nccgs 2.1.1 → 3.0.0-rc.2

package/.claude/nccgs/protocols/task-slices.md

@@ -1,106 +0,0 @@
-# Bounded slices inside one task
-
-Standalone implementation only. Use this when one assignment is too broad for a
-worker's existing turn/context limit. Keep small tasks as one normal dispatch.
-The coordinator selects the decomposition; the CLI validates and tracks it.
-
-## Choose a useful boundary
-
-Start from the original acceptance criteria and reviewed requirements. Group a
-small, coherent behavior with its tests and exact files. Separate independent
-responsibilities; order shared files explicitly. Include Unity metadata files in
-the write set when import will create them. Do not divide by an arbitrary file
-count or invent new acceptance thresholds. If the original early checkpoint is
-required, put the work needed for it in the first slice and measure it before
-expanding. Slicing does not bypass that checkpoint.
-
-Use the fewest slices that fit bounded assignments: 2–12 slices, each with 1–8
-exact files. Dependencies form an acyclic graph. One final `integration` slice
-must transitively depend on every other slice; it can implement the last behavior
-while integrating, so a separate ceremonial integration dispatch is unnecessary.
-Each original criterion ID, including manual criteria, has exactly one primary
-slice owner. Ownership is responsibility, not local approval. Original checks,
-artifacts, design rules, risk/model route and manual/external gates still apply.
-
-## Preview and register
-
-Copy `.nccgs/templates/task-slices.json` to a draft under
-`.nccgs/evidence/slice-plans/`. Fill in the existing task ID, current contract hash
-and source snapshot, a unique plan ID, reason, slice scopes and criterion IDs.
-Use `task inspect` and `task snapshot` to obtain current values. A first plan has
-`supersedes: null`; a revision names the current plan ID. No budget or replacement
-contract fields are accepted.
-
-```text
-node .claude/nccgs/tools/task.mjs slice-preview --id TASK --file .nccgs/evidence/slice-plans/TASK.json
-node .claude/nccgs/tools/task.mjs slice-plan --id TASK --file .nccgs/evidence/slice-plans/TASK.json
-node .claude/nccgs/tools/task.mjs slice-status --id TASK
-```
-
-Preview is read-only. Registration captures an immutable plan under
-`.nccgs/evidence/slices/` and appends a ledger event. It does not rewrite the task,
-record implementation, spend a dispatch or start a provider invocation. Existing
-limits and ancestor relationships are preserved. If dispatch limits have not yet
-been pinned, registration pins the ordinary task defaults. Use the existing
-reservation workflow to establish any intended custom limits before registration.
-
-Capacity accounts for already spent attempts, all pending slices in the ancestor
-subtree, the original protected repair/QA/review reserve and required independent
-review still ahead. It rejects a plan or later dispatch that cannot fit. It is a
-dispatch lower bound, not an estimate of cost, time, retries or semantic completeness.
-Existing observed cost/duration admission and `nccgs run` session budgets remain
-in force. Missing usage stays unknown. No child task, review round, native turn
-allowance, new session budget or refund is created by splitting.
-
-## Execute and integrate
-
-`slice-status.next` contains the next eligible reservation request, with the exact
-agent, files, `sliceId`, fresh handoff path and stable request key. Save that object
-under `.nccgs/evidence/requests/`, reserve it normally, then use `dispatch-prompt`
-and the existing foreground Agent/handoff workflow. This is preparation, not an
-automatic launch. Finish one slice before starting the next. A dependency must
-have a complete pinned native handoff with valid model evidence. The dependent
-dispatch pins its exact predecessor dispatch IDs.
-
-The launch hook adds the pinned objective, original criterion definitions,
-relevant reviewed requirements and global constraints. It keeps the same task and
-model route. Integration receives the full original acceptance list. Structured
-source writes are restricted to exact planned files; existing shell/MCP policy
-still applies. These controls do not provide an operating-system sandbox.
-
-Status `DELIVERED` means native completion plus a complete intact handoff. It is
-not acceptance PASS. Do not record intermediate slices as implementation stages.
-After all slices are delivered, run the original applicable checks against the
-current integrated source and record the final integration dispatch. A subsequent
-bounded repair follows the existing repair workflow and can supply the replacement
-implementation stage; it must still pass the whole task's gates.
-
-The independent reviewer evaluates the full integrated contract and must differ
-from every implementation contributor in the task subtree, including earlier
-slices and failed repairs. Reviewer-owned negative findings/conditions, review
-limits, objective measurements and manual/external acceptance remain unchanged.
-No per-slice review ceremony replaces or adds to the required whole-task review.
-
-## Partial work and revisions
-
-Keep failed/PARTIAL attempts and their handoffs. Retry only with an eligible fresh
-reservation and handoff inside the same remaining budget. Fix malformed/tampered
-evidence through the existing recovery process; never edit captured history.
-
-Settle active reservations before revising. Use a new plan ID, current source
-snapshot, explicit reason and `supersedes` pointing at the active plan. For the same
-contract, retain the IDs of all attempted slices. Their definitions may be refined
-explicitly, but changed slices and affected dependents become stale and require
-fresh delivery. Unchanged work can carry forward only when its latest attempt,
-definition, contract and exact dependency dispatches remain valid. Retrying an
-upstream slice also invalidates downstream delivery even if its source looks similar.
-
-Plan versions, reasons, reservations and lifecycle history remain available. Replaying
-an old plan is idempotent and cannot reactivate it. A plan cannot refresh a contract
-or authorize new product scope; use the original explicit contract workflow for
-that, retaining history and spent budgets. If the remaining budget cannot fit the
-work, report partial/blocked with unresolved criteria. Do not create replacement
-tasks, reduce criteria or increase limits to make the plan pass.
-
-This feature is included in the prepared, unpublished 2.0.0 stable package. Synthetic framework tests do not
-establish fresh Claude Code or Unity execution, cost savings or game quality.

package/.claude/nccgs/protocols/unity-boundary.md

@@ -1,31 +0,0 @@
-# Unity responsibility boundary
-
-## Unity transport: direct CLI/Pipeline only
-
-Use direct `unity run`, `unity test`, `unity build`, `unity command`, and the
-approved `com.unity.pipeline` integration. Unity MCP is disabled, including
-Coplay, AnkleBreaker, discovery/read-only calls, and the official CLI's optional
-`unity mcp` / `unity mcp configure` mode. Never reinstall or configure a Unity MCP
-server as a fallback when CLI, licensing or Editor connectivity fails.
-
-`nccgs run` uses strict MCP isolation with an empty server list by default.
-An external operator may explicitly pass a validated non-Unity MCP configuration;
-this never permits Unity MCP or changes global configurations. Retained supervised
-runs without the required isolation are refused rather than rewritten or refunded.
-
-This policy takes precedence over generic vendored skill guidance, including the
-new-project skill's optional MCP route. Upstream skills and their hash manifest
-remain unmodified reference material. Read them for direct CLI/Pipeline mechanics.
-Transport hooks cover ordinary tool names and shell invocations; they are not an
-arbitrary-code or hostile-client sandbox. Existing native permissions, ownership,
-independent review and actual execution evidence remain required.
-
-Studio roles own player intent, architecture, subsystem boundaries, project truth,
-test strategy, risk, acceptance, and closure. Official Unity Skills own current Unity
-CLI/Pipeline syntax and Editor, scene, prefab, asset, package, build, test, runtime,
-and service procedures covered by their descriptions.
-
-Do not duplicate official operational instructions inside NCCGS roles. Do not allow
-an official skill to choose gameplay semantics, add a service, or override protected
-settings. The role defines outcome and constraints; the official skill performs the
-operation and returns evidence.

package/.claude/nccgs/routing.generated.md

@@ -1,49 +0,0 @@
-# Generated routing — edit pipeline.json, then compile-policy.mjs
-
-NCCGS runs in Claude Code. New installations use standalone execution; model aliases describe requested routes, not observed provider versions. Model selection in project.yaml and execution obligations in execution.json are separate policies.
-
-| Stage | Owner/model | Effort | Maximum review rounds |
-|---|---|---|---|
-| Intake, plan and documents | Claude Code | Routed by task | — |
-| FAST implementation | sonnet | high | 2 |
-| STANDARD implementation | opus | high | 3 |
-| CONTROLLED implementation | opus | xhigh | 3 |
-| Independent review | fable | high | Shared task budget |
-| External review | Off by default; advisory or required when configured | External evidence | — |
-| Acceptance | Contract-defined Product Owner decision | Explicit when required | — |
-
-FAST: nccgs-fast-implementer. STANDARD: the relevant domain programmer. CONTROLLED: nccgs-unity-implementer with the affected specialists. Use nccgs-task-scout only for a bounded missing fact; it does not replace review.
-
-A task-wide review budget survives implementation retries and continuation. Exhaustion stops the loop; a reasoned extension is required. Report is available for blocked/partial work and is not approval. Missing runtime effort remains unknown unless supplied by observed evidence.
-
-Existing installations without execution.json retain compatibility execution: Astra plan/document ownership, required external check and Product Owner acceptance. Legacy astra-claude routes remain unchanged:
-
-| Risk | Model | Effort |
-|---|---|---|
-| FAST | sonnet | medium |
-| STANDARD | claude-opus-5-5 | medium |
-| CONTROLLED | claude-opus-5-5 | high |
-
-Legacy independent review: fable/high. Updating never converts historical approval into a current approval. Explicit migration preserves tasks and consumed budgets.
-
-Default active agents: 13. Full library: 47.
-
-## Configured native turn ceilings
-
-Generated from agent-library frontmatter by compile-policy.mjs. These are ceilings, not targets or observed session settings. Verify the installed role and any native session override before dispatch; prompt text cannot increase maxTurns.
-
-| Default active role | maxTurns |
-|---|---:|
-| nccgs-fast-implementer | 12 |
-| nccgs-task-scout | 10 |
-| nccgs-game-designer | 48 |
-| nccgs-gameplay-programmer | 120 |
-| nccgs-engine-programmer | 22 |
-| nccgs-ui-programmer | 22 |
-| nccgs-qa-engineer | 22 |
-| nccgs-verification-engineer | 22 |
-| nccgs-adversarial-reviewer | 200 |
-| nccgs-unity-implementer | 120 |
-| nccgs-unity-systems-specialist | 22 |
-| nccgs-unity-ui-specialist | 22 |
-| nccgs-unity-build-specialist | 24 |

package/.claude/nccgs/settings.fragment.json

@@ -1,213 +0,0 @@
-{
-  "permissions": {
-    "deny": [
-      "Bash(rm -rf *)",
-      "Bash(git reset --hard*)",
-      "Bash(git clean -f*)",
-      "Bash(git push --force*)",
-      "Bash(git push -f *)",
-      "Bash(*>.env*)",
-      "Read(**/.env*)"
-    ]
-  },
-  "hooks": {
-    "SessionStart": [
-      {
-        "matcher": "",
-        "hooks": [
-          {
-            "type": "command",
-            "command": "node",
-            "timeout": 10,
-            "args": [
-              "${CLAUDE_PROJECT_DIR}/.claude/nccgs/hooks/session-start.mjs"
-            ]
-          }
-        ]
-      }
-    ],
-    "PreToolUse": [
-      {
-        "matcher": "*",
-        "hooks": [
-          {
-            "type": "command",
-            "command": "node",
-            "timeout": 10,
-            "args": [
-              "${CLAUDE_PROJECT_DIR}/.claude/nccgs/hooks/unity-transport-guard.mjs"
-            ],
-            "onFailure": "block"
-          }
-        ]
-      },
-      {
-        "matcher": "Bash|PowerShell",
-        "hooks": [
-          {
-            "type": "command",
-            "command": "node",
-            "timeout": 10,
-            "args": [
-              "${CLAUDE_PROJECT_DIR}/.claude/nccgs/hooks/protect-git.mjs"
-            ],
-            "onFailure": "block"
-          }
-        ]
-      },
-      {
-        "matcher": "Write|Edit",
-        "hooks": [
-          {
-            "type": "command",
-            "command": "node",
-            "timeout": 10,
-            "args": [
-              "${CLAUDE_PROJECT_DIR}/.claude/nccgs/hooks/protect-write.mjs"
-            ],
-            "onFailure": "block"
-          }
-        ]
-      },
-      {
-        "matcher": "*",
-        "hooks": [
-          {
-            "type": "command",
-            "command": "node",
-            "timeout": 30,
-            "args": [
-              "${CLAUDE_PROJECT_DIR}/.claude/nccgs/hooks/closure-guard.mjs",
-              "pre"
-            ],
-            "onFailure": "block"
-          }
-        ]
-      }
-    ],
-    "PreCompact": [
-      {
-        "matcher": "",
-        "hooks": [
-          {
-            "type": "command",
-            "command": "node",
-            "timeout": 10,
-            "args": [
-              "${CLAUDE_PROJECT_DIR}/.claude/nccgs/hooks/pre-compact.mjs"
-            ]
-          }
-        ]
-      }
-    ],
-    "PostCompact": [
-      {
-        "matcher": "",
-        "hooks": [
-          {
-            "type": "command",
-            "command": "node",
-            "timeout": 10,
-            "args": [
-              "${CLAUDE_PROJECT_DIR}/.claude/nccgs/hooks/post-compact.mjs"
-            ]
-          }
-        ]
-      }
-    ],
-    "SubagentStart": [
-      {
-        "matcher": "",
-        "hooks": [
-          {
-            "type": "command",
-            "command": "node",
-            "timeout": 5,
-            "args": [
-              "${CLAUDE_PROJECT_DIR}/.claude/nccgs/hooks/agent-audit.mjs",
-              "start"
-            ],
-            "onFailure": "block"
-          }
-        ]
-      }
-    ],
-    "SubagentStop": [
-      {
-        "matcher": "",
-        "hooks": [
-          {
-            "type": "command",
-            "command": "node",
-            "timeout": 5,
-            "args": [
-              "${CLAUDE_PROJECT_DIR}/.claude/nccgs/hooks/agent-audit.mjs",
-              "stop"
-            ],
-            "onFailure": "block"
-          }
-        ]
-      }
-    ],
-    "Stop": [
-      {
-        "matcher": "",
-        "hooks": [
-          {
-            "type": "command",
-            "command": "node",
-            "timeout": 10,
-            "args": [
-              "${CLAUDE_PROJECT_DIR}/.claude/nccgs/hooks/session-stop.mjs"
-            ]
-          }
-        ]
-      },
-      {
-        "matcher": "",
-        "hooks": [
-          {
-            "type": "command",
-            "command": "node",
-            "timeout": 30,
-            "args": [
-              "${CLAUDE_PROJECT_DIR}/.claude/nccgs/hooks/closure-guard.mjs",
-              "stop"
-            ]
-          }
-        ]
-      }
-    ],
-    "PostToolUseFailure": [
-      {
-        "matcher": "Agent|Task",
-        "hooks": [
-          {
-            "type": "command",
-            "command": "node",
-            "timeout": 30,
-            "args": [
-              "${CLAUDE_PROJECT_DIR}/.claude/nccgs/hooks/runtime-observation.mjs",
-              "failure"
-            ]
-          }
-        ]
-      }
-    ],
-    "PostToolUse": [
-      {
-        "matcher": "Agent|Task",
-        "hooks": [
-          {
-            "type": "command",
-            "command": "node",
-            "timeout": 30,
-            "args": [
-              "${CLAUDE_PROJECT_DIR}/.claude/nccgs/hooks/runtime-observation.mjs"
-            ]
-          }
-        ]
-      }
-    ]
-  }
-}