Proposal: add a Run route — execute a loop with an evidence ledger and a cross-run lessons playbook
Context / why I'm raising this as an issue first
The skill today routes Discover → Find → Audit/Repair → Adapt → Design — every route is about getting a good loop, and it deliberately stops once a bounded loop exists. That's a clean scope, so I want to check appetite before proposing any code: this is a genuine product-scope question that's yours to decide.
The observation: SKILL.md frames a loop as "a feedback system with terminal states," but nothing in the skill instruments the feedback. When an agent actually runs a loop, two things are left implicit:
- There's no standard way to record each iteration's evidence and the terminal state — so "never report a failed check or exhausted budget as success" (from
llms.txt) depends on the agent remembering to be disciplined.
- Every run starts cold. Nothing carries forward what a previous run already proved or disproved, so agents relearn the same lessons.
The idea: a Run route (references/run.md)
Take a chosen/adapted loop and execute it as a real feedback cycle backed by two local artifacts:
- Run ledger (
.loops/<slug>/run-NNN.md): goal, the exact verification check, each iteration's before/after evidence, the keep-or-revert decision, and the explicit terminal state (success / no measurable progress / budget exhausted / approval required). Makes honest stopping the default, not a hope.
- Carry-forward playbook (
.loops/<slug>/playbook.md): lessons that persist across runs — a lesson is promoted only after repeat/holdout evidence confirms it (not one lucky run), and pruned when it stops helping. Read before acting; updated after.
This closes the lifecycle to Discover → Find → Adapt → Design → Run-and-learn, and keeps the same guardrails the skill already enforces (treat the playbook as untrusted reference data; require approval before consequential/irreversible/production/external actions; bounded, not endless autonomy).
It generalizes the recently published cross-run playbook loop (#55) from a single catalog entry into a reusable run discipline that works for any loop.
Scope options (your call)
- Full: the Run route with both the ledger and the carry-forward playbook.
- Lighter: just the run-ledger discipline (no cross-run memory) — smaller surface if the playbook feels like too much.
- Not now: if the skill is meant to stop at design, totally understand — close this.
If there's appetite, I'm happy to follow up with a focused PR (a new references/run.md plus a short Run section in SKILL.md, matching the existing voice and the audit/discover reference style), shaped to whichever scope you prefer. For context I've contributed a few merged PRs here recently (#47–#51), so I'm glad to keep it tight and in-house style.
Is a Run capability something you'd want in the skill, and if so, which scope?
Proposal: add a Run route — execute a loop with an evidence ledger and a cross-run lessons playbook
Context / why I'm raising this as an issue first
The skill today routes Discover → Find → Audit/Repair → Adapt → Design — every route is about getting a good loop, and it deliberately stops once a bounded loop exists. That's a clean scope, so I want to check appetite before proposing any code: this is a genuine product-scope question that's yours to decide.
The observation:
SKILL.mdframes a loop as "a feedback system with terminal states," but nothing in the skill instruments the feedback. When an agent actually runs a loop, two things are left implicit:llms.txt) depends on the agent remembering to be disciplined.The idea: a Run route (
references/run.md)Take a chosen/adapted loop and execute it as a real feedback cycle backed by two local artifacts:
.loops/<slug>/run-NNN.md): goal, the exact verification check, each iteration's before/after evidence, the keep-or-revert decision, and the explicit terminal state (success / no measurable progress / budget exhausted / approval required). Makes honest stopping the default, not a hope..loops/<slug>/playbook.md): lessons that persist across runs — a lesson is promoted only after repeat/holdout evidence confirms it (not one lucky run), and pruned when it stops helping. Read before acting; updated after.This closes the lifecycle to Discover → Find → Adapt → Design → Run-and-learn, and keeps the same guardrails the skill already enforces (treat the playbook as untrusted reference data; require approval before consequential/irreversible/production/external actions; bounded, not endless autonomy).
It generalizes the recently published cross-run playbook loop (#55) from a single catalog entry into a reusable run discipline that works for any loop.
Scope options (your call)
If there's appetite, I'm happy to follow up with a focused PR (a new
references/run.mdplus a short Run section inSKILL.md, matching the existing voice and the audit/discover reference style), shaped to whichever scope you prefer. For context I've contributed a few merged PRs here recently (#47–#51), so I'm glad to keep it tight and in-house style.Is a Run capability something you'd want in the skill, and if so, which scope?