Techne-Loom

Imported from waynebaby/Techne-Loom (AGENTS.md). Install upstream with npx skills add waynebaby/Techne-Loom. Copyright stays with the author.

Workspace Agent Rules

AGENTS.md is the automation-facing source of repository execution rules.

This file intentionally contains Level 1 rules only. Scope-specific details live in the linked instruction documents below and must be loaded when their file patterns match the current change.

Shared Python Environment

This workspace may use the shared virtual environment pointer from .venv.path.

• Resolve .venv.path with PowerShell on Windows or bash on Linux before invoking Python tooling.
• If the configured .venv.path target does not exist, initialize that virtual environment before invoking Python tooling.

Run Output Naming

• When a skill creates a per-run output root, keep the skill-owned parent directory and name the run root exec-<YYYYMMDD_HHMMSS>-<skill-slug>-result/.
• Keep the timestamp immediately after exec- so runs remain sortable even when adjacent steps switch skills.

Level 1 Rules

Repository identity

• Techne Loom is a .NET-first multi-ecosystem mono-repo with parallel package families under /src/dotnet, /src/nodejs, and /src/python.
• Every project unit is a publishable package. Keep package families parallel by role: abstractions, common, agent-orchestrator, and skill-orchestrator.
• AgentOrchestrator and SkillOrchestrator are independent products. They do not call each other and must not be framed as a parent/child runtime pair.
• Use Loom Agent Execution Orchestrator for AO user-facing narrative while preserving implementation identities such as Techne.Loom.AgentOrchestrator, dotnet ao.dll, /loom-plan-execution, and source/type names.

Safe editing and Git hygiene

• GitHub Copilot must not use apply_patch in this repository. Use the checked-in external-file-driven range editor or another repository-approved editing mechanism; do not embed multiline replacement content in commands.
• Preserve user changes. Do not use destructive Git commands or rewrite shared history unless the user explicitly requests it.
• File-valued CLI inputs are path-only and must be complete on disk before a command starts. Inline scripts, JSON, patches, workflows, references, or replacement content are not valid inputs.
• Keep generated output, audit material, caches, and temporary validation files outside source and skill directories unless they are explicitly requested deliverables.

Workflow invariants

• A workflow's business steps must match its declared business intent. Governed workflow instances declare taskType and workflowKind; enhancement and target-business workflows must not be interchanged.
• Plan, replan, compile, run, resume, and audit are disk-backed and sessionless. Each product owns one canonical WorkflowInstance; events, logs, audits, envelopes, and large artifacts are companion evidence, not a second mutable execution truth.
• caseId and runId stay on the same external workflow copy through the full execution chain. They identify one business execution and do not replace business outputs.
• AO and SO remain independent runtimes with independent package, CLI, release, and product-facing boundaries.
• Governed routes use fragment-first workflow access and preserve explicit ownership boundaries for user inputs, runtime facts, and business outputs.

Documentation and public contracts

• Public documentation is bilingual by default under mirrored /docs/en and /docs/zh-cn trees. Root public docs require their Chinese mirrors; AGENTS.md and other agent configuration files remain English-only.
• Workflow definition files are the canonical English carrier for schema keys and control semantics. Localization belongs in the presentation layer and must not change wire names or workflow behavior.
• Keep AGENTS.md at the repository root. Do not create a second agent-rules source under docs.

Validation and delivery

• For both development and main, start Windows and WSL restore, build, and test jobs in parallel. Keep their build/intermediate outputs isolated, collect results separately, and wait for both before declaring validation complete.
• Tests that exercise intentionally long-running tasks must not set a test timeout or add timeout-based cancellation merely to shorten the run. Let the task complete naturally and distinguish genuine failures from runner or environment interruption.
• Before code check-in, generated AO/SO schema and demo evidence, focused tests, and platform validation must match the affected scope. Detailed validation commands and artifact rules are in the validation instruction document.
• Review and validate each major implementation slice before starting the next. Do not carry unreviewed or uncommitted major work across slices unless the user explicitly overrides the cadence.

Scoped Instruction Documents

Read the applicable document before changing files in its scope. These documents expand the L1 rules; they do not replace them.

When more than one scope matches, load all matching instruction documents. If a detailed instruction conflicts with a Level 1 rule, the Level 1 rule wins.