@mriekki/project-forge 2.2.0 → 4.0.0

package/assets/generated/core/core/skills/grill-me/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: grill-me
-description: Continuously challenge requirements, architecture, plans, implementation assumptions, or work-item coverage without editing files.
+description: Continuously challenge requirements, architecture, plans, implementation assumptions, or work-item coverage, with optional session-level reconciliation of resolved decisions into permitted authorities.
 argument-hint: "<work-item ID, external key, document, plan, or topic>"
 user-invocable: true
 disable-model-invocation: true
@@ -8,70 +8,108 @@ disable-model-invocation: true
 
 # Grill Me
 
-Purpose: improve reasoning by challenging the highest-value uncertainty without adding process overhead.
+Purpose: improve reasoning by challenging the highest-value uncertainty without adding process overhead or forcing the developer to choose a separate documentation-aware grilling command.
 
-## Rules
+Invocation begins read-only.
 
-- Read `.github/project-forge-workspace.json` when present and only the minimum relevant context.
-- Ask one material question at a time.
-- Explain briefly why the question matters.
-- Challenge vague answers, hidden scope, contradictions, and weak assumptions.
-- Recommend an answer or 2–3 options with tradeoffs when useful.
-- Do not edit files.
-- Do not invoke subagents.
-- Do not create artifacts merely because grilling occurred.
-- Do not claim external Jira/Issue content was read unless it was actually available.
+## 1. Resolve authority and minimum context
 
-Challenge areas may include:
-- missing acceptance behavior,
-- cross-work-item/Epic gaps,
-- domain/business rules,
-- lifecycle/state transitions,
-- validation/failure behavior,
-- data integrity,
-- concurrency/transactions,
-- destructive/reversible actions,
-- architecture assumptions,
-- security/operational risk,
-- testability.
+Read `.github/project-forge-workspace.json` when present and only the minimum relevant context.
 
-## Cadence
+Determine when applicable:
+- the authoritative work-item source;
+- which local documents are authoritative, supplemental, read-only, or not used;
+- relevant accepted ADR/design rules;
+- whether an external write integration is actually available and permitted.
 
-```text
-QUESTION
-<one focused question>
+Never assume Product/Feature Briefs exist. Never assume `Jira Epic == Feature Brief`. Do not claim external Jira/Issue content was read unless it was actually available.
 
-WHY IT MATTERS
-<short explanation>
+## 2. Continuous grilling
 
-RECOMMENDATION
-<recommended answer/options when useful>
-```
+Ask one material question at a time. Explain briefly why the question matters. Challenge vague answers, hidden scope, contradictions, and weak assumptions.
+
+Challenge areas may include:
+- missing acceptance behavior;
+- cross-work-item/Epic gaps;
+- domain/business rules;
+- lifecycle/state transitions;
+- validation/failure behavior;
+- data integrity;
+- concurrency/transactions;
+- destructive/reversible actions;
+- architecture assumptions;
+- security/operational risk;
+- accessibility/UX implications when relevant;
+- testability.
+
+When the purpose is to expose or strengthen the human's reasoning, ask open-ended first. For a bounded decision with a small credible option set, follow `INTERACTIVE_SESSION_GUIDE.md`: use host-native single-select structured choice UI for mutually exclusive alternatives when available, multi-select only when several selections are valid together, and preserve free-text/`Other / specify` when choices may be incomplete.
 
-After the human answers:
+Recommend at most one answer/option only when current evidence or a clear engineering tradeoff supports it. A recommendation is advice, not an accepted decision.
+
+Cadence:
 
 ```text
-analyze/challenge the answer
+question
+→ human answer
+→ challenge/analyze the answer
 → explain implications when useful
-→ ask the next material question
+→ reconcile the resolved decision when the session is authorized to do so
+→ next material question
 ```
 
 Do not ask `Do you want to continue?` after each answer. The human already entered a continuous grilling mode.
 
-Stop only when:
-- the human says stop/done/pause,
-- no material issues remain,
-- a true scope/decision/action gate appears,
+## 3. Optional decision reconciliation
+
+Grilling does not require document writes. When resolved decisions would benefit from durable reconciliation, present one session-level **single-select structured choice** through the host UI when available. In VS Code use the native question surface such as `vscode_askQuestions`.
+
+Offer:
+- `Continue grilling only` `[READ ONLY]` — default when no durable update is needed;
+- `Reconcile resolved decisions during this session` `[MODIFIES DOCS]`;
+- `Stop` `[NO ACTION]`.
+
+Use a numbered text list only when structured choice UI is unavailable.
+
+Selecting `Reconcile resolved decisions during this session` authorizes only focused updates to the permitted current work-item documents for this active session. Do not ask again after every answer.
+
+When a decision is resolved:
+- update an existing permitted local authority/supplement only when this session has document-write authorization;
+- if Jira/GitHub/another external system is authoritative and no permitted write integration is available, accumulate a concise `PROPOSED EXTERNAL WORK-ITEM UPDATE`;
+- if a permitted external write integration exists, require an explicit `[EXTERNAL WRITE]` choice before writing externally;
+- never create a competing local source merely to store an answer.
+
+For local-document models, common homes are:
+- brief: high-level scope/outcome/major rule;
+- combined requirements: detailed user/engineering behavior and testable scenarios;
+- decisions: resolved work-item decision/basis;
+- ADR: only a meaningful cross-cutting architecture decision.
+
+Pause reconciliation when an answer would materially change higher-level approved scope, contradict another authority/ADR, or require a new cross-cutting decision. Continue read-only grilling when appropriate.
+
+## 4. Boundaries
+
+Do not:
+- implement production code;
+- invoke subagents;
+- create artifacts merely because grilling occurred;
+- infer approval from an answer;
+- silently write local or external authorities without the applicable session/write authorization.
+
+Stop when:
+- the human says stop/done/pause;
+- no material issues remain;
+- a true scope/decision/action gate appears;
 - the human switches workflow.
 
-## Natural completion
+## 5. Natural completion
 
 Summarize once:
-- decisions strengthened,
-- unresolved issues,
+- decisions strengthened;
+- documents or proposed external updates changed, if any;
+- unresolved issues/conflicts;
 - recommended next workflow.
 
 Typical next workflows:
-- `/grill-with-docs` to reconcile resolved decisions,
-- `/refine-requirements` for systematic detailed coverage,
-- `/start-work` when the item is ready.
+- `/refine-requirements` for systematic detailed coverage;
+- `/start-work` when the item is ready;
+- normal review/approval when a material requirement gate remains.

package/assets/generated/core/core/skills/progress/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: progress
-description: "Read-only developer-facing progress view across POC creation, this workspace instance, work-item/delivery state, and exact project-defined environment capabilities."
+description: "Read-only developer-facing progress view across Project creation, this workspace instance, work-item/delivery state, and exact project-defined environment capabilities."
 argument-hint: "[optional work-item ID or environment]"
 user-invocable: true
 disable-model-invocation: true