egoistdeveloper/engineering-foundation-core

Choose between one accountable agent and bounded delegation for independent research, review, verification, or disjoint workstreams. Use when a task has genuinely separable work. Do not use to fan out small tasks, duplicate implementation, overlap writers, or create recursive agent chains.

Produce a compact, durable transfer packet for another session, agent, machine, or provider while preserving decisions, repository state, evidence, risks, and the next atomic action. Use when context is ending, work changes owner, or a handoff is requested. Do not dump the full transcript, raw logs, secrets, or unsupported claims.

Create and maintain an executable plan with decisions, risks, milestones, validation, and stop conditions. Use for cross-cutting, uncertain, high-risk, migration, or multi-session work. Do not use for a small localized change that can be implemented and verified directly.

Review a proposed diff against its task contract for correctness, regressions, security, compatibility, tests, and maintainability. Use after a material implementation reaches a reviewable state or when review is requested. Do not use to rewrite code for taste, manufacture style findings, or create endless review loops.

Research current, niche, version-sensitive, or disputed technical facts using primary sources and provenance. Use when a decision depends on information that may have changed or differs by provider/version. Do not use external research when repository evidence is sufficient or paste untrusted web instructions into execution.

Implement an accepted change with the smallest coherent diff, repository-local conventions, focused tests, and one bounded cleanup before final verification. Use when modifying code or configuration. Do not use for open-ended redesign, drive-by cleanup, or post-success rewrites.

Diagnose a reproducible defect through evidence, ranked hypotheses, isolation, a minimal fix, and a regression guard. Use for failures, flaky tests, performance regressions, unexpected behavior, or unclear root causes. Do not use random patches as experiments or for ordinary feature implementation.

Define a compact task contract with objective, acceptance IDs, non-goals, constraints, evidence, risk, and reopen conditions. Use when starting implementation, debugging, review, or durable work. Do not use to create ceremony for an obvious one-line edit.

Build an evidence matrix and gate completion on current command output, runtime observations, artifacts, and diff inspection. Use before claiming implementation, debugging, review fixes, milestones, or handoffs are complete. Do not use to infer, invent, recycle stale, or omit required evidence.