chati-dev 4.3.1 → 4.4.1

package/README.md

@@ -128,8 +128,10 @@ The system saves your full session state (pipeline position, current agent, deci
 |---------|--------------|
 | **Specialized Agents** | Each agent has a defined mission, success criteria, and handoff protocol. Not one AI trying to do everything |
 | **Agent Teams** | Planning Team (Detail + Architect + UX + QA-Planning) and Build Team (Dev + QA-Implementation) with Wave-Based Pipeline. Teams work in parallel with peer-to-peer communication, cross-review, and intra-team quality assurance. No single agent bottleneck |
-| **Visual QA** | QA-Visual agent validates the visual experience: screenshots at 5 scroll positions, hover states, GSAP/Lenis verification, brandbook compliance. Powered by Playwright. |
-| **Premium Animations** | GSAP + ScrollTrigger + Lenis as mandatory stack for UI projects. Every page has scroll reveals, micro-interactions, and smooth scrolling. |
+| **Visual QA** | QA-Visual scores the rendered UI against an 8-dimension rubric (layout, typography, motion, interaction, color, spacing, accessibility, performance) with an anti-slop catalog and a spring-motion veto (linear easing and `transition: all` are rejected). Screenshots at multiple scroll positions, hover states, and brandbook compliance, powered by Playwright. |
+| **Premium Motion** | Premium motion is required for UI projects, with cubic-bezier easing and full reduced-motion support. The visual gate is animation-stack aware: the motion-premium presets use GSAP + Lenis, the saas-dashboard preset uses framer-motion. |
+| **Scaffold Presets** | Premium starting points applied at build time: motion-premium and motion-premium-3d for animated showcase sites, saas-dashboard for a token-driven app shell (sidebar, sortable data table, KPI cards, a Cmd+K command palette). |
+| **Execution Kinds** | A task declares how it executes. Deterministic dev tasks (lint, tests, file and schema checks) run as zero-LLM scripts through whitelisted executors instead of spending an agent turn. Feature-flagged, off by default, fully back-compatible. |
 | **Multi-CLI Architecture** | Choose your AI provider at install time: Claude, Gemini, or Codex. Each agent gets the optimal model for that provider |
 | **Quality Gates** | Every phase is validated before moving forward, including QA-Visual for visual quality. 3-tier verdicts: APPROVED, NEEDS_REVISION, or BLOCKED |
 | **Fault Vector Protocol** | Root Layer Routing classifies bugs by origin (INTENT/SPEC/CODE/DEFER). Evidence-Bound Verdicts, Decision Trail, and Echo Detection prevent cyclical rework |
@@ -137,7 +139,7 @@ The system saves your full session state (pipeline position, current agent, deci
 | **Session Lock** | Once activated, you stay inside the system. No accidentally "falling out" into generic AI mode |
 | **Multi-Terminal** | Autonomous agents run in parallel. Planning Team (Detail + Architect + UX + QA-Planning) and Build Team (Dev + QA-Implementation) coordinate via peer-to-peer mailbox |
 | **Memory System** | The system learns from mistakes. Gotchas are captured automatically and recalled when relevant |
-| **Execution Profiles** | Three profiles: explore (read-only), guided (default), autonomous (gate >= 95%). Includes safety net and circuit breaker |
+| **Execution Mode** | Two modes: interactive (default) and autonomous (gate >= 95%). Safety reversion to interactive on stuck loop, quality drop, scope creep, error cascade, or user override (Article XVII) |
 | **IDE-Agnostic** | Works with Claude Code, VS Code, Cursor, Gemini CLI, Codex CLI, and AntiGravity |
 | **4 Languages** | Interface supports English, Portuguese, Spanish, and French. Artifacts are always generated in English |
 | **Supply Chain Security** | Every file is cryptographically signed (Ed25519). Tampered packages are blocked on install |
@@ -190,17 +192,16 @@ Prompts are piped via stdin for all providers. You can override individual agent
 
 ---
 
-## Execution Profiles
+## Execution Mode
 
-Three profiles control how much autonomy agents have:
+Two modes control how much autonomy agents have:
 
-| Profile | Behavior | When to use |
-|---------|----------|-------------|
-| **explore** | Read-only. Agents analyze but don't modify files | Understanding a new codebase |
-| **guided** | Default. Agents propose changes, you approve | Normal development workflow |
+| Mode | Behavior | When to use |
+|------|----------|-------------|
+| **interactive** | Default. Agents propose changes, you approve | Normal development workflow |
 | **autonomous** | Agents execute without confirmation (quality gates >= 95%) | Trusted pipelines with high quality scores |
 
-The system starts in `guided` mode. Transition to `autonomous` requires both QA gates scoring >= 95%. A safety net with 5 triggers (stuck loop, quality drop, scope creep, error cascade, user override) automatically reverts to guided mode when needed.
+The system starts in `interactive` mode. Transition to `autonomous` requires both QA gates scoring >= 95%. Safety reversion takes the pipeline back to interactive on a stuck loop, quality drop, scope creep, error cascade, or user override (Article XVII).
 
 ---
 
@@ -274,13 +275,13 @@ Three systems operate transparently behind the pipeline:
 
 | System | What it does |
 |--------|-------------|
-| **Context Engine (PRISM)** | Injects the right context at the right time. 6 layers of context (from constitution rules down to keyword-matched domain rules). Uses Progressive Reinforcement: as context degrades, injection budgets INCREASE (1.5% to 5.0% of provider context window). Provider-aware for Claude (200K), Gemini (1M), and Codex (128K). |
+| **Context Engine (PRISM)** | Injects the right context at the right time. 6 layers of context (from constitution rules down to keyword-matched domain rules). Uses Progressive Reinforcement: as context degrades, injection budgets INCREASE (1.5% to 5.0% of provider context window). Provider-aware for Claude (1M), Gemini (1M), and Codex (128K). |
 | **Memory System (RECALL)** | Remembers decisions, gotchas, and lessons across sessions. Organized into 4 sectors: what happened (episodic), what we know (semantic), how we do things (procedural), and what we learned (reflective). |
 | **Decision Engine (COMPASS)** | Before creating something new, checks if a similar component already exists. Decides whether to reuse, adapt, or create from scratch. Keeps a registry of all project entities. |
 
 ### Constitution
 
-The system is governed by a **23-article Constitution** that enforces agent behavior, quality standards, security, and system integrity:
+The system is governed by a **25-article Constitution** that enforces agent behavior, quality standards, security, and system integrity:
 
 - **Agent Governance** - Every agent has a defined mission, scope, and success criteria
 - **Quality Standards** - Minimum 95% score on quality gates. 3-tier verdicts (APPROVED / NEEDS_REVISION / BLOCKED)
@@ -288,10 +289,14 @@ The system is governed by a **23-article Constitution** that enforces agent beha
 - **Mode Governance** - Planning mode can't modify project code. Build mode has full access
 - **Session Lock** - Once activated, all messages route through the orchestrator
 - **Model Governance** - Each agent runs on its designated model, enforced by the CLI adapter
-- **Execution Profiles** - Explore, guided, and autonomous modes with safety net and circuit breaker
+- **Execution Mode** - Interactive (default) and autonomous modes, with safety reversion to interactive (Article XVII)
 - **Multi-CLI** - Provider-agnostic architecture with adapter pattern and automatic model mapping
 - **Agent Teams** - Real-time peer communication between agents with cross-review, per-task QA, and sealed-bid quality assurance
 - **Fault Vector Protocol** - Root Layer Routing, Evidence-Bound Verdicts, Decision Trail, and Echo Detection for adversarial quality
+- **Anti-Loop Protocol** - the 3-Strike Rule and mandatory diagnosis mode stop an agent from editing the same file in circles
+- **Reasoning Tier** - per-task cognitive depth (standard / deep / deliberate) with deterministic escalation
+- **Model-Harness Parity** - a re-audit advisory when the active model family changes, so calibration does not drift silently
+- **Execution Kinds** - a task declares how it runs (agent, deterministic script, specialist, or human gate); deterministic dev work can run with no model call (feature-flagged, off by default)
 
 ---
 

package/bin/chati.js

@@ -141,6 +141,22 @@ async function main() {
         console.log(`  Framework sync skipped: ${err.message}`);
       }
 
+      // 2.55. Refresh .claude/settings.json (merge-safe) so hook registrations
+      // added in newer versions reach existing installs. Without this, projects
+      // installed before a hook was added to the generator never receive it
+      // (reasoning-escalator and git-push-authority were missing for everyone).
+      try {
+        const { writeClaudeSettingsWithMerge } = await import('../src/installer/core.js');
+        const { existsSync: ex } = await import('fs');
+        const settingsPath = join(targetDir, '.claude', 'settings.json');
+        if (ex(join(targetDir, '.claude'))) {
+          writeClaudeSettingsWithMerge(settingsPath);
+          console.log('  Claude settings refreshed (hooks re-registered).');
+        }
+      } catch (err) {
+        console.log(`  Settings refresh skipped: ${err.message}`);
+      }
+
       // 2.6. Re-bundle CLI (Bug 4 part 3, v4.3.1+) — refreshes .chati.dev/_cli/
       // and writes version.json so chati-router.js mismatch detection clears
       // on the next invocation.
@@ -363,7 +379,7 @@ async function main() {
       console.log('  - Specialized agents across DISCOVER, PLAN, BUILD, DEPLOY phases');
       console.log('  - 6 workflow blueprints');
       console.log('  - 6 templates');
-      console.log('  - Constitution (23 Articles + Preamble)');
+      console.log('  - Constitution (25 Articles + Preamble)');
       console.log('  - Dashboard TUI');
       console.log('  - Upgrade system with migrations');
       console.log('  - 6 IDE support');
@@ -439,7 +455,13 @@ Telemetry:
       const keyFlag = args.find(a => a.startsWith('--key='));
       const keyArg = keyFlag ? keyFlag.replace('--key=', '') : null;
       const { runActivate } = await import('../src/license/commands.js');
-      await runActivate(targetDir, keyArg);
+      const result = await runActivate(targetDir, keyArg);
+      // runActivate returns { ok, cancelled?, error? } and never exits the
+      // process itself (keeps it unit-testable). Translate to an exit code
+      // here at the CLI boundary: failures other than user-cancel exit 1.
+      if (result && result.ok === false && !result.cancelled) {
+        process.exit(1);
+      }
       break;
     }
 
@@ -474,7 +496,7 @@ Telemetry:
 
       const scaffoldSrc = (() => {
         try { return resolveScaffoldSource(targetDir); }
-        catch (e) { console.error(`Error: ${e.message}`); process.exit(1); }
+        catch (e) { console.error(`Error: ${e.message}`); process.exit(1); return undefined; }
       })();
 
       if (!sub || sub === 'list') {
@@ -545,8 +567,10 @@ Telemetry:
       const VALID_ORCHESTRATE_SUBS = new Set([
         'next', 'advance', 'init', 'status', 'validate-handoff',
         'deviation', 'exit', 'providers', 'detect-flow', 'backlog',
-        'qa-plan-score', 'qa-impl-score', 'scan',
+        'qa-plan-score', 'qa-impl-score', 'qa-visual-score', 'scan',
         'spawn-team', 'team-status', 'team-dissolve',
+        'wait-for-capture', 'wave-status', 'assess-coverage',
+        'wait-for-license', 'doctor', 'run-script-task',
       ]);
       if (!VALID_ORCHESTRATE_SUBS.has(subCommand)) {
         console.error(`Error: Unknown orchestrate subcommand "${subCommand}".`);

package/framework/agents/build/dev.md

@@ -806,6 +806,8 @@ Score = criteria met / total criteria
 Threshold: >= 95% per task (minimum 9/10)
 ```
 
+This self-validation score is the Dev agent's own per-task done-check (it decides when a task is finished and ready to hand off). It is NOT the phase gate. The BUILD phase only advances when the separate QA-Implementation agent's verdict clears its threshold (enforced in the orchestrator pipeline manager). Never wire this self-score into a phase-transition decision; the separated QA verdict is always the gate.
+
 ---
 
 ## Intelligence Integration
@@ -822,7 +824,9 @@ Before implementing each task:
 After completing each task:
 1. If a new gotcha was discovered -> append to gotchas.yaml
 2. If a successful pattern was used -> append to patterns.yaml
-3. Update confidence.yaml with execution results
+3. Update confidence.yaml with execution results. When you write measured
+   values, set that agent's `calibration_status: calibrated` and
+   `last_calibrated` to today; never leave a measured number as `seed`.
 ```
 
 ---

package/framework/agents/discover/greenfield-wu.md

@@ -308,13 +308,28 @@ post-qa-planning gate fires when both align.
 
 ### motion-premium triggers (2D premium)
 Record `motion-premium` as a candidate when ALL of the following are true:
-- The stack includes **Next.js** (any version) OR **React** (any version).
+- The stack includes **Next.js** (any version). These presets ship Next.js templates, so a plain React or React+Vite stack is not a match.
 - The stack includes **Tailwind CSS** (any version — v3 or v4).
 - The stack includes at least one premium animation library: **GSAP**,
   **Lenis**, or **Framer Motion**. A user-described intent that clearly
   implies these (e.g. "smooth scroll pinning", "scroll-reveal storytelling",
   "page-scroll narrative") also counts — document the phrase as evidence.
 
+### saas-dashboard triggers (app-shell / data-dense)
+Record `saas-dashboard` as a candidate when ALL of the following are true:
+- The stack includes **Next.js** (any version). These presets ship Next.js templates, so a plain React or React+Vite stack is not a match.
+- The stack includes **Tailwind CSS** (any version).
+- The project is an **application or internal tool**, not a marketing or
+  showcase site. Any of these signals counts (document the phrase as evidence):
+  dashboard, admin panel, SaaS app, internal tool, analytics, data table,
+  KPI or metrics view, sidebar navigation, app shell, settings pages, command
+  palette (Cmd+K), CRUD over records, back office.
+
+saas-dashboard is the app-shell counterpart to the motion-premium presets. In
+practice it is mutually exclusive with them (an animated marketing site is not
+a data dashboard), but recording both is harmless: the auto-gate fires the
+highest-confidence candidate first and the user chooses.
+
 ### motion-premium-3d triggers (R3F + Three.js, workstream 3)
 Record `motion-premium-3d` as a candidate when ALL of the following are true:
 - The stack includes **Next.js** (any version).

package/framework/agents/plan/tasks.md

@@ -85,12 +85,43 @@ Task Definition:
   - Priority: critical | high | medium | low
   - Size: XS (<1h) | S (1-2h) | M (2-4h) | L (4-8h) | XL (8h+ -> split!)
   - Dependencies: [T{x}.{y}, ...]
+  - Execution Kind: agent (default) | script | specialist | human  (see Step 2c)
   - Acceptance Criteria (Given-When-Then):
     - Given {initial context/state}
     - When {action is performed}
     - Then {expected outcome, verifiable}
 ```
 
+### Step 2c: Execution Kind (deterministic tasks run as scripts)
+
+Most tasks are `agent`: an LLM implements them. Some, though, are fully
+deterministic and waste an agent turn. When the `execution_kinds` feature is on,
+assign the right kind so the orchestrator can run deterministic work as a
+zero-LLM script. When the feature is off, every task runs as `agent` and this
+step is a no-op, so it is always safe to record the kind.
+
+Decision tree (top to bottom, first match wins):
+
+1. **human** - a critical, irreversible, legal, financial, or product-judgment
+   decision a person must own (for example "confirm the production migration",
+   "approve the pricing copy"). Set `execution_ref` to a one-line description of
+   the decision. A person decides; no automation.
+2. **script** - a deterministic dev operation with a single correct outcome and
+   no design judgment: run the linter or the test suite, verify build output
+   exists, validate a JSON file. Set `execution_ref` to an executor from
+   `chati.dev/executors/registry.yaml` with its argument, for example
+   `npm-script:test`, `npm-script:lint`, `path-check:dist,build`, or
+   `json-validate:package.json`. The acceptance criterion is the executor's pass
+   condition (exit 0, paths present, valid JSON). If no executor fits, it is not
+   a script task.
+3. **specialist** - creative work that carries a fixed methodology the agent must
+   follow exactly (reserved; today this resolves to `agent`).
+4. **agent** (default) - anything that needs implementation, design, or judgment.
+
+Be conservative: when in doubt between script and agent, choose agent. A script
+task that needs any interpretation is a defect. Only the registered executors
+are valid `script` refs; never invent an executor id.
+
 ### Step 2b: Animation Inventory → Tasks (1:1 mapping, REQUIRED when inventory exists)
 
 Before producing the final task list, check for `artifacts/4-UX/animation-inventory.md`. If it exists, it is the SOURCE OF TRUTH for every animation-related task in the build. The brand-architect locked this artifact at the `advance ux` gate; your job is to convert it into tasks faithfully.

package/framework/agents/plan/ux-brand-architect.md

@@ -157,6 +157,8 @@ Premium means:
 - Stagger animations on any repeated elements (cards, list items, grid children)
 - Page load MUST have an entrance animation sequence (not just content appearing)
 
+**Design toward the QA-Visual rubric.** The canonical quality bar is the 8-dimension rubric in `chati.dev/data/qa-rubrics.yaml` under `visual_rubric`. Its per-dimension premium markers are your targets; its slop markers are forbidden. The tokens, modular type scale, semantic color system, spacing scale, and motion language you define here are exactly what QA-Visual scores against, so set them to clear the premium band on every dimension. Layout, typography, color, and motion carry the highest weights. Concrete musts the rubric enforces: a base-unit spacing system (4px or 8px, no off-scale values), a modular type scale (a consistent ratio, not round-number sizes), a documented semantic color layer (no raw hex literals, accent restraint), cubic-bezier easing tied to each transition's physical metaphor (never `linear` or bare `ease`, never `transition: all`), and full interactive-state coverage. Avoid the anti-slop catalog tells (generic saturated-blue primary, metronomic identical section padding, the symmetric centered hero with no compositional interest).
+
 The 5 benchmark sites (landonorris.com, iertqa.com, toptier.relats.com, refractweb.com, magic5.ro) are the FLOOR, not the ceiling.
 
 If `visual_quality: functional` → reduce to CSS transitions only, no GSAP requirement.
@@ -379,10 +381,27 @@ Append a `## Scaffold Signals` section to your handoff file
   evidence:
     - "Named 3D reference site"
     - "Inventory row mapping to scaffold/motion-premium-3d/*"
+
+- preset: saas-dashboard
+  candidate: true
+  confidence: 0.0 to 1.0
+  evidence:
+    - "App-shell intent (dashboard, admin panel, SaaS app, internal tool)"
+    - "Brief names data tables, KPIs, sidebar navigation, or a command palette"
 ```
 
-Emit both entries when both presets apply — the post-qa-planning gate
-fires them in sequence after the user applies or skips the first.
+Emit each applicable entry. saas-dashboard is the app-shell counterpart and is
+mutually exclusive with the motion presets in practice (a data dashboard is not
+an animated marketing site); emit it instead of the motion presets when the
+brief describes an application rather than a showcase. The post-qa-planning gate
+fires the candidates in sequence after the user applies or skips the first.
+
+`confidence` scoring for `saas-dashboard` (gate threshold 0.7):
+- 1.0 — the brief explicitly calls the product a dashboard, admin panel, or
+  SaaS app AND names at least two app-shell elements (data tables, KPIs,
+  sidebar nav, command palette, settings).
+- 0.7 to 0.9 — clear application intent with one or two app-shell elements.
+- below 0.7 — ambiguous; a content or marketing site. Do NOT emit.
 
 `confidence` scoring for `motion-premium` (gate threshold 0.7):
 - 1.0 — multiple named premium refs (Oryzo + Norris + Relats, etc.) AND

package/framework/agents/plan/ux-component-engineer.md

@@ -17,6 +17,8 @@ You are the **Component Engineer**, a specialist sub-agent of the UX Manager. Yo
 
 Define HOW components behave, move, and comply with accessibility standards. Build the component hierarchy (Atomic Design), discover reusable components from 21st.dev, validate WCAG 2.2 + APCA compliance, and define the Motion System. Every component must have full state coverage and every animation must be GPU-accelerated with reduced-motion alternatives.
 
+**Design toward the QA-Visual rubric.** Your work is scored against the `motion`, `interaction`, and `accessibility` dimensions of `chati.dev/data/qa-rubrics.yaml` (`visual_rubric`). Targets the rubric enforces: every transition uses a cubic-bezier curve matched to its physical metaphor (deceleration for entering, acceleration for leaving), with a duration scale tied to element size (micro-interactions 80-150ms, component state 200-350ms, page-level <= 500ms); never `linear`, never bare `ease`, never `transition: all`. Every interactive element has four visually distinct states (resting, hover, active, disabled) and a high-contrast focus ring (>= 2px width, >= 2px offset). Accessibility is a hard gate: text meets WCAG 2.1 AA contrast at rendered size, focus order matches visual order, headings form a strict hierarchy with no skipped levels, and inputs never use placeholder as the sole label. A failing accessibility dimension caps the whole visual score at 70.
+
 ---
 
 ## On Activation
@@ -74,6 +76,14 @@ Use the BEST patterns found as the BASIS for this project's animation specs.
 Do NOT invent animation patterns from scratch - ALWAYS reference proven, production implementations.
 Adapt patterns to the project's design tokens (colors, fonts, spacing from Brand Architect).
 
+> Governance scope: the sites above (and shadcn/ui below) are external DISCOVERY
+> sources this agent reads at design time to ground specs in proven patterns.
+> They are tools, like any component library or icon set, not a lineage. The
+> authorship rule governs the SHIPPED output: it must be original work in the
+> project's own tokens and voice. When adapting a fetched component, reimplement
+> it against the project tokens and never carry over its branding, attribution,
+> or source-comment headers.
+
 **Component Library Strategy** (verified, scrapable sources):
 
 Primary: Shadcn/ui (MIT license, React + Radix UI + Tailwind + cva)

package/framework/agents/quality/qa-implementation.md

@@ -697,7 +697,7 @@ Criteria (binary pass/fail):
 16. Decision Trail checked pre-flight and updated post-correction (Article XXII)
 
 Score = criteria met / total criteria
-Threshold: >= 95% (15/16 minimum)
+Threshold: >= 95% (15/16 minimum). Canonical value: chati.dev/data/qa-rubrics.yaml (thresholds.qa-implementation); this number must match that file.
 ```
 
 ---

package/framework/agents/quality/qa-planning.md

@@ -373,7 +373,7 @@ Criteria (binary pass/fail):
 10. Correction loops executed for all issues (or escalated with justification)
 
 Score = criteria met / total criteria
-Threshold: >= 95% (10/10 minimum, no criteria may fail)
+Threshold: >= 95% (10/10 minimum, no criteria may fail). Canonical value: chati.dev/data/qa-rubrics.yaml (thresholds.qa-planning); this number must match that file.
 ```
 
 ---

package/framework/agents/quality/qa-visual.md

@@ -43,6 +43,12 @@ Validate that the implemented UI meets visual quality standards: animations are
 
 ---
 
+## Scoring rubric (canonical)
+
+The visual score is computed against the 8-dimension rubric in `chati.dev/data/qa-rubrics.yaml` under `visual_rubric`. That file is the single source of truth: the dimensions (layout, typography, motion, interaction, color, spacing_composition, accessibility, performance), their weights, the premium and slop markers, the measurable and judged signals, the `weak_dimension_cap`, the `anti_slop_catalog`, and the `scoring_protocol` all live there. Do NOT invent ad-hoc criteria. Score every dimension against the rubric and follow its `scoring_protocol`: scoring order (accessibility first, it carries the hardest cap), evidence standard, anti-slop catalog penalties, the weak-gateway cap (84 when Playwright is unavailable), the weighted-total formula, and correction-loop routing. Apply `weak_dimension_cap` before the weighted total. A measurable-signal finding MUST cite the computed value; a judged-signal finding MUST cite a screenshot filename (Article XXII evidence standard). The em-dash penalty and `visual-qa.js` `motion_errors` (the spring-motion veto) feed the motion and typography dimensions.
+
+---
+
 ## What This Agent Validates
 
 | Check | Tool |
@@ -304,33 +310,30 @@ If no reference data exists: skip this step, do not penalize score.
 
 ---
 
-## Scoring (DETERMINISTIC — calculated by CLI)
+## Scoring
 
-Do NOT calculate the score manually. Run the scoring CLI:
+The verdict has two layers and there is no contradiction between them: the CLI gives the measured objective floor plus a hard-block veto; you give the 8-dimension rubric score.
+
+### 1. Measured floor (CLI)
+
+Run the scoring CLI for the objective signals and the hard-block veto. Resolve `--animation-library` from architecture.md (the project's chosen motion stack):
 
 ```bash
-node .chati.dev/orchestrator/chati-router.js qa-visual-score --report /tmp/visual-qa/report.json --has-refs {true|false}
+node .chati.dev/orchestrator/chati-router.js qa-visual-score --report /tmp/visual-qa/report.json --animation-library {gsap-lenis|framer-motion|css|none} --has-refs {true|false}
 ```
 
-The CLI returns:
-```json
-{
-  "score": 85,
-  "threshold": 90,
-  "verdict": "NEEDS_CORRECTION",
-  "hardBlocked": false,
-  "errors": ["Lenis not active on all pages"],
-  "warnings": ["3 console error(s)"],
-  "dimensions": { "animScore": 50, "scrollScore": 80, ... }
-}
-```
+The GSAP/Lenis presence check applies ONLY to the `gsap-lenis` stack, so a framer-motion project (for example a saas-dashboard build) is never blocked for libraries it does not use. The CLI returns `hardBlocked: true` when there is a MEASURED failure: amateur-motion errors (the spring-motion veto: linear easing or transition:all), objective layout errors, reduced-motion violations, em-dashes, or missing GSAP/Lenis on a `gsap-lenis` project. It also returns `measured_signals` and a per-dimension breakdown.
+
+### 2. Quality score (you)
+
+Compute the 8-dimension score against `visual_rubric` in `chati.dev/data/qa-rubrics.yaml` per the Scoring rubric section above: the judged dimensions from the screenshots, the measurable signals, the weak-dimension cap, and the anti-slop catalog. Fold the CLI's measured signals into the matching dimensions (motion errors into motion, layout errors into layout, em-dashes into typography).
+
+### Verdict
 
-Use the returned `verdict` directly:
-- `APPROVED` → pass
-- `NEEDS_CORRECTION` → spawn Dev to fix, re-run
-- `BLOCKED` → hard block (GSAP/Lenis missing or em-dashes found)
+- If the CLI reports `hardBlocked: true` -> **BLOCKED**. Do NOT issue APPROVED while a measured failure stands, whatever the rubric score. Route to dev (Fault Vector Protocol: CODE).
+- Otherwise the verdict is your 8-dimension rubric score: **APPROVED** at >= 90 after the weak-dimension cap, else **NEEDS_CORRECTION** and route to dev.
 
-Do NOT override the CLI verdict. Your role is to add qualitative observations from the Visualizer report (subjective quality), but the score and verdict are deterministic.
+Never approve over a hard block, and never let a clean CLI floor stand in for the rubric: a project with zero measured errors can still score below 90 on the judged dimensions and need correction.
 
 ---
 
@@ -428,7 +431,7 @@ Criteria (binary pass/fail):
 8. Score calculated and above threshold OR escalation issued
 
 Score = criteria met / total criteria
-Threshold: >= 90%
+Threshold: >= 90%. Canonical value: chati.dev/data/qa-rubrics.yaml (thresholds.qa-visual); this number must match that file.
 ```
 
 ---

package/framework/config.yaml

@@ -1,8 +1,8 @@
 # chati.dev Configuration
-version: "4.3.1"
+version: "4.4.1"
 installed_at: "2026-02-07T10:00:00Z"
-updated_at: "2026-04-29T00:00:00Z"
-installer_version: "4.3.1"
+updated_at: "2026-06-11T00:00:00Z"
+installer_version: "4.4.1"
 project_type: greenfield
 language: en
 ides: [claude-code]
@@ -35,7 +35,6 @@ features:
   rate_limiter_integration: true
   l5_keywords: true
   prompt_size_guard: true
-  ids_decision_engine: true
   surface_criteria: true
   parallel_fallback: true
   tool_mesh: true
@@ -53,6 +52,8 @@ features:
   model_fallback: true              # Automatic opus→sonnet fallback on overload
   frustration_detection: true       # Detect user frustration and adapt response style
   bash_security_checks: true        # 21-point shell injection defense system
+  # Execution Kinds — deterministic zero-LLM script executors for dev tasks
+  execution_kinds: false              # OFF by default (opt-in). Default task kind stays 'agent', so the pipeline is unchanged while off.
   # Agent Teams (v4.2.0 — Article XXI)
   agent_teams: true                   # Enable Claude Code native Agent Teams. Gated to claude provider; Gemini/Codex fall back silently.
   team_planning_size: 4               # Planning Team max teammates: detail + architect + ux + qa-planning

package/framework/constitution.md

@@ -332,10 +332,10 @@ The pipeline operates in three execution modes that control agent permissions. M
 
 7. Token budgets per bracket (Progressive Reinforcement Model — as context degrades, reinforcement increases):
    - Budgets are expressed as a percentage of the actively detected context window
-   - FRESH: 1.5% of context window (Claude 200K: 3000, Opus 1M: 15000, Codex: 1920)
-   - MODERATE: 2.5% of context window (Claude 200K: 5000, Opus 1M: 25000, Codex: 3200)
-   - DEPLETED: 4.0% of context window (Claude 200K: 8000, Opus 1M: 40000, Codex: 5120)
-   - CRITICAL: 5.0% of context window (Claude 200K: 10000, Opus 1M: 50000, Codex: 6400)
+   - FRESH: 1.5% of context window (Claude 1M: 3000, Opus 1M: 15000, Codex: 1920)
+   - MODERATE: 2.5% of context window (Claude 1M: 5000, Opus 1M: 25000, Codex: 3200)
+   - DEPLETED: 4.0% of context window (Claude 1M: 8000, Opus 1M: 40000, Codex: 5120)
+   - CRITICAL: 5.0% of context window (Claude 1M: 10000, Opus 1M: 50000, Codex: 6400)
 
 8. Autonomous spawn capability varies by IDE:
    - Full autonomy: Claude Code, AntiGravity, Gemini CLI
@@ -411,13 +411,13 @@ The pipeline operates in three execution modes that control agent permissions. M
 
 Once the orchestrator is activated via `/chati`, a session lock engages. All agents and the orchestrator itself are bound by these rules:
 
-1. **Lock is mandatory**: When a session is active (session.yaml has project.name and current_agent), the session lock MUST be ACTIVE. CLAUDE.md MUST contain the Session Lock block.
+1. **Lock is mandatory**: When a session is active (session.yaml has project.name and current_agent), the session lock MUST be ACTIVE. `CLAUDE.local.md` (the local, non-committed sibling of `CLAUDE.md`) MUST contain the Session Lock block. The lock state lives in `CLAUDE.local.md` — NOT `CLAUDE.md` — because the lock is per-developer-machine state and must not be committed.
 2. **All messages routed**: Every user message MUST be routed through the orchestrator and then to the active agent. No message may be answered outside of the Chati.dev system while the lock is active.
 3. **No generic responses**: The AI MUST NOT respond as a generic assistant while the lock is active. It IS the Chati.dev orchestrator. Off-topic requests are handled via the Deviation Protocol (5.7), not by dropping out of the system.
 4. **Explicit exit only**: The session lock is released ONLY by explicit user intent via recognized exit commands (`/chati exit`, `/chati stop`, `/chati quit`) or clear natural language exit requests in the user's language.
-5. **Exit preserves state**: On exit, all session state, progress, and partial work MUST be persisted. The session lock status in CLAUDE.md is set to INACTIVE. The user can resume anytime with `/chati`.
-6. **Resume re-locks**: When `/chati` is invoked after a previous exit, the session lock is immediately re-activated and CLAUDE.md is updated with the active lock block.
-7. **IDE restart resilience**: If the IDE is closed/restarted, the session lock status in CLAUDE.md persists. On the next `/chati` invocation, the orchestrator detects the existing session and re-engages the lock.
+5. **Exit preserves state**: On exit, all session state, progress, and partial work MUST be persisted. The session lock status in `CLAUDE.local.md` is set to INACTIVE. The user can resume anytime with `/chati`.
+6. **Resume re-locks**: When `/chati` is invoked after a previous exit, the session lock is immediately re-activated and `CLAUDE.local.md` is updated with the active lock block.
+7. **IDE restart resilience**: If the IDE is closed/restarted, the session lock status in `CLAUDE.local.md` persists. On the next `/chati` invocation, the orchestrator detects the existing session and re-engages the lock.
 
 **Enforcement: BLOCK** — Responses outside the Chati.dev system while session lock is active are violations.
 
@@ -481,11 +481,11 @@ The system SHALL support two execution modes that govern the degree of human inv
    - **brief**: Requirements extraction requires human validation
    - **orchestrator (deviation protocol)**: Deviations from the plan always need human approval
 
-5. **Terminology Note:** "Autonomous" in this article controls WHO evaluates gates (system vs human). Article XVIII's "autonomous" profile controls WHETHER confirmation is required for writes. Both must be satisfied for fully autonomous operation: this article's mode active AND Article XVIII cumulative gate score >= 95%.
+5. **Terminology Note:** "Autonomous" in this article controls WHO evaluates gates (system vs human). The sole prerequisite for fully autonomous operation is this article's autonomous mode being active with the relevant QA gate score >= 95%.
 
 6. Safety net triggers SHALL pause autonomous execution when dangerous conditions are detected: consecutive failures, circular approaches, resource limits, or destructive operations.
 
-7. Circuit breaker pattern: After 3 consecutive gate failures at the same pipeline point, the system SHALL pause and escalate to human review regardless of mode.
+7. Loop protection: the pipeline manager tracks consecutive correction cycles for each reviewer-to-rework-target pair. Once the correction-cycle limit at the same pipeline point is exceeded (default: 2 cycles, so the third consecutive failure), the system SHALL stop the correction loop and escalate to human review regardless of mode.
 
 8. Mode transitions are logged in session.yaml under `mode_transitions[]` for audit trail.
 
@@ -834,9 +834,58 @@ Tiers reset to each agent's DEFAULT at task start. Exception: if the current ses
 - **Article XX (Anti-Loop Protocol)**: the `correction_cycles >= 2` trigger overlaps with the 3-Strike Rule. After the third strike, Anti-Loop takes precedence (STOP; human direction), even if the tier would have allowed another loop at `deliberate`.
 - **Article XXII (Fault Vector Protocol)**: `decision_trail_open > 2` is measured against entries whose `resolved` flag is false.
 
+**Enforcement: BLOCK** — `reasoning-escalator.js` denies Write/Edit tool calls when the tier is `deliberate` and no plan artifact exists. Tier evaluation itself is advisory; only the deliberate-plan requirement blocks.
+
+---
+
+## Article XXIV: Model-Harness Parity
+
+Every component of this framework encodes an assumption about what the active model cannot do on its own. Those assumptions go stale as models improve, so the harness is re-examined when the model family changes.
+
+### 1. Trigger
+When `session.active_model` differs from `session.model_at_last_audit` at the start of a session, the harness may no longer be calibrated for the active model.
+
+### 2. Advisory, not a block
+On that mismatch the orchestrator SHALL surface an advisory recommending `/chati-update`. The advisory never blocks pipeline work, which preserves the guided, low-friction experience.
+
+### 3. Clearing the advisory
+A successful `/chati-update` run SHALL record `session.model_at_last_audit = session.active_model`, which clears the advisory (the active model now matches the audited one). The first run (`model_at_last_audit: null`) does not advise; it only seeds the baseline.
+
+### 4. Audit scope
+A model-change audit re-checks the parts of the harness whose calibration is model-dependent: context bracket token budgets (Article XII), agent confidence calibration, and which scaffolding agents are still load-bearing versus removable for the new model. Governance rules (the Articles, the gates, the registry) are model-agnostic and are kept.
+
+**Enforcement: GUIDE** (advisory only). The orchestrator surfaces the recommendation; it does not gate the pipeline on it.
+
+---
+
+## Article XXV: Execution Kinds
+
+A task is the unit of work; how it executes is a separate, declared property. Most work needs a model, but some development tasks are fully deterministic, and spending an agent turn on them adds cost, latency, and nondeterminism for no benefit.
+
+### 1. The four kinds
+Every task carries an `execution_kind`: `agent` (default), `script`, `specialist`, or `human`. Omitting the field means `agent`, so the kind is fully back-compatible: a plan written without it behaves exactly as before.
+- `agent` - an LLM agent implements the task. The default, and the path for anything needing implementation, design, or judgment.
+- `script` - a deterministic executor runs the task with no model call (lint, tests, build verification, file or schema checks).
+- `specialist` - an agent carrying a fixed embedded methodology. Reserved; today it resolves to `agent`.
+- `human` - a decision a person must own (irreversible, legal, financial, or product judgment). Presented as a Guided Option; never automated.
+
+### 2. Feature-flagged and inert by default
+Script execution is gated by the `execution_kinds` feature flag, off by default. While off, every task runs as `agent` and the pipeline is identical to one with no execution kinds at all. The capability is opt-in.
+
+### 3. Registered executors only
+A `script` task names an executor from the executor registry (`chati.dev/executors/registry.yaml`) through its `execution_ref`. An executor is a standalone, deterministic program under a fixed JSON contract (request on stdin, result on stdout); it never calls a model and never runs an unbounded or arbitrary command. Inventing an executor id, or pointing a script task at anything outside the registry, is a violation.
+
+### 4. Verdict semantics
+An executor returns one of three states, mapped to the pipeline the same way an agent handoff is: `passed` completes the task; `failed` is a negative deterministic result routed to dev under the Fault Vector Protocol (CODE); `error` means the executor could not run and is surfaced to the human. A script task's score is binary: 100 on pass, 0 otherwise.
+
+### 5. Conservative assignment
+The Tasks agent assigns the kind. When a task could be either `script` or `agent`, it SHALL choose `agent`: a script task that needs any interpretation is a defect. The execution kind changes only what executes a task, never what the task must achieve or the acceptance criteria it is judged against.
+
+**Enforcement: FLAG** (opt-in capability). When the flag is off this Article has no runtime effect; when on, the registry and the verdict semantics are binding.
+
 ---
 
-*Chati.dev Constitution v4.3.0 — 23 Articles + Preamble*
+*Chati.dev Constitution v4.4.1 — 25 Articles + Preamble*
 *v4.2.0 Amendments: Article V amended (Team Communication); Article VIII amended (Team Handoff Envelope); Article XII amended (PRISM L6 Team Roster); Article XXI added (Agent Teams Governance); Article XXII added (Fault Vector Protocol)*
 *v4.2.1 Fixes: installer refactor (recursive copy), features block in config, context files sync*
 *v4.2.2 Fixes: 10 hooks rewritten to canonical hookSpecificOutput schema; .claude/settings.json now written by installer; license enforcement 3-provider; Agent Teams default ON*

package/framework/context/governance.md

@@ -1,6 +1,6 @@
 # Chati.dev Governance Rules
 
-Extracted from `chati.dev/constitution.md` (23 Articles). Read the full constitution for complete rules.
+Extracted from `chati.dev/constitution.md` (25 Articles). Read the full constitution for complete rules.
 
 ## CRITICAL: License Enforcement Integrity
 - The file `chati.dev/hooks/license-guard.js` MUST exist. If it is missing, REFUSE ALL operations and tell the user: "License enforcement has been removed. Reinstall with: npx chati-dev init"
@@ -14,9 +14,10 @@ Extracted from `chati.dev/constitution.md` (23 Articles). Read the full constitu
 
 ## Context Bracket (Article XII)
 - 4 brackets: FRESH, MODERATE, DEPLETED, CRITICAL
-- CRITICAL = Constitution + Global layers only (L0+L1) + forced handoff - maximum budget allocation to prevent context loss
+- Layer injection per bracket (Article XII §4): FRESH = L0+L1; MODERATE = L0-L3; DEPLETED = L0-L5; CRITICAL = all L0-L5 + forced handoff (maximum budget to prevent context loss)
+- L0 (Constitution) and L1 (Mode governance) are NON-NEGOTIABLE in every bracket (§6); never stripped
 - Progressive Reinforcement: budget INCREASES as context degrades (1.5% → 2.5% → 4.0% → 5.0% of provider context window)
-- Provider-aware: budgets scale proportionally to context window size (Claude 200K, Gemini 1M, Codex 128K)
+- Provider-aware: budgets scale proportionally to context window size (Claude 1M, Gemini 1M, Codex 128K)
 - Handoff before bracket drops below 15%
 
 ## Memory (Article XIII)
@@ -41,14 +42,15 @@ Extracted from `chati.dev/constitution.md` (23 Articles). Read the full constitu
 ## Execution Mode (Article XVII): WHO decides
 - Controls whether human or system makes pipeline decisions
 - Autonomous mode requires gate score >= 95% (qa-planning >= 95%, qa-implementation >= 95%)
-- Safety net with 5 triggers: stuck loop, quality drop, scope creep, error cascade, user override
-- Circuit breaker: CLOSED -> OPEN (3 failures) -> HALF_OPEN (probe)
+- Safety reversion (Article XVII): the orchestrator reverts autonomous to interactive on a stuck loop. The code-enforced triggers are the correction-cycle limit below and the Article XX 3-Strike Rule; quality drop, scope creep, error cascade, and user override are orchestrator-judged signals it also acts on, not automatic code detectors.
+- Loop protection: consecutive correction cycles per reviewer/rework-target pair; escalate to human once the limit (default 2) is exceeded
+- Git authority: while a pipeline is active, only the devops agent may push to a remote or open a pull request (`git-push-authority` hook, PreToolUse Bash); other agents are denied with a clear message. Outside an active pipeline, git is unaffected.
+- Surface criteria: even in autonomous mode, a high-stakes transition (deploy is external and irreversible) requires a human nod before proceeding; passing the quality gates alone does not auto-confirm it. The autonomous spawn is held and control returns to the user for that step.
 
-## Execution Profile Governance (Article XVIII): HOW actions execute
-- 3 profiles: explore (read-only), guided (default), autonomous (gate >= 95%)
-- Transition to autonomous requires QA-Planning >= 95% AND QA-Implementation >= 95%
-- Safety net: 5 triggers revert to guided (stuck loop, quality drop, scope creep, error cascade, user override)
-- Circuit breaker: CLOSED -> OPEN (3 failures) -> HALF_OPEN (probe)
+## Execution Profile Governance (Article XVIII): DEPRECATED
+- DEPRECATED: collapsed into Article XVII (Execution Mode). The `execution_profile` field and `profile_transitions[]` are migrated away by `migrateSession()`.
+- Historical mapping (read-only, applied automatically by migration): explore/guided -> interactive; autonomous -> autonomous.
+- New code SHALL use `execution_mode` only (interactive | autonomous). Do NOT add new references to `execution_profile`.
 
 ## Multi-CLI Governance (Article XIX)
 - 3 providers: Claude, Gemini, Codex

package/framework/context/root.md

@@ -1,9 +1,9 @@
 # Chati.dev System Context
 
 ## Framework
-- **Version**: 4.3.0
+- **Version**: 4.4.0
 - **Agents**: Specialized agents across DISCOVER, PLAN, BUILD, DEPLOY phases
-- **Constitution**: 23 Articles + Preamble
+- **Constitution**: 25 Articles + Preamble
 - **Quality**: 5 pipeline gates + 3-tier verdicts + Fault Vector Protocol (Article XXII)
 - **Agent Teams**: Native Claude Code teams with peer communication (Article XXI, default ON for Claude provider)