design-loop

Files (2)

• SKILL.md
• openai.yaml · agents/

SKILL.md (first lines)

Use the [handoff contract](../../references/design-loop-contract.md). Inputs: endgoal, starting workspace/artifacts, output directory and optional N, criteria, constraints, domains and scoped model/effort choices. Reuse context and extend it for new work.

## Bounds

`N` caps attempted cycles, including the first PoC and failures. Require a positive integer; default to 3 and state it. Apply core's iteration bounds; brainstorming is a cycle's first work. Run through N unless the caller stops, a resource bound is reached, required capability or authorization is missing, or explicit stop-on-goal applies. Goal attainment alone does not shorten the run.

Choose stage assignments under core execution rules, preserving dependencies, independent testing and caller limits.

## Run

1. Establish or resume the checkpoint, state N and bounds, and preserve the initial baseline. Start each cycle from the last accepted state. Target the smallest working PoC first; later cycles use accumulated decisions and observations, including failures.
2. Record a new attempt before work, or resume the active one. Apply [brainstorm-research](../brainstorm-research/SKILL.md) with request, baseline and observations, then [design-increment](../design-increment/SKILL.md) to its result.
3. Create an isolated baseline copy; apply [implement-increment](../implement-increment/SKILL.md) with the design. Freeze the returned candidate; pass it, the baseline and evaluation plan to [test-increment](../test-increment/SKILL.md).
4. Apply [analyze-iteration](../analyze-iteration/SKILL.md) to all cycle evidence, including failures/gaps. Check its recommendation; record adopt/retain, exact accepted state and next-cycle observations. Checkpoint each returned stage; use the rules below for incomplete stages.
5. Stop at the bound or declared early-stop condition. Further work requires a new or extended caller request.

Brainstorm-research calls two leaves; testing supplies one independent comparison round. There is no additional final review or repair loop.

## Failure and resumption

On failure, retain evidence and mark dependent stages unexecuted. Run scheduled analysis on the partial handoff if possible; otherwise record missing analysis and retain the baseline. Failed attempts count. Continue only if another bounded cycle can progress; persistent capability/authorization blocks end the run with a gap. Corrections belong to later cycles within N.

For no justified change, mark implementation/testing inapplicable, analyze the evidence and record retain. The attempt counts; normal early-stop rules apply.

To resume, read checkpoint and attempt-start records and verify state/artifact identities. Continue the same attempt's first unfinished stage, reusing valid work and rejoining live assignments before replacements. Changed inputs make affected downstream evidence stale; repeating completed stages requires a new or extended caller run.

## Return

Return exact accepted state, usage, initial-to-final changes/evidence, completed/attempted counts, decisions, gaps and checkpoint path. Keep retained candidates as evidence. For requested integration, verify the checkout still matches its preserved starting state; report conflicts instead of overwriting newer work.

From DanMcInerney/orchflows, MIT.

More from DanMcInerney/orchflows