd=r.cost, ) ] --- src/jaz/hooks/builtin/replay.py

External dependencies and evidence limits

The package uses Python and dependencies declared in pyproject.toml, including Jinja2, HTTPX, Tenacity, beartype, typing-extensions, pathspec and LiteLLM. Optional integrations include OpenTelemetry, Transformers and the GitHub Copilot SDK. Only JAZ's integration code is frozen; dependency source, provider services, user credentials, tool backends and model weights are not registered evidence. This prevents claims about their internal enforcement, uptime or model performance. SRC-1, pyproject.toml and src/jaz/llm/_litellm.py.

This repository describes itself as a generated distribution:

This repository is generated: each release publishes a cleaned copy of the package, so changes are not made here. Contributor setup and code conventions live with the source. --- README.md

The tracked tree contains no tests, evaluation harness, benchmark datasets, run traces or contributor-source tree. References in documentation/comments to an eval harness, examples, experiments or other source-tree consumers do not make those artifacts available at this pin. This prevents independent assessment of benchmark outcomes, optimization/training effectiveness and behaviors of unshipped applications. It does not exclude the console or shipped replay tools. Available package files not yet read are inspection work, not unavailable evidence. No other repository or moving web documentation is admitted.

Source register