nccgs 1.1.1 → 1.3.0

package/.claude/skills/story-readiness/SKILL.md

@@ -6,9 +6,10 @@ allowed-tools: Read, Glob, Grep, Bash
 
 # Story Readiness
 
-Read `.nccgs/project.yaml`, `.nccgs/state.md`, and the active context packet when present. Preserve Product Owner authority and unrelated work.
+Read models.profile in .nccgs/project.yaml; if absent, use the defaultProfile in
+.claude/nccgs/pipeline.json. Load exactly one procedure:
 
-Remain read-only. Check approved objective, stable requirements, acceptance criteria, governing decisions, dependencies, ownership, protected paths, compatibility, testability, and unresolved semantics. Verify the context packet source revisions. Return READY, READY WITH RISKS, or BLOCKED with exact missing inputs and the smallest next action. Do not demand documents that add no implementation value.
-
-Return outcome, evidence, affected records, blockers or decisions, residual risk, and the exact next workflow. Do not claim DONE unless this is the `closure` skill and every configured gate passes.
+- astra-claude: [Astra/Claude procedure](references/astra.md).
+- quality, balanced, inherit: [Legacy procedure](references/legacy.md).
 
+Do not load or merge the other branch. Unknown profiles are configuration errors.

package/.claude/skills/story-readiness/references/astra.md

@@ -0,0 +1,17 @@
+# Story Readiness — astra-claude
+
+Read [the pipeline protocol](../../../nccgs/protocols/astra-claude.md) and
+[generated routing](../../../nccgs/routing.generated.md). Use Astra's supplied plan,
+requirements, and exact context references. Never substitute a legacy procedure.
+
+Domain objective: Checks whether a feature or story is safe and specific enough to implement, including requirements, decisions, dependencies, ownership, tests, and Product Owner gates.
+
+Work only inside the supplied plan and assigned write set. Activate an optional
+specialist with the agents command only when its domain is required. Keep one writer
+per path and use official Unity Skills for Editor/assets/runtime operations.
+
+This procedure is read-only: return evidence-backed findings and decisions to Astra
+or the implementer. Supplemental analysis does not replace the mandatory Fable gate.
+
+Use [task gate commands](../../../nccgs/protocols/task-gates.md) for evidence and
+state transitions. Report outcome, evidence, pending decisions, and next owner.

package/.claude/skills/story-readiness/references/legacy.md

@@ -0,0 +1,7 @@
+# Story Readiness
+
+Read `.nccgs/project.yaml`, `.nccgs/state.md`, and the active context packet when present. Preserve Product Owner authority and unrelated work.
+
+Remain read-only. Check approved objective, stable requirements, acceptance criteria, governing decisions, dependencies, ownership, protected paths, compatibility, testability, and unresolved semantics. Verify the context packet source revisions. Return READY, READY WITH RISKS, or BLOCKED with exact missing inputs and the smallest next action. Do not demand documents that add no implementation value.
+
+Return outcome, evidence, affected records, blockers or decisions, residual risk, and the exact next workflow. Do not claim DONE unless this is the `closure` skill and every configured gate passes.

package/.claude/skills/test/SKILL.md

@@ -1,16 +1,15 @@
 ---
 name: test
 description: Plans or executes requirement-based Unity verification using automated tests, Unity CLI/Pipeline, runtime, build, regression, profiling, and explicit Product Owner gates. Use to prove behavior or reproduce failures without changing intended semantics.
-allowed-tools: Read, Glob, Grep, Write, Edit, Bash, Task, AskUserQuestion
+allowed-tools: Read, Glob, Grep, Write, Edit, Bash, Agent, AskUserQuestion
 ---
 
 # Test
 
-Read policy, context packet, requirements, decisions, implementation, existing tests, and baseline. Classify each target as objective verification, regression, or subjective validation.
+Read models.profile in .nccgs/project.yaml; if absent, use the defaultProfile in
+.claude/nccgs/pipeline.json. Load exactly one procedure:
 
-Use `/qa-plan` when coverage is not obvious. Choose the strongest proportional evidence described in the [evidence protocol](../../nccgs/protocols/evidence.md). Use official Unity Skills and `unity-cli` for Unity operations.
-
-Run existing relevant checks first. Add deterministic tests only when the surrounding request authorizes implementation; remain read-only for a standalone audit. Capture environment, commands/actions, counts, skips, baseline comparison, runtime/build state, console errors, and regressions.
-
-Use an independent verification role for STANDARD/CONTROLLED work. Verdicts are PASS, FAIL, NOT PROVEN, PO GATE, or N/A with reason. Do not mark the feature DONE; hand evidence to `/evidence-review` and `/closure`.
+- astra-claude: [Astra/Claude procedure](references/astra.md).
+- quality, balanced, inherit: [Legacy procedure](references/legacy.md).
 
+Do not load or merge the other branch. Unknown profiles are configuration errors.

package/.claude/skills/test/references/astra.md

@@ -0,0 +1,20 @@
+# Test — astra-claude
+
+Read [the pipeline protocol](../../../nccgs/protocols/astra-claude.md) and
+[generated routing](../../../nccgs/routing.generated.md). Use Astra's supplied plan,
+requirements, and exact context references. Never substitute a legacy procedure.
+
+Domain objective: Plans or executes requirement-based Unity verification using automated tests, Unity CLI/Pipeline, runtime, build, regression, profiling, and explicit Product Owner gates. Use to prove behavior or reproduce failures without changing intended semantics.
+
+Work only inside the supplied plan and assigned write set. Activate an optional
+specialist with the agents command only when its domain is required. Keep one writer
+per path and use official Unity Skills for Editor/assets/runtime operations.
+
+For code changes: task init with Astra's plan, bind the session, implement and test,
+record implementation evidence, request independent review, fix and re-review as
+needed, then task check --target report. Only after that check passes write and
+record the report for Astra. Stop at AWAITING_ASTRA_CHECK. Astra checking and the
+Product Owner's explicit acceptance happen outside the Claude implementation session.
+
+Use [task gate commands](../../../nccgs/protocols/task-gates.md) for evidence and
+state transitions. Report outcome, evidence, pending decisions, and next owner.

package/.claude/skills/test/references/legacy.md

@@ -0,0 +1,9 @@
+# Test
+
+Read policy, context packet, requirements, decisions, implementation, existing tests, and baseline. Classify each target as objective verification, regression, or subjective validation.
+
+Use `/qa-plan` when coverage is not obvious. Choose the strongest proportional evidence described in the [evidence protocol](../../../nccgs/protocols/evidence.md). Use official Unity Skills and `unity-cli` for Unity operations.
+
+Run existing relevant checks first. Add deterministic tests only when the surrounding request authorizes implementation; remain read-only for a standalone audit. Capture environment, commands/actions, counts, skips, baseline comparison, runtime/build state, console errors, and regressions.
+
+Use an independent verification role for STANDARD/CONTROLLED work. Verdicts are PASS, FAIL, NOT PROVEN, PO GATE, or N/A with reason. Do not mark the feature DONE; hand evidence to `/evidence-review` and `/closure`.

package/.claude/skills/ui-review/SKILL.md

@@ -1,14 +1,15 @@
 ---
 name: ui-review
 description: Performs a read-only review of Unity UI behavior, architecture, navigation, responsiveness, localization, accessibility, performance, and UX conformance.
-allowed-tools: Read, Glob, Grep, Bash, Task
+allowed-tools: Read, Glob, Grep, Bash, Agent
 ---
 
 # Ui Review
 
-Read `.nccgs/project.yaml`, `.nccgs/state.md`, and the active context packet when present. Preserve Product Owner authority and unrelated work.
+Read models.profile in .nccgs/project.yaml; if absent, use the defaultProfile in
+.claude/nccgs/pipeline.json. Load exactly one procedure:
 
-Read approved UX flows, UI requirements, platform/input policy, accessibility targets, implementation, scenes/assets, and tests. Pair UX and Unity UI lenses without conflating them. Inspect state ownership, navigation, focus, errors, scaling, layout expansion, binding, lifecycle, allocations, and runtime evidence. Report prioritized findings and subjective PO scenarios; do not redesign or edit.
-
-Return outcome, evidence, affected records, blockers or decisions, residual risk, and the exact next workflow. Do not claim DONE unless this is the `closure` skill and every configured gate passes.
+- astra-claude: [Astra/Claude procedure](references/astra.md).
+- quality, balanced, inherit: [Legacy procedure](references/legacy.md).
 
+Do not load or merge the other branch. Unknown profiles are configuration errors.

package/.claude/skills/ui-review/references/astra.md

@@ -0,0 +1,17 @@
+# Ui Review — astra-claude
+
+Read [the pipeline protocol](../../../nccgs/protocols/astra-claude.md) and
+[generated routing](../../../nccgs/routing.generated.md). Use Astra's supplied plan,
+requirements, and exact context references. Never substitute a legacy procedure.
+
+Domain objective: Performs a read-only review of Unity UI behavior, architecture, navigation, responsiveness, localization, accessibility, performance, and UX conformance.
+
+Work only inside the supplied plan and assigned write set. Activate an optional
+specialist with the agents command only when its domain is required. Keep one writer
+per path and use official Unity Skills for Editor/assets/runtime operations.
+
+This procedure is read-only: return evidence-backed findings and decisions to Astra
+or the implementer. Supplemental analysis does not replace the mandatory Fable gate.
+
+Use [task gate commands](../../../nccgs/protocols/task-gates.md) for evidence and
+state transitions. Report outcome, evidence, pending decisions, and next owner.

package/.claude/skills/ui-review/references/legacy.md

@@ -0,0 +1,7 @@
+# Ui Review
+
+Read `.nccgs/project.yaml`, `.nccgs/state.md`, and the active context packet when present. Preserve Product Owner authority and unrelated work.
+
+Read approved UX flows, UI requirements, platform/input policy, accessibility targets, implementation, scenes/assets, and tests. Pair UX and Unity UI lenses without conflating them. Inspect state ownership, navigation, focus, errors, scaling, layout expansion, binding, lifecycle, allocations, and runtime evidence. Report prioritized findings and subjective PO scenarios; do not redesign or edit.
+
+Return outcome, evidence, affected records, blockers or decisions, residual risk, and the exact next workflow. Do not claim DONE unless this is the `closure` skill and every configured gate passes.

package/.claude/skills/work/SKILL.md

@@ -1,36 +1,15 @@
 ---
 name: work
 description: Runs end-to-end Unity feature, bug-fix, refactor, tooling, content-integration, or other concrete project work. Routes by domain and risk, composes specialist roles, verifies independently, synchronizes records, and closes configured gates.
-allowed-tools: Read, Glob, Grep, Write, Edit, Bash, Task, AskUserQuestion, WebSearch, WebFetch
+allowed-tools: Read, Glob, Grep, Write, Edit, Bash, Agent, AskUserQuestion, WebSearch, WebFetch
 ---
 
 # Work
 
-Complete the approved outcome rather than stopping at a plan.
+Read models.profile in .nccgs/project.yaml; if absent, use the defaultProfile in
+.claude/nccgs/pipeline.json. Load exactly one procedure:
 
-## Route
-
-Read project policy and state, inspect Git status, and classify FAST, STANDARD, or CONTROLLED using [classification](references/classification.md). Identify affected domains before selecting roles from `../../nccgs/studio.json`.
-
-- FAST normally stays in the main session or uses one specialist.
-- STANDARD uses a context packet, an accountable implementer, and independent verification.
-- CONTROLLED adds the relevant decision owner and adversarial review.
-
-Read the [orchestration protocol](../../nccgs/protocols/orchestration.md) for multi-role work. Do not build a ceremonial management chain.
-
-## Contract
-
-Use [feature contracts](references/feature-contracts.md). Create or refresh a context packet through `/context-pack` for STANDARD or CONTROLLED work. Resolve Product Owner decisions that affect semantics, scope, canon, compatibility, dependencies, or irreversibility; continue safe independent nodes while decisions are pending.
-
-## Execute
-
-Build a dependency graph with one writer per path. Route distinct domains to their specialists. Use official Unity Skills for operation mechanics and `unity-cli` for connected Editor, tests, builds, and runtime evidence. Agents receive the context packet path and bounded ownership, not the entire chat or document set.
-
-Fix in-scope failures, inspect the final diff, and preserve unrelated work.
-
-## Verify and close
-
-Use [verification](references/verification.md) and the shared [evidence protocol](../../nccgs/protocols/evidence.md). Map objective evidence to requirements, run proportional regression, and separate PO gates. CONTROLLED work receives `/review` or `/evidence-review` from an independent role.
-
-Synchronize state and technical records, then run `/closure`. Report classification, roles/models used, changed files, evidence, blockers, PO gates, residual risk, and final state.
+- astra-claude: [Astra/Claude procedure](references/astra.md).
+- quality, balanced, inherit: [Legacy procedure](references/legacy.md).
 
+Do not load or merge the other branch. Unknown profiles are configuration errors.

package/.claude/skills/work/references/astra.md

@@ -0,0 +1,20 @@
+# Work — astra-claude
+
+Read [the pipeline protocol](../../../nccgs/protocols/astra-claude.md) and
+[generated routing](../../../nccgs/routing.generated.md). Use Astra's supplied plan,
+requirements, and exact context references. Never substitute a legacy procedure.
+
+Domain objective: Runs end-to-end Unity feature, bug-fix, refactor, tooling, content-integration, or other concrete project work. Routes by domain and risk, composes specialist roles, verifies independently, synchronizes records, and closes configured gates.
+
+Work only inside the supplied plan and assigned write set. Activate an optional
+specialist with the agents command only when its domain is required. Keep one writer
+per path and use official Unity Skills for Editor/assets/runtime operations.
+
+For code changes: task init with Astra's plan, bind the session, implement and test,
+record implementation evidence, request independent review, fix and re-review as
+needed, then task check --target report. Only after that check passes write and
+record the report for Astra. Stop at AWAITING_ASTRA_CHECK. Astra checking and the
+Product Owner's explicit acceptance happen outside the Claude implementation session.
+
+Use [task gate commands](../../../nccgs/protocols/task-gates.md) for evidence and
+state transitions. Report outcome, evidence, pending decisions, and next owner.

package/.claude/skills/work/references/closure.md

@@ -48,3 +48,13 @@ and next candidate. Git and closure records hold history.
 Technical records may be synchronized autonomously. Canonical player-facing design
 changes require an accepted PO decision. If code and canon disagree, fix or escalate;
 never rewrite canon silently.
+
+## Mandatory astra-claude gates
+
+For astra-claude, lean mode does not remove Fable review, the reviewed report,
+Astra checking, or explicit Product Owner acceptance. Require Fable PASS on the
+final snapshot, then the report, then Astra PASS, then Product Owner ACCEPTED on
+that same snapshot. Claude records factual progress/evidence and hands off at
+AWAITING_ASTRA_CHECK. Astra owns final documentation and closure. Missing or stale
+sign-offs block DONE; changes require renewed review and downstream acceptance.
+Only an explicit Product Owner policy change can alter this sequence.

package/.claude/skills/work/references/feature-contracts.md

@@ -30,3 +30,7 @@ debugging.
 Requirement IDs belong in `.nccgs/requirements.yaml` only for player-visible
 contracts, technical invariants, acceptance criteria, and cross-system dependencies.
 Do not assign IDs to private class structure or every documentation sentence.
+
+Under astra-claude, Astra authors and revises these contracts and context packets.
+Claude consumes them, records newly discovered facts in evidence, and requests
+revisions from Astra instead of rewriting planning or canonical documents.

package/.claude/skills/work/references/legacy.md

@@ -0,0 +1,29 @@
+# Work
+
+Complete the approved outcome rather than stopping at a plan.
+
+## Route
+
+Read project policy and state, inspect Git status, and classify FAST, STANDARD, or CONTROLLED using [classification](../references/classification.md). Identify affected domains before selecting roles from `../../nccgs/studio.json`.
+
+- FAST normally stays in the main session or uses one specialist.
+- STANDARD uses a context packet, an accountable implementer, and independent verification.
+- CONTROLLED adds the relevant decision owner and adversarial review.
+
+Read the [orchestration protocol](../../../nccgs/protocols/orchestration.md) for multi-role work. Do not build a ceremonial management chain.
+
+## Contract
+
+Use [feature contracts](../references/feature-contracts.md). Create or refresh a context packet through `/context-pack` for STANDARD or CONTROLLED work. Resolve Product Owner decisions that affect semantics, scope, canon, compatibility, dependencies, or irreversibility; continue safe independent nodes while decisions are pending.
+
+## Execute
+
+Build a dependency graph with one writer per path. Route distinct domains to their specialists. Use official Unity Skills for operation mechanics and `unity-cli` for connected Editor, tests, builds, and runtime evidence. Agents receive the context packet path and bounded ownership, not the entire chat or document set.
+
+Fix in-scope failures, inspect the final diff, and preserve unrelated work.
+
+## Verify and close
+
+Use [verification](../references/verification.md) and the shared [evidence protocol](../../../nccgs/protocols/evidence.md). Map objective evidence to requirements, run proportional regression, and separate PO gates. CONTROLLED work receives `/review` or `/evidence-review` from an independent role.
+
+Synchronize state and technical records, then run `/closure`. Report classification, roles/models used, changed files, evidence, blockers, PO gates, residual risk, and final state.

package/README.md

@@ -5,27 +5,33 @@ specialization of a full game-development team with project-aware governance,
 bounded orchestration, configurable model routing, official Unity Skills, objective
 evidence, and strict closure.
 
-Version 1.1.1 contains:
+Version 1.3.0 contains:
 
-- 45 studio agents across leadership, design, programming, Unity, content, quality,
-  and release;
+- 10 active core agents and 35 optional specialist roles in a 45-agent library;
 - 36 composable NCCGS workflows with eight primary entrypoints;
 - 11 path-scoped project rules;
-- 8 cross-platform Node lifecycle/safety hooks;
+- 11 cross-platform Node lifecycle/safety hook commands;
 - 22 unmodified official Unity Skills with a SHA-256 manifest;
 - brownfield migration, context packets, evidence records, closure modes, model
-  profiles, installer backup, and behavioral framework tests.
+  profiles, installer backup, and behavioral framework tests;
+- executable task gates, runtime observations, generated routing from one policy,
+  and separate Astra/legacy workflow instructions.
+
+Read the [NCCGS 1.3 operating guide](docs/NCCGS-1.3.md) for setup and evidence schemas.
+Install the npm release with the commands below, or use this checkout's
+`scripts/cli.mjs` when working from source.
 
 ## Operating architecture
 
 ```text
 Product Owner
-  -> Claude Code primary orchestrator
-  -> NCCGS policy, context packet, risk and model router
-  -> specialized studio roles (bounded to one delegation level)
-  -> official Unity Skills
-  -> Unity CLI/Pipeline and Editor
-  -> independent evidence, review, and closure
+  -> Astra: planning, contracts, and canonical documents
+  -> Claude Opus 5.5 / xhigh: implementation and tests
+  -> Fable / high: independent review (Opus fixes -> Fable re-review)
+  -> reviewed report to Astra
+  -> Astra: plan/evidence check and document updates
+  -> Product Owner: final acceptance
+  -> DONE when all gates pass
 ```
 
 NCCGS does not simulate a hierarchy for ceremony. It selects specialized roles when
@@ -53,6 +59,11 @@ npx nccgs@latest update
 npx nccgs@latest doctor
 ```
 
+To adopt the Astra/Opus/Fable pipeline in an existing project, add
+`--model-profile astra-claude` to both update commands. Without this option, updates
+preserve the project's current profile. Publishing a new release does not modify
+Unity projects automatically; run the update command in each project.
+
 The current directory is the installation target. `--project <directory>` remains
 available for automation, but is not needed for ordinary use.
 
@@ -73,28 +84,51 @@ Select a model profile during installation when needed:
 
 ```powershell
 npx nccgs@latest update --model-profile quality
+npx nccgs@latest update --model-profile astra-claude
 npx nccgs@latest update --model-profile balanced
 npx nccgs@latest update --model-profile inherit
 ```
 
-Open a new Claude Code session after first-time adoption of an existing project and
-run `/migrate-project` once. It performs review, plan, apply, and verify as one
-CONTROLLED operation, pausing only for a real Product Owner decision or unsafe side
-effect. Ordinary framework updates report `No project migration required` and need
-no Claude migration command.
+For `astra-claude`, Astra first prepares the migration plan and maps existing project
+obligations into `.nccgs/gates.json`. Claude then runs `/migrate-project` against that
+plan. Legacy profiles retain their combined review/plan/apply/verify workflow.
+Ordinary framework updates need no migration when the installer reports none.
 
 ## Model routing
 
-| Profile | Lightweight | Standard | Deep | Leadership |
-|---|---|---|---|---|
-| `quality` (default) | Haiku | Sonnet | Opus | Opus |
-| `balanced` | Haiku | Sonnet | Opus | Sonnet |
-| `inherit` | Session model | Session model | Session model | Session model |
+| Profile | Claude implementation/support | Independent reviewer | Effort |
+|---|---|---|---|
+| `astra-claude` (default) | Pinned `claude-opus-5-5` | `fable` | Implementation `xhigh`; review `high` |
+| `quality` | Haiku lightweight; Sonnet standard; Opus deep/leadership | Opus | Session |
+| `balanced` | As quality, with Sonnet leadership | Opus | Session |
+| `inherit` | Session model | Session model | Session |
+
+In `astra-claude`, Astra owns planning and canonical documents outside Claude Code.
+All Claude coding/support roles use Opus 5.5/xhigh; planning roles are consultants
+only. Every code change, including FAST work, requires independent Fable review
+before Claude produces the final report for Astra. Astra checks it, updates documents,
+and requests explicit Product Owner acceptance. Claude does not declare DONE itself.
+See [the complete handoff protocol](.claude/nccgs/protocols/astra-claude.md).
+
+The configurator sets agent model/effort and the project's main session settings.
+The canonical source is `.claude/nccgs/pipeline.json`; `npm run policy:build`
+generates routing outputs and `npm run policy:check` detects drift. Optional roles
+are activated with `nccgs agents --enable <role>` and preserved through updates.
+No silent model/effort fallback is accepted for this profile. Existing projects keep
+their current profile on ordinary updates; select `--model-profile astra-claude`
+explicitly to adopt this chain. Legacy profiles remain available by explicit choice.
+Opus 5.5 requires Claude Code 2.1.280 or later. Fable is an alias: record its actual
+resolved model and effort from runtime evidence, not configuration alone.
+
+To use a source checkout directly, run its CLI with the target Unity project,
+first with `--dry-run`:
+
+```powershell
+node "D:\path\to\NCCGS\scripts\cli.mjs" update --project "D:\path\to\UnityProject" --model-profile astra-claude --dry-run
+```
 
-Role selection happens before model selection. Risk may escalate a task, but a title
-alone does not justify Opus. If a pinned model is unavailable or conflicts with the
-current context mode, project policy can allow the primary session to continue using
-the inherited model and report the fallback.
+Use `install` for first adoption. Remove `--dry-run` to apply. NCCGS prepares a
+file-backed report for Astra; it does not automatically send messages between apps.
 
 Reconfigure an installed project:
 

package/UPGRADING.md

@@ -2,6 +2,43 @@
 
 Always inspect a dry run and preserve project policy.
 
+## 1.3.0: executable gates and focused agent activation
+
+Read the [1.3 guide](docs/NCCGS-1.3.md). Routing now comes from
+`.claude/nccgs/pipeline.json`; generated files are checked for drift. Only ten core
+agents are active by default; 35 additional roles are available through `agents
+--enable`. Project customizations are preserved for explicit migration review.
+
+Astra must map existing checks, closure artifacts and manual decisions into
+`.nccgs/gates.json`, referencing the current project policy hash. A 1.2 policy with
+embedded `profile_contracts` needs migration to the canonical `pipeline_policy`
+reference. Do not discard project requirements during this migration.
+
+Task records now enforce implementation -> independent review -> report -> Astra
+check -> Product Owner acceptance. Missing observed effort blocks the gate; requested
+settings alone do not prove runtime behavior. Each workflow loads only its selected
+Astra or legacy branch. The legacy instructions later in this document apply only
+to explicitly selected legacy profiles.
+
+## Astra / Opus / Fable routing
+
+New installations default to `astra-claude`: Astra plans and owns documents;
+Opus 5.5/xhigh implements; Fable/high reviews; Claude hands the reviewed report to
+Astra; Astra checks it; the Product Owner accepts before DONE. See the
+[handoff protocol](.claude/nccgs/protocols/astra-claude.md).
+
+Existing projects keep their profile unless explicitly changed. In each Unity
+project, run `npx nccgs@latest update --model-profile astra-claude --dry-run`, inspect
+it, then remove `--dry-run`. Source checkouts can use `node <NCCGS>/scripts/cli.mjs`
+instead of `npx nccgs@latest`. Publishing alone does not update installed projects.
+Use `install` instead of `update` for a new installation. Require Claude Code
+2.1.280+ and access to Opus 5.5 and Fable. Review any
+preserved agent/settings overrides so the observed model and effort match policy.
+The configurator preserves unrelated shared settings and backs up replacements;
+existing sessions and local/environment overrides still require a runtime check.
+
+## Earlier update behavior
+
 ```powershell
 cd "D:\path\to\UnityProject"
 npx nccgs@latest update --dry-run

package/VERSION

@@ -1 +1 @@
-1.1.1
+1.3.0

package/docs/ARCHITECTURE.md

@@ -11,8 +11,8 @@ distinct decision surface, failure mode, artifact, or verification lens.
 1. **Kernel** — constitution, authority, project policy, ownership, state, closure.
 2. **Router** — domain/risk classification, model profile, context packet, execution
    graph and write ownership.
-3. **Studio roles** — 45 leadership, lead, specialist, Unity, quality, and release
-   agents with shared handoff contracts.
+3. **Studio roles** — ten active core agents, with 35 optional roles in a 45-agent
+   library. Activation choices survive updates; customized files are preserved.
 4. **Workflow modules** — 36 composable skills with eight user-facing entrypoints.
 5. **Unity operations** — unmodified official Unity Skills and Unity CLI/Pipeline.
 6. **Automation** — Node hooks for recovery, audit and safety.
@@ -20,7 +20,9 @@ distinct decision surface, failure mode, artifact, or verification lens.
 
 ## Bounded orchestration
 
-The primary Claude Code session is the sole orchestrator. It constructs a DAG,
+The primary Claude Code session is the sole Claude subagent orchestrator. Under the
+default astra-claude profile, Astra owns planning/documents outside that session;
+Claude executes within Astra's plan. It constructs an implementation DAG,
 selects the smallest complete role set, passes a context packet, assigns one writer
 per path, and merges handoffs. Studio agents do not recursively delegate.
 
@@ -40,14 +42,16 @@ require reconstructing truth from chat history.
 
 ## Model architecture
 
-Each catalog role has a capability class: lightweight, standard, deep, or leadership.
-An installation profile compiles that class into agent frontmatter. The package
-defaults to quality but supports balanced and inherit profiles without duplicating
-agent prompts.
+Each catalog role has a capability class: lightweight, standard, deep, leadership,
+or review. `pipeline.json` is the routing source; the compiler generates catalog
+routes and agent frontmatter, and validation rejects drift. The default
+astra-claude profile uses claude-opus-5-5/xhigh for Claude work and fable/high for
+independent review. It also configures the main project session. Legacy quality,
+balanced, and inherit profiles remain available without duplicating agent prompts.
 
-The router selects role/domain first and model second. Risk escalation is evidence
-based. Project policy records fallback behavior so model availability failure does
-not become an unexplained task blocker.
+Legacy profiles support policy-directed fallback. The astra-claude profile instead
+blocks a required stage when its model/effort cannot be confirmed; it never labels
+an inherited model's output as Fable review or Opus 5.5 implementation.
 
 ## Unity boundary
 
@@ -57,9 +61,22 @@ SHA-256 manifest makes upstream modifications visible.
 
 ## Evidence and closure
 
-Implementation and verification are separate work nodes. STANDARD work normally has
-independent verification; CONTROLLED work adds adversarial review. Evidence is mapped
-to requirements and source revision.
+For astra-claude, every code change follows Opus implementation -> Fable review ->
+report to Astra -> Astra check -> explicit Product Owner acceptance. Opus fixes
+findings and Fable re-reviews each changed snapshot. These gates apply to FAST and
+lean work as well. `task.mjs` enforces this sequence using snapshot hashes, immutable
+artifact references, runtime observations and explicit external decisions. A changed
+snapshot invalidates previous review/acceptance. Re-recording an earlier stage clears
+downstream stages while retaining history. Hooks supplement the task gate; they are
+not a security sandbox. There is no automatic cross-app transport.
+
+Requested model settings and observed runtime values are separate fields. Native
+events can report the resolved model and fallback history; absent observed effort
+stays unknown until supporting runtime evidence is imported.
+
+Every workflow dispatcher loads exactly one `references/astra.md` or
+`references/legacy.md`. Risk affects the extent of verification, never the mandatory
+Fable gate in the Astra profile. Legacy profiles retain their risk-based review.
 
 One canonical closure record links existing project artifacts. `lean`, `linked`, and
 `strict` policies determine what blocks DONE. Waivers are item-specific and retain