These are exploratory directions. The specification defines supported behavior; Python owns orchestration, and hosts own isolation and durable records. The philosophy guides design decisions, and design drivers record application needs, findings, and evidence criteria. Each investigation should identify the uncertainty to resolve and evidence that would change a decision. Record findings with their scenario and evaluation evidence, then revise or retire the investigation here.
Can existing hooks support semantic checks with understandable failure and recursion behavior? Test small examples combining Natural findings, deterministic checks, and an application policy at an execution boundary. Start with existing function composition and oversight hooks. Investigate nested Natural checks, inherited policy, unintended hook recursion, rejection, and checker failures before proposing new runtime primitives. Distinguish applying a check from establishing the correctness of its judgment.
Where does Nighthawk reduce construction and revision burden, and where does it add complexity? Use the Markdown knowledge-management scenario to identify concrete composition gaps. Keep its detailed design, knowledge collection, and operational rules in application-owned sources. Bring back small examples, findings, and contract requirements, with an explicit verification status. Compare proposed abstractions with research, business, and simulator-backed workflows to assess generality. Evaluation and simulation systems remain connected through application functions and libraries.
Can agents reuse and improve functions from earlier tasks while preserving accepted behavior? Use the construction and reuse scenario to assess discovery, contracts, context, and failure feedback across representative tasks. Compare construction and revision success, assessment of results and effects, and overall complexity with ordinary Python and model libraries.
Do semantic checks for documentation and development workflows produce useful findings at acceptable review burden? Assess false positives and missed violations before adopting them in the development process. Treat development automation and public library features as separate decisions.
Explore concise summaries of task-relevant working state through the context builder. Keep live values accessible independently of their summaries.
Extend descriptions of opaque domain values while preserving the inspection contract: bounded presentation, provenance, and live-value access.
Evaluate whether model-specific guidance improves judgment quality enough to justify additional prompt and maintenance cost.
Provide conventions for limiting include paths and template access within authorized Natural sources, whether repository-managed or generated at runtime.
Use the in-task revision scenario to evaluate reusable helpers for host authorization and revision evidence under the trusted-program contract. Evaluate whether reusable helpers reduce the work of maintaining revision evidence without prescribing an application's policy or storage. Exercise embedded execution without function interception, helper-only revisions, and changes of execution context. Start with existing boundaries and small examples; apply the Core evidence criteria to any unmet execution need.
Develop host/backend integration guidance for process isolation and permissions.
Align trace fields for attempts, decisions, and state transitions across resilience primitives. Keep telemetry optional and avoid redundant events.
Explore reusable host-selected policy defaults that compose with per-function overrides. Preserve explicit retry boundaries and require recovery decisions after partial effects.
Provide examples of host-managed execution records and application snapshots, with explicit deduplication and recovery decisions.
Explore explicit serialization of host data and correlation metadata for multi-process execution.
Investigate propagating selected context-variable rebindings back from bridge threads while preserving shared-value and expression-lifetime boundaries.
Improve diagnosis and handling of failures involving structured output and backend tools.
Evaluate native document transport behind an explicit backend capability.
Define acceptance criteria and integration examples for additional backends.
Explore shared skill metadata that adapters translate into each backend's native format.
- Which task-specific guidance improves use of inspection and discovery?
- Which debugging views make model decisions easier to assess alongside Python effects?