nccgs 2.1.0 → 3.0.0-rc.1

package/CHANGELOG.md

@@ -1,200 +1,69 @@
 # Changelog
 
-## 2.1.0 — Unreleased
-
-- Handle the final LOW diagnostic finding: empty, incomplete, unreadable or
-  disappearing admission locks refuse with an actionable message, preserving
-  the original error cause. Keep ownership checks and explicit retry unchanged.
-
-- Resolve the follow-up review's two low-severity observations: distinguish
-  admission contention from active uninstall, and record session recovery as an
-  immutable pending receipt plus a separate success/failure outcome. Missing or
-  malformed outcomes remain unconfirmed; prior attempts and reviews stay intact.
-
-- Resolve the first independent maintenance review's four findings: add explicit
-  hash-bound operator recovery for crashed launcher leases with immutable receipts;
-  retain unknown process-state refusal and all task/job/run ownership checks.
-  Serialize run/install/agent/migration/policy admissions against uninstall,
-  cancel pre-mutation refusals without stranded receipts, and return backup IDs
-  even when empty-directory cleanup produces warnings. Historical negative review
-  and packages stay unchanged; the corrected archive requires fresh qualification.
-
-- Add `nccgs uninstall --project DIR`, a read-only `--dry-run` and guarded
-  `--rollback ID`. Remove unchanged inventoried framework files and exact NCCGS
-  hooks/imports; preserve customized files, shared Claude preferences, game source
-  and the full `.nccgs` policy/task/review/evidence/budget history.
-- Save every removed/replaced byte in a project-local uninstall receipt before
-  changing files. Refuse unsafe paths, ambiguous custom hooks, active/unverified
-  work and rollback over later user edits; attempt restoration on write failure.
-- Register `nccgs run` sessions so uninstallation cannot race a new managed
-  launch. An interrupted uninstall blocks launch/update until recovery. No process
-  is killed and no task, acceptance requirement, review or budget is reset.
-- The published 2.0.0 archive remains unchanged. This feature receives local
-  framework and installation tests; it is not a new Claude Code/Unity benchmark.
-  See [uninstall and recovery](docs/UNINSTALL.md).
-
-## 2.0.0 — 2026-10-28
-
-Published to npm on 2026-10-28. The following notes describe its original
-preparation scope; the exact publication and native qualification have separate
-frozen receipts. No new version is published by preparing 2.1.0.
-
-- Publish existing external operator Editor start, graceful shutdown command and
-  settled stop receipts into immutable reviewer-readable project bundles. Preserve
-  original bytes, hashes and ownership, reject inconsistent/uncertain/forced stops,
-  and identify historical imports without rewriting sealed tasks or verdicts.
-- Keep RC4 handoff, scene guard, source controls, process management, independent
-  review, acceptance and total-budget enforcement unchanged. No new game gate or
-  process permission is introduced.
-- Record scoped exact-RC4 native qualification: real dirty-scene refusal, new repair,
-  automatic PARTIAL handoff with no continuation, all 28 criteria/385 assertions,
-  independent PASS and normal closure. The original LOW evidence-location finding
-  remains in that sealed review; its addendum is assessed separately.
-- Stable archive validation and independent maintenance review are recorded in
-  release receipts. They do not claim another full Unity task through these bytes.
-
-
-## 2.0.0-rc.4 — 2026-10-28
-
-Prepared candidate; unpublished. No npm tag or stable promotion.
-
-- Settle a protected implementation work phase with a deterministic PARTIAL
-  report when current implementation and measured checks are ready for review.
-  A prose-only response no longer spends another continuation solely on handoff.
-  Preserve independent verdicts, all acceptance gates, history and original caps.
-- Pin an atomic loaded-scene precondition to new Unity repair reservations.
-  Dirty named, additive or untitled scenes block the operation without auto-save
-  or discard. Retain failed attempts and old reservations; no implicit retry.
-- Reserve final independent review money/time within the original root budget.
-  Share the work pool across design, slices, repair and verification. Admit
-  unknown worker cost only through verified, exclusive native session coverage;
-  missing session usage fails closed, and resumed cumulative cost is never summed
-  twice. The prior native reserve case qualified its disclosed new-script and
-  asset repair path with 28 criteria/385 assertions and independent final PASS.
-- RC4-specific handoff/scene changes receive framework tests and operator Unity
-  checks. A fresh native task through this exact archive remains pending.
-
-- Qualify the full component lifecycle on a fresh Unity project through actual
-  Claude Code: reviewed design, two dependent implementation slices from API
-  scaffolds, first-import verification, independent review, a disclosed fault,
-  typed repair and independent re-review. All 24 original criteria (349 assertions)
-  pass; the real failed PlayMode attempt remains recorded. The combined design and
-  implementation phases used 7/12 dispatches, three independent reviews and
-  USD 10.214066/12 within the original 75-minute envelope. Installed source guards,
-  process ownership, role limits, criteria and history remained unchanged. This
-  known-test component qualification does not establish broad game productivity.
-- Add opt-in exact `unityGeneratedFiles` declarations before task initialization
-  so Unity's first import can create IDE/local metadata without a false source
-  mutation failure. Pin the declaration as a canonical input and retain output
-  before/after hashes in managed receipts. Keep all undeclared files tracked;
-  reject source paths, folders, globs and canonical overlaps. Existing policies,
-  historical attempts, process ownership, review and budgets remain unchanged.
-- Settle cold Unity compilation without waiting for the qualified Roslyn server's
-  idle keepalive. After the command root exits and no other worker remains, request
-  graceful shutdown only from the verified compiler pipe inside the owned Windows
-  Job Object. Match the tested compiler/host binary hashes; unknown builds retain
-  the original wait/timeout behavior. Still drain every descendant and preserve the
-  command exit code, deadlines, independent review, criteria, history and budgets.
-- Avoid Unity CLI's two-minute idle auth broker in managed verification command
-  trees using the supported process-local `UNITY_NO_AUTH_BROKER=1` setting.
-  Preserve OS process ownership, descendant draining, failure/timeout semantics,
-  account settings, independent review, acceptance criteria and spent budgets.
-- Add bounded task slices with preview, pinned plans, dependency tracking and
-  focused worker prompts inside one original task. Final integration retains all
-  original criteria/checks and a reviewer independent of every contributor.
-- Protect aggregate dispatch capacity, original repair/QA/review reserves and
-  review limits. Preserve plan revisions, partial attempts and dependency history;
-  retries do not reset limits or create child/session budgets. Cost/time admission
-  stays unchanged.
-- Qualify the two-slice and existing-file `unity-code` repair path with actual
-  Claude Code workers and Unity tests on 2026-10-27. The controlled regression
-  retains all 24 criteria, failed evidence, both independent reviews and original
-  total budgets. This known-component case does not qualify large-game productivity,
-  new-file imports, Editor asset operations or the exact published RC archive.
-- Add opt-in schema-2 Unity repair packets: exact code edits/creation or pinned
-  Editor asset operations. Preserve schema-1 limits and historical reservations.
-- Bound Editor operations to one project, fixed positional arguments, finite
-  foreground calls and durable single-attempt receipts. Check source scope and
-  existing GUIDs; import new metadata before completion. Uncertain operations
-  retain ownership until externally reconciled, without refunds or automatic retry.
-- Unity repair still requires original measured checks, independent review and
-  manual/external acceptance. Synthetic tests do not qualify native Unity delivery.
-- Synchronize architecture and orchestration guidance with standalone defaults;
-  generate native role ceilings from agent-library frontmatter alongside routing.
-- Extend `task dispatch-prompt` to prepare registered implementation, analysis and
-  review assignments with safe PARTIAL handoff templates. Repair packets keep
-  their existing prompt and permissions.
-- Add read-only `task handoff-validate` for a worker's own declared handoff and
-  coordinator inspection, so schema errors can be corrected before completion.
-- Add coordinator `task record-review` to import a completed reviewer's exact
-  pinned result through the existing record gate. Review independence, objective
-  criteria, external/manual gates, task history and spent budgets are preserved.
-- The scoped native qualifications above predate the final RC4 handoff/scene
-  additions. Their frozen source, tests and historical evidence remain separate.
-
-## 2.0.0-rc.3 — 2026-10-07
-
-Published-1.3.0 upgrade correction on next; latest remains unchanged.
-
-- Fixed false overrides when a role generated by a 1.3.0 profile becomes a managed
-  file in 2.0. Existing generated fingerprints now identify unchanged agent files;
-  later user edits and explicit overrides still require review.
-- RC2 qualification exposed the issue in quality, balanced and inherit profiles.
-  The astra-claude path was unaffected. Original failed results are retained.
-- Release qualification uses the published 1.3.0 archive and the exact candidate
-  archive across four model profiles, owner settings and open task history,
-  override preservation, dry runs, explicit execution migration, rollback and
-  refusal of changed configuration/backups. See docs/UPGRADE-QUALIFICATION.md.
-- A new task executed by Claude Code through this exact release remains pending;
-  this correction does not promote stable or change the frozen game evaluation.
-
-## 2.0.0-rc.2 — 2026-10-07
-
-Release documentation follow-up on next; latest remains unchanged.
-
-- Unified current version labels and exact-version npm install, doctor, run and migration commands.
-- Replaced the obsolete Vietnamese Astra/Fable-default guide with standalone RC instructions.
-- Separated file update, explicit execution migration and guarded rollback; documented that rollback is not a full package downgrade.
-- Corrected active role counts and current workflow/project-policy guidance.
-- Added a release-status matrix and visible historical labels instead of hidden comments.
-- Retained RC1, frozen evaluation evidence and internal history. No runtime JavaScript, game source, model routing or acceptance threshold changed.
-- Packaging validation covers public contents, local documentation links, version consistency and fresh install/doctor. Published-1.3.0 upgrade qualification and a fresh exact-RC task remain pending.
-
-## 2.0.0-rc.1 — 2026-10-07
-
-Release channel: next. This candidate does not replace the stable/latest tag.
-
-### Candidate scope since npm 1.3.0
-
-- Standalone Claude Code execution with separate model, execution and external-review policies.
-- Reviewed design handoffs, pinned contracts, measured acceptance criteria and reviewer-owned verdicts.
-- Durable worker reservations, path ownership, protected repair/QA/review capacity,
-  verification jobs and bounded continuation without refunding spent attempts.
-- Explicit execution migration, preserved legacy obligations, backup and guarded rollback.
-- Reviewer capacity of up to 200 native turns; valid finding enums and standalone
-  command syntax documented to address observed closeout failures.
-- Portable public operating principles replace project-specific internal memory
-  in the package. The original development history is archived outside the package.
-
-### Validation scope
-
-The maintained framework regression recorded 183 PASS, 0 FAIL and 6 Windows
-symlink-fixture SKIP. An operator-assisted native Claude Code delivery V2 case
-completed 24 acceptance criteria,137 EditMode tests,52 PlayMode tests and an
-independent unconditional PASS. Failed attempts and evidence limitations were
-retained. This is not a clean autonomous benchmark or a stable-release claim.
-
-The game case ran a frozen framework with a session-only reviewer turn-capacity
-override. This RC packages the maintained source. Upgrade qualification from
-published 1.3.0 and a fresh end-to-end task using this exact RC are pending.
-Try a disposable project or backed-up copy; see UPGRADING.md.
-# Stable preparation follow-up — 2.0.0 (unpublished)
-
-- Correct remaining current-version RC labels while retaining historical release
-  records and their original qualification scope.
-- Add opt-in, correlated native-result transport for read-only analysis/review
-  handoffs when the native client requires returned text. Copies exact output;
-  existing review, acceptance, source, process and budget gates still apply.
-- Failed audit attempts and their consumed allowances remain in their original
-  history. New qualification and publication require their own explicit evidence.
+## 3.0.0-rc.1 — 2026-11-05 — phase 5: clean-up and release candidate
 
+- Windows: process trees run inside a Job Object through `koffi` (optional dependency; `npx nccgs` installs it). A timed-out batch Editor dies with every descendant, including re-parented ones; `unity open`'s Editor is never touched because kill-on-close is used only for `unity test`/`build`. Without koffi the taskkill /T + ppid walk stays. `doctor` reports which layer is active; `scripts/check-job-object.mjs` proves it on a machine.
+- Unity Skills re-synced from upstream `ba28c24` (2026-11-04): all 22 vendored skills updated, 13 new ones added (including `unity-pipeline`, `create-2d-physics`, `2d-pixel-perfect`, `manage-sprite-atlas`, `create-tile-palette-or-rule-tiles`); the commit is pinned in the manifest, so update notices work.
+- README rewritten around daily use (under 120 lines). `unity-cli-notes.md` now overlays the official `unity-pipeline` skill instead of replacing it.
+- Everything from alpha.2 to alpha.8 stays as it was; see below.
+
+## 3.0.0-alpha.8 — 2026-11-05 — phase 4: game capability
+
+- Slice plan (`runtime/task/plan.mjs`): `.nccgs/design/slice.json` holds ordered task contracts with `dependsOn`; `task plan [--next|--show id|--validate file]`, `task open --plan [id]` opens the next ready contract (flags override). Skills `/brainstorm` (brief with options), `/map-systems` (systems map + validated plan), `/vertical-slice` (runs the plan through `/work`, stops at the first partial), `/unity-foundation` (asmdefs, test assemblies, input, pipeline as a measured task).
+- Playtest with telemetry (`runtime/unity/playtest.mjs`, `task verify --kind playtest --spec`): one `script eval` presses the keys on a timeline or pattern, samples C# probes at 10 Hz, stops on `stopWhen`, writes CSV; artifact has `summary` (final/min/max, first change of booleans = time of death) and `samples`. `/playtest` skill and the `nccgs-playtester` agent (judgement + options from the artifact). `protocols/playtest.md`.
+- Asset check (`runtime/unity/assets.mjs`, `task verify --kind assets --spec`): reads `TextureImporter` fields through the Editor and lists mismatches against the spec; `/art-bible` and `/asset-spec` skills.
+- Playtest spec also takes `reset` (C# statement run at schedule time, e.g. reload the scene, since the schedule lands 2–3 s after Play), `settleS`, and `rules[]` (reactive input: press a key whenever a C# condition holds, with hold and cooldown) — a metronome never lines up with hazards, a rule plays like a person. Samples carry `gameTime`; the final sample is taken at the stop condition so a death is in the data; the game's log lines during the run are kept in `console.entries`.
+- `/work` opens from the plan when one exists and routes feel questions to playtests.
+
+## 3.0.0-alpha.7 — 2026-11-05 — phase 3: reports
+
+- Token cost per task (`runtime/task/cost.mjs`): assistant messages in the Claude Code transcripts between `openedAt` and `closedAt`, deduplicated by message id, split by agent (coordinator / reviewer / …, from SubagentStart events) and by model; subagent transcripts under `<session>/subagents/` included. Hooks now record `transcript_path`; older sessions are found by session id under `~/.claude/projects`. USD only with a `pricing` row in project.json.
+- `task close` writes `<id>.html` beside `<id>.md`: self-contained (screenshots embedded as data URIs, `gameTimeS` captions), criteria, tests per run, changes with scope marks, findings with open/resolved state, cost table. `task report --id X` re-renders both from the ledger.
+- `.nccgs/project.html` (`task project`, `nccgs status`, `/status`): tiles (closed/success, open, latest tests, open findings H/M/L, tokens, est. cost), task table with report links, latest frame per task, open findings sorted by severity, recent closes from state.md.
+- `nccgs release --from <ref|date> --to <ref|date|now> [--name]`: changelog section from the tasks closed in the range, grouped by outcome, written to `.nccgs/releases/`.
+- `/work` ends with the report path and its three lead numbers; `/status` rewrites project.html first.
+
+## 3.0.0-alpha.6 — 2026-11-04
+
+From the RUN-3 review notes.
+
+- Timed observe records `screens[].gameTimeS` (`Time.timeSinceLevelLoad` at each capture, written by the Editor to `screens/<id>/timing.json`), `scheduledAtGameTime` and `scheduleDelayMs`, so "500 ms" is no longer mistaken for 0.5 s of play.
+- `observe` records `matchLog`, the pattern `logMatches` counted, and counts only entries logged after Play started (console cursor taken before `editor.playmode start`), so a console full of earlier runs no longer inflates the count.
+- `observe` turns Run In Background on for the run (`editor.playmode set_run_in_background`) and restores it after, so an unfocused Editor does not freeze Play Mode.
+- Notes: `job wait --on_met capture` as the state-triggered alternative to time offsets.
+
+## 3.0.0-alpha.5 — 2026-11-04
+
+Found by RUN-2: both screenshots showed the runner on the ground because a 0.75 s jump is over before `capture screenshot` returns (each Pipeline call is ~0.9 s). The person had to confirm the jump live.
+
+- `observe --shots <ms,...> [--hold-key <ms>]` (timed mode): one `script eval` schedules the key down, key up and every `ScreenCapture` inside the Editor on `EditorApplication.update`, so frames land at exact offsets after the press. Verified on 6000.3.17f1: apex frame at 380 ms shows the runner airborne, 1200 ms shows it landed, one `[Runner] jump`.
+- `task close --decide id=pass` (also accepted|yes|ok; fail|rejected|no) — only `accepted|rejected` worked before.
+- Report lists every test run by name (editmode and playmode), not just the last one.
+
+## 3.0.0-alpha.4 — 2026-11-04
+
+Found by RUN-1 (first game task). Three runtime fixes were made by Claude Code inside the project during the task, verified against the live Editor, and are ported here unchanged; the rest follow its review findings.
+
+- `task verify --kind tests --name X` writes `<id>-X.json` (before: always `<id>-tests.json`, so a PlayMode criterion could never point at its own artifact).
+- Live PlayMode tests run through `test run --async_tests` + `test status` polling (a synchronous run is dropped when Play Mode reloads the domain). A Pipeline-level failure (`result.success:false`, cancelled, timeout) is `ok:false`, and a completed run with 0 tests is `ok:false` with `NO_TESTS_RAN`.
+- `observe`: `--keys` are pressed once before frame 1 (before: before every frame, so `--keys Space,Space --frames 2` was 4 presses); `--keys-every-frame` restores the old behaviour. `--match-log <regex>` records `logMatches` and `keyPresses`, so a criterion can assert one log line per press instead of "log count >= 2".
+- `files.volatile` gains `ProjectSettings/SceneTemplateSettings.json` (rewritten by the Editor when a scene is opened).
+- Report: screenshots listed once even when observe ran more than once.
+- `unity-cli-notes.md`: `simulate_key` delivers down+up in one Input System update; poll-based `wasPressedThisFrame` misses it, an `InputAction` does not. Legacy Input is not driven.
+
+## 3.0.0-alpha.3 — 2026-11-04
+
+Found by the first real `/work` run (task T-RUNNER on Unity 6000.3.17f1): the reviewer reported that tests and console evidence sat on an older snapshot than its own review, and the report listed `ProjectSettings/ShaderGraphSettings.asset` as an edit outside the contract. Nobody had edited it; the Editor re-saves it when entering Play Mode, and alpha.2 fingerprinted files by mtime.
+
+- Snapshots now hash file content (`runtime/core/snapshot.mjs`). Files are rehashed only when size or mtime changed (`.nccgs/cache/hash-cache.json`); files above `files.hashMaxBytes` (8 MB) are fingerprinted by size+mtime. Each snapshot's per-file map is kept under `.nccgs/cache/snapshots/<sha>.json`, and `changedFiles` diffs maps (added, modified, removed) instead of scanning mtimes. Tasks opened by alpha.2 fall back to the mtime scan.
+- `files.volatile` (project.json): Editor-churned files excluded from snapshots. Default: `ProjectSettings/ShaderGraphSettings.asset`, `ProjectSettings/BurstAotSettings_*.json`, `Assets/**/*.orig`.
+- Report: `.meta` files are collapsed into one line and count as in scope when their asset is; a stray `.meta` is still named.
+- `task review --resolve <index|title words> --reason "..."` marks a MEDIUM/LOW finding resolved (by whom, why, on which snapshot). HIGH/BLOCKER can only be resolved by a new verdict. The report shows every finding as `(open)` or `(resolved by …: reason)`.
+- `task close --outcome success` also refuses when the latest review on the current snapshot is not PASS.
+- `guard-write` blocks hand edits under `.nccgs/cache/`.
+- `/work` skill and `protocols/evidence.md` describe the resolve step.
+- `task verify --kind console` with no Editor open now records `summary.reason` ("no Editor open for this project…") instead of a silent null ground truth (seen on T-SPEED6).
+
+## 3.0.0-alpha.2 — 2026-11-03
+
+First end-to-end run on a live Editor: install, doctor, `task open/verify/review/close`, report.

package/LICENSE

@@ -1,6 +1,6 @@
 MIT License
 
-Copyright (c) 2026 Donchitos
+Copyright (c) 2026 LacBird
 
 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal

package/README.md

@@ -1,191 +1,82 @@
-# NCCGS — New Claude Code Game Studio
-
-**2.1.0** is the unpublished next package for standalone game-development
-orchestration in **Claude Code**. It adds reversible project uninstallation;
-the published **2.0.0** archive remains unchanged. Local preparation moves no npm tags.
-It retains handoff automation, bounded task slicing and Unity repair, protected
-final-review allocation, managed Unity settlement and exact first-import output
-declarations. New repair steps block unsaved loaded scenes before execution.
-See [handoff operations](docs/HANDOFF-AUTOMATION.md), [bounded repair](docs/BOUNDED-REPAIR.md),
-[task slicing](docs/TASK-SLICING.md), [uninstallation](docs/UNINSTALL.md) and [release notes](CHANGELOG.md).
-
-Claude Code handles intake, planning, specialist dispatch, testing, independent
-review and reporting. External review can be configured separately. This major
-version introduces standalone defaults and explicit execution-policy migration
-from the older compatibility workflow; existing task obligations remain preserved.
-
-## Validation and limits
-
-The frozen exact RC4 archive was exercised through actual Claude Code on a
-controlled Unity scene repair: dirty-scene refusal preserved unsaved content,
-a fresh worker repaired the target, deterministic PARTIAL handoff required no
-continuation, all five checks and 28 criteria/385 assertions passed, and a fresh
-independent reviewer returned unconditional PASS before normal task closure.
-This is scoped workflow evidence, not broad game-quality or productivity proof.
-
-Version 2.0.0 keeps those runtime guards and adds a receipt publisher so external
-Editor stop evidence is available inside the reviewer's project. It does not
-operate processes or alter acceptance. See [Editor evidence](docs/EDITOR-SESSION-EVIDENCE.md).
-The exact 2.0.0 archive subsequently passed its scoped native Unity qualification
-and was published on 2026-10-28. Frozen release receipts retain its tests and
-independent reviews. The new 2.1.0 uninstaller receives local filesystem/CLI tests;
-that does not claim another Claude Code/Unity run. The first independent maintenance
-review required four fixes; this candidate adds those fixes and awaits qualification
-of its exact archive, including a fresh independent review. Earlier verdicts remain
-unchanged in their separate receipts.
-Frozen history remains intact.
-See [release status](docs/RELEASE-STATUS.md), [upgrading](UPGRADING.md),
-[Vietnamese quick start](docs/HUONG-DAN-MIGRATE-VA-SU-DUNG.md) and [decisions](MEMORY.md).
-
-## What changed
-
-Standalone `/design`, `/design-review` and `/story-readiness` now load the
-[design lifecycle](.claude/nccgs/protocols/design-lifecycle.md). New installs require
-a reviewed design handoff before dependent feature implementation; criteria and
-their source/threshold authority are pinned across that handoff. The fixed
-`task design-validate` command supports plan tasks without granting arbitrary
-shell execution. Existing policies and frozen evaluations retain their obligations;
-see [upgrading](UPGRADING.md). Schema validation is not semantic or gameplay acceptance.
-
-
-
-NCCGS includes a [shared Ponytail package](.claude/nccgs/packages/ponytail/README.md)
-for technical planning, implementation, supplementary complexity review and known
-code limits. The coordinator and agents load it by assignment through their common
-contracts. It borrows selected guidance from DietrichGebert/Ponytail under MIT;
-the [original license](.claude/nccgs/packages/ponytail/LICENSE) and
-[attribution/change notice](.claude/nccgs/packages/ponytail/NOTICE.md) ship with every
-installation. It preserves full design and QA obligations; no separate Ponytail
-plugin or global hooks are installed. Broad comparative effectiveness remains unmeasured.
-
-- Separate execution policy, model profile and external review (`off`, `advisory`,
-  `required`); project updates preserve existing obligations.
-- Task contracts for `implement`, `audit` and `plan`, with conservative
-  FAST/STANDARD/CONTROLLED classification, reasons, inputs and acceptance gates.
-- Durable dispatch reservations, path ownership checks, two-worker ceiling,
-  independent reviewer slot, ancestor quotas and recovery without budget refunds.
-- Final evidence tied to task, source, contract, runtime invocation and raw source.
-  Missing model/effort/usage is explicit; requested settings do not prove execution.
-- Reports for success, partial, blocked and failed outcomes; strict successful
-  closure, optional external receipts and explicit acceptance when configured.
-- Execution migration with dry-run, archived configuration and guarded rollback.
-
-The studio retains 47 roles (13 active by default, including Game Designer), 36 NCCGS workflows, scoped
-rules, lifecycle hooks and unmodified official Unity Skills. Enable specialists
-when their expertise resolves a concrete problem; do not dispatch the whole studio.
-
-## Try the local 2.1.0 package
-
-Requirements: Claude Code and Node.js 18+. Unity CLI/Pipeline is needed for live
-Unity operations. Run against a disposable Unity project first:
-
-Use the locally prepared archive and its accompanying SHA-256 manifest:
-
-```powershell
-$NccgsArchive = "D:\path\to\nccgs-2.1.0.tgz"
-npx --yes --package "$NccgsArchive" nccgs install --project "D:\path\to\UnityProject" --dry-run
-npx --yes --package "$NccgsArchive" nccgs install --project "D:\path\to\UnityProject"
-npx --yes --package "$NccgsArchive" nccgs doctor --project "D:\path\to\UnityProject"
+# NCCGS 3
+
+A Unity-first evidence loop for Claude Code. A task is done when the game has run and been looked at, not when the code has been written.
+
+```
+brief → options for the user → contract → edit → run (tests, Play Mode, screenshots, console, playtest)
+      → look → fix or ask → independent review → close → one-page report with real token cost
 ```
 
-Before 2.1.0 publication, use this local archive. Published nccgs@2.0.0 does not
-include the uninstall command. Use update for an existing installation only after reviewing its
-dry run and backup. See [the tested upgrade matrix](docs/UPGRADE-QUALIFICATION.md) and its limits.
-The installer preserves customized files and unrelated Claude settings, records
-overrides and backs up replacements. Do not use --force as an ordinary update path.
+**3.0.0-rc.1.** Built and verified on Unity 6000.3 and 6000.6 with Unity CLI 1.0.0-beta.14 and com.unity.pipeline 0.9, Claude Code 2.1.3xx, Windows. The sample game (a 2D endless runner) went from an empty scene to a vertical slice in four tasks with this loop; its reports are the reference for what a closed task looks like.
+
+## Requirements
+
+Node 20+, Claude Code 2.1.295+, Unity 6.0+ with `com.unity.pipeline` in the project (`unity pipeline install --project-path <project>`), the Unity CLI on PATH. Windows, macOS and Linux; on Windows the optional `koffi` dependency gives Job Object process control (installed automatically by `npx`).
 
-New installs default to standalone with `claude-standalone`. Existing installations
-without execution.json remain compatibility. To migrate explicitly:
+## Install
 
-```powershell
-npx --yes --package "$NccgsArchive" nccgs migrate-execution --project "D:\path\to\UnityProject" --to standalone --model-profile claude-standalone --dry-run
+```
+npx nccgs install --project D:\Games\MyGame --dry-run
+npx nccgs install --project D:\Games\MyGame
+npx nccgs doctor  --project D:\Games\MyGame
+npx nccgs run     --project D:\Games\MyGame
 ```
 
-Inspect the dry run, then remove `--dry-run`. Omitting `--model-profile` preserves
-existing model routes. See [upgrading](UPGRADING.md) for rollback and historical
-task handling. Updating files never converts an old approval into a new one.
+`install` adds `CLAUDE.md` (kept if you have one), `.claude/skills/` (10 NCCGS skills + 35 official Unity Skills), `.claude/agents/nccgs-*.md`, `.claude/nccgs/` (runtime, protocols, Ponytail) and `.nccgs/` (project config, tasks, evidence, reports). It merges hooks into `.claude/settings.json` without touching your other settings. `update` replaces managed files you did not change and records the ones you did; `uninstall` removes managed files, keeps `.nccgs/` and backs up what it removes. `run` launches Claude Code with the permission mode and MCP policy NCCGS expects (`acceptEdits` by default; `--autonomy full` for `bypassPermissions`).
 
-## Operating flow
+## Daily use
 
-Start a task session with native permissions already selected:
+Open the project in the Editor (screenshots and live tests need it; without it tests run in a batch Editor pinned to the project's own version), run `nccgs run`, then:
 
-```powershell
-npx --yes --package "$NccgsArchive" nccgs run --project "D:\path\to\UnityProject"
-```
+| Skill | When | Output |
+| --- | --- | --- |
+| `/work` | any change to code, scenes, prefabs or assets | a closed task with `.nccgs/reports/<id>.md` + `.html` |
+| `/test` | run tests and observe without changing anything | measurements under `.nccgs/evidence/` |
+| `/status` | where are we | `.nccgs/project.html` and ten lines of status |
+| `/brainstorm` → `/map-systems` → `/vertical-slice` | a new game or a multi-task feature | `brief.md`, `systems.md`, `slice.json`, then one `/work` per planned task |
+| `/unity-foundation` | a fresh project, or tests cannot run | asmdefs, test assemblies, input, pipeline as a measured task |
+| `/playtest` | does it feel right, tuning, difficulty | telemetry artifact, the playtester's judgement, your decision |
+| `/art-bible`, `/asset-spec` | before and after art | visual rules; import settings checked against them |
 
-New installations default to task autonomy `full`. Existing installations keep
-their setting; use `update --autonomy full` to select it explicitly. For a resumed
-session, append `-- --resume SESSION_ID`. If Claude is not on PATH, pass
-`--claude "C:\path\to\claude.exe"` before the separator. Model and tool settings
-are inherited. Ordinary Bash, report writing and task work outside the project
-need no repeated permission; explicit deny rules and integrity/destructive-action
-guards remain effective. See [task autonomy](.claude/nccgs/protocols/task-autonomy.md).
-
-```text
-User request -> bounded contract + mode + risk
-            -> reserved Claude Code worker(s)
-            -> targeted evidence + independent internal review
-            -> honest report of outcome and unresolved findings
-            -> configured external/manual/acceptance gates -> DONE
-```
+The loop refuses to lie: `task close --outcome success` is rejected when a criterion fails, the source changed after the last verify, or the review is missing; `partial` and `blocked` are always allowed and reported as such.
+
+## What a task carries
+
+- **Contract**: files in scope, criteria as assertions on artifacts `task verify` produces (`eq | gte | lte | exists | contains | range | image`), rigor (`prototype`: no review, 6 fix rounds · `production`: one review, 4 · `release`: two reviews, 2), a manual criterion for anything only a person can judge.
+- **Evidence**, written only by the runtime: `tests` (EditMode/PlayMode, live Pipeline or batch), `observe` (Play Mode, keys, screenshots at exact milliseconds with `--shots`, console, performance), `console`, `playtest` (scripted keys or reactive rules, probes sampled at 10 Hz), `assets` (importer settings vs the art bible), `command`.
+- **Review** by a fresh-context reviewer whose verdict must name the screenshots it looked at; `PASS` cannot carry an unresolved HIGH/BLOCKER; MEDIUM/LOW are closed with `task review --resolve ... --reason ...` and the report shows who closed what.
+- **Report**: criteria, tests per run, changes with scope marks, findings, screenshots with game time, and the token cost per agent and model read from the Claude Code transcripts (USD only when `pricing` in `.nccgs/project.json` has a row). `nccgs status` rewrites `project.html`; `nccgs release --from <tag|date> --to <tag|date|now>` writes a changelog from the tasks closed in the range.
 
-| Workflow | Outcome |
-|---|---|
-| `/work` | Implement a feature, fix, refactor, tool or content change |
-| `/design` | Plan design and requirements |
-| `/audit` | Examine the project without source edits |
-| `/migrate-project` | Review and apply authorized project adoption changes |
-| `/test` | Verify declared requirements |
-| `/review` | Independent review |
-| `/release` | Prepare a verified release candidate |
-| `/status` | Read state, evidence, blockers and next action |
-
-FAST uses Sonnet/medium; STANDARD uses Opus/medium; CONTROLLED uses Opus/high.
-Independent review uses Opus/high. These aliases are requested routes. Real Claude
-model resolution and provider access must be verified in the target session.
-Historical `astra-claude` routes remain available under compatibility execution.
-
-## Remove NCCGS when it is no longer needed
-
-Close Claude Code sessions, then preview and apply from an external terminal:
-
-```powershell
-npx --yes --package "$NccgsArchive" nccgs uninstall --project "D:\path\to\UnityProject" --dry-run
-npx --yes --package "$NccgsArchive" nccgs uninstall --project "D:\path\to\UnityProject"
+## Modes
+
+`FAST`, `BALANCE` (default) and `PERFORMANCE` set model and effort per agent (coordinator, implementer, reviewer, fast-fix, scout, playtester). They are read-only; copy one and edit the copy:
+
+```
+npx nccgs mode create --from PERFORMANCE --name mine --project .
+npx nccgs mode set --name mine --agent reviewer --model opus --effort xhigh --project .
+npx nccgs mode use mine --project .
 ```
 
-This also supports an existing 2.0.0 installation without updating it first.
-Game source, customized files and `.nccgs` history/budgets remain. The command
-removes unchanged managed files and exact framework hooks/imports, saves a backup
-and refuses active or uncertain work. Use `uninstall --rollback ID` to restore
-that backup when no later edits conflict. Shared permission/model settings stay
-because their ownership was not recorded by earlier installers.
-See [details and recovery](docs/UNINSTALL.md).
-
-## Maintainer verification and limits
-
-These commands run from a source/package checkout. They are not extra steps for
-users installing with npx, and do not initiate a Claude Code game run.
-
-```text
-npm run policy:build
-npm run policy:check
-npm run validate
-npm test
-node scripts/benchmark-v15.mjs --out PATH_TO_JSON
-npm pack --dry-run
+`PERFORMANCE` asks for Fable for review and falls back to Opus/xhigh when the account does not have it; the report records which model reviewed. `nccgs doctor --check-models` probes the current mode's models.
+
+## Unity facts the runtime relies on
+
+Verified live, kept in `.claude/nccgs/protocols/unity-cli-notes.md` and `docs/EVIDENCE-VERIFICATION.md`: Pipeline command names are discovered at runtime, never hard-coded; `--editor-version` is always pinned from `ProjectVersion.txt`; entering Play Mode drops the Pipeline server for ~3 s on 6.3 and the runtime waits; PlayMode tests run async and are polled; `simulate_key` delivers key-down and key-up in one input update, so input is read through `InputAction`s; every Pipeline round trip costs ~0.9 s, so anything that must happen within a second of a key press is scheduled inside the Editor by one `script eval`; Run In Background is switched on for every run so an unfocused Editor does not freeze.
+
+## Layout
+
 ```
+.claude/nccgs/runtime/   task CLI, Unity wrapper, hooks (managed)
+.claude/nccgs/protocols/ evidence.md, review.md, playtest.md, unity-cli-notes.md
+.claude/nccgs/ponytail/  engineering.md, complexity-review.md (MIT, adapted)
+.claude/skills/          NCCGS skills + official Unity Skills (do not edit; `nccgs update --skills-only`)
+.nccgs/                  project.json, modes.json, state.md, design/, tasks/, evidence/, reports/, project.html
+```
+
+## Updating
 
-Generated catalogs, workflow wrappers and agents come from pipeline.json and the
-agent library. Do not patch generated model pins directly. Official Unity Skills
-remain the source for generic Editor/asset mechanics.
+`nccgs doctor --check-updates` reports newer NCCGS, Unity Skills, Unity CLI and Pipeline versions; nothing is applied on its own. `nccgs update` refreshes managed files; `--skills-only` refreshes just the vendored Unity Skills (pinned to an upstream commit in `unity-skills/manifest.json`).
 
-Reservations and direct edit hooks enforce declared path ownership. Arbitrary
-implementation shell/MCP tools are not filesystem-sandboxed; review actual changes.
-Background or uncertain workers retain slots until final correlated evidence or
-external reconciliation. Worker telemetry has explicit coverage and excludes
-unobserved coordinator/human/provider activity. It is not account-wide billing.
+## Licenses
 
-See [the Claude Code pilot and benchmark protocol](docs/NCCGS-1.5-PILOT.md),
-[the standalone workflow](.claude/nccgs/protocols/standalone.md), and
-[the 1.4.1 integrity guide](docs/NCCGS-1.4.1.md).
+MIT for NCCGS (LacBird). Vendored Unity Skills: Unity Companion License (see `THIRD_PARTY_NOTICES.md`). Ponytail material: MIT, DietrichGebert.

package/THIRD_PARTY_NOTICES.md

@@ -1,38 +1,10 @@
 # Third-party notices
 
-## Ponytail
+## Official Unity Skills (vendored, unmodified)
+The directories listed in `unity-skills/manifest.json` are an unmodified snapshot of https://github.com/Unity-Technologies/skills, licensed under the Unity Companion License for Unity-dependent projects (https://unity3d.com/legal/licenses/unity_companion_license). They are installed into `.claude/skills/` and must not be edited there; NCCGS carries its own guidance in `.claude/nccgs/protocols/`.
 
-The shared package under `.claude/nccgs/packages/ponytail/` borrows and adapts
-engineering, complexity-review and deliberate-simplification guidance from
-**Ponytail by DietrichGebert**, v4.10.3, commit
-`c982cd411abb53323c4baa1baa3c2f020b8d0b08`.
+## Ponytail (adapted)
+`template/ponytail/engineering.md` and `complexity-review.md` adapt material from Ponytail by DietrichGebert, v4.10.3, commit c982cd411abb53323c4baa1baa3c2f020b8d0b08, MIT License. The complete license and the attribution/change notice ship in `template/ponytail/LICENSE` and `NOTICE.md`.
 
-Source: https://github.com/DietrichGebert/ponytail
-
-Copyright (c) 2026 DietrichGebert. Licensed under the **MIT License**. The complete
-original [license](.claude/nccgs/packages/ponytail/LICENSE) and
-[attribution with adaptation details](.claude/nccgs/packages/ponytail/NOTICE.md)
-are included and installed with the shared package. Retain them on redistribution.
-This is an NCCGS adaptation, not the unmodified Ponytail plugin.
-
-## Claude Code Game Studios
-
-NCCGS was informed by the workflow, traceability, brownfield adoption, and QA
-ideas in Donchitos/Claude-Code-Game-Studios. The supplied source is MIT licensed.
-NCCGS replaces its multi-engine agent hierarchy rather than redistributing that
-hierarchy verbatim.
-
-Source: https://github.com/Donchitos/Claude-Code-Game-Studios
-
-## Unity Skills
-
-The directories identified in `.claude/nccgs/unity-skills-manifest.json` are an
-unmodified snapshot of Unity Technologies' Unity Skills. Unity Skills are
-copyright Unity Technologies and are governed by the Unity Companion License for
-Unity-dependent projects.
-
-Source: https://github.com/Unity-Technologies/skills
-
-License: https://unity3d.com/legal/licenses/unity_companion_license
-
-Do not remove this notice when redistributing NCCGS with the vendored snapshot.
+## Prior work
+NCCGS 3 is a fresh implementation informed by NCCGS 2.x (LacBird) and by Donchitos/Claude-Code-Game-Studios (MIT). No code is copied from either.