@leing2021/super-pi 0.23.0 → 0.23.2

package/README.md

@@ -10,22 +10,6 @@ pi install npm:@leing2021/super-pi
 
 ---
 
-## Why Super Pi
-
-Using a bare AI agent to write code has three common failure modes:
-
-1. **Builds before thinking** — you finish and realize it's not what you wanted
-2. **Loses context on interruption** — close the terminal, lose the progress
-3. **Repeats the same mistakes** — every session starts from zero
-
-Super Pi's answers:
-
-- **Forces clarity before action** — not a form to fill out, but an AI that interrogates you like a YC partner demanding specific evidence
-- **Auto-resume from checkpoints** — restart after interruption, skip completed work, continue from where you left off
-- **Auto-compounds experience** — every solved problem becomes a searchable knowledge card; next time, the agent finds and reuses it
-
----
-
 ## The Five-Step Loop
 
 ```
@@ -33,360 +17,109 @@ Super Pi's answers:
     think          plan      build      review      learn
 ```
 
-Each step has a dedicated skill + tool pair. Not just prompts — structured toolchains.
+| Skill | Does | Core Tool |
+|-------|------|-----------|
+| **01-brainstorm** | YC-style interrogation, three modes (Startup/Builder/CE) | `brainstorm_dialog` |
+| **02-plan** | RED→GREEN→REFACTOR, incremental updates, optional CEO Review | `plan_diff` |
+| **03-work** | Parallel execution, checkpoint resume, strict TDD | `ce_subagent`, `ce_parallel_subagent` |
+| **04-review** | Auto-assigned reviewers, structured findings, browser QA | `review_router` |
+| **05-learn** | Pattern extraction → searchable knowledge cards | `pattern_extractor` |
+| **06-next** | Next-step recommendation + full status report | `workflow_state` |
+| **07-worktree** | Isolated git worktree development | `worktree_manager` |
+| **08-help** | Phase 1 skill explainer and usage guide | — |
 
-### New: Stage model routing
+### Model & Thinking Routing
 
-Configure once in `.pi/settings.json`:
+Configure in `.pi/settings.json`:
 
 ```json
 {
   "modelStrategy": {
-    "01-brainstorm": "claude-sonnet-4-20250514",
-    "02-plan": "claude-opus-4-20250115",
-    "03-work": "claude-sonnet-4-20250514",
-    "04-review": "claude-sonnet-4-20250514",
-    "05-learn": "claude-haiku-4-20250414",
-    "default": "claude-sonnet-4-20250514"
+    "01-brainstorm": "anthropic/claude-sonnet-4-20250514",
+    "02-plan": "anthropic/claude-opus-4-20250115"
+  },
+  "thinkingStrategy": {
+    "01-brainstorm": "high",
+    "02-plan": "high",
+    "03-work": "medium"
   }
 }
 ```
 
-How it works:
-- Model switching is handled automatically by the ce-core extension `input` hook — no manual `/model` needed.
-- When you type `/skill:01-brainstorm` through `/skill:05-learn`, the extension reads `modelStrategy[stage]` (or `modelStrategy.default`) and switches before the skill runs.
-- Supported formats: full reference (`"anthropic/claude-opus-4-1"`) or bare model id (`"claude-opus-4-1"`, reuses current provider).
-- Every stage prints a `📊 Pipeline Status` block with `Current / Output / Next`.
-- A `Switched model for <stage>: <provider>/<model>` notification appears when the model changes.
-
-Quick example:
-1. Run `/skill:01-brainstorm` — model auto-switches to the configured brainstorm model
-2. Approve the design
-3. Run `/skill:02-plan` — model auto-switches to the configured plan model
-4. Continue through each stage — model switches automatically at each step
-
----
-
-## What Each Step Does
-
-### 01-brainstorm: Think First
-
-Not "describe your requirements." Three modes for three scenarios:
-
-| Mode | For | What it does |
-|------|-----|-------------|
-| **Startup Diagnostic** | Startup ideas, new products | Six YC-style forcing questions, pushed until you produce specific evidence (not "people are interested" — "who would freak out if this disappeared tomorrow?") |
-| **Builder Mode** | Side projects, hackathons | Only focused on building the coolest thing. If you accidentally mention revenue, it auto-upgrades to Startup Diagnostic |
-| **CE Brainstorm** | Adding features to existing projects | Multi-round dialog to clarify scope boundaries, generates a structured requirements doc |
-
-All three modes run a **premise challenge** (are your assumptions valid?) and **alternatives generation** (at least one minimal + one ideal approach) before you're allowed to move on.
-
-### 02-plan: Plan Well
-
-Breaks requirements into implementation units, each following strict **RED → GREEN → REFACTOR** (no production code without a failing test first).
-
-**Incremental updates**: Requirements changed? `plan_diff` detects added/removed/modified units and patches the plan instead of rewriting it.
-
-**CEO Review (optional)**: After planning, you can request a CEO Review. It challenges your plan using Bezos reversible-decision frameworks, Munger inversion, Jobs subtraction, forces alternative approaches, and draws error maps. Like having a demanding CTO review your proposal for free.
-
-### 03-work: Build Right
-
-**Parallel execution**: `task_splitter` uses a Union-Find algorithm to analyze file dependencies, feeds conflict-free units to `ce_parallel_subagent` for concurrent execution.
-
-**Checkpoint resume**: After each unit, a checkpoint is saved. Interrupted? Next startup auto-loads, skips completed work, continues from the breakpoint. Failed? `fail` records the error → `retry` suggests a recovery strategy (timeout? extend timeout. Permission issue? check permissions first. Code error? fix then retry).
-
-**Strict TDD**: Run failing test → write minimal implementation → test passes → refactor. Every step requires command output as evidence. No skipping.
-
-### 04-review: Review Thoroughly
-
-**Structured code review**: `review_router` auto-assigns reviewer personas based on diff metadata (changed payment code? bring in the security reviewer). Review discipline is technical evaluation, not theater — every finding must cite specific code, YAGNI checks, no performative agreement.
-
-**Browser QA (optional)**: Uses `agent-browser` to open your app, click through pages, screenshot bugs, fix by severity, up to 3 auto-fix iterations. Can auto-generate regression tests. Like having a QA engineer run acceptance tests.
-
-### 05-learn: Compound Learnings
-
-`pattern_extractor` scans existing artifacts, extracts and categorizes patterns. Turns "the pitfall we hit this time" into a YAML-tagged solution card in `docs/solutions/`.
-
-Two-level storage: project-specific → inside the project; cross-project → `~/.pi/agent/docs/solutions/` globally searchable.
-
-Next time `02-plan` or `04-review` runs, a grep-first search strategy automatically retrieves relevant past experience.
-
----
-
-## Technical Architecture
-
-### 8 Skills (workflow nodes)
-
-| Skill | One-liner | Core Tool |
-|-------|-----------|-----------|
-| `01-brainstorm` | Deep requirements mining in three modes | `brainstorm_dialog` |
-| `02-plan` | Break into units, TDD gates, incremental updates | `plan_diff` |
-| `03-work` | Parallel execution, checkpoint resume, error recovery | `session_checkpoint`, `task_splitter`, `ce_parallel_subagent` |
-| `04-review` | Persona-routed review + live browser testing | `review_router` |
-| `05-learn` | Pattern extraction → knowledge card compounding | `pattern_extractor` |
-| `06-next` | What to do next + full status report | `workflow_state`, `session_history` |
-| `07-worktree` | Git worktree isolated development | `worktree_manager` |
-| `08-help` | Usage guide | — |
-
-
-### 14 Tools + 2 Helpers (underlying capabilities)
-
-| Tool | What it does |
-|------|-------------|
-| `task_splitter` | Union-Find algorithm analyzes file dependencies, auto-groups parallel-safe units |
-| `session_checkpoint` | JSON-persisted checkpoints with save/load/fail/retry operations |
-| `plan_diff` | Incremental plans: compare detects diffs, patch applies changes |
-| `ce_parallel_subagent` | `Promise.allSettled`-style CE parallel skill-based subagent execution with context slimming |
-| `review_router` | Auto-assign reviewer personas from diff metadata |
-| `pattern_extractor` | Extract and categorize patterns from artifacts |
-| `brainstorm_dialog` | Multi-round dialog state machine (start → refine × N → summarize) |
-| `session_history` | Cross-session execution history recording and querying |
-| `workflow_state` | Scan docs/ and .context/ to summarize workflow state |
-| `worktree_manager` | Full git worktree lifecycle management |
-| `artifact_helper` | Artifact path resolution and directory creation |
-| `ask_user_question` | Structured user prompts (choices / free input) |
-| `ce_subagent` | CE serial skill-based subagent chain with depth guard and context control |
-| `context_handoff` | Cross-stage context handoffs with evidence-first templates (save/load/latest/status) |
-| `subagent-depth-guard` | Helper: env-based recursion depth tracking (prevents runaway nesting) |
-| `async-mutex` | Helper: serializes `process.env` mutation for concurrency-safe child process spawning |
-
-### Compatibility with pi-subagents
+Model and thinking level switch automatically — no manual `/model` needed.
 
-super-pi's CE skill-router tools use a dedicated namespace to avoid runtime tool-name conflicts:
+### pi-subagents Compatibility
 
-- **CE tools**: `ce_subagent` and `ce_parallel_subagent` — for CE pipeline skill execution
-- **Generic `subagent`**: reserved for third-party agent extensions like [pi-subagents](https://www.npmjs.com/package/pi-subagents)
-
-Both extensions can be installed simultaneously without tool collision. super-pi manages the CE pipeline (brainstorm → plan → work → review → learn); pi-subagents provides general-purpose multi-agent execution. No configuration needed — install both and each keeps its own tool namespace.
-
----
-
-## Token Cost
-
-New conversation overhead: **~2,600 tokens** (1.3% of Claude Sonnet 4's 200K context).
-
-| Component | Tokens | When loaded |
-|-----------|--------|-------------|
-| 8 skill registrations | ~490 | Every conversation (fixed) |
-| 14 tool registrations | ~2,055 | Every conversation (fixed) |
-| Hooks & filters | 0 | Runtime interception, zero prompt cost |
-| Single skill trigger | ~1,000–4,000 | On-demand via `read` |
-| Rules minimal (2 files) | ~900 | Before plan/work |
-| Rules + language (7 files) | ~2,600 | Before work with specific language |
-
-| vs bare Pi | vs global rules injection | vs super-pi |
-|-----------|----------------------|------------|
-| No rules | All rules loaded upfront | Progressive on-demand |
-| No output filtering | No output filtering | Auto-compress (bash ~65–98%, read ~30–60%) |
-| No TDD gate | No TDD gate | Hard gate prevents rework |
-| 0 tokens | ~5,000–36,000 tokens | **~2,600 tokens** |
-
-Single `npm install` output filtered once pays for the entire overhead. Full evaluation → [`docs/token-cost-evaluation.md`](docs/token-cost-evaluation.md)
-
----
-
-## Code Scale
-
-~2800 lines of TypeScript implementing 14 registered tools + 2 helpers, 23 Markdown reference files + 79 rule files driving 8 skills, 175 tests covering all tool logic.
-
-Not a heavy framework. Each tool has a single responsibility, each skill works independently, and together they form a complete workflow.
+CE skill tools use a dedicated namespace (`ce_subagent`, `ce_parallel_subagent`) to avoid conflicts with third-party extensions like [pi-subagents](https://www.npmjs.com/package/pi-subagents). Both can coexist without configuration.
 
 ---
 
 ## Quick Start
 
-### New idea
-
 ```
 You: I want to build a tool that helps indie devs find users
 
-→ Auto-enters 01-brainstorm, YC-style interrogation
-→ Generates docs/brainstorms/2026-04-18-find-users-requirements.md
-→ Recommends next: 02-plan
+→ 01-brainstorm: YC-style interrogation → docs/brainstorms/requirements.md
+→ 02-plan: RED→GREEN→REFACTOR units → docs/plans/plan.md
+→ 03-work: parallel execution, checkpoint resume
+→ 04-review: structured findings, optional browser QA
+→ 05-learn: knowledge compounding
 
 You: continue
-
-→ 02-plan breaks into units, optional CEO Review
-→ Generates docs/plans/2026-04-18-find-users-plan.md
-
-You: continue
-
-→ 03-work parallel execution, checkpoint resume
-→ 04-review code review + optional browser QA
-→ 05-learn knowledge compounding
+→ Next skill recommended automatically via /skill:06-next
 ```
 
-### Adding a feature
-
+**After interruption:**
 ```
-You: I want to add user authentication to the project
-
-→ 01-brainstorm CE mode, multi-round dialog: OAuth2? JWT? MFA?
-→ Requirements doc → 02-plan → 03-work → 04-review → 05-learn
-```
-
-### Resume after interruption
-
-```
-You: /skill:03-work docs/plans/auth-plan.md
-
+You: /skill:03-work docs/plans/plan.md
 → Auto-loads checkpoint, skips completed units, resumes from breakpoint
 ```
 
-### Requirements changed
+---
 
-```
-You: Requirements changed, need to add SSO support
+## Token Cost
 
-→ 02-plan uses plan_diff to detect changes, patches incrementally
-```
+New conversation overhead: **~2,600 tokens** (1.3% of 200K context).
 
-### Check progress anytime
+| Component | Tokens |
+|-----------|--------|
+| 8 skill registrations | ~490 |
+| System prompt (skills) | ~1,400 |
+| Skill inlining (per invocation) | ~500-800 |
 
-```
-You: /skill:06-next
+Progressive loading: only needed skills loaded on-demand.
 
-→ Scans all artifacts, shows progress + recommends next step
-```
+Full evaluation → [`docs/token-cost-evaluation.md`](docs/token-cost-evaluation.md)
 
 ---
 
-## Generated File Structure
+## Generated Structure
 
 ```
 your-project/
 ├── docs/
-│   ├── brainstorms/                  # Requirements (01-brainstorm)
-│   ├── plans/                        # Execution plans (02-plan)
-│   └── solutions/                    # Knowledge cards (05-learn)
+│   ├── brainstorms/      # Requirements
+│   ├── plans/             # Execution plans
+│   └── solutions/          # Knowledge cards
 └── .context/
     └── compound-engineering/
-        ├── checkpoints/              # Breakpoint files (session_checkpoint)
-        ├── dialogs/                  # Dialog state (brainstorm_dialog)
-        └── history/                  # Execution history (session_history)
-```
-
-**Recommendation: commit everything to git** — these files are the project's traceable memory.
-
-### Progressive Rule Loading
-
-Built-in coding rules live under `rules/` in the package. Rules are loaded **progressively** by each skill — never all at once, only what the current task needs.
-
-**How it works:**
-
-```
-system prompt (30 tokens: skill name + description)
-  → skill SKILL.md (~200 tokens: loading decision tree)
-    → specific rule files via read tool (on-demand, 900–2600 tokens)
+        ├── checkpoints/   # Breakpoint files
+        ├── dialogs/       # Dialog state
+        └── history/       # Execution history
 ```
 
-Three skills auto-trigger rule loading at their entry points:
-
-| Skill | Rules pre-loaded |
-|-------|-----------------|
-| `02-plan` | `common/` rules + language detection + matching language rules (e.g. `rules/typescript/`) |
-| `03-work` | `common/` rules + language detection + matching language rules + `web/` if frontend |
-| `04-review` | `common/code-review.md` + language detection + matching language rules + `web/` if frontend |
-
-Language detection uses file heuristics (see `skills/references/language-detection.md`): `tsconfig.json` → TypeScript, `Cargo.toml` → Rust, `go.mod` → Go, etc.
-
-**Rule precedence** (when layers overlap on the same topic):
-
-```
-language-specific  >  web  >  common
-```
-
-No rules are loaded when you brainstorm, check status, or do non-code tasks. Zero waste.
-
-#### Included rule layers
-
-| Layer | Files | When loaded |
-|-------|-------|------------|
-| `common/` | 11 files (includes `naming.md`) | Always (baseline for all tasks) |
-| `typescript/`, `python/`, `cpp/`, `csharp/`, `dart/`, `golang/`, `java/`, `kotlin/`, `perl/`, `php/`, `rust/`, `swift/` | 5 files each | When the task touches that language |
-| `web/` | 7 files (includes `design-quality.md`, `performance.md`) | When frontend/browser is relevant |
-
-#### Customizing rules for your project
-
-Two rule sources exist, with project-level taking priority:
-
-| Source | Location | Survives `pi update`? |
-|--------|----------|----------------------|
-| **Project-level** | `{your-project-root}/rules/` | ✅ Yes |
-| Package-level | Inside `node_modules/@leing2021/super-pi/rules/` | ❌ No |
-
-To customize, create a `rules/` directory in your project root. Skills check it first — if a file exists there, it overrides the package default for that topic.
-
-**Add a language** — create a new directory with the 5 standard topics:
-
-```bash
-mkdir rules/elixir
-touch rules/elixir/{coding-style,testing,patterns,security,hooks}.md
-```
-
-Each file should start with:
-
-```markdown
-> This file extends [common/xxx.md](../common/xxx.md) with Elixir-specific content.
-```
-
-**Remove unused languages** — just delete the directory:
-
-```bash
-rm -rf rules/perl rules/cpp  # don't need these? remove them
-```
-
-**Tweak a rule** — edit the `.md` file directly:
-
-```bash
-# Override testing conventions for your team
-vim rules/common/testing.md
-
-# Override for a specific language
-vim rules/typescript/testing.md
-```
-
-**Add a new topic** — create a new `.md` in the appropriate layer:
-
-```bash
-# Common topic
-vim rules/common/api-design.md
-
-# Language-specific override
-vim rules/python/api-design.md
-```
-
-Skills will pick up any `.md` file in `rules/` — no configuration needed. If a language directory exists, it's available for loading. If it's gone, it's simply never loaded.
+**Commit everything to git** — these files are the project's traceable memory.
 
 ---
 
-## Design Philosophy & Acknowledgements
-
-**80% planning and review, 20% execution.**
-
-The goal isn't making AI write code faster — it's making AI think before writing, review after writing, and compound learnings after reviewing. Speed comes from fewer rewrites, not from skipping steps.
-
-The following projects directly inspired this work:
+## Architecture
 
-- **[everything-claude-code](https://github.com/affaan-m/everything-claude-code)** (162K★) → Parallel subagent orchestration, checkpoint resume, continuous learning
-- **[superpowers](https://github.com/obra/superpowers)** (161K★) → Strict TDD gates, design checklists, review discipline
-- **[gstack](https://github.com/garrytan/gstack)** (78K★) → YC-style forcing questions, CEO Review frameworks, browser QA
-- **[compound-engineering-plugin](https://github.com/EveryInc/compound-engineering-plugin)** (14K★) → Five-step loop, knowledge compounding backbone
+- **8 skills** with dedicated tools
+- **14 tools** + 2 helpers
+- **~2800 lines** TypeScript, **175 tests**
+- **Progressive rule loading** — only what each task needs
 
-Not a fork. Not a wrapper. Methodologies extracted and rebuilt with Pi's native tool + skill system.
-
----
-
-## Best Practices
-
-| Tip | Why |
-|-----|-----|
-| Start with 01-brainstorm | Whatever the scenario, thinking first never hurts |
-| Use 07-worktree for big features | Isolated dev, no impact on main branch |
-| Use CEO Review on plans | Like having a demanding CTO review for free |
-| Use browser QA for acceptance | Code review can't catch layout breaks and blank screens |
-| Don't panic on interruption | Next 03-work auto-resumes from checkpoint |
-| Use 06-next when lost | One glance shows where you are and what to do next |
+Rules in `rules/` (11 common + language-specific). Project-level overrides take priority.
 
 ---
 
@@ -394,8 +127,6 @@ Not a fork. Not a wrapper. Methodologies extracted and rebuilt with Pi's native
 
 See [CHANGELOG.md](./CHANGELOG.md) for full version history.
 
----
-
 ## Repository
 
 - **GitHub**: https://github.com/leing2021/super-pi
@@ -406,4 +137,4 @@ See [CHANGELOG.md](./CHANGELOG.md) for full version history.
 ```bash
 bun test
 npm publish --dry-run
-```
+```

package/extensions/ce-core/index.ts

@@ -31,15 +31,46 @@ const PIPELINE_STAGE_KEYS = new Set([
   "05-learn",
 ])
 
-interface ModelStrategySettings {
+interface StrategySettings {
   modelStrategy?: Record<string, string>
+  thinkingStrategy?: Record<string, string>
 }
 
-async function readProjectSettings(cwd: string): Promise<ModelStrategySettings | null> {
-  const settingsPath = path.join(cwd, ".pi", "settings.json")
+/**
+ * Read settings from two locations:
+ * 1. Project-level: {cwd}/.pi/settings.json (highest priority)
+ * 2. Global-level: ~/.pi/agent/settings.json (fallback)
+ *
+ * Project-level takes precedence; global-level is used as fallback.
+ */
+async function readSettings(cwd: string): Promise<StrategySettings | null> {
+  // Try project-level first
+  const projectPath = path.join(cwd, ".pi", "settings.json")
   try {
-    const content = await readFile(settingsPath, "utf8")
-    return JSON.parse(content) as ModelStrategySettings
+    const content = await readFile(projectPath, "utf8")
+    const projectSettings = JSON.parse(content) as StrategySettings
+    // If project has modelStrategy or thinkingStrategy, use it
+    if (projectSettings.modelStrategy || projectSettings.thinkingStrategy) {
+      return projectSettings
+    }
+  } catch {
+    // Project settings not found, continue to global
+  }
+
+  // Fallback to global-level
+  const globalPath = path.join(process.env.HOME || "~", ".pi", "agent", "settings.json")
+  try {
+    const content = await readFile(globalPath, "utf8")
+    return JSON.parse(content) as StrategySettings
+  } catch {
+    // Global settings not found either
+  }
+
+  // Try ~/.pi/settings.json as another fallback
+  const altGlobalPath = path.join(process.env.HOME || "~", ".pi", "settings.json")
+  try {
+    const content = await readFile(altGlobalPath, "utf8")
+    return JSON.parse(content) as StrategySettings
   } catch {
     return null
   }
@@ -295,6 +326,11 @@ const contextHandoffParams = Type.Object({
   artifacts: Type.Optional(Type.Record(Type.String(), Type.Optional(Type.String()), { description: "Artifact paths (requirements, plan, checkpoint, proof)" })),
   handoffMarkdown: Type.Optional(Type.String({ description: "Custom handoff markdown content" })),
   handoffPath: Type.Optional(Type.String({ description: "Specific handoff file path to load" })),
+  currentTruth: Type.Optional(Type.Array(Type.String(), { description: "Known true statements validated during session" })),
+  invalidatedAssumptions: Type.Optional(Type.Array(Type.String(), { description: "Assumptions proven wrong during session" })),
+  openDecisions: Type.Optional(Type.Array(Type.String(), { description: "Pending decisions that affect next steps" })),
+  recentlyAccessedFiles: Type.Optional(Type.Array(Type.String(), { description: "Files recently read or edited (defaults to activeFiles)" })),
+  compressionRisk: Type.Optional(Type.Array(Type.String(), { description: "Context compression risks to watch for" })),
 })
 
 const patternExtractorParams = Type.Object({
@@ -319,47 +355,61 @@ export default function ceCoreExtension(pi: ExtensionAPI) {
       return { action: "continue" as const }
     }
 
-    const settings = await readProjectSettings(ctx.cwd)
+    const settings = await readSettings(ctx.cwd)
     const modelStrategy = settings?.modelStrategy
-    if (!modelStrategy) {
-      return { action: "continue" as const }
-    }
-
-    const targetModelRef = modelStrategy[stageKey] ?? modelStrategy.default
-    if (!targetModelRef) {
-      return { action: "continue" as const }
-    }
-
-    const parsed = parseModelRef(targetModelRef, ctx.model?.provider)
-    if (!parsed) {
-      if (ctx.hasUI) {
-        ctx.ui.notify(`Invalid modelStrategy entry for ${stageKey}: ${targetModelRef}`, "warning")
-      }
-      return { action: "continue" as const }
-    }
-
-    if (ctx.model?.provider === parsed.provider && ctx.model?.id === parsed.id) {
-      return { action: "continue" as const }
-    }
-
-    const model = ctx.modelRegistry.find(parsed.provider, parsed.id)
-    if (!model) {
-      if (ctx.hasUI) {
-        ctx.ui.notify(`Configured model not found for ${stageKey}: ${targetModelRef}`, "warning")
+    const thinkingStrategy = settings?.thinkingStrategy
+
+    // Model switching
+    if (modelStrategy) {
+      const targetModelRef = modelStrategy[stageKey] ?? modelStrategy.default
+      if (targetModelRef) {
+        const parsed = parseModelRef(targetModelRef, ctx.model?.provider)
+        if (parsed) {
+          // Skip if already using the same model
+          if (ctx.model?.provider !== parsed.provider || ctx.model?.id !== parsed.id) {
+            const model = ctx.modelRegistry.find(parsed.provider, parsed.id)
+            if (model) {
+              const switched = await pi.setModel(model)
+              if (switched) {
+                if (ctx.hasUI) {
+                  ctx.ui.notify(`Switched model for ${stageKey}: ${model.provider}/${model.id}`, "info")
+                }
+              } else {
+                if (ctx.hasUI) {
+                  ctx.ui.notify(`No API key for ${stageKey}: ${model.provider}/${model.id}`, "warning")
+                }
+              }
+            } else if (ctx.hasUI) {
+              ctx.ui.notify(`Model not found for ${stageKey}: ${targetModelRef}`, "warning")
+            }
+          }
+        } else if (ctx.hasUI) {
+          ctx.ui.notify(`Invalid modelStrategy for ${stageKey}: ${targetModelRef}`, "warning")
+        }
       }
-      return { action: "continue" as const }
     }
 
-    const switched = await pi.setModel(model)
-    if (switched) {
-      if (ctx.hasUI) {
-        ctx.ui.notify(`Switched model for ${stageKey}: ${model.provider}/${model.id}`, "info")
+    // Thinking level switching
+    if (thinkingStrategy) {
+      const targetThinking = thinkingStrategy[stageKey] ?? thinkingStrategy.default
+      if (targetThinking) {
+        const levelMap: Record<string, "low" | "medium" | "high"> = {
+          low: "low",
+          medium: "medium",
+          high: "high",
+          "0": "low",
+          "1": "medium",
+          "2": "high",
+        }
+        const normalized = levelMap[targetThinking.toLowerCase()] ?? "medium"
+        const currentLevel = ctx.getThinkingLevel?.() ?? "medium"
+        if (currentLevel !== normalized) {
+          ctx.setThinkingLevel?.(normalized)
+          if (ctx.hasUI) {
+            ctx.ui.notify(`Switched thinking level for ${stageKey}: ${normalized}`, "info")
+          }
+        }
       }
-      return { action: "continue" as const }
-    }
-
-    if (ctx.hasUI) {
-      ctx.ui.notify(`No API key available for configured model: ${model.provider}/${model.id}`, "warning")
     }
 
     return { action: "continue" as const }
@@ -713,6 +763,11 @@ export default function ceCoreExtension(pi: ExtensionAPI) {
         artifacts: params.artifacts as Record<string, string | undefined> | undefined,
         handoffMarkdown: params.handoffMarkdown,
         handoffPath: params.handoffPath,
+        currentTruth: params.currentTruth,
+        invalidatedAssumptions: params.invalidatedAssumptions,
+        openDecisions: params.openDecisions,
+        recentlyAccessedFiles: params.recentlyAccessedFiles,
+        compressionRisk: params.compressionRisk,
       })
 
       return {