Loopforge

Loopforge

把模糊工作流锻造成可复用的 AI Agent 工作流循环规范。

Route

Choose one path:

• Extract: convert notes, transcripts, prompts, artifacts, or history into a loop spec.
• Forge: create a new loop spec from the user's intended recurring workflow.
• Doctor: audit and minimally repair an existing loop spec.
• Package: prepare a loop spec for reuse or publication.

Do not handle:

• official published loop search; send that to the official Loop Library skill or catalog
• direct loop execution; hand off to the runtime
• pure /goal writing; use a goal skill when available
• general systems-thinking feedback-loop analysis

Intake

Ask one high-leverage question by default; ask up to three only when answers materially change the spec. Default only when grounded or safe no-op. Ask for permissions, security, production writes, external messages, private data, and validation.

Required Reading

• For boundary decisions, read references/loop-vs-goal-sop-checklist.md.
• For spec fields, read references/loop-spec-v1.md.
• For repair work, read references/loop-doctor.md.
• For examples or pattern help, read references/case-patterns.md and references/templates.md.

Output Contract

Return one:

• a concise prompt-mode loop the user can copy
• a full loop-spec-v1 YAML or JSON spec, followed by a copy-ready Loop Prompt companion unless the user asks for machine-only output
• a doctor verdict: Ready, Repair needed, or Not actually a loop
• a package checklist for reuse or publication

Every spec must satisfy references/loop-spec-v1.md: identity, trigger/exclusion, inputs, authority, state, round, next_action_rule, working_signal, acceptance_gate, terminal states, outputs, provenance, and optional execution_handoff.

When outputting YAML/JSON, also finish with the module-based Loop Prompt from references/templates.md. YAML/JSON is the source; the prompt is the copy/use entry.

Validation

For YAML/JSON specs, run:

python3 scripts/lint_loop_spec.py <spec-file>

Use --profile template for reusable templates with declared placeholders. Use --profile runnable when the loop is ready to run and placeholders must be gone.

If the linter fails, fix the spec before presenting it as ready.