loop-hardening-contract
Harden repeated delivery loops with evidence-based harness stripping, immutable runtime-path sprint packets, and broken-window revalidation without automatic revert. Use for Cardputer, browser, CI, runtime, or production work split into repeated sprints.
Install
npx skills add https://github.com/markoblogo/abvx-agent-skills/tree/main/skills/loop-hardening-contract
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install markoblogo-abvx-agent-skills@llmmart
git clone https://github.com/markoblogo/abvx-agent-skills.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole markoblogo/abvx-agent-skills collection as a plugin from our marketplace. Git is the plain clone.
Skill manifest
Loop Hardening Contract
Use this contract after loop-readiness-review when repeated delivery needs tighter execution evidence. It is a review and packet format, not a runner or scheduler.
1. Harness Stripping
Remove harness complexity only through an isolated, one-component experiment:
- Record the component, its intended failure mode, and its operating cost.
- Capture a representative baseline for completion, verification, cost, and failures.
- Disable exactly one non-safety component between runs, never during a live run.
- Repeat the same bounded case and classify
REMOVE,RESTORE,HARMFUL, orINCONCLUSIVE. - Retain the evidence and human settlement decision.
Never strip sandboxing, budgets, stop conditions, human gates, root verification, or protected-data boundaries.
2. Runtime-Path Sprint Packet
Freeze this packet before execution:
sprint_idand version;- one concrete deliverable;
- 3–7 observable acceptance predicates;
- the actual runtime path to exercise;
- explicit out-of-scope items;
- owned files or write boundary;
- stop conditions and budget;
- verification commands, device/browser steps, and expected evidence.
The executor cannot accept its own result. A mid-run requirement change creates a new packet version with a reason and explicit human confirmation.
3. Broken-Window Revalidation
Before a new sprint, re-run the latest accepted result through its real runtime path:
STILL_GREEN: the accepted predicates still pass;REOPENED: a previously accepted predicate now fails;INCONCLUSIVE: the required runtime or evidence is unavailable.
On REOPENED, stop new work, retain the evidence, reopen the result, and propose a bounded repair. Do not auto-revert. A revert may erase useful subsequent work or hide the causal change and therefore requires an inspected, human-approved settlement.
Compilation is not Cardputer device proof. Unit tests are not browser or production proof when the accepted claim names those runtime paths.
Output
Return:
- sprint packet version;
- revalidation status and evidence;
- harness experiment and classification, if any;
- root verification result;
- settlement:
accept,revise,repair,restore, orinconclusive; - remaining risks.
Guardrails
- Keep the contract opt-in and bounded; do not install Loopkit or add scheduling.
- Do not auto-revert, auto-commit, auto-merge, deploy, release, or publish.
- Do not use source inspection as runtime-path evidence.
- Root verifies integration; executor prose is evidence, not acceptance.
- Prefer a one-shot checklist when the work is not genuinely repeated.
Pair With
loop-readiness-reviewfor readiness level, state, isolation, and human gates.bounded-evaluationfor representative fixtures and comparison rubrics.bug-evidence-protocolfor captured red-to-green command evidence.git-native-context-contractcpatrecords for recurring failure prevention.bounded-orchestration-contractwhen execution is delegated across disjoint file scopes.
Provenance
Adapted from Archive228/loopkit sprint, broken-window, and harness-evolution ideas. The ABVX version removes the autonomous runner, transcript scraping, self-declared completion, automatic rollback, broad installation, and unpinned runtime dependencies.
Files (abvx-agent-skills)
-
agents
-
openai.yaml 232 B
Loading…
-
-
SKILL.md 3.6 KB
--- name: loop-hardening-contract description: Harden repeated delivery loops with evidence-based harness stripping, immutable runtime-path sprint packets, and broken-window revalidation without automatic revert. Use for Cardputer, browser, CI, runtime, or production work split into repeated sprints. license: MIT metadata: abvx_status: experimental abvx_origin: adapted --- # Loop Hardening Contract Use this contract after `loop-readiness-review` when repeated delivery needs tighter execution evidence. It is a review and packet format, not a runner or scheduler. ## 1. Harness Stripping Remove harness complexity only through an isolated, one-component experiment: 1. Record the component, its intended failure mode, and its operating cost. 2. Capture a representative baseline for completion, verification, cost, and failures. 3. Disable exactly one non-safety component between runs, never during a live run. 4. Repeat the same bounded case and classify `REMOVE`, `RESTORE`, `HARMFUL`, or `INCONCLUSIVE`. 5. Retain the evidence and human settlement decision. Never strip sandboxing, budgets, stop conditions, human gates, root verification, or protected-data boundaries. ## 2. Runtime-Path Sprint Packet Freeze this packet before execution: - `sprint_id` and version; - one concrete deliverable; - 3–7 observable acceptance predicates; - the actual runtime path to exercise; - explicit out-of-scope items; - owned files or write boundary; - stop conditions and budget; - verification commands, device/browser steps, and expected evidence. The executor cannot accept its own result. A mid-run requirement change creates a new packet version with a reason and explicit human confirmation. ## 3. Broken-Window Revalidation Before a new sprint, re-run the latest accepted result through its real runtime path: - `STILL_GREEN`: the accepted predicates still pass; - `REOPENED`: a previously accepted predicate now fails; - `INCONCLUSIVE`: the required runtime or evidence is unavailable. On `REOPENED`, stop new work, retain the evidence, reopen the result, and propose a bounded repair. Do not auto-revert. A revert may erase useful subsequent work or hide the causal change and therefore requires an inspected, human-approved settlement. Compilation is not Cardputer device proof. Unit tests are not browser or production proof when the accepted claim names those runtime paths. ## Output Return: - sprint packet version; - revalidation status and evidence; - harness experiment and classification, if any; - root verification result; - settlement: `accept`, `revise`, `repair`, `restore`, or `inconclusive`; - remaining risks. ## Guardrails - Keep the contract opt-in and bounded; do not install Loopkit or add scheduling. - Do not auto-revert, auto-commit, auto-merge, deploy, release, or publish. - Do not use source inspection as runtime-path evidence. - Root verifies integration; executor prose is evidence, not acceptance. - Prefer a one-shot checklist when the work is not genuinely repeated. ## Pair With - `loop-readiness-review` for readiness level, state, isolation, and human gates. - `bounded-evaluation` for representative fixtures and comparison rubrics. - `bug-evidence-protocol` for captured red-to-green command evidence. - `git-native-context-contract` `cpat` records for recurring failure prevention. - `bounded-orchestration-contract` when execution is delegated across disjoint file scopes. ## Provenance Adapted from `Archive228/loopkit` sprint, broken-window, and harness-evolution ideas. The ABVX version removes the autonomous runner, transcript scraping, self-declared completion, automatic rollback, broad installation, and unpinned runtime dependencies. -
SKILL_CARD.md 1.8 KB
Loading…
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.