Use when the user explicitly selects this chain. First readExecution and recovery, executing along the same SDK root Run and the current sub-Run.
Responsibilities of the original Skills
-
grill: load the fixed original grill-me text; it only requires running a grilling session and has no file output.
Execute immediately: Use the Skill tool to load the grilling skill. Skipping this step is prohibited.
After the skill is loaded, interview according to the original design tree/frontier rules. The sub-Agent investigates environmental facts; actual decisions are asked to the user and answers awaited. Once the frontier is empty, ask the user to confirm shared understanding. Keep the real session and return the session summary to the SDK grill Action; save the understanding confirmation through that sub-Run's confirm-understanding Wait. This phase does not create briefs or other business documents and does not borrow Native's phase artifacts.
-
to-spec: read the fixed original to-spec text, synthesize the confirmed session and code facts, and do not interview again on decisions already made. Generate the first formal document with the original template's seven chapters, confirm the test seams, and publish to the local tracker's .scratch/grill-test/spec.md. Save the actual confirmation through confirm-spec.
-
to-tickets: read the fixed original to-tickets text, present the tracer-bullet slices, granularity, and blocking edges, and after obtaining the user's decision, write one ticket per file according to the original template into .scratch/grill-test/issues/, using the ready-for-agent label. tickets.json is only an index this application saves to check the actual files and dependencies; it is not an artifact of the original Skill. Save the actual confirmation through confirm-tickets.
-
implement: after the root Run's authorize-implementation is approved, read the fixed original implement text. Implement according to the approved tickets, verify at the agreed seams using tdd, run the relevant checks and the final complete business tests, conduct the Standards/Spec dual-axis sub-Agent review according to the original code-review text, and commit only locally in the isolated project. implementation.json is the application's execution record and is not called an artifact of the original Skill.
The original files of grill-me, to-spec, to-tickets, and implement contain disable-model-invocation; when the host cannot nest calls, read the current package's fixed original text and truthfully record the file loads; grilling, tdd, and code-review must be loaded through the actual Skill tool when available. Do not modify the original Skills' frontmatter or body.
Acceptance and recovery
Inputs are goal, acceptance[{id,description}], and a testCommand array; the local tracker is chosen by the user in advance. All business checks use the original test commands; verify's machine execution port actually runs the tests, and the verifier only reads the results. The independent Eval separately runs the original tests; task completion cannot be replaced by a report or string.
The approved Spec and tickets remain read-only during the implementation phase. When a real approval or Skill capability is missing, stop at the current step and preserve the original Run; for running/unknown, first check the host and actual results. Use only the current Action/Wait identity and do not fabricate session answers, approvals, or success states. Only when the actual SDK root Run is completed, the business tests pass, and the review evidence is verifiable may the business task be reported as complete.
Credentials remain only in the execution environment. Fixed application, SDK state, and installation directories are maintained by the official Runtime. This application does not authorize remote Issues, PRs, pushes, releases, or real messages.
All four stages retain the original Skill's execution responsibility within the SDK subprocesses; the root Run's authorize-implementation still confirms the implementation scope before the implement subprocess starts. The application Guard only allows actual stage artifacts of claimed sub-Actions; grill has no business file scope.
SDK Execution Protocol
Fixed combination: e0a776e8c9940ecf07ee97c8399d4500bdaa6889e717966487c075ef33a19fa5 Runtime: 0.5.0; base flow: standalone.
Start the following flow with comet runtime dispatch --application-file <this directory>/application.json --project-root <project> --request <temporary JSON>:
- grill-test@3
- grill-me@1
- to-spec@1
- to-tickets@1
- implement@1
For query and recovery, use --application grill-test with the original Run ID, first inspecting the original Action; preserve attempt, inputHash, and claimToken. For unknown executions, verify the result first. Approvals are submitted only along the Runtime's current Wait.
When the current Action returns skillWork or the current Wait returns waitSkillWork, read the actual name of skill.files["SKILL.md"] item by item; if there is no name, determine it according to the host naming rules of that fixed directory, and block if it cannot be determined. skill.id is only a logical binding; the application name or alias cannot be used in place of the actual Skill name.
Execute immediately: Use the Skill tool to load the skill. Skipping this step is prohibited.
After the skill is loaded, verify that the actual loaded source corresponds to the current skill.root, and verify that the complete content digest of SKILL.md, scripts, and resources matches skill.contentHash; the same name cannot substitute for the fixed version. If the host cannot load that fixed version, block and preserve the original Action. After verification passes, the Action retains actionId, attempt, and inputHash; the Wait retains waitId and proposalHash, and the wait point cannot be treated as an Action claim. Pass the current input, scope, and fixed guidance, execute the work, and submit the actual result; wait-point guidance is not equivalent to a user decision or machine-check evidence, and approvals still submit only the user decision for the current Wait.
See application.json for the flow and bindings, and installation.json for installation and recovery identity. A completion string cannot replace actual Schema, candidate, or artifact checks; after a check rejects, preserve the scene, correct the actual artifacts, and continue the original action. This package passes checks through compositional structure and actual registration implementation; real host, model execution, and full business acceptance must be recorded separately.
See artifactContracts and application rules in application.json for the required artifacts per stage. Before reporting succeeded, check the actual files and content conditions; when sha256From is present, return the corresponding actual SHA-256, otherwise outcome.artifacts provides uri/sha256. The SDK execution port collects file evidence. Re-verify the dependencies of the original Action before downstream startup and completion; when files change, preserve the scene and re-check according to the declared repair steps.
Read before executionapplication rules. Rules are generated together with the flow and write declarations; host integration is deployed through the formal installation preview.