docs: replace remaining Mermaid text diagrams with PlantUML PNG images

docs: replace Mermaid text diagram with PlantUML PNG image
docs(superpowers): local-first LLM routing design spec (#197 )
2026-07-23 17:09:16 -05:00 · 2026-07-23 17:03:30 -05:00 · 2026-05-16 16:18:54 -05:00 · 2026-05-14 09:51:03 -05:00 · 2026-05-14 09:50:50 -05:00 · 2026-05-14 09:48:26 -05:00
34 changed files with 2694 additions and 180 deletions
@@ -154,6 +154,34 @@ jobs:
          - Vision
          - Voice
    steps:
+      # Grogu is a shared workstation, not an ephemeral VM. Interactive
+      # Xcode test sessions, debugging, or aborted `swift test` runs can
+      # leave orphan xctest/swift-test/swift-driver processes around for
+      # hours, holding Mach services + sockets + ports that a new
+      # WKWebView mount or test-target launch then blocks on (observed:
+      # AgentCore hung 34min on develop CI 25824607025 with 6 zombie
+      # xctests aged 22-24h from prior interactive use).
+      #
+      # Age filter (>= 1 hour): Grogu now has SIBLING runners
+      # (Grogu-jarvis, -2, -3) so concurrent CI jobs can be in flight
+      # simultaneously. CI jobs complete in <10 min, so a 1-hour cutoff
+      # safely reaps interactive-use orphans without killing concurrent
+      # CI work on sibling runners. `ps -o etime` formats: `MM:SS`
+      # (<1h, skip), `HH:MM:SS` (>=1h, kill), `DD-HH:MM:SS` (days, kill).
+      # github-hosted VMs start clean, so this step is gated to self-hosted.
+      - name: Reap orphan xctest/swift-test processes older than 1h (self-hosted only)
+        if: runner.environment == 'self-hosted'
+        run: |
+          age_filter='/(xctest|swift-test|swift-driver)$/ && !/awk/ && (match($2, /-/) || match($2, /[0-9]+:[0-9]+:[0-9]+/)) {print $1}'
+          /bin/ps -A -o pid,etime,comm \
+            | awk "$age_filter" \
+            | xargs -r -I{} kill -TERM {} 2>/dev/null || true
+          sleep 2
+          /bin/ps -A -o pid,etime,comm \
+            | awk "$age_filter" \
+            | xargs -r -I{} kill -9 {} 2>/dev/null || true
+          echo "Surviving xctest/swift-test processes (any age — includes concurrent CI work on sibling runners):"
+          /bin/ps -A -o pid,etime,comm | grep -E '(xctest|swift-test|swift-driver)$' || echo "  (none)"
      - uses: actions/checkout@v4
      # Project targets macOS 26 Tahoe + Swift 6.2/6.3. Xcode 16 (Swift 6.0)
      # cannot parse `mlx-audio-swift` v0.1.2's `swift-tools-version:6.2`,
@@ -403,3 +431,51 @@ jobs:
          fi
          git tag -a "$TAG" -m "Automated tag from CI for $GH_REF"
          git push origin "$TAG"
+      # GitHub Actions GITHUB_TOKEN cannot trigger downstream workflows. The
+      # separate `release.yml` workflow listens for `push: tags:` events but
+      # never sees a CI-pushed tag — observed on v0.1.0-dev.50 (tag created,
+      # Release object missing, had to be manually created). Create the
+      # Release object directly here in the same job so the Releases page
+      # reflects CI-pushed tags without a manual step. `release.yml` stays
+      # as a fallback for manually-pushed tags (which CAN trigger workflows).
+      - name: Create GitHub Release
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          TAG: ${{ steps.version.outputs.tag }}
+          PRERELEASE: ${{ steps.version.outputs.prerelease }}
+          RUN_URL: ${{ format('https://github.com/{0}/actions/runs/{1}', github.repository, github.run_id) }}
+        run: |
+          set -euo pipefail
+          if ! git rev-parse "refs/tags/$TAG" >/dev/null 2>&1; then
+            echo "Tag $TAG not present locally; nothing to release." >&2
+            exit 0
+          fi
+          if gh release view "$TAG" >/dev/null 2>&1; then
+            echo "Release $TAG already exists — skipping create (idempotent)."
+            exit 0
+          fi
+          prev=$(git tag --list 'v*' --sort=-v:refname | grep -v "^${TAG}$" | head -n 1 || true)
+          notes_file=$(mktemp)
+          {
+            echo "## Changes"
+            echo
+            if [ -n "$prev" ]; then
+              git log --pretty='format:- %s (%h)' "${prev}..${TAG}"
+            else
+              git log --pretty='format:- %s (%h)' -20
+            fi
+            echo
+            echo
+            echo "## Verification"
+            echo
+            echo "- CI build: $RUN_URL"
+            echo "- Boundary gates, webview, app-build, 13/13 SPM packages: all green on this commit"
+          } > "$notes_file"
+          flag=()
+          if [ "$PRERELEASE" = "true" ]; then
+            flag+=("--prerelease")
+          fi
+          gh release create "$TAG" \
+            --title "$TAG" \
+            --notes-file "$notes_file" \
+            "${flag[@]}"
@@ -0,0 +1,62 @@
+# OpenSSF Scorecard — security supply-chain posture for KofTwentyTwo/Jarvis.
+#
+# Issue #178. Scorecard scores 18 security checks (Branch-Protection, Code-Review,
+# Pinned-Dependencies, SAST, Token-Permissions, etc.) and publishes results as
+# SARIF to GitHub code-scanning. A weekly schedule keeps the score fresh; we
+# also re-run on every push to develop so significant posture regressions show
+# up promptly. The badge in README.md links back to the public scorecard page.
+#
+# Permissions are the canonical Scorecard set:
+#   - id-token: write          OIDC for publish-results
+#   - contents: read           read repo tree
+#   - security-events: write   upload SARIF to code-scanning
+#
+# This workflow runs on github-hosted ubuntu-latest. Self-hosted Grogu is not
+# used here -- Scorecard is short, network-bound, and benefits from a clean
+# ephemeral environment.
+#
+# Workflow consumes no untrusted user inputs (no github.event.* interpolation
+# in run: blocks), so it is safe from workflow-injection patterns.
+
+name: OpenSSF Scorecard
+
+on:
+  schedule:
+    # Mondays at 06:00 UTC. Off-peak relative to weekday push spikes.
+    - cron: "0 6 * * 1"
+  push:
+    branches: [develop]
+  workflow_dispatch:
+
+permissions: read-all
+
+jobs:
+  analysis:
+    name: Scorecard analysis
+    runs-on: ubuntu-latest
+    permissions:
+      id-token: write
+      contents: read
+      security-events: write
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+        with:
+          persist-credentials: false
+
+      - name: Run Scorecard
+        uses: ossf/scorecard-action@v2.4.0
+        with:
+          results_file: results.sarif
+          results_format: sarif
+          # publish_results: true uploads anonymized scores to the OpenSSF
+          # public dataset, which is what powers the README badge URL
+          # (api.scorecard.dev). Safe for a public repo; would need
+          # reconsidering if the repo were ever made private.
+          publish_results: true
+
+      - name: Upload SARIF to code-scanning
+        uses: github/codeql-action/upload-sarif@v3
+        with:
+          sarif_file: results.sarif
@@ -740,10 +740,11 @@ final class AppDelegate: NSObject, NSApplicationDelegate {

        // 4.5 Bus wiring — construct the `WebviewBridge` around the HUD's
        //     `WKWebView`, install the `WKUserScript` at document-start in
-        //     the isolated `JarvisBusWorld`, start the single-writer
-        //     `HudStateCoordinator`, and load the R3F bundle's
-        //     `index.html` so the handshake kicks off. See `installBus`
-        //     for the seven-step sequence.
+        //     the page world (the shared `WKContentWorld.page` used by
+        //     Injection.js and the HUD bundle alike since `fb41c5f`),
+        //     start the single-writer `HudStateCoordinator`, and load the
+        //     R3F bundle's `index.html` so the handshake kicks off. See
+        //     `installBus` for the seven-step sequence.
        installBus()

        // 5. Keychain probe. Banner suppression: when the API key is
@@ -64,7 +64,7 @@ The top-level shape is fixed and agreed. Push back with specific reasoning only
  - New tokenizer produces ~35% more tokens than Opus 3.x for the same text. Instrument token budgets accordingly; cap `tool_result` content at 8 KB.
  - Cache TTL default is 5 minutes ephemeral. For 1-hour TTL, pass `ttl: "1h"` explicitly on `cache_control` blocks **AND** include the request header `anthropic-beta: extended-cache-ttl-2025-04-11`. Without the beta header, 1h is silently ignored. Verify via DevOverlay watching `cache_creation_input_tokens` vs `cache_read_input_tokens` per turn.
  - Tool-use and extended thinking both stream; the SSE parser must handle `content_block_start` with `input_json_delta` for tool args. Close on `message_stop`, not `message_delta`; swallow `ping`; route `thinking_delta` to its own case; treat `stop_reason: "refusal"` as first-class; emit `partial_tool_use_at_disconnect` on mid-delta termination.
- **Ollama caveat:** as of April 2026, Qwen 3 / 3.5 / Gemma 4 tool-calling is broken in Ollama — confirmed via live issue tracker (ollama/ollama#14493 Qwen 3.5 27B non-functional; #14601 Qwen3 malformed tool defs via `/api/chat`; #14745 qwen3.5:9b print-not-execute; #15315 Gemma 4 tool parser). Root cause: Ollama's renderer/parser maps Qwen 3.5 through a Hermes-style JSON pipeline, but the model was trained on Qwen3-Coder XML format; unclosed `<think>` tags corrupt multi-turn. **Known-good local tool-calling baseline: `qwen2.5-coder:32b`.** Qwen3 stays disabled as opt-in until upstream closes these issues. Llama 4 with the `llama4_pythonic` parser is worth testing but not yet trusted. Document the local model in use alongside every eval run.
+- **Ollama caveat:** as of April 2026, Qwen 3 / 3.5 / Gemma 4 tool-calling is broken in Ollama — confirmed via live issue tracker (ollama/ollama#14493 Qwen 3.5 27B non-functional; #14601 Qwen3 malformed tool defs via `/api/chat`; #14745 qwen3.5:9b print-not-execute; #15315 Gemma 4 tool parser). Root cause: Ollama's renderer/parser maps Qwen 3.5 through a Hermes-style JSON pipeline, but the model was trained on Qwen3-Coder XML format; unclosed `<think>` tags corrupt multi-turn. **Known-good local tool-calling baseline: `qwen2.5-coder:32b-instruct-q8_0`** (the variant pinned by `ModelID.qwen25coder32b`; vanilla `qwen2.5-coder:32b` is not used). Qwen3 stays disabled as opt-in until upstream closes these issues. Llama 4 with the `llama4_pythonic` parser is worth testing but not yet trusted. Document the local model in use alongside every eval run.
 - **Ollama transport gotcha:** `/api/chat` (native NDJSON) emits `tool_calls` on the chunk **preceding** the `done: true` terminator, not with it. Decoder must read `tool_calls` whenever seen and never gate on `done`. Separate decoders for `/api/chat` vs `/v1/chat/completions` (OpenAI-compat SSE with atomic `tool_calls`).
 - **Tool-choice discipline:** `LLMProvider.stream(..., toolChoice:)` is mandatory. Cap-recovery "one more call" sets `.none` (Anthropic: `{type: "none"}`; Ollama: drop `tools` array entirely). Eval asserts zero `.toolUseRequested` on recovery call.

@@ -87,7 +87,7 @@ The top-level shape is fixed and agreed. Push back with specific reasoning only
 **Memory stack:**
 - **Store:** SQLite with **FTS5** (keyword search) + **sqlite-vec** extension (vector search), single file under `~/Library/Application Support/Jarvis/`.
 - **Embeddings:** `nomic-embed-text` via Ollama (local, free, 768-dim).
- **Extraction:** mem0's `ADD / UPDATE / NOOP` pattern — after each turn, a local extractor (Qwen 2.5-Coder 32B via Ollama) decides what facts to commit/update/ignore. **Fully local; no user data ever leaves the machine for memory.**
+- **Extraction:** mem0's `ADD / UPDATE / NOOP` pattern — after each turn, a local extractor (`qwen3.6:latest` via Ollama, per commit `8d0a959`) decides what facts to commit/update/ignore. **Fully local; no user data ever leaves the machine for memory.**
 - **Temporal validity:** record `valid_from` / `valid_to` on facts (Zep/Graphiti style) so "Sarah works at Acme" can be superseded without deleting history.
 - **Scope tiers:** `UserDefaults` for trivial prefs, JSON for tool configs/feature flags, SQLite for everything conversational.
 - Secrets (API keys) → macOS **Keychain**, never plaintext, never checked in.
@@ -209,6 +209,7 @@ bash scripts/check-corpus-secrets.sh
 bash scripts/check-embedding-dim-literal.sh
 bash scripts/check-install-order.sh
 bash scripts/check-no-evaluate-javascript.sh
+bash scripts/check-no-jarvis-bus-world-references.sh
 bash scripts/check-no-leftover-stubs.sh
 bash scripts/check-no-modal-presentation.sh
 bash scripts/check-no-null-voice-adapters.sh
@@ -227,7 +228,7 @@ Codesign + entitlement verification: `bash scripts/verify-entitlements.sh --pre-

 ```bash
 JARVIS_REAL_MODELS=1 swift test --package-path packages/Memory --filter MemoryRegressionCorpusTests
-# requires: local Ollama with nomic-embed-text + qwen2.5-coder:32b + loadable vec0.dylib
+# requires: local Ollama with nomic-embed-text + qwen3.6:latest (extractor) + qwen2.5-coder:32b-instruct-q8_0 (agent fallback) + loadable vec0.dylib

 JARVIS_REAL_CAMERA=1 swift test --package-path packages/Vision --filter CameraCaptureRealHardwareTests
 # requires: real AVCaptureDevice + .authorized TCC
@@ -236,7 +237,7 @@ JARVIS_REAL_MODELS=1 swift test --package-path packages/Voice --filter OrpheusTT
 # requires: ~6 GB Orpheus weights downloaded; produces empirical TTFA measurement
 ```

-Eval harness (Anthropic + local-model eval matrix; corpus in `.planning/evals/`) — pinned to `qwen2.5-coder:32b` for the local-model lane.
+Eval harness (Anthropic + local-model eval matrix; corpus in `.planning/evals/`) — pinned to `qwen2.5-coder:32b-instruct-q8_0` for the local-model lane (matches `ModelID.qwen25coder32b`).

 ### Self-hosted CI runner (Grogu)

@@ -250,6 +251,7 @@ Eval harness (Anthropic + local-model eval matrix; corpus in `.planning/evals/`)
 - **Cache:** the `actions/cache@v4` step is gated to `runner.environment == 'github-hosted'`. Self-hosted relies on the persistent `_work/` dir on HD-2 (and SwiftPM's natural `.build/` reuse) instead. Github-hosted caches bake absolute `/Users/runner/...` paths into `.build/artifacts/onnxruntime.xcframework` and other XCFrameworks, which Grogu cannot resolve — restricting the cache action keeps these caches strictly per-environment.
 - **Runner keychain** (`~/Library/Keychains/jarvis-runner.keychain-db`): created once on Grogu with password `grogu-jarvis-ci` (also stored as repo secret `JARVIS_RUNNER_KEYCHAIN_PASSWORD`). The LaunchAgent has no UI session, so `SecItemAdd` against a locked login keychain returns `errSecInteractionNotAllowed (-25308)`. CI unlocks `jarvis-runner.keychain-db`, prepends it to the user search list, AND sets it as the user's default keychain before `swift test`. The default-keychain step is necessary because on the legacy file-based macOS keychain, `SecItemAdd` writes to the user's *default* keychain (not the first writable in the search list). **Side effects on interactive use:** (1) `jarvis-runner.keychain-db` sits first in your user search list, and (2) it is also your default keychain — so `SecItemAdd` calls *without* an explicit keychain target write there instead of `login.keychain-db`. Apps that target login keychain explicitly (most do) are unaffected. To undo: `security default-keychain -d user -s ~/Library/Keychains/login.keychain-db && security list-keychains -d user -s ~/Library/Keychains/login.keychain-db`.
 - **When Grogu reboots or sleeps**, the LaunchAgent restarts on login but jobs queued during downtime will sit waiting. For long offline windows, push to a branch other than develop or temporarily revert `runs-on:` to `macos-15`.
+- **Process hygiene:** the workflow's first step on self-hosted is a `Reap orphan xctest/swift-test processes older than 1h` cleanup. Grogu is a shared workstation — interactive Xcode test runs, aborted `swift test`, or debugger sessions can leak `xctest` / `swift-test` / `swift-driver` processes that hold Mach services + sockets and cause subsequent CI jobs to deadlock (observed: AgentCore hung 34 min on a develop run when 6 such zombies aged 22–24h were present). The age filter (≥1h) keeps concurrent CI work on sibling runners safe (jobs complete in <10 min).

 # CLAUDE.md

@@ -8,6 +8,7 @@
 [![Issues](https://img.shields.io/github/issues/KofTwentyTwo/Jarvis)](https://github.com/KofTwentyTwo/Jarvis/issues)
 [![Platform: macOS 26](https://img.shields.io/badge/platform-macOS%2026%20Tahoe-blue)](https://www.apple.com/macos/)
 [![Swift 6](https://img.shields.io/badge/swift-6.0-orange)](https://swift.org)
+[![OpenSSF Scorecard](https://api.scorecard.dev/projects/github.com/KofTwentyTwo/Jarvis/badge)](https://scorecard.dev/viewer/?uri=github.com/KofTwentyTwo/Jarvis)

 Jarvis is a hybrid native + embedded-web macOS app that runs as an ambient
 menu-bar presence. A Swift host owns the operating system (mic, camera,
@@ -53,46 +54,7 @@ view, the Bus is the only legal way they talk. Subsystems live in
 single-purpose SPM packages; **18 boundary-gate scripts** keep the
 isolation honest.

-```mermaid
-flowchart TB
-    User[User<br/>voice / text / hotkey]
-
-    subgraph Swift["Swift host (AppKit + SwiftUI)"]
-        AppDel[AppDelegate<br/>install DAG + lifecycle]
-        Orch[AgentOrchestrator<br/>tool-loop + streaming]
-        Bridge[WebviewBridge<br/>typed JSON Bus]
-        Voice[Voice<br/>wake / STT / TTS]
-        Vision[Vision<br/>camera + frame-attach]
-        Memory[Memory<br/>SQLite + vec + mem0]
-        MCP[MCPRuntime<br/>helpers + in-proc tools]
-        Replay[Replay<br/>turn-by-turn audit log]
-        Boot[BootHealth<br/>NOT FAKED probes]
-    end
-
-    LLM["LLMProvider<br/>Anthropic Opus 4.7 / Ollama"]
-
-    subgraph Web["WKWebView (React + R3F)"]
-        Ring[Particle ring]
-        Chat[Chat panel]
-        DevOv[Dev Overlay]
-    end
-
-    User --> Voice
-    User --> Bridge
-    AppDel --> Vision
-    AppDel --> Orch
-    AppDel --> Voice
-    AppDel --> Boot
-    Voice --> Orch
-    Orch <--> LLM
-    Orch --> MCP
-    Orch --> Memory
-    Orch --> Vision
-    Orch --> Replay
-    Orch --> Bridge
-    Boot --> Bridge
-    Bridge <-->|Bus protocol| Web
-```
+![Jarvis 2-Process Architecture Topology](docs/architecture.png)

 The deep architecture doc is [`ARCHITECTURE.md`](ARCHITECTURE.md) (subsystem
 deep-dives, anti-patterns, navigation cheat sheet). The v1.0 typed-API
@@ -135,21 +97,7 @@ asserts the spawn-line ordering AND that `voiceInstallTask` awaits
 `agentInstallTask?.value` (a strictly stronger guarantee than spawn-order
 alone).

-```mermaid
-flowchart LR
-    Boot([applicationWillFinishLaunching]) --> Vision[installVision]
-    Boot --> Memory[installMemory]
-    Boot --> MCP[installMCP]
-    Boot --> Replay[installReplay]
-    Vision --> Agent[installAgent]
-    Memory --> Agent
-    MCP --> Agent
-    Replay --> Agent
-    Agent --> Voice[installVoice]
-    Voice --> Probes[registerBootHealthProbes]
-    Probes --> Bridge[Webview handshake]
-    Bridge --> Ready([HUD ready])
-```
+![Jarvis Install Order Sequence](docs/install_order.png)

 ## Build & run

@@ -215,21 +163,7 @@ subscribers that need it. If a subsystem is degraded, the orchestrator
 **injects a system-prompt preamble** so the agent doesn't promise to
 remember things the memory store can't actually persist.

-```mermaid
-flowchart LR
-    In[User input<br/>text / voice transcript] --> Orch[AgentOrchestrator]
-    Health[BootHealth.rollup] -->|degradation preamble| Orch
-    Orch --> Provider[LLMProvider.stream]
-    Provider --> Events[(streaming events<br/>token / thinking / tool-use / stop)]
-    Events --> Bcast[Broadcaster<br/>single consumer]
-    Bcast --> Tx[Transcript]
-    Bcast --> Mem[Memory extractor]
-    Bcast --> Dev[DevOverlay]
-    Bcast --> Voice[Voice / TTS]
-    Bcast --> Bus[WebviewBridge → HUD]
-    Orch --> Tools[MCP tool calls]
-    Tools --> Provider
-```
+![Jarvis Agent Turn Lifecycle](docs/turn_lifecycle.png)

 ## Boot-health probes

@@ -240,32 +174,7 @@ failures turn the menu-bar icon **red**, raise a non-dismissible HUD
 banner, and inject the degradation preamble into the agent's system
 prompt.

-```mermaid
-flowchart LR
-    M[MemoryBootHealthProbe<br/>vec0 + FTS5 reachable]
-    A[AnthropicBootHealthProbe<br/>keychain has API key]
-    O[OllamaBootHealthProbe<br/>service reachable + model loaded]
-    Vc[VoiceBootHealthProbe<br/>audio graph alive]
-    Vi[VisionBootHealthProbe<br/>AVCapture authorized]
-    Mc[MCPBootHealthProbe<br/>helper apps + in-proc tools]
-    R[ReplayBootHealthProbe<br/>WAL writable]
-    W[WebviewBootHealthProbe<br/>handshake completed]
-    Roll[BootHealthOrchestrator.rollup]
-    Icon[Menu-bar icon<br/>red on critical]
-    Banner[Non-dismissible banner]
-    Pre[Agent system-prompt<br/>degradation preamble]
-    M --> Roll
-    A --> Roll
-    O --> Roll
-    Vc --> Roll
-    Vi --> Roll
-    Mc --> Roll
-    R --> Roll
-    W --> Roll
-    Roll --> Icon
-    Roll --> Banner
-    Roll --> Pre
-```
+![Jarvis Boot-Health Probe Ladder](docs/boot_health.png)

 ## Project culture

@@ -0,0 +1,124 @@
+# Handoff — 2026-05-14: multi-runner shipped, v0.2 ready
+
+**Session window:** 2026-05-13 afternoon (Grogu single runner setup) → 2026-05-14 (multi-runner expansion + first auto-Release)
+**Branch:** `develop`, clean
+**HEAD at write time:** `774e60a0` — `ci(actions): create GitHub Release directly in publish-tag job (#187)`
+**Open PR count:** 2 (#185, #191 — both with shepherds running)
+**Repo visibility:** PUBLIC
+
+---
+
+## TL;DR
+
+- **3-runner self-hosted CI on Grogu is live** — `Grogu-jarvis`, `-2`, `-3`. Matrix parallelism observed (≥3 SPM jobs concurrent). Develop CI green in ~5 min vs. ~13 min single-runner.
+- **Auto-Release flow proven** — `v0.1.0-dev.64` has a real GitHub Release object, created by the new `gh release create` step in `publish-tag`. No more manual `gh release create` needed.
+- **Brainstorm → spec → plan → execute discipline established** via the Superpowers skills. Multi-runner work has a committed spec + plan under `docs/superpowers/{specs,plans}/`. Pattern to follow for v0.2 onward.
+- **Memory test flake** surfaced: `MemoryWiringEndToEndTests.testMemoryEnqueueReceivesNonNilUserAndAssistantText` fails intermittently on PR branches but passes on develop. Hypothesis: race in `TurnTranscriptStore` actor under concurrent-runner load. Pre-existing; not from any 2026-05-13/14 code change.
+
+---
+
+## What shipped today
+
+### v0.1 mop-up + first dev tag
+- **PR #184** (squash `5fc2d3fa`) — Ollama model baseline reconciled to `qwen3.6:latest` (extractor) + `qwen2.5-coder:32b-instruct-q8_0` (agent fallback). Closes #58 (last v0.1 issue).
+- **v0.1.0-dev.50** — first release tag of the project. Release object created manually (workflow couldn't trigger downstream).
+- **v0.1.0-dev.64** — second release tag. Release object created **automatically** via the new inline `gh release create` step in `publish-tag`.
+
+### Self-hosted runner saga
+- **PR #181** (`eef17252`) — Moved `swift-packages` matrix to `Grogu-jarvis`. DEVELOPER_DIR Xcode-select. Fork-PR gate.
+- **PR #183** (`81ce64c3`) — Orphan-process reaper at job start (closes #182).
+- **PR #189** (`4cdb3560`) — Reaper gets a `>=1h` age filter so sibling runners on Grogu don't kill each other.
+- **PR #187** (`774e60a0`) — `publish-tag` step creates the Release object directly instead of relying on `release.yml` triggering on tag push (GITHUB_TOKEN can't trigger downstream workflows).
+- **PR #190** (`a26f49fd`) — Multi-runner spec + plan committed to `docs/superpowers/`.
+- **PR #185** — open, AgentCore #23+#24 fix. Currently has a Memory-flake CI failure under shepherd watch.
+- **PR #191** — open, CLAUDE.md inventory update. Same Memory flake situation.
+
+### Grogu state
+- 3 runners installed at `~/actions-runner-jarvis{,-2,-3}/`, each with `_work` symlinked to `/Volumes/HD-2/jarvis-runner-work{,-2,-3}/`.
+- LaunchAgents managed by `~/Library/LaunchAgents/actions.runner.KofTwentyTwo-Jarvis.Grogu-jarvis{,-2,-3}.plist`.
+- Keychain workflow: `~/Library/Keychains/jarvis-runner.keychain-db`, password in repo secret `JARVIS_RUNNER_KEYCHAIN_PASSWORD`. CI unlocks + makes default + prepends to search list.
+- Side effect: `jarvis-runner.keychain-db` is your user's default keychain after CI runs. Undo command in CLAUDE.md.
+
+---
+
+## Active background processes at handoff
+
+| Task ID | What | Expected exit |
+|---|---|---|
+| `bny76q2ad` | Final shepherd watching PR #185 + #191 | When both PRs are handled (merged or marked `ci-failed`), or 60-min timeout |
+
+**Log location:** `/tmp/final-shepherd.log`
+
+Reading the log will tell you what the shepherd did. If it marked both PRs `ci-failed` because of the Memory flake, see "Memory flake disposition" below.
+
+---
+
+## Memory flake disposition (what to do when you pick up)
+
+`MemoryWiringEndToEndTests.testMemoryEnqueueReceivesNonNilUserAndAssistantText` fails with:
+```
+XCTAssertEqual failed: ("0") is not equal to ("1") - enqueue was not called — BLOCKER-1 regression
+XCTAssertEqual failed: ("nil") is not equal to ("Optional("hi")") - userText empty/nil — Plan 4's submit-time append did not propagate via TurnTranscriptStore
+XCTAssertEqual failed: ("nil") is not equal to ("Optional("hello there")") - assistantText empty/nil — Plan 1's transcript subscriber did not propagate via TurnTranscriptStore
+```
+
+**Evidence it's a flake, not a regression:**
+- PR #191 doesn't touch AgentCore at all (only `CLAUDE.md`) yet hits the same exact failure.
+- Develop's latest CI (`25832341662`, on `774e60a0`) passed Memory in 45s.
+- Both #185 and #191 are branched off slightly different develop SHAs — the test passes on the latest develop but fails on these branch HEADs after forward-merge.
+
+**Recommended next steps:**
+1. Open a new issue: `bug,severity:medium,area:memory` — "MemoryWiringEndToEndTests intermittent failure under concurrent runner load". Describe the `TurnTranscriptStore` race hypothesis.
+2. Either rerun the failed jobs (CI may pass on next attempt) OR merge both PRs as-is since branch protection has 0 required checks and the test passes reliably on develop.
+3. Long-term: file the test as known-flaky in v0.2 milestone.
+
+---
+
+## v0.2 Wave 1 (next session start)
+
+Tiny / doc-only / deterministic fixes — safe to land in a single session each:
+
+| Issue | What | Effort | Why first |
+|---|---|---|---|
+| **#19** | MemoryStore.swift stale Phase 7 hedge doc-comments | ~5 min | Pure docs, zero risk |
+| **#45** | Bus 'JarvisBusWorld' docstring drift (20+ docstrings) | ~15 min | Pure docs, mechanical |
+| **#46** | Stale webview bundle artifacts in `App/Resources/webview/assets/` (~50MB) | ~10 min | File deletes |
+| **#37** | STTBackendSelector snake_case vs camelCase mismatch | ~15 min | Single rename |
+| **#178** | Add OpenSSF Scorecard workflow | ~15 min | One new `.github/workflows/` file |
+| **#47** | WebviewBridge.sendRaw NaN→null JSON guard | ~20 min | One-line bug fix |
+
+Each gets brainstorm → spec → plan via the Superpowers skills (specs/plans templates established in `docs/superpowers/`). Brief specs (a few paragraphs) — these are "issue body IS the design" cases.
+
+## v0.2 Wave 2 / Wave 3 — wait
+
+Larger items that need real diagnostic work or design decisions:
+- #18 MemoryReplaySink error propagation — needs care
+- #23 + #24 already in PR #185 (held)
+- #44 HUD audioLevel consumer rewire — MED, design choice
+- #43 sessionHistory dead-letter — overlaps M-7, defer
+- #176 vitest 4.x major bump — needs eval
+- #26 confirmation panel multi-monitor — needs real hardware
+- Any M-0..M-7 issue — strategic, gets its own spec → plan
+
+---
+
+## Key references
+
+- **Spec template:** `docs/superpowers/specs/2026-05-13-multi-runner-grogu-design.md`
+- **Plan template:** `docs/superpowers/plans/2026-05-13-multi-runner-grogu.md`
+- **Previous handoff:** `docs/HANDOFF-2026-05-13-v0.1-shipped.md` (now superseded for the multi-runner story but still describes v0.1 mop-up)
+- **Roadmap:** `.planning/ROADMAP-2026-05-12.md`
+- **Live state:** GitHub Issues + Milestones at https://github.com/KofTwentyTwo/Jarvis
+- **CI workflow:** `.github/workflows/ci.yml` — has the age-filtered cleanup + inline release create
+- **CLAUDE.md `Self-hosted CI runner (Grogu)`** section — operational reference for the 3-runner setup
+
+## What this handoff deliberately doesn't relitigate
+
+- The Memory flake's root cause — open a v0.2 bug and move on; develop's CI is reliably green so it's not blocking.
+- Whether to merge PR #185 / PR #191 with a known-flaky failure — branch protection allows it; user judgment.
+- Wave 2/3 v0.2 issues — they need design conversations; pick them up after Wave 1.
+- M-0..M-7 — strategic, gets its own session.
+
+---
+
+*Written 2026-05-14 at the natural session-end after multi-runner expansion shipped. Next session should start with `/clear`, read this doc, then read the latest `/tmp/final-shepherd.log` to see what the background shepherd did, then pick up Wave 1.*
@@ -0,0 +1,374 @@
+# Multi-Runner Expansion on Grogu — Implementation Plan
+
+> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
+
+**Goal:** Scale Grogu from 1 to 3 self-hosted GitHub Actions runners so the 13-package SwiftPM matrix parallelizes, cutting develop CI wall-clock from ~13 min serial to ~5 min.
+
+**Architecture:** Two new LaunchAgent runners (`Grogu-jarvis-2`, `Grogu-jarvis-3`), each in its own `~/actions-runner-jarvis-{2,3}/` directory with `_work/` symlinked to `/Volumes/HD-2/jarvis-runner-work-{2,3}/`. All three runners share labels `[self-hosted, macOS, ARM64, jarvis]` so matrix jobs dispatch to whichever is free.
+
+**Tech Stack:** GitHub Actions self-hosted runner v2.334.0 (osx-arm64), `gh` CLI for token + verification, macOS LaunchAgents for per-user service management.
+
+**Spec:** `docs/superpowers/specs/2026-05-13-multi-runner-grogu-design.md`
+
+**Related issues:** #180 (single-runner setup), #182 (zombie cleanup), #186 (auto-release), #188 (age-filter — precondition).
+
+---
+
+## File Structure
+
+This plan is primarily operational; the only repo file modified is `CLAUDE.md` (runner inventory documentation).
+
+| Change | Path | Responsibility |
+|---|---|---|
+| **System (new)** | `~/actions-runner-jarvis-2/` | Runner-2 agent binary tree (already pre-staged) |
+| **System (new)** | `~/actions-runner-jarvis-2/_work` → `/Volumes/HD-2/jarvis-runner-work-2/` | Symlink for build artifacts (already pre-staged) |
+| **System (new)** | `~/Library/LaunchAgents/actions.runner.KofTwentyTwo-Jarvis.Grogu-jarvis-2.plist` | LaunchAgent service plist (created by `svc.sh install`) |
+| **System (new)** | `~/actions-runner-jarvis-3/` | Runner-3 agent binary tree (already pre-staged) |
+| **System (new)** | `~/actions-runner-jarvis-3/_work` → `/Volumes/HD-2/jarvis-runner-work-3/` | Symlink (already pre-staged) |
+| **System (new)** | `~/Library/LaunchAgents/actions.runner.KofTwentyTwo-Jarvis.Grogu-jarvis-3.plist` | LaunchAgent plist |
+| **Repo (modify)** | `CLAUDE.md` | Runner inventory section updated to list all 3 runners |
+
+The pre-staging step (download + extract agent + create `_work` symlinks for runner 2 and 3) was performed earlier in the session. Task 2 step 1 verifies that pre-stage rather than redoing it.
+
+---
+
+## Task 1: Verify precondition — PR #189 merged to develop
+
+**Files:**
+- Read-only check: `origin/develop` git log
+
+PR #189 (age-filter zombie reaper) MUST be on `develop` before either new runner starts accepting jobs. The current cleanup step kills any `xctest`/`swift-test`/`swift-driver` regardless of age; concurrent runners' cleanup steps would mutually kill each other's in-flight tests.
+
+- [ ] **Step 1: Fetch latest develop**
+
+```bash
+git fetch origin develop
+```
+
+Expected: no output, or "Updating <hash>..<hash>".
+
+- [ ] **Step 2: Verify the age-filter commit is on develop**
+
+```bash
+git log origin/develop --grep="age-filter zombie reaper" --oneline | head -3
+```
+
+Expected: at least one line containing `676c54d ci(actions): age-filter zombie reaper (>= 1h) for multi-runner safety` (the exact SHA prefix may differ post-squash-merge).
+
+If empty: PR #189 is not yet merged. Stop the plan here and re-run from Task 1 once it is. Do NOT proceed to Task 2 — starting concurrent runners against an un-filtered cleanup step will trigger mutual xctest kill.
+
+---
+
+## Task 2: Register Grogu-jarvis-2
+
+**Files:**
+- System: `~/actions-runner-jarvis-2/` (pre-staged binaries — verify only)
+- System: `~/actions-runner-jarvis-2/_work` symlink (pre-staged — verify only)
+- System (new): `~/actions-runner-jarvis-2/.runner` (created by `config.sh`)
+- System (new): `~/Library/LaunchAgents/actions.runner.KofTwentyTwo-Jarvis.Grogu-jarvis-2.plist` (created by `svc.sh install`)
+
+- [ ] **Step 1: Verify pre-stage**
+
+```bash
+test -x ~/actions-runner-jarvis-2/config.sh && \
+  readlink ~/actions-runner-jarvis-2/_work | grep -q '/Volumes/HD-2/jarvis-runner-work-2$' && \
+  echo "OK: pre-stage intact" || echo "FAIL"
+```
+
+Expected: `OK: pre-stage intact`.
+
+If FAIL: re-run the pre-stage block from earlier in the session — `mkdir -p` the dirs on HD-2, download `actions-runner-osx-arm64-2.334.0.tar.gz`, extract into `~/actions-runner-jarvis-2/`, `ln -s /Volumes/HD-2/jarvis-runner-work-2 ~/actions-runner-jarvis-2/_work`.
+
+- [ ] **Step 2: Get a fresh registration token**
+
+```bash
+TOKEN=$(gh api -X POST /repos/KofTwentyTwo/Jarvis/actions/runners/registration-token -q '.token')
+test -n "$TOKEN" && echo "Token acquired (len=${#TOKEN})" || echo "FAIL"
+```
+
+Expected: `Token acquired (len=33)` or similar non-zero length. Token expires in 1 hour; do not pause between this step and step 4.
+
+If FAIL: check `gh auth status` — the active account must have `workflow` scope on the KofTwentyTwo/Jarvis repo.
+
+- [ ] **Step 3: Configure runner-2 (unattended)**
+
+```bash
+cd ~/actions-runner-jarvis-2
+./config.sh \
+  --url https://github.com/KofTwentyTwo/Jarvis \
+  --token "$TOKEN" \
+  --name "Grogu-jarvis-2" \
+  --labels "jarvis" \
+  --work "_work" \
+  --unattended \
+  --replace 2>&1 | grep -v -i 'token' | tail -10
+```
+
+Expected last lines include `√ Connected to GitHub`, `√ Runner successfully added`, `√ Settings Saved.`. The `grep -v token` strips any line containing the literal token string from the logged output as a defense-in-depth measure (the token itself is short-lived but should not be left in scrollback).
+
+- [ ] **Step 4: Install + start as LaunchAgent**
+
+```bash
+cd ~/actions-runner-jarvis-2
+./svc.sh install
+./svc.sh start
+./svc.sh status
+```
+
+Expected last block:
+```
+status actions.runner.KofTwentyTwo-Jarvis.Grogu-jarvis-2:
+/Users/james.maes/Library/LaunchAgents/actions.runner.KofTwentyTwo-Jarvis.Grogu-jarvis-2.plist
+
+Started:
+<PID> 0 actions.runner.KofTwentyTwo-Jarvis.Grogu-jarvis-2
+```
+
+(PID will vary.)
+
+- [ ] **Step 5: Verify runner-2 is online via GitHub API**
+
+```bash
+gh api /repos/KofTwentyTwo/Jarvis/actions/runners \
+  -q '.runners[] | select(.name=="Grogu-jarvis-2") | {name, status, busy, labels: [.labels[].name]}'
+```
+
+Expected:
+```json
+{"busy":false,"labels":["self-hosted","macOS","ARM64","jarvis"],"name":"Grogu-jarvis-2","status":"online"}
+```
+
+If `status` is anything other than `online`, check `~/Library/Logs/actions.runner.KofTwentyTwo-Jarvis.Grogu-jarvis-2/stdout.log` for the runner-side error and stop the plan.
+
+---
+
+## Task 3: Register Grogu-jarvis-3
+
+Identical sequence to Task 2 with `-2` replaced by `-3` everywhere. Per the spec's sequential-validation principle, do NOT start runner-3 until runner-2 is confirmed `status=online, busy=false`.
+
+**Files:**
+- System: `~/actions-runner-jarvis-3/` (pre-staged — verify only)
+- System: `~/actions-runner-jarvis-3/_work` symlink (pre-staged — verify only)
+- System (new): `~/actions-runner-jarvis-3/.runner` (created by `config.sh`)
+- System (new): `~/Library/LaunchAgents/actions.runner.KofTwentyTwo-Jarvis.Grogu-jarvis-3.plist`
+
+- [ ] **Step 1: Verify pre-stage**
+
+```bash
+test -x ~/actions-runner-jarvis-3/config.sh && \
+  readlink ~/actions-runner-jarvis-3/_work | grep -q '/Volumes/HD-2/jarvis-runner-work-3$' && \
+  echo "OK: pre-stage intact" || echo "FAIL"
+```
+
+Expected: `OK: pre-stage intact`.
+
+- [ ] **Step 2: Get a fresh registration token**
+
+```bash
+TOKEN=$(gh api -X POST /repos/KofTwentyTwo/Jarvis/actions/runners/registration-token -q '.token')
+test -n "$TOKEN" && echo "Token acquired (len=${#TOKEN})" || echo "FAIL"
+```
+
+Expected: `Token acquired (len=33)` or similar.
+
+- [ ] **Step 3: Configure runner-3 (unattended)**
+
+```bash
+cd ~/actions-runner-jarvis-3
+./config.sh \
+  --url https://github.com/KofTwentyTwo/Jarvis \
+  --token "$TOKEN" \
+  --name "Grogu-jarvis-3" \
+  --labels "jarvis" \
+  --work "_work" \
+  --unattended \
+  --replace 2>&1 | grep -v -i 'token' | tail -10
+```
+
+Expected last lines: `√ Connected to GitHub`, `√ Runner successfully added`, `√ Settings Saved.`
+
+- [ ] **Step 4: Install + start as LaunchAgent**
+
+```bash
+cd ~/actions-runner-jarvis-3
+./svc.sh install
+./svc.sh start
+./svc.sh status
+```
+
+Expected: `Started: <PID> 0 actions.runner.KofTwentyTwo-Jarvis.Grogu-jarvis-3`.
+
+- [ ] **Step 5: Verify runner-3 is online**
+
+```bash
+gh api /repos/KofTwentyTwo/Jarvis/actions/runners \
+  -q '.runners[] | select(.name=="Grogu-jarvis-3") | {name, status, busy, labels: [.labels[].name]}'
+```
+
+Expected:
+```json
+{"busy":false,"labels":["self-hosted","macOS","ARM64","jarvis"],"name":"Grogu-jarvis-3","status":"online"}
+```
+
+- [ ] **Step 6: Confirm all 3 runners visible**
+
+```bash
+gh api /repos/KofTwentyTwo/Jarvis/actions/runners \
+  -q '.runners | length, [.runners[].name] | sort'
+```
+
+Expected (exactly):
+```
+3
+["Grogu-jarvis","Grogu-jarvis-2","Grogu-jarvis-3"]
+```
+
+If the count is anything other than 3, stop and investigate which runner failed to register.
+
+---
+
+## Task 4: Acceptance — observe matrix parallelism
+
+**Files:** none.
+
+**Premise:** PR #185 (`fix/agent-cache-text-blocks-23-24`) was cancelled earlier in this session with one previously-failed SPM job. Rerun it to get a CI run with the new runner topology in effect. Watch for ≥3 SPM jobs in `in_progress` simultaneously.
+
+- [ ] **Step 1: Identify the PR #185 run id**
+
+```bash
+RUN_ID=$(gh run list --branch fix/agent-cache-text-blocks-23-24 --workflow=CI --limit 1 --json databaseId -q '.[0].databaseId')
+echo "PR #185 latest run: $RUN_ID"
+```
+
+Expected: a numeric run id, e.g. `25827802988`.
+
+- [ ] **Step 2: Trigger a full rerun**
+
+```bash
+gh run rerun "$RUN_ID" --repo KofTwentyTwo/Jarvis
+sleep 5
+gh run view "$RUN_ID" --json status,conclusion -q '"\(.status) / \(.conclusion // "(none)")"'
+```
+
+Expected: `queued / (none)` or `in_progress / (none)`. (Using `gh run rerun` without `--failed` reruns the whole workflow so we observe a fresh matrix dispatch across all runners.)
+
+- [ ] **Step 3: Wait for matrix dispatch, then sample concurrent jobs**
+
+Wait ~60-90 seconds for the webview/boundary jobs to complete and the SPM matrix to start dispatching.
+
+```bash
+sleep 90
+gh run view "$RUN_ID" --json jobs -q '[.jobs[] | select(.name | startswith("SPM ")) | select(.status == "in_progress") | .name] | "in_progress count: \(length)\nnames: \(.)"'
+```
+
+Expected: `in_progress count: 3` (or higher if the matrix has dispatched more quickly than expected) followed by three SPM job names.
+
+If `in_progress count: 1`: GitHub may have only just started dispatching. Wait another 60 seconds and re-sample. If still 1 after 5 minutes, investigate — possibly one or both new runners are offline. Cross-check with `gh api /repos/.../actions/runners -q '.runners[].busy'` (should show ≥2 `true`).
+
+- [ ] **Step 4: Cross-verify on Grogu**
+
+```bash
+gh api /repos/KofTwentyTwo/Jarvis/actions/runners \
+  -q '[.runners[] | select(.busy == true) | .name] | "busy runners: \(length)\nnames: \(.)"'
+```
+
+Expected: `busy runners: 3` and the three runner names. (If 2, the matrix is partially dispatched; if 1, only the original runner is being used — investigate.)
+
+- [ ] **Step 5: Wait for the run to terminate**
+
+```bash
+gh run watch "$RUN_ID" --repo KofTwentyTwo/Jarvis --exit-status
+echo "Exit: $?"
+```
+
+Expected: process exits with 0 and a `✓ run completed with 'success'` message. Wall-clock from trigger to completion should be approximately 5–7 minutes (vs the prior ~13 min single-runner baseline). If the run conclusion is `failure`, inspect failed jobs via `gh run view "$RUN_ID" --log-failed` to determine whether the failure is multi-runner-related (e.g., test-suite race) or pre-existing flakiness (e.g., the AgentCore deadlock seen earlier).
+
+---
+
+## Task 5: Update CLAUDE.md with the new runner inventory
+
+**Files:**
+- Modify: `CLAUDE.md` (the "Self-hosted CI runner (Grogu)" section)
+
+- [ ] **Step 1: Create a branch**
+
+```bash
+git fetch origin develop
+git checkout -b docs/multi-runner-inventory origin/develop
+```
+
+Expected: `Switched to a new branch 'docs/multi-runner-inventory'`.
+
+- [ ] **Step 2: Locate the runner section in CLAUDE.md**
+
+```bash
+grep -n "Where it lives\|Service:\|jarvis-runner.keychain" CLAUDE.md | head -3
+```
+
+Expected: three line numbers pointing into the `### Self-hosted CI runner (Grogu)` section.
+
+- [ ] **Step 3: Replace the "Where it lives" + "Service" bullets with multi-runner equivalents**
+
+Edit the two bullets via the `Edit` tool. The OLD text:
+
+```
+- **Where it lives:** `~/actions-runner-jarvis/` on Grogu. `_work/` symlinked to `/Volumes/HD-2/jarvis-runner-work/` (1.3 TB) to keep boot-disk pressure off.
+- **Service:** launchd LaunchAgent `actions.runner.KofTwentyTwo-Jarvis.Grogu-jarvis` (`~/Library/LaunchAgents/`). Manage with `cd ~/actions-runner-jarvis && ./svc.sh {status,start,stop}`. Only runs while user is logged in.
+```
+
+The NEW text:
+
+```
+- **Where they live:** three runners on Grogu — `~/actions-runner-jarvis/` (Grogu-jarvis), `~/actions-runner-jarvis-2/` (Grogu-jarvis-2), `~/actions-runner-jarvis-3/` (Grogu-jarvis-3). Each `_work/` symlinked to `/Volumes/HD-2/jarvis-runner-work{,-2,-3}/` (1.3 TB available, ~10 GB peak per runner) to keep boot-disk pressure off.
+- **Service:** one launchd LaunchAgent per runner — `actions.runner.KofTwentyTwo-Jarvis.Grogu-jarvis{,-2,-3}.plist` in `~/Library/LaunchAgents/`. Manage individually with `cd ~/actions-runner-jarvis-N && ./svc.sh {status,start,stop}` (omit `-N` for runner 1). All three only run while user is logged in.
+- **Why three:** the 13-package SPM matrix parallelizes across runners. Single-runner serial = ~13 min CI; three-runner parallel = ~5 min. Grogu has 24 cores / 128 GB RAM; 3 concurrent SwiftPM cold builds peak at ~24 cores + ~24 GB, well within budget. See `docs/superpowers/specs/2026-05-13-multi-runner-grogu-design.md`.
+```
+
+- [ ] **Step 4: Stage + commit**
+
+```bash
+git add CLAUDE.md
+git diff --cached --stat
+git commit -m "docs(claude-md): runner inventory updated for 3-runner Grogu setup
+
+Replaces single-runner 'Where it lives' / 'Service' bullets with the
+3-runner inventory (Grogu-jarvis, -2, -3). Adds a brief 'Why three'
+bullet documenting the parallelism win and links to the spec.
+
+Refs #180"
+```
+
+Expected output: `[docs/multi-runner-inventory <sha>] docs(claude-md): runner inventory updated...` and `1 file changed`.
+
+- [ ] **Step 5: Push + open PR**
+
+```bash
+git push -u origin docs/multi-runner-inventory
+gh pr create --base develop --head docs/multi-runner-inventory \
+  --title "docs(claude-md): runner inventory updated for 3-runner Grogu setup" \
+  --body "Updates CLAUDE.md's \`Self-hosted CI runner\` section to reflect the new 3-runner Grogu topology (Grogu-jarvis, -2, -3) post-multi-runner-expansion. No behavior change — documentation only.
+
+Spec: \`docs/superpowers/specs/2026-05-13-multi-runner-grogu-design.md\`
+Plan: \`docs/superpowers/plans/2026-05-13-multi-runner-grogu.md\`
+
+Refs #180"
+```
+
+Expected: the new PR URL is printed. Note the PR number; the runner-side CI for this PR is itself an acceptance test (it will exercise the new 3-runner topology).
+
+---
+
+## Definition of done
+
+- All steps above are checked off.
+- `gh api /repos/KofTwentyTwo/Jarvis/actions/runners` returns 3 runners, all `status=online`.
+- The retriggered PR #185 CI run terminates with `conclusion=success` and showed ≥3 SPM jobs `in_progress` simultaneously during execution.
+- `docs/multi-runner-inventory` PR is open (or merged) updating CLAUDE.md.
+
+## Self-review (run by author before handoff)
+
+- ✅ **Spec coverage:** every component in the spec is covered by a task — runner registration (Tasks 2+3), acceptance validation (Task 4), doc update (Task 5), precondition (Task 1).
+- ✅ **No placeholders:** every step has concrete commands and expected output.
+- ✅ **Type consistency:** runner names (`Grogu-jarvis-N`), labels (`jarvis`), plist paths, and `_work` symlink targets are spelled identically across tasks.
+- ✅ **Spec out-of-scope items not present:** no warm-cache work, no monitoring/alerting, no per-runner labels.
+- ✅ **Rollback path documented in spec, referenced as needed.**
@@ -0,0 +1,133 @@
+# Multi-runner expansion on Grogu — design
+
+**Date:** 2026-05-13
+**Status:** approved (brainstorm phase, awaiting implementation plan)
+**Workstream:** A — local CI capacity scaling
+**Related issues:** #180 (single-runner setup), #182 (zombie cleanup), #186 (auto-release), #188 (age-filter)
+
+## Problem
+
+Develop CI wall-clock is bottlenecked on a single self-hosted runner (`Grogu-jarvis`) serializing the 13-package SwiftPM matrix. Each cold matrix job takes 1–2 minutes, so a full CI run is 13–15 minutes of serial work plus app-build on github-hosted in parallel. Grogu's hardware (24 physical cores, 128 GB RAM, 1.3 TB free on HD-2) is heavily under-utilized — single-runner load is ~5% of capacity.
+
+Adding sibling runners would parallelize the matrix and roughly cut CI wall-clock by the runner count. Three runners is the recommended target: comfortably within hardware capacity (~24 cores used at peak when all three are mid-SwiftPM-build) without contending with interactive developer work.
+
+## Goal
+
+Three self-hosted runners on Grogu (`Grogu-jarvis`, `Grogu-jarvis-2`, `Grogu-jarvis-3`), each with its own `_work/` directory on HD-2, each registered as a LaunchAgent under james.maes's user session, all sharing labels `[self-hosted, macOS, ARM64, jarvis]` so GitHub Actions dispatches matrix jobs to whichever runner is free.
+
+**Acceptance bar (minimum):** a green develop CI run after registration where `gh run view --json jobs` shows ≥3 SPM matrix jobs in `in_progress` state simultaneously. Warm-cache persistence and runner-health monitoring are explicit non-goals for this phase — they fall to follow-up tickets.
+
+## Components
+
+| Runner | Agent dir | `_work` symlink target | LaunchAgent plist | GitHub name |
+|---|---|---|---|---|
+| 1 (existing) | `~/actions-runner-jarvis/` | `/Volumes/HD-2/jarvis-runner-work/` | `actions.runner.KofTwentyTwo-Jarvis.Grogu-jarvis.plist` | `Grogu-jarvis` |
+| 2 (new) | `~/actions-runner-jarvis-2/` | `/Volumes/HD-2/jarvis-runner-work-2/` | `actions.runner.KofTwentyTwo-Jarvis.Grogu-jarvis-2.plist` | `Grogu-jarvis-2` |
+| 3 (new) | `~/actions-runner-jarvis-3/` | `/Volumes/HD-2/jarvis-runner-work-3/` | `actions.runner.KofTwentyTwo-Jarvis.Grogu-jarvis-3.plist` | `Grogu-jarvis-3` |
+
+All three runners share the same label set; jobs targeting `runs-on: [self-hosted, macOS, jarvis]` will dispatch to any free runner.
+
+Agent binaries are v2.334.0 (osx-arm64) — already extracted into the new runner dirs as a pre-staging step; only `config.sh` + `svc.sh install/start` remain.
+
+## Precondition
+
+PR #189 (age-filter for the orphan-process cleanup step) **must be on `develop`** before either new runner starts accepting jobs. The current cleanup step kills ANY `xctest`/`swift-test`/`swift-driver` process at job start; with concurrent runners, sibling cleanup steps would mutually destroy each other's in-flight tests. PR #189 narrows the kill set to processes ≥1 hour old — concurrent CI jobs (typically <10 min) are safe.
+
+The gating sequence is enforced operationally, not via code: don't run `./svc.sh start` on runner 2 or 3 until `git log origin/develop` shows the age-filter merge.
+
+## Sequence — register one runner
+
+For each `N` in `{2, 3}`:
+
+1. Obtain a fresh registration token (1-hour expiry):
+   ```
+   gh api -X POST /repos/KofTwentyTwo/Jarvis/actions/runners/registration-token
+   ```
+2. Configure:
+   ```
+   cd ~/actions-runner-jarvis-N
+   ./config.sh \
+     --url https://github.com/KofTwentyTwo/Jarvis \
+     --token "$TOKEN" \
+     --name "Grogu-jarvis-N" \
+     --labels "jarvis" \
+     --work "_work" \
+     --unattended \
+     --replace
+   ```
+3. Install as LaunchAgent + start:
+   ```
+   ./svc.sh install
+   ./svc.sh start
+   ```
+4. Verify online via GitHub API:
+   ```
+   gh api /repos/KofTwentyTwo/Jarvis/actions/runners -q '.runners[] | select(.name=="Grogu-jarvis-N") | {status, busy, labels}'
+   ```
+   Expected: `status=online, busy=false, labels=[self-hosted, macOS, ARM64, jarvis]`.
+
+Sequential ordering (runner 2 fully online before configuring runner 3) makes any misconfiguration easy to bisect.
+
+## Acceptance validation
+
+After both runners are online:
+
+1. Retrigger a stuck or cancelled CI run via `gh run rerun <id>` (PR #185 already has a failed AgentCore job that can serve as the test bed).
+2. While the run is in progress, run:
+   ```
+   gh run view <id> --json jobs -q '.jobs[] | select(.status == "in_progress") | .name'
+   ```
+   Expected: at least 3 SPM matrix job names listed simultaneously, e.g., `SPM Bus`, `SPM Memory`, `SPM Voice`.
+3. Cross-verify on Grogu:
+   ```
+   gh api /repos/KofTwentyTwo/Jarvis/actions/runners -q '.runners[] | {name, busy}'
+   ```
+   Expected: at least 3 runners with `busy=true` during peak matrix execution.
+4. Wait for the run to terminate. Confirm `conclusion = success` (no cross-runner interference caused failures).
+
+## Failure modes & mitigations
+
+| Mode | Probability | Mitigation |
+|---|---|---|
+| Mutual xctest kill from un-filtered cleanup | High (without #189) | PR #189 precondition — age-filter at ≥1h. Don't start new runners until merged. |
+| HD-2 disk pressure | Negligible | 1.3 TB free; 3 `_work` dirs × ~10 GB peak = 30 GB. |
+| CPU/RAM saturation | Low | 3 concurrent SwiftPM builds peak at ~24 cores + ~24 GB. Within 24/128 budget. Headroom for interactive work. |
+| Keychain races (`unlock-keychain`) | Negligible | `security unlock-keychain` is idempotent; default-keychain set persistently for the user. |
+| Shared SwiftPM global cache races | Low | `~/Library/Caches/org.swift.swiftpm` is mostly read-after-populate. If contention surfaces, the per-package `actions/checkout clean: true` wipes per-job state cleanly. |
+| Test port conflicts | Low | Jarvis tests use process-local Mach IPC (WKWebView) and per-process temp DBs (SQLite). No fixed-port servers found in audit. |
+| One runner goes offline silently | Medium | Surfaces only at the next CI run (other runners pick up jobs). No alerting — explicit follow-up scope. |
+| Runner agent auto-update breaks during long-running job | Low | Unlikely on a single-job timescale; runner updates happen between jobs. |
+
+## Rollback
+
+Each runner is independently revertible.
+
+- **Pause a runner** (keeps registration): `cd ~/actions-runner-jarvis-N && ./svc.sh stop`
+- **Deregister a runner** (free the slot):
+  - On Grogu: `./svc.sh uninstall`
+  - Via GitHub: `gh api -X DELETE /repos/KofTwentyTwo/Jarvis/actions/runners/<id>` (or `./config.sh remove --token $(gh api -X POST /repos/.../actions/runners/remove-token)`)
+- **Revert to single runner**: stop + uninstall + deregister runners 2 and 3. Workflow unchanged (the `runs-on` selector targets the shared label set, not specific runners).
+
+No code or workflow changes are required to roll back, which is the strongest possible safety property for an infrastructure change.
+
+## Out of scope
+
+- Warm-cache persistence (`actions/checkout clean: false` or custom DerivedData paths). Saves ~3 min per CI run but adds workflow complexity. Follow-up issue.
+- Runner-health monitoring / alerting. Currently relies on `gh api .../actions/runners` snapshots and manual log inspection on Grogu. Follow-up issue.
+- Per-runner-distinct labels (e.g., `runner-2`, `runner-3`). Default identical labels maximize matrix parallelism; per-runner labels would only matter for targeting a specific machine, which we don't need.
+- Scaling beyond 3 runners. Hardware can support more, but matrix size (13 jobs) caps useful parallelism at ~4 runners. Cross that bridge only if matrix grows.
+
+## Definition of done
+
+- `gh api /repos/KofTwentyTwo/Jarvis/actions/runners` shows three runners, all `status=online`.
+- A retriggered CI run shows ≥3 SPM jobs `in_progress` simultaneously at peak.
+- The retriggered run terminates `success` (no failures attributable to multi-runner setup).
+- `CLAUDE.md` updated with the new runner inventory and operational notes.
+- Spec document committed and merged.
+
+## References
+
+- #180 — original single-runner setup
+- #182 — zombie cleanup
+- #188 — age-filter for multi-runner safety (precondition)
+- `~/.claude/plugins/cache/claude-plugins-official/superpowers/5.1.0/skills/brainstorming` — process this design followed
@@ -0,0 +1,151 @@
+# Local-first LLM routing — design
+
+**Status:** draft 2026-05-14, awaiting user approval
+**Authors:** James + Claude (Opus 4.7)
+**Scope:** flip the chat-turn provider default from Anthropic to local Ollama, with reactive escalation to Claude Opus on detectable Ollama failure.
+
+## Goal
+
+Default every conversation turn to local Ollama (`qwen2.5-coder:32b-instruct-q8_0`). Escalate to Anthropic Claude Opus 4.7 only when local fails in a way the runtime can detect. The user sees one coherent reply per turn, with a small badge if escalation fired.
+
+## Why now
+
+The codebase is already structurally ready:
+
+- `LLMProvider` protocol is symmetric between Anthropic + Ollama (same `AsyncThrowingStream<LLMEvent>` shape, same `ToolChoice.none` semantics).
+- `AgentOrchestrator` is already config-driven via `providerFactory(perTurn.resolvedProvider)` (line 311).
+- A mid-conversation escalation hook already exists — line 793 retries on `streamTruncated` by calling `providerFactory(updated.resolvedProvider)`.
+- Memory extraction is wired to Ollama separately and is not affected by chat-turn routing.
+
+This is mostly a policy + UX change, not a new abstraction.
+
+## Approach
+
+Reactive fallback only. No predictive routing, no router-model classifier, no self-reported escalation tool. The orchestrator tries Ollama first; if it fails in one of an enumerated set of detectable ways, it retries the turn once on Anthropic, emits an escalation event for the HUD, and surfaces the answer to the user.
+
+User picked this approach because:
+
+1. The existing escalation hook is reactive — extending what's there beats inventing a new layer.
+2. Reactive fallback produces evidence — every escalation is a real-world data point for any future predictive routing.
+3. Zero added latency on the happy path; no extra LLM call to classify the turn.
+
+## Spec
+
+### 1. Default provider flip
+
+- `default-config.json:9` changes from `"anthropic"` to `"ollama"`.
+- `PerTurnSnapshot.provider` already exists as the policy carrier.
+- Ollama target model is `qwen2.5-coder:32b-instruct-q8_0` (the tool-calling baseline pinned by `ModelID.qwen25coder32b`).
+- Anthropic remains a first-class provider and remains selectable.
+
+### 2. Escalation trigger taxonomy
+
+The runtime fires escalation on, and only on, these detectable failure modes (closed enum `OllamaFailureKind`):
+
+| Trigger | Detection |
+|---|---|
+| `streamTruncated` | Existing — already wired at `AgentOrchestrator.swift:793`. |
+| `malformedToolCall` | `tool_calls` payload fails to decode against `ToolUseRequest`. Known Qwen 3.x footgun per CLAUDE.md. |
+| `unknownTool` | Model invokes a tool name not in the registry. |
+| `refusal` | `stopReason: .refusal` event arrives. |
+| `connectionFailure` | URLSession error before any tokens received (Ollama down, port refused). |
+| `emptyResponse` | `messageStop` with zero `textDelta` and zero `toolUseRequested`. |
+
+Explicit non-triggers (do not escalate):
+
+- Partial answer that "feels weak"
+- Semantic hallucinations the runtime can't detect
+- Tool that returned an error (tool failure ≠ model failure)
+- Slow streaming (no SLA-based escalation)
+
+These are quality problems the runtime cannot judge. Out of scope for this design.
+
+### 3. Escalation contract
+
+```swift
+struct EscalationDecision: Sendable, Codable {
+    let from: ProviderSelection           // .ollama
+    let to: ProviderSelection             // .anthropic
+    let reason: OllamaFailureKind
+    let firedAt: Date
+}
+```
+
+`AgentOrchestrator` enforces **at most one escalation per turn**. If the post-escalation Anthropic call also fails, the orchestrator surfaces the error to the user — it does not fall back further. (A second failure means the cloud provider is also down; pretending otherwise is dishonest.)
+
+### 4. Bus event for the HUD badge
+
+New outbound event:
+
+```swift
+case escalated(EscalationDecision)
+```
+
+Emitted once when fallback fires. HUD renders an unobtrusive badge inline with the affected message: "Escalated to Opus — Ollama returned malformed tool call" (or whichever `OllamaFailureKind` fired). The badge stays with that specific message; subsequent messages are visually clean.
+
+The badge is HUD-side only — it does not affect the chat message content, persistence, or replay log shape.
+
+### 5. Per-turn vs sticky
+
+**Per-turn.** Next user message starts fresh on Ollama. Reasoning: one Ollama failure on a Python question doesn't mean the next "what time is it" should go to Opus.
+
+User can manually pin a conversation to Anthropic via Settings. Manual pin overrides escalation logic; in that mode, no escalation badges fire because no escalation can fire.
+
+### 6. Settings UX
+
+Settings → Provider section gets one new control:
+
+- **Provider preference:** `Local-first (recommended)` / `Anthropic only` / `Ollama only`
+- `Local-first` is the new default; corresponds to `provider: "ollama"` plus escalation enabled.
+- `Anthropic only` pins to Anthropic, escalation moot.
+- `Ollama only` pins to Ollama AND disables escalation — Ollama failures surface as errors. Useful for hard-offline / privacy-strict mode.
+
+The existing `default-config.json` `provider:` field continues to drive the runtime; new logic decides whether escalation is wired based on a sibling `escalationEnabled` field that defaults to `true` only when `provider == "ollama"`.
+
+### 7. Testing
+
+Three eval lanes added to the Harness package:
+
+- **Reactive fallback fires correctly** — fixture: Ollama returns `malformedToolCall`; assert orchestrator switches to Anthropic mid-turn AND emits one `escalated` bus event.
+- **Escalation budget honored** — fixture: Ollama fails AND Anthropic also fails on the same turn; assert orchestrator surfaces a structured error, does not loop, emits exactly one `escalated` event.
+- **Pinned mode disables escalation** — config: `provider=ollama, escalationEnabled=false` + Ollama failure → surface error directly, zero `escalated` events.
+
+No new fixture corpus needed; leverages existing `FixtureKind` cases in `MockLLMProvider`.
+
+### 8. Non-goals
+
+- No predictive routing.
+- No "stream Anthropic continuing Ollama's partial output" — clean restart only.
+- No automatic re-Ollama after Anthropic finishes — once escalated, the rest of the turn stays on Anthropic.
+- No memory extraction changes; that pipeline is independent and stays on Ollama.
+- No changes to non-chat surfaces (TTS, MCP, voice loop) — they have their own provider stories.
+- No new GitHub issue label or milestone; this is a v0.5+ migration item that the user wants brought forward.
+
+### 9. Effort estimate
+
+Single PR, ~3-4 hours:
+
+| Subtask | Estimate |
+|---|---|
+| `default-config.json` flip + Settings UI control | 30 min |
+| Orchestrator escalation logic extension + budget tracking | 60 min |
+| Bus event + HUD badge component | 60 min |
+| Harness tests (3 lanes) | 60 min |
+| Docs + comments | 15 min |
+
+## Success criteria
+
+- Default fresh-launch experience uses Ollama for chat turns.
+- A turn that fails on Ollama via any `OllamaFailureKind` automatically retries on Anthropic and the user sees a single coherent reply with a badge.
+- A turn that fails on both providers shows a structured error, not an infinite retry.
+- A user with `Ollama only` mode sees the Ollama error directly with no escalation.
+- All existing tests pass; three new escalation tests pass.
+- Memory extraction is unaffected.
+- The chat surface still streams tokens at acceptable latency (no perceptible regression).
+
+## Open questions for the user
+
+None blocking — design is concrete enough to implement. If anything below is wrong, push back before the spec is committed.
+
+- Should the badge wording differ per `OllamaFailureKind`, or just say "Escalated to Opus"? Current design: differentiate, because the reason is small additional information that helps the user understand when Ollama is struggling.
+- Should the orchestrator emit a single `bus.escalated` per turn, or stream per-attempt diagnostics? Current design: one per turn, keeps the bus contract simple.
@@ -0,0 +1,114 @@
+# v0.2 Wave 1 sequencing and execution — design
+
+**Status:** approved 2026-05-14 (autonomous execution authorized)
+**Authors:** brainstorm with James + Claude (Opus 4.7)
+**Scope:** Six tiny v0.2 issues (#19, #37, #45, #46, #47, #178) — sequenced into three PRs.
+
+## Goal
+
+Land all six Wave 1 issues against `develop` in three coherent PRs, in an order that minimises CI exposure to the known `MemoryWiringEndToEndTests` flake and keeps each PR's narrative reviewable as a single story.
+
+## Why three PRs, not six and not two
+
+Six individual PRs (the handoff's implicit default) maximises history granularity but burns 6× the SPM CI matrix and exposes us to the Memory flake on every branch. Two giant PRs maximise throughput but mix surfaces (build script + new workflow + bundle rebuild + Swift code) into a single review and a single revert unit.
+
+Three PRs grouped by surface area give us:
+
+- 3× CI runs instead of 6×.
+- Each PR has one coherent narrative (docs ⋄ infra ⋄ code).
+- Each PR can be reverted independently without untangling unrelated work.
+
+## Composition
+
+### PR 1 — "Doc drift cleanup" (`docs/v0.2-wave-1-doc-drift`)
+
+**Issues:** #19 (MemoryStore stale Phase-7 doc-comments) + #45 (Bus JarvisBusWorld drift across 20+ docstrings post `fb41c5f`).
+
+**Why grouped:** both are descriptive/textual changes to documentation that no longer matches the code. #45 already requires a `@jarvis/bus` bundle rebuild and a new grep-gate, so it warrants doc-class company rather than mixing with real code fixes. Both touch package-internal documentation only; no behavioural change.
+
+**Deliverables:**
+
+- `packages/Memory/Sources/Memory/MemoryStore.swift` — scrub "Plan 07-02 wires", "Plan 07-03 will replace", "future ops plan" hedges. Keep the 2026-05-11 comment at line 464–468 intact (it is correct).
+- Bus comments renamed across `WebviewBridge.swift`, `Injection.js`, `webview/packages/bus/src/bridge.ts`, `webview/packages/hud/src/main.tsx`, `App/AppDelegate.swift` — replace "JarvisBusWorld" / "WKContentWorld('JarvisBusWorld')" / cross-world-attach language with "page world" / "shared world" / "defensive double-install guard".
+- `bridge.ts:60–99` rationale rewritten — the duck-type `isJarvisBusShape` guard is now defensive against double-install (both `Injection.js`'s stub and the bundle's `installJarvisBus` running in the same page world), not cross-world attach. Comment must say this without misleading; do **not** delete the guard itself.
+- Bundle rebuilt: `webview/packages/bus/dist/bridge.{js,d.ts}` regenerated via `pnpm --filter @jarvis/bus build` (or equivalent).
+- `@jarvis/bus` `package.json` patch-version bumped to reflect comment changes in dist (matches existing project convention of patch-bumping when the published comments shift).
+- New grep-gate: `scripts/check-no-jarvis-bus-world-references.sh` — fails if any tracked file outside `.planning/`, `docs/`, or git-history references contains the literal `JarvisBusWorld`. Wired into the boundary-gates list in CI and in `CLAUDE.md`.
+
+**Closes:** #19, #45.
+
+### PR 2 — "Build hygiene" (`ci/v0.2-wave-1-build-hygiene`)
+
+**Issues:** #46 (stale webview bundle artifacts) + #178 (OpenSSF Scorecard workflow).
+
+**Why grouped:** both touch repo-level infrastructure (`scripts/` and `.github/workflows/`) and neither touches Swift packages, so the heavy SPM matrix runs but with low chance of new test failures. Quick CI; safe to land first.
+
+**Deliverables:**
+
+- `scripts/build-webview.sh` — before the existing `rsync -a ... dist/ -> App/Resources/webview/` at line 88, add an explicit cleanup of `App/Resources/webview/assets/*.js`, `*.js.map`, `*.css` (preserving `bus-harness.html`, the directory's `.gitignore`, and any other already-excluded files). Easiest mechanism: pass `--delete-excluded` plus an `--exclude` for each preserve target, or `rm -f` the asset glob before rsync. Either works; preference is `rm -f` of the explicit asset glob because it's grep-able.
+- `App/Resources/webview/assets/` — delete the 10 stale `index-*.js` files and their `.map` siblings in the same commit so the directory is clean post-merge.
+- `.github/workflows/scorecard.yml` — new file. Trigger `schedule: cron: "0 6 * * 1"` (weekly Mondays) + `push: branches: [develop]` + `workflow_dispatch`. Permissions: `id-token: write`, `contents: read`, `security-events: write`. Uses `ossf/scorecard-action@v2.4.0` (or latest at execution time). Upload SARIF to GitHub code-scanning.
+- `README.md` — add OpenSSF Scorecard badge near the top (under the title or alongside existing badges). Badge URL pattern is the standard `https://api.scorecard.dev/projects/github.com/KofTwentyTwo/Jarvis/badge`.
+
+**Closes:** #46, #178.
+
+### PR 3 — "Small code fixes" (`fix/v0.2-wave-1-code-fixes`)
+
+**Issues:** #37 (STTBackendSelector snake_case vs call-site camelCase) + #47 (WebviewBridge NaN→null JSON guard for `audioLevel`).
+
+**Why grouped:** both are genuine one-shot code fixes with a unit test each, in different packages (Voice/App + Bus). Landing them together gives one Swift-package-test cycle for both fixes.
+
+**Deliverables for #37:**
+
+- Decision: standardise on **camelCase** at call sites (matches existing identifier style elsewhere in the codebase). Update the constants in `STTBackendSelector` to camelCase string values, OR delete the unused constants in favour of an enum if call sites can adopt cheaply. Master agent picks the smaller diff after a quick read.
+- Single rename or constant-shape change; ensure `App/AppDelegate.swift:2028` (and any other call sites a grep reveals) is consistent.
+- Existing tests must still pass; if none cover the constant identity, add a small one in `packages/Voice/Tests/` asserting the chosen shape.
+
+**Deliverables for #47:**
+
+- First verify #47 is still latent or has become active post-#42: grep for the audioLevel consumer in `webview/packages/hud/` and check whether real decode errors are firing in any logged scenario. Either way, ship the clamp.
+- `packages/Bus/Sources/Bus/WebviewBridge.swift` — in `sendRaw` (or the `BusOutbound.audioLevel.encode` path, whichever is the cleanest place), replace non-finite `rms` floats with `0.0` before encoding. Inline fix only; do not refactor `AudioLevelEmitter` simultaneously (separate concern).
+- New unit test: `BusOutboundEncodingTests` (or extend the nearest existing suite) asserts `BusOutbound.audioLevel(rms: .nan)` decodes JS-side with `rms == 0.0`. Use the existing harness convention (`Mock*` only if recording, `Stub*` for canned data, `Fake*` for stateful) per CLAUDE.md.
+
+**Closes:** #37, #47.
+
+## Landing order
+
+1. **PR 2 first** — lowest blast radius, lightest CI footprint, validates the multi-runner pipeline on something we can revert in seconds if it breaks.
+2. **PR 1 second** — purely descriptive, no behavioural change. CI surface is large (Bus + Memory packages, full webview build), but new tests should not fail because no logic moves.
+3. **PR 3 last** — real code touches. By the time we get here, the Memory flake is the only plausible failure mode, so if a job goes red we know it is the flake.
+
+## Memory-flake disposition
+
+Per `docs/HANDOFF-2026-05-14-multi-runner-shipped.md`: `MemoryWiringEndToEndTests` flakes intermittently on PR branches under concurrent-runner load. Branch protection has zero required checks, so flake-induced red is not blocking.
+
+Master agent policy:
+
+1. If a PR's CI fails on Memory only, rerun the failed jobs once.
+2. If they pass on rerun, merge.
+3. If they fail a second time, leave the PR open with the Memory failure visible and open `bug,severity:medium,area:memory` issue documenting the test as known-flaky in v0.2. Do not merge without user sign-off.
+4. If any *non-Memory* check fails, stop and report back — do not rerun, do not merge.
+
+## Non-goals
+
+- Do not touch PRs #185, #191, #192. They are tracked in a separate disposition.
+- Do not open or address Wave 2/3 issues (#18, #23/#24, #43, #44, #176, #26, M-series epics).
+- Do not refactor `AudioLevelEmitter` while fixing #47.
+- Do not change branch protection settings.
+- Do not bypass pre-commit hooks (`--no-verify`).
+- Do not force-push.
+- Do not rebase #185/#191/#192 onto the new tip of develop.
+
+## Success criteria
+
+- Issues #19, #37, #45, #46, #47, #178 all closed via merge.
+- `develop` HEAD's CI is green after all three merges.
+- `App/Resources/webview/assets/` contains only the active bundle (one `index-*.js` + `.map` + `.css`) after PR 2 merges.
+- OpenSSF Scorecard workflow appears in Actions → Workflows and has at least one successful run (or a clearly-named permission failure to address in a follow-up).
+- Grep for `JarvisBusWorld` outside `.planning/`, `docs/`, and committed history returns nothing.
+- `swift test --package-path packages/Bus` passes including the new NaN-clamp test.
+- No new test doubles violate the `Mock*`/`Stub*`/`Fake*` convention.
+
+## Effort budget
+
+~90 min of coding across the three PRs, plus CI wait time. Master agent should target completion in 90 min wall clock; if it falls behind, it must report partial progress rather than rush.
@@ -43,6 +43,20 @@ public struct CacheHints: Sendable, Equatable {
    /// Once the system prompt grows past the threshold (e.g. memory
    /// hydration adds context, presence enrichment lengthens the prompt)
    /// the hint resumes automatically.
+    ///
+    /// **Current production state (as of v0.1):** `selfAwarePreamble` is
+    /// ~700 chars and the orchestrator is constructed with an `.empty`
+    /// initial context, so the threshold is never crossed on a healthy
+    /// boot — `eligibleForSystemPrompt` returns `nil` on every turn and
+    /// the prompt cache is effectively dormant. This is intentional
+    /// cost-acceptance: padding the preamble to ≥4096 chars would
+    /// require real prompt-engineering work and risks bleeding
+    /// irrelevant context into the model. Cache hits resume naturally
+    /// once Memory M-4 (v0.8 milestone) lands `searchFacts`-driven
+    /// hydration that grows the prompt past 1024 tokens organically;
+    /// see #23 for the audit context. Until then, every turn pays full
+    /// input-token cost — a ~30–80% cost reduction is on the table once
+    /// M-4 ships.
    public static func eligibleForSystemPrompt(
        _ systemPromptText: String,
        ttl: CacheTTL = .extended1h
@@ -508,6 +508,15 @@ public actor AgentOrchestrator {
        // for the current pass. Reset on each pass restart (initial entry +
        // T1→T2 escalation re-enter via `continue outer`).
        var assistantTextSoFar: String = ""
+        // #24: text accumulated since the most recent toolUse append (or
+        // since the start of the round if no toolUse yet). Distinct from
+        // `assistantTextSoFar` which stays cumulative for D-02 vision
+        // routing on endTurn. This buffer feeds the assistant message
+        // history alongside each toolUse block so the model sees its own
+        // narration ("Let me check..." → tool_use(get_time)) on subsequent
+        // round-trips; previously the text was dropped entirely from
+        // history and the model would deny ever saying it.
+        var assistantTextSinceLastFlush: String = ""
        // D-02: escalation gets ONE shot. The second low-confidence outcome
        // stays on T1 (the user got the response we have). This mirrors the
        // AGENT-09 retry budget but is independent of it.
@@ -575,6 +584,10 @@ public actor AgentOrchestrator {
                        // can hand it to VisionRouter.evaluatePostResponse.
                        // Reset on `continue outer` from the escalation arm.
                        assistantTextSoFar += s
+                        // #24: also accumulate into the per-tool-use flush
+                        // buffer so the next toolUseRequested attaches the
+                        // preceding narration to its assistant message.
+                        assistantTextSinceLastFlush += s
                        await replayLog.record(.textDelta(s), for: currentTurnId)
                        await events.send(.tokenDelta(turnId: currentTurnId, text: s))

@@ -601,9 +614,29 @@ public actor AgentOrchestrator {
                        // history so providers that require the original
                        // tool_use round-trip (Anthropic) can resolve the
                        // tool_use_id when the next tool_result lands.
-                        messages.append(LLMMessage(role: .assistant, content: [
-                            .toolUse(id: req.id, name: req.name, argsJSON: req.argsJSON),
-                        ]))
+                        //
+                        // #24: bundle any text the model emitted since the
+                        // last flush — Opus 4.7 commonly narrates ("Let me
+                        // check the time...") before calling a tool. Without
+                        // this the text was dropped from history and the
+                        // model would deny ever saying it on a follow-up.
+                        // Trim whitespace-only text to avoid emitting empty
+                        // text blocks (which Anthropic rejects with a 400).
+                        var assistantContent: [LLMMessage.ContentBlock] = []
+                        let pendingText = assistantTextSinceLastFlush
+                            .trimmingCharacters(in: .whitespacesAndNewlines)
+                        if !pendingText.isEmpty {
+                            assistantContent.append(.text(pendingText))
+                        }
+                        assistantContent.append(
+                            .toolUse(id: req.id, name: req.name, argsJSON: req.argsJSON)
+                        )
+                        messages.append(LLMMessage(role: .assistant, content: assistantContent))
+                        // Consumed — the next toolUseRequested in the same
+                        // round (rare but possible) gets only the text that
+                        // came BETWEEN the two tool_use blocks. D-02's
+                        // cumulative `assistantTextSoFar` is untouched.
+                        assistantTextSinceLastFlush = ""

                        do {
                            let resultData = try await toolDispatcher.dispatch(toolUse: req)
@@ -0,0 +1,210 @@
+import XCTest
+import AgentCore
+import Config
+import Replay
+@testable import AgentOrchestrator
+
+/// **#24 — assistant text blocks alongside tool_use must not be dropped.**
+///
+/// Opus 4.7 frequently narrates before invoking a tool ("Let me check the
+/// time for you..." → `tool_use(get_time)`). Before this fix, the
+/// orchestrator appended only the `.toolUse` block to `messages`, so on
+/// the next round-trip the model had no record of saying anything — a
+/// user follow-up like "what did you say before checking the time?"
+/// would get a denial.
+///
+/// These tests assert the assistant message at the round-trip boundary
+/// contains BOTH the preceding text and the tool_use block. D-02's
+/// cumulative `assistantTextSoFar` is preserved separately; only the
+/// per-flush buffer is consumed.
+@MainActor
+final class AssistantTextWithToolUseTests: XCTestCase {
+
+    var tempHome: TempReplayHome!
+
+    override func setUp() async throws {
+        try await super.setUp()
+        tempHome = TempReplayHome.make()
+    }
+
+    override func tearDown() async throws {
+        tempHome?.cleanup()
+        try await super.tearDown()
+    }
+
+    private func buildOrchestrator(
+        provider: MockLLMProvider,
+        dispatcher: any ToolDispatcher
+    ) async throws -> AgentOrchestrator {
+        let replay = try ReplayLog(databaseURL: tempHome.dbURL)
+        let session = try await replay.beginSession(appVersion: "test", buildSHA: "deadbeef")
+        return AgentOrchestrator(
+            configStore: makeConfigStore(),
+            providerFactory: { _ in provider },
+            toolDispatcher: dispatcher,
+            replayLog: replay,
+            sessionId: session,
+            systemPrompt: "you are jarvis",
+            availableTools: []
+        )
+    }
+
+    private func collectUntilIdle(orch: AgentOrchestrator, timeout: TimeInterval = 5.0) async {
+        let deadline = Date().addingTimeInterval(timeout)
+        for await event in orch.events {
+            if case .stateChange(.idle) = event { return }
+            if Date() > deadline { return }
+        }
+    }
+
+    // MARK: - ATU-1
+
+    /// Single text-then-tool sequence: model emits "Let me check..." then
+    /// `tool_use(get_time)`. The assistant message in the second provider
+    /// call must contain a `.text` block AND a `.toolUse` block.
+    func test_ATU1_textBeforeToolUseBundledIntoSameAssistantMessage() async throws {
+        let toolReq = ToolUseRequest(id: "tu1", name: "get_time", argsJSON: Data("{}".utf8))
+        let mock = MockLLMProvider(scripts: [
+            .init(events: [
+                .messageStart(LLMMessageStart(messageId: "m1", model: "x", usagePrefix: nil)),
+                .textDelta("Let me check the time for you... "),
+                .toolUseRequested(toolReq),
+                .stopReason(.toolUse),
+                .messageStop,
+            ]),
+            .init(events: [
+                .messageStart(LLMMessageStart(messageId: "m2", model: "x", usagePrefix: nil)),
+                .textDelta("It's 12:34."),
+                .stopReason(.endTurn),
+                .messageStop,
+            ]),
+        ])
+        let dispatcher = StubToolDispatcher(dispatch: { _ in Data("12:34".utf8) })
+
+        let orch = try await buildOrchestrator(provider: mock, dispatcher: dispatcher)
+        _ = await orch.submit(.text("what time"))
+        await collectUntilIdle(orch: orch)
+
+        let calls = await mock.getRecordedCalls()
+        XCTAssertEqual(calls.count, 2, "expected 2 provider calls (tool-use loop)")
+
+        // Find the assistant message containing the tool_use block in the
+        // second call's history.
+        let secondCall = calls[1].messages
+        guard let assistantWithTool = secondCall.first(where: { msg in
+            msg.role == .assistant && msg.content.contains(where: { block in
+                if case .toolUse(id: "tu1", _, _) = block { return true }
+                return false
+            })
+        }) else {
+            return XCTFail("no assistant message with the tool_use block found in second call")
+        }
+
+        // Block ordering: text first, then tool_use.
+        let blockKinds = assistantWithTool.content.map { block -> String in
+            switch block {
+            case .text: return "text"
+            case .toolUse: return "toolUse"
+            case .toolResult: return "toolResult"
+            }
+        }
+        XCTAssertEqual(blockKinds, ["text", "toolUse"],
+            "assistant block sequence must be [text, toolUse] — got \(blockKinds)")
+
+        // The text content matches what the model streamed (trimmed).
+        if case .text(let txt) = assistantWithTool.content[0] {
+            XCTAssertEqual(txt, "Let me check the time for you...",
+                "text block must contain the streamed narration (trimmed)")
+        } else {
+            XCTFail("expected first block to be .text")
+        }
+    }
+
+    // MARK: - ATU-2
+
+    /// If the model emits a tool_use with NO preceding text (Opus
+    /// occasionally does this — straight to tool), the assistant message
+    /// must contain ONLY the toolUse block (no empty `.text("")` block,
+    /// which Anthropic rejects with a 400).
+    func test_ATU2_noPrecedingTextOnlyToolUseBlock() async throws {
+        let toolReq = ToolUseRequest(id: "tu2", name: "get_time", argsJSON: Data("{}".utf8))
+        let mock = MockLLMProvider(scripts: [
+            .init(events: [
+                .messageStart(LLMMessageStart(messageId: "m1", model: "x", usagePrefix: nil)),
+                .toolUseRequested(toolReq),
+                .stopReason(.toolUse),
+                .messageStop,
+            ]),
+            .init(events: [
+                .messageStart(LLMMessageStart(messageId: "m2", model: "x", usagePrefix: nil)),
+                .textDelta("done"),
+                .stopReason(.endTurn),
+                .messageStop,
+            ]),
+        ])
+        let dispatcher = StubToolDispatcher(dispatch: { _ in Data("12:34".utf8) })
+
+        let orch = try await buildOrchestrator(provider: mock, dispatcher: dispatcher)
+        _ = await orch.submit(.text("time"))
+        await collectUntilIdle(orch: orch)
+
+        let calls = await mock.getRecordedCalls()
+        guard let assistantWithTool = calls[1].messages.first(where: { msg in
+            msg.role == .assistant && msg.content.contains(where: { block in
+                if case .toolUse(id: "tu2", _, _) = block { return true }
+                return false
+            })
+        }) else {
+            return XCTFail("expected assistant message with tool_use block")
+        }
+
+        XCTAssertEqual(assistantWithTool.content.count, 1,
+            "no preceding text — assistant message must contain only the toolUse block, got \(assistantWithTool.content.count) blocks")
+        if case .toolUse = assistantWithTool.content[0] {
+            // ok
+        } else {
+            XCTFail("expected sole content block to be .toolUse")
+        }
+    }
+
+    // MARK: - ATU-3
+
+    /// Whitespace-only preceding text must NOT be emitted as a `.text`
+    /// block (Anthropic rejects empty/whitespace-only text blocks). The
+    /// implementation trims whitespace before testing for emptiness.
+    func test_ATU3_whitespaceOnlyTextNotEmitted() async throws {
+        let toolReq = ToolUseRequest(id: "tu3", name: "get_time", argsJSON: Data("{}".utf8))
+        let mock = MockLLMProvider(scripts: [
+            .init(events: [
+                .messageStart(LLMMessageStart(messageId: "m1", model: "x", usagePrefix: nil)),
+                .textDelta("   \n\t  "),  // whitespace only
+                .toolUseRequested(toolReq),
+                .stopReason(.toolUse),
+                .messageStop,
+            ]),
+            .init(events: [
+                .messageStart(LLMMessageStart(messageId: "m2", model: "x", usagePrefix: nil)),
+                .stopReason(.endTurn),
+                .messageStop,
+            ]),
+        ])
+        let dispatcher = StubToolDispatcher(dispatch: { _ in Data("ok".utf8) })
+
+        let orch = try await buildOrchestrator(provider: mock, dispatcher: dispatcher)
+        _ = await orch.submit(.text("hi"))
+        await collectUntilIdle(orch: orch)
+
+        let calls = await mock.getRecordedCalls()
+        guard let assistantWithTool = calls[1].messages.first(where: { msg in
+            msg.role == .assistant && msg.content.contains(where: { block in
+                if case .toolUse(id: "tu3", _, _) = block { return true }
+                return false
+            })
+        }) else {
+            return XCTFail("expected assistant message with tool_use block")
+        }
+
+        XCTAssertEqual(assistantWithTool.content.count, 1,
+            "whitespace-only preceding text must be dropped, leaving only the toolUse block")
+    }
+}
@@ -159,7 +159,20 @@ extension BusOutbound: Codable {
            try container.encode(text, forKey: .text)
        case .audioLevel(let rms):
            try container.encode(Discriminator.audioLevel, forKey: .type)
-            try container.encode(rms, forKey: .rms)
+            // Issue #47: JSON has no canonical encoding for `Float.nan` or
+            // `±Float.infinity`. Foundation's `JSONEncoder` defaults to
+            // throwing for non-finite floats (`JSONEncoder.NonConformingFloat-
+            // EncodingStrategy.throw`), which would crash the per-frame bus
+            // path the moment `AudioLevelEmitter` produces a NaN — easy to
+            // hit on a CoreAudio over-/underflow or a buffer of all-zero
+            // samples. Clamp non-finite to 0.0 inline at the encode site so
+            // the rms float in the wire JSON is always a finite, ring-shader-
+            // safe number. The HUD-side consumer (`webview/packages/hud/src/
+            // bus/client.ts`) currently no-ops on audioLevel; the clamp is
+            // defensive against the bug becoming active once the consumer is
+            // wired up.
+            let safe = rms.isFinite ? rms : 0.0
+            try container.encode(safe, forKey: .rms)
        case .toolCallStart(let id, let name, let argsPreview):
            try container.encode(Discriminator.toolCallStart, forKey: .type)
            try container.encode(id.uuidString.lowercased(), forKey: .id)
@@ -1,6 +1,7 @@
-// Injected at document-start in WKContentWorld("JarvisBusWorld").
-// Installs window.jarvisBus with the minimum surface needed to complete
-// the handshake.
+// Injected at document-start in the page world (the default
+// `WKContentWorld.page` since `fb41c5f` retired the prior isolated named
+// world; see WebviewBridge.swift). Installs window.jarvisBus with the
+// minimum surface needed to complete the handshake.
 //
 // IMPORTANT — protocolVersion drift hazard: the full TS bundle's installer
 // (Vy() in the minified output) shape-checks window.jarvisBus and SKIPS
@@ -6,8 +6,9 @@ import JarvisLogging
 /// Main-actor-bound bridge between Swift and the webview's JS layer.
 ///
 /// Responsibilities:
-/// - Install a `WKScriptMessageHandlerWithReply` in an isolated content world
-///   (`JarvisBusWorld`) so page scripts cannot observe or post to it.
+/// - Install a `WKScriptMessageHandlerWithReply` on the page world
+///   (`WKContentWorld.page`, the default since `fb41c5f`) so page scripts
+///   in the loaded HUD bundle can address it via `window.webkit.messageHandlers.jarvisBus`.
 /// - Inject a `WKUserScript` at `.atDocumentStart` in the same world so
 ///   `window.jarvisBus` exists before any page script runs (pitfall 3).
 /// - Decode inbound JSON into `BusInbound` and either (a) handle the handshake
@@ -123,6 +123,44 @@ final class CodableRoundTripTests: XCTestCase {
        )
    }

+    /// Issue #47: `BusOutbound.audioLevel` encodes a `Float`, and Foundation's
+    /// default `JSONEncoder.NonConformingFloatEncodingStrategy` is `.throw`.
+    /// A non-finite `rms` (NaN from a CoreAudio over-/underflow, or ±∞ from a
+    /// pathological emitter buffer) would crash the per-frame bus path inside
+    /// `WebviewBridge.sendRaw`. The encode site in `BusOutbound.swift` clamps
+    /// non-finite floats to `0.0`; this test pins the contract so the clamp
+    /// can't silently regress.
+    func test_audioLevel_nanIsClampedToZero() throws {
+        let encoded = try BusCoder.makeEncoder().encode(BusOutbound.audioLevel(rms: .nan))
+        let decoded = try BusCoder.makeDecoder().decode(BusOutbound.self, from: encoded)
+        guard case .audioLevel(let rms) = decoded else {
+            XCTFail("expected .audioLevel after round-trip, got \(decoded)")
+            return
+        }
+        XCTAssertEqual(rms, 0.0, "NaN rms must be clamped to 0.0 at encode time (#47)")
+        XCTAssertTrue(rms.isFinite, "post-clamp rms must be finite")
+    }
+
+    /// Companion to `test_audioLevel_nanIsClampedToZero` — covers the ±∞
+    /// edge cases. CoreAudio drivers don't normally emit infinities, but the
+    /// encode-time clamp is sourced from `Float.isFinite`, which covers both
+    /// NaN and ±∞ via one check; locking the contract here prevents a future
+    /// refactor from narrowing the guard to NaN-only.
+    func test_audioLevel_infinityIsClampedToZero() throws {
+        let encoder = BusCoder.makeEncoder()
+        let decoder = BusCoder.makeDecoder()
+
+        for value in [Float.infinity, -Float.infinity] {
+            let encoded = try encoder.encode(BusOutbound.audioLevel(rms: value))
+            let decoded = try decoder.decode(BusOutbound.self, from: encoded)
+            guard case .audioLevel(let rms) = decoded else {
+                XCTFail("expected .audioLevel for \(value)")
+                continue
+            }
+            XCTAssertEqual(rms, 0.0, "±∞ rms (\(value)) must be clamped to 0.0 (#47)")
+        }
+    }
+
    func test_roundTrip_toolCallStart() throws {
        let id = UUID(uuidString: "550e8400-e29b-41d4-a716-446655440000")!
        try assertOutboundRoundTrips(
@@ -15,7 +15,7 @@ import WebKit
 /// **Why this exists** — Phase D-Phase-2 / Phase F1 of
 /// `.planning/AUDIT-AND-FIX-PLAN.md`. The 2026-05-02 audit-and-fix
 /// session uncovered a `WKContentWorld` isolation bug
-/// (`Injection.js` ran in `WKContentWorld(name: "JarvisBusWorld")`
+/// (`Injection.js` ran in a prior isolated named content world
 /// while the bundle ran in the page world; `window.jarvisBus`
 /// properties were per-world isolated; bundle's `attachBus()` saw
 /// `undefined` and created a SECOND fresh bus object; two stubs, two
@@ -127,13 +127,13 @@ final class RealWKWebViewIntegrationTests: XCTestCase {

    /// **The test that would have caught the WKContentWorld bug.**
    ///
-    /// On the bug branch, `Injection.js` ran in
-    /// `WKContentWorld(name: "JarvisBusWorld")` while
-    /// `WebviewBridge.sendRaw` invoked `callAsyncJavaScript` against
-    /// the same named world — but the message *handler* was registered
-    /// against `WKContentWorld.page`, so JS-side `postMessage` calls
-    /// from the named-world stub never reached Swift. Handshake hung
-    /// at `.sentHello`, eventually timed out with an `NSAlert`.
+    /// On the bug branch, `Injection.js` ran in the prior named
+    /// content world while `WebviewBridge.sendRaw` invoked
+    /// `callAsyncJavaScript` against the same named world — but the
+    /// message *handler* was registered against `WKContentWorld.page`,
+    /// so JS-side `postMessage` calls from the named-world stub never
+    /// reached Swift. Handshake hung at `.sentHello`, eventually timed
+    /// out with an `NSAlert`.
    ///
    /// On the fixed branch (fb41c5f), Injection.js auto-acks `hello`
    /// at the stub layer in the unified `.page` world, the bridge's
@@ -78,8 +78,8 @@ final class WebviewBridgeOutboundTests: XCTestCase {
            .decode(BusOutbound.self, from: Data((payload ?? "").utf8))
        XCTAssertEqual(decoded, .hudState(.idle))
        // F-A2-01: production WebviewBridge uses `WKContentWorld.page` (the
-        // default since `fb41c5f`); the prior `JarvisBusWorld` named-world
-        // assertion was retired in that commit but this test wasn't updated.
+        // default since `fb41c5f`); the prior named-world assertion was
+        // retired in that commit but this test wasn't updated until #45.
        XCTAssertEqual(call.contentWorld, .page)
    }

@@ -160,11 +160,14 @@ final class WebviewBridgeTests: XCTestCase {
        // Re-adding a handler with the same (contentWorld, name) pair would
        // raise `NSInvalidArgumentException`. The constructor succeeding is
        // itself evidence of registration. As a double-check, `remove` on a
-        // registered handler must succeed without throwing.
+        // registered handler must succeed without throwing. We remove from
+        // `WKContentWorld.page` because that is the default world
+        // `WebviewBridge.init` registers the handler against since `fb41c5f`
+        // unified bus and bundle into the same page world.
        let controller = webView.configuration.userContentController
        controller.removeScriptMessageHandler(
            forName: "jarvisBus",
-            contentWorld: WKContentWorld.world(name: "JarvisBusWorld")
+            contentWorld: .page
        )
        // No XCTAssert needed — the remove call succeeding on a registered
        // handler is the assertion.
@@ -12,7 +12,7 @@ import Replay

 /// Test seam over `Replay.ReplayLog`. The Memory module deliberately does NOT
 /// depend on a concrete `ReplayLog` instance — `AppDelegate.installMemory`
-/// (Plan 07-06) hands a wrapper through `setReplayLog`. Tests inject a
+/// hands a wrapper through `setReplayLog` at startup. Tests inject a
 /// mock conforming to this protocol.
 public protocol MemoryReplaySink: Sendable {
    func record(_ event: ReplayEvent, forTriggerTurnId triggerTurnId: Int64)
@@ -34,8 +34,10 @@ public protocol MemoryReplaySink: Sendable {
 /// memory store the app refuses to launch. This eliminates the silent
 /// dormancy that hid B-02 / B-04-class substrate failures.
 ///
-/// The actor is the only writer. Plan 07-02 wires the bounded extraction
-/// channel; Plan 07-03 wires the search path + ReplayEvent emission.
+/// The actor is the only writer. The bounded extraction channel writes
+/// here via `applyOp`; the hybrid (FTS5 + vec0) search path reads through
+/// `MemoryReadStore` conformance and emits `ReplayEvent.memoryRetrieval`
+/// via `recordRetrieval`.
 public actor MemoryStore {

    public static let logChannel = "memory"
@@ -331,11 +333,11 @@ public actor MemoryStore {
        sink.record(.memoryMutation(bytes), forTriggerTurnId: triggerTurnId)
    }

-    // MARK: - Read path (Plan 07-03 — MEM-07 + TEXT-03 + D-02 + D-05)
+    // MARK: - Read path (MEM-07 + TEXT-03 + D-02 + D-05)

    /// SOLE production emission site for ReplayEvent.memoryRetrieval (D-05).
-    /// Mirrors recordMemoryMutation from 07-02. Best-effort; if no replay
-    /// sink has been injected, the call is a silent no-op.
+    /// Mirrors `recordMemoryMutation` on the write path. Best-effort; if no
+    /// replay sink has been injected, the call is a silent no-op.
    public func recordRetrieval(_ ref: FactRef, triggerTurnId: Int64) {
        guard let sink = replayLog else { return }
        let stamp: Int64 = ref.timestamp == 0
@@ -538,13 +540,13 @@ public actor MemoryStore {
    }
 }

-// MARK: - HybridSearch + SessionHistory conformance (Plan 07-03 Task 2)
+// MARK: - HybridSearch + SessionHistory conformance

 extension MemoryStore: MemoryReadStore {}
 extension MemoryStore: SessionHistoryReading {}
 extension OllamaEmbeddingClient: EmbeddingProviding {}

-// MARK: - Plan 07-06 / Regression corpus test seams
+// MARK: - Regression corpus test seams
 //
 // These small read-only seams are exercised exclusively by the env-gated
 // MemoryRegressionCorpusTests suite (JARVIS_REAL_MODELS=1). They wrap raw
@@ -12,7 +12,7 @@ import OllamaProvider
 //
 // Gated by JARVIS_REAL_MODELS=1 — skipped in CI. Runs interactively against
 // a real local Ollama with both `nomic-embed-text` (embedding) and
-// `qwen2.5-coder:32b` (extractor) pulled.
+// `qwen3.6:latest` (extractor, per commit 8d0a959) pulled.
 //
 // Run interactively on the developer's Apple Silicon Mac:
 //   JARVIS_REAL_MODELS=1 swift test --package-path packages/Memory --filter MemoryRegressionCorpusTests
@@ -37,7 +37,7 @@ final class MemoryRegressionCorpusTests: XCTestCase {
    private func skipIfNotRealModels() throws {
        try XCTSkipUnless(
            ProcessInfo.processInfo.environment["JARVIS_REAL_MODELS"] == "1",
-            "Set JARVIS_REAL_MODELS=1 to run memory regression corpus (requires local Ollama with nomic-embed-text + qwen2.5-coder:32b, plus a loadable vec0.dylib)"
+            "Set JARVIS_REAL_MODELS=1 to run memory regression corpus (requires local Ollama with nomic-embed-text + qwen3.6:latest, plus a loadable vec0.dylib)"
        )
    }

@@ -28,6 +28,16 @@ public protocol STTWarningLogger: Sendable {
 public enum STTBackendSelector {

    // MARK: - Valid backend strings
+    //
+    // Canonical backend identifiers. Snake_case is the wire shape stored in
+    // `PerTurnSnapshot.stt.backend`. Call sites (`AppDelegate.sttBackend`)
+    // SHOULD reference these constants rather than the bare string literals;
+    // `STTBackendSwitchTests.testB5_canonicalConstantsRoundTripThroughMake`
+    // locks the round-trip contract by asserting both that the constants
+    // map to the right `STTProvider` AND that camelCase string literals
+    // (the pre-2026-05-12 anti-pattern from Issue #29 / #37) fall through
+    // to the default branch with a warning. Do not rename without updating
+    // `PerTurnSnapshot` and the regression test in lockstep.

    public static let backendSpeechAnalyzer = "speech_analyzer"
    public static let backendWhisperKit = "whisperkit"
@@ -77,6 +77,16 @@ pnpm --filter @jarvis/hud build

 DEST="$REPO_ROOT/App/Resources/webview"
 mkdir -p "$DEST"
+mkdir -p "$DEST/assets"
+
+# Issue #46: Vite emits content-hashed asset names (e.g. `index-CXPb0I5z.js`),
+# so the active bundle filename changes on every dependency or source edit.
+# Without explicit cleanup, stale chunks accumulate in `assets/` because we
+# don't pass `--delete` to rsync below. Sweep the asset glob (JS + maps + CSS)
+# before the rsync drops the fresh bundle. `bus-harness.html`, `index.html`,
+# and the directory's `.gitignore` live one level up (in $DEST), not in
+# $DEST/assets, so they are not at risk from this rm.
+rm -f "$DEST"/assets/*.js "$DEST"/assets/*.js.map "$DEST"/assets/*.css "$DEST"/assets/*.css.map

 echo "[build-webview] rsync dist/ -> $DEST"
 # --exclude bus-harness.html: keep 02-03's committed copy intact (parity
@@ -85,6 +95,7 @@ echo "[build-webview] rsync dist/ -> $DEST"
 # --exclude .gitignore: preserve the App/Resources/webview/.gitignore that
 #   lets index.html stay source-committed while assets/ regenerates.
 # We do NOT pass --delete: other vendored assets coexist in this dir.
+# (The explicit `rm -f` above handles stale `assets/` chunks; see #46.)
 rsync -a \
      --exclude=bus-harness.html \
      --exclude=.gitignore \
@@ -0,0 +1,54 @@
+#!/usr/bin/env bash
+#
+# check-no-jarvis-bus-world-references.sh
+#
+# Issue #45 boundary gate.
+#
+# The named `WKContentWorld("JarvisBusWorld")` was retired in `fb41c5f` —
+# the bus message handler and `Injection.js` user script now both register
+# against `WKContentWorld.page` (the default page world). Old doc-comments
+# and test fixtures hung onto the name for months afterward; #45 closed
+# the cleanup.
+#
+# This gate prevents new code or comments from re-introducing the literal
+# string `JarvisBusWorld` outside the historical reference corpora:
+#
+#   - `.planning/`              archived plans / audit reports / handoffs
+#   - `docs/`                   spec documents that explain the migration
+#   - `node_modules/`           vendored deps (out of repo control anyway)
+#   - `.build/`, build/         SwiftPM + Xcode derived data
+#   - `webview/packages/*/dist/` TypeScript build output (regenerated from src)
+#   - this script itself        we have to reference the literal to grep for it
+#   - `.git/`                   git internals
+#
+# Fails CI if any other tracked file contains the literal string.
+
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+REPO_ROOT="$(dirname "$SCRIPT_DIR")"
+
+cd "$REPO_ROOT"
+
+# Use `git grep` so we automatically respect .gitignore (so dist/ etc. that
+# aren't tracked don't trip the gate even if they exist on disk locally).
+# `--cached` would miss new working-tree files; default mode greps the
+# working tree filtered by tracked-or-untracked-not-ignored, which is what
+# we want before a commit.
+matches=$(git grep --no-color -n "JarvisBusWorld" -- \
+    ':!.planning/' \
+    ':!docs/' \
+    ':!scripts/check-no-jarvis-bus-world-references.sh' \
+    ':!webview/packages/*/dist/' \
+    || true)
+
+if [ -n "$matches" ]; then
+    echo "[check-no-jarvis-bus-world-references] FAIL: tracked source still references the retired 'JarvisBusWorld' named content world." >&2
+    echo "" >&2
+    echo "$matches" >&2
+    echo "" >&2
+    echo "[check-no-jarvis-bus-world-references] Rewrite these to 'page world' / 'shared world' / 'defensive double-install guard' — see #45 and the JSDoc in webview/packages/bus/src/bridge.ts for the canonical phrasing." >&2
+    exit 1
+fi
+
+echo "[check-no-jarvis-bus-world-references] PASS (no tracked source references the retired named world)"
@@ -1,7 +1,7 @@
 {
  "name": "@jarvis/bus",
  "private": true,
-  "version": "0.0.0",
+  "version": "0.0.1",
  "type": "module",
  "main": "dist/index.js",
  "types": "dist/index.d.ts",
@@ -57,27 +57,28 @@ export interface InstallOptions {
 * - `send(msg)` JSON-encodes and posts to `webkit.messageHandlers.jarvisBus`.
 *   Callers await the returned Promise to observe Swift's reply.
 *
- * ## WKContentWorld attach-to-existing path (H-02)
+ * ## Defensive double-install guard (H-02)
 *
- * When the HUD bundle runs in the page's default JS world (via
- * `<script type="module" src="…">`), `webkit.messageHandlers.jarvisBus` is
- * NOT visible — Swift registered the handler `in: JarvisBusWorld` and
- * WebKit isolates message handlers per content world. However,
- * `window.jarvisBus` is a property on the (cross-world-shared) `window`
- * object, so the bus instance installed by `Injection.js` inside
- * `JarvisBusWorld` IS observable from the default world.
+ * Both `Injection.js` (a `WKUserScript` injected at `.atDocumentStart`) and
+ * this bundle's `installJarvisBus()` run in the SAME page world — the
+ * default `WKContentWorld.page` since `fb41c5f` retired the prior
+ * isolated named world. Injection.js fires first and stamps a
+ * minimal `window.jarvisBus` with `protocolVersion`, `_handler`, plus the
+ * `send` / `receive` / `onOutbound` surface needed for the handshake.
+ * The bundle's installer fires later (top-level side-effect of `main.tsx`),
+ * and if we naively overwrote `window.jarvisBus` we would orphan
+ * Injection.js's queued early messages and any handshake state already
+ * captured against the original closure.
 *
- * If we detect a pre-existing `window.jarvisBus` with the expected
- * `send` + `onOutbound` + `receive` shape, we attach our options to it
- * and return it verbatim rather than overwriting it. This routes every
- * outbound `send()` through the JarvisBusWorld-captured message handler
- * reference and keeps `onOutbound` hooked to the same `_handler` slot the
- * Injection.js bus exposes. Without this branch the default-world bundle
- * replaces `window.jarvisBus` with a bus whose `send()` cannot reach
- * `webkit.messageHandlers.jarvisBus` and every outbound frame — including
- * the first `uiReady` — is lost.
+ * The shape check below (`isJarvisBusShape`) is therefore a defensive
+ * double-install guard inside one page world, not a cross-world attach.
+ * When we detect a pre-existing `window.jarvisBus` with the canonical
+ * `send` + `onOutbound` + `receive` surface, we return early and leave
+ * Injection.js's instance in place. Outbound `send()` calls still reach
+ * `window.webkit.messageHandlers.jarvisBus` because that handler lives on
+ * the same page world too.
 *
- * bus-harness.html (Phase 2) loads its marker script in a page where
+ * `bus-harness.html` (Phase 2) loads its marker script in a page where
 * Injection.js has installed the same bus; the harness never called `send()`
 * so the regression wasn't observed until Phase 3. For that path this
 * function still returns the pre-installed bus, which is exactly what the
@@ -93,10 +94,11 @@ export function installJarvisBus(options: InstallOptions = {}): () => void {
      console.error("[bus] decode failed:", error, raw);
    });

-  // H-02: attach-to-existing path. Injection.js in JarvisBusWorld installs a
-  // minimal bus with the same surface; if it's already on window we MUST NOT
-  // replace it (that would orphan webkit.messageHandlers.jarvisBus, which is
-  // only visible inside JarvisBusWorld).
+  // H-02: defensive double-install guard. Injection.js runs first (at
+  // document-start in the same page world) and stamps a minimal bus with the
+  // same surface; if it's already on `window`, we MUST NOT replace it — doing
+  // so would orphan Injection.js's queued early messages and captured handshake
+  // state. See the class doc above for the full rationale.
  const existing: unknown =
    typeof window !== "undefined"
      ? (window as unknown as { jarvisBus?: unknown }).jarvisBus
@@ -156,7 +158,7 @@ export function installJarvisBus(options: InstallOptions = {}): () => void {
 /**
 * Duck-type guard for a pre-existing `window.jarvisBus`. Matches the shape
 * installed by `packages/Bus/Sources/Bus/Resources/Injection.js` at
- * document-start in JarvisBusWorld. Intentionally narrow: if the shape
+ * document-start in the page world. Intentionally narrow: if the shape
 * doesn't match (e.g. a left-over test stub), we fall through to the
 * fresh-install branch and overwrite it.
 */
@@ -74,14 +74,13 @@ describe("installJarvisBus", () => {
    expect(b).toHaveBeenCalledOnce();
  });

-  // H-02: the HUD bundle (`main.tsx`) runs in the default JS world, but
-  // Swift registered `webkit.messageHandlers.jarvisBus` in JarvisBusWorld,
-  // so a default-world `send()` cannot reach the message handler. The
-  // Injection.js-installed bus in JarvisBusWorld CAN reach it (it captured
-  // the reference at document-start inside that world), and because
-  // `window.jarvisBus` is a cross-world-shared property the default world
-  // observes the Injection.js bus. Therefore `installJarvisBus()` must
-  // attach to — not replace — a pre-existing bus.
+  // H-02: both `Injection.js` (`WKUserScript` at document-start) and this
+  // bundle's `installJarvisBus()` run in the SAME page world since `fb41c5f`
+  // retired the prior isolated named content world. Injection.js fires first
+  // and stamps a minimal `window.jarvisBus`; if `installJarvisBus()` naively
+  // overwrote it, the early handshake state and queued messages captured
+  // against the original closure would be orphaned. Therefore the bundle's
+  // installer must attach to — not replace — a pre-existing bus.
  describe("H-02 attach-to-existing (pre-installed window.jarvisBus)", () => {
    it("returns the pre-existing bus unchanged when one is already installed", () => {
      const preExistingSend = vi.fn().mockResolvedValue(undefined);
@@ -96,9 +95,10 @@ describe("installJarvisBus", () => {
      // @ts-expect-error — test shim: seed a pre-installed bus shape matching Injection.js
      globalThis.window = {
        jarvisBus: preExisting,
-        // Note: webkit is intentionally absent — in real JarvisBusWorld the
-        // pre-existing bus captured webkit at install time; the default world
-        // never sees webkit. The fix MUST NOT touch the pre-existing bus.
+        // Note: webkit is intentionally absent — Injection.js's pre-existing
+        // bus captures a reference to `window.webkit` at install time, and the
+        // shape guard must trust the pre-existing bus to route through that
+        // captured reference rather than re-resolving `webkit` here.
      };

      installJarvisBus();
@@ -110,7 +110,7 @@ describe("installJarvisBus", () => {
      expect(window.jarvisBus.onOutbound).toBe(preExistingOnOutbound);
    });

-    it("routes send() through the pre-existing bus's send (reaches JarvisBusWorld-captured handler)", async () => {
+    it("routes send() through the pre-existing bus's send (reaches Injection.js-captured handler)", async () => {
      const preExistingSend = vi.fn().mockResolvedValue({ ok: true });
      const preExisting = {
        protocolVersion: "2.0.0",
@@ -5,15 +5,15 @@ import { attachBus } from './bus/client'
 import './theme/tokens.css'

 // `attachBus()` runs in the page's default JS world (this module loads via
-// `<script type="module" src="…">` in index.html). Swift installed the real
-// `window.jarvisBus` inside `JarvisBusWorld` via Injection.js — WebKit
-// content worlds isolate `webkit.messageHandlers`, so the default-world
-// bundle can't reach them directly. `window.jarvisBus` IS visible though,
-// because WebKit keeps `window` as one cross-world object (only *variables*
-// are world-scoped). The H-02 fix in `@jarvis/bus` detects the pre-existing
-// `window.jarvisBus` and attaches to it instead of replacing it, which
-// routes every outbound `send()` back through the JarvisBusWorld-captured
-// message handler reference. See `packages/bus/src/bridge.ts` for details.
+// `<script type="module" src="…">` in index.html). Swift's
+// `WebviewBridge.installScriptHandlers` (since `fb41c5f`) registers the
+// message handler on the same `WKContentWorld.page`, and Injection.js — also
+// in the page world — has already stamped a minimal `window.jarvisBus` at
+// document-start. `installJarvisBus()` here detects that pre-existing
+// instance via a duck-type shape guard and returns early, leaving the
+// Injection.js bus untouched so its captured handshake state and queued
+// early messages are preserved. See `packages/bus/src/bridge.ts` for the
+// full defensive-double-install rationale.
 attachBus()

 const rootEl = document.getElementById('root')
Author	SHA1	Message	Date
James Maes	7c4bb248d5	docs: replace remaining Mermaid text diagrams with PlantUML PNG images	2026-07-23 17:09:16 -05:00
James Maes	29b6cc9d01	docs: replace Mermaid text diagram with PlantUML PNG image	2026-07-23 17:03:30 -05:00
james.maesandClaude Opus 4.7	034a115009	docs(superpowers): local-first LLM routing design spec (#197 ) * docs(superpowers): local-first LLM routing design spec Flip default chat-turn provider from Anthropic to local Ollama with reactive escalation to Opus on detectable failure. Per-turn escalation budget, clean-restart UX with visible badge. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * docs(superpowers): local-first LLM routing implementation plan 10-task plan with TDD steps, exact file paths, and code blocks for each step. Self-reviewed for spec coverage and type consistency. Ready for subagent-driven execution. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>	2026-05-16 16:18:54 -05:00
james.maesandClaude Opus 4.7	5ebe510627	docs: scrub Phase-7 + JarvisBusWorld doc drift (#19 , #45 ) (#195 ) * docs(memory): scrub Phase-7 hedge language from MemoryStore (#19) Audit 2026-05-12 flagged five doc-comment references that mis-described the implementation state. The cited plans (07-02, 07-03, 07-06) shipped months ago — the bounded extraction channel, the hybrid FTS5 + vec0 read path, the ReplayEvent emission, the regression corpus test seams, and `installMemory`'s replay-sink wiring are all live, not "wires" or "will replace" promises. Five touchpoints rewritten to describe the code as it is: - `MemoryReplaySink` doc: drop the "Plan 07-06" plan reference; the seam is just used at startup. - `MemoryStore` actor doc: replace the "Plan 07-02 wires / Plan 07-03 wires" sentence with the actual contract (single writer; hybrid search through `MemoryReadStore`; `recordRetrieval` emits `memoryRetrieval`). - Read-path MARK: drop the "Plan 07-03" plan label; keep the requirement-ID list (MEM-07 / TEXT-03 / D-02 / D-05) since those are still useful for traceability to the audit corpus. - `recordRetrieval` doc: replace "Mirrors recordMemoryMutation from 07-02" with "Mirrors recordMemoryMutation on the write path." - Test-seams MARK: drop "Plan 07-06" prefix; describe by the test target (JARVIS_REAL_MODELS=1 corpus suite) instead. The 2026-05-11 D-5/D-6 markers and the "Boot policy (changed 2026-05-11)" doc are preserved unchanged — they document current behaviour, not historical hedges. `swift build --package-path packages/Memory` clean. Closes #19 Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * docs(bus): scrub JarvisBusWorld drift + add grep gate (#45) `fb41c5f` (2026-05-02) unified `Injection.js` and the HUD bundle into the default `WKContentWorld.page` and retired the prior isolated `JarvisBusWorld` named world. The doc-comments and test fixtures across 20+ lines kept describing the bug-era topology — pre-existing-bus attach "because cross-world isolation", message-handler-only-visible-in-named- world, etc. — for months afterward. Result: a new reader following the docs would conclude the bus uses two worlds when it actually uses one. Five rationale rewrites: - `WebviewBridge.swift` class doc — handler is on the page world, not an isolated named world. Page scripts in the loaded HUD bundle address it via `window.webkit.messageHandlers.jarvisBus`. - `Injection.js` header — names the page world explicitly; one-line nod to `fb41c5f` for the historical context. - `bridge.ts` JSDoc + `isJarvisBusShape` rationale — recast the H-02 attach-to-existing path as a "defensive double-install guard" inside ONE page world. Injection.js fires first at document-start; the bundle installer fires later as a top-level `main.tsx` side-effect; the guard prevents the bundle from clobbering Injection.js's queued early messages and captured handshake state. The duck-type `isJarvisBusShape` CODE is unchanged — only the comment narrative was misleading. - `webview/packages/hud/src/main.tsx` — drop the cross-world-isolation framing; describe the single-world reality. - `App/AppDelegate.swift` install comment — page world, not isolated. Test-side updates (same scrub, plus one latent test bug surfaced): - `WebviewBridgeTests.test_scriptMessageHandlerIsRegisteredUnderConfigured- Name`: was removing from `WKContentWorld.world(name: "JarvisBusWorld")` while the bridge registered against `.page` — silent no-op since `removeScriptMessageHandler` doesn't throw on absent. Now removes from `.page` (the world the bridge actually uses), making the "registration-is-evidence-of-success" assertion meaningful again. - `WebviewBridgeOutboundTests`: comment-only. - `RealWKWebViewIntegrationTests`: comment-only (forensic narrative kept, literal world name replaced with "prior isolated named content world"). - `webview/packages/bus/tests/bridge.test.ts`: H-02 describe-block + one it-name rewritten to match the new defensive-double-install rationale. Infrastructure: - `scripts/check-no-jarvis-bus-world-references.sh`: new boundary gate. Greps for the literal `JarvisBusWorld` in tracked files; excludes `.planning/`, `docs/`, `webview/packages//dist/`, and this script itself. Wired into `CLAUDE.md`'s gate listing; the CI `boundary-gates` job auto-discovers all `scripts/check-.sh`, so no workflow edit is needed. - `webview/packages/bus/package.json`: 0.0.0 → 0.0.1 (per spec — comment-only changes still bump patch so consumers can pin). - `webview/packages/bus/dist/bridge.{js,d.ts}`: rebuilt via `pnpm --filter @jarvis/bus build`; NOT committed because `dist/` is not tracked (the build phase regenerates on each xcodebuild invocation). Verified locally: - `swift build --package-path packages/Bus` clean - `pnpm --filter @jarvis/bus test` — 37/37 pass - `bash scripts/check-no-jarvis-bus-world-references.sh` — PASS - `bash scripts/check-bus-protocol-version.sh` — PASS (v2.3.0) - `bash scripts/check-bus-harness-parity.sh` — PASS Closes #45 Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>	2026-05-14 09:51:03 -05:00
james.maesandClaude Opus 4.7	a121d79174	fix: STT backend constant casing + audioLevel NaN guard (#37 , #47 ) (#196 ) * docs(voice): pin STT backend constant identity contract (#37) The runtime mismatch flagged by #37 was already fixed at the call sites in `App/AppDelegate.swift:1631-1642` and `:2562-2575` — both routes now reference `STTBackendSelector.backendSpeechAnalyzer` / `backendWhisperKit` (snake_case) instead of the pre-2026-05-12 camelCase string literals. The regression test `STTBackendSwitchTests.testB5_canonicalConstantsRoundTripThroughMake` already locks both halves of the contract: 1. The snake_case constants round-trip through `make(backend:)` without hitting the unknown-backend default branch. 2. The pre-fix camelCase shape (`"whisperKit"`) DOES land in the default branch and DOES emit an unknown-backend warning. So the #37 close action is just to make that contract self-documenting at the declaration site rather than relying on readers finding the test. Add a doc-comment to the two `public static let backend` constants that explains the wire shape, names the regression test, references Issues #29 + #37, and warns against renaming without updating `PerTurnSnapshot` and the test in lockstep. No code change; `swift test --package-path packages/Voice` continues to pass (26 tests). Closes #37 Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> fix(bus): clamp non-finite audioLevel rms to 0.0 at encode (#47) `BusOutbound.audioLevel` carries a `Float rms` that `AudioLevelEmitter` produces per audio frame (~100 Hz once the consumer is wired up). Foundation's `JSONEncoder.NonConformingFloatEncodingStrategy` defaults to `.throw`, so a single NaN — easy to land via a CoreAudio over-/underflow or an all-zero buffer fed through certain RMS shapes — would throw inside `WebviewBridge.sendRaw` and tear down the per-frame bus path. The HUD-side consumer in `webview/packages/hud/src/bus/ client.ts:135` is currently a no-op (the ring shader is still on the synthetic-sine animation from Plan 03-03), so the bug is LATENT, not actively firing — but the clamp is cheap, defensive, and pins the contract before the consumer activates. Inline clamp at the encode site (`BusOutbound.encode(to:)` switch arm for `.audioLevel`): replace non-finite floats with `0.0` via `rms.isFinite ? rms : 0.0` immediately before `container.encode(_:)`. No refactor of `AudioLevelEmitter` — the per-spec call here is defensive at the serialization boundary, where every encoder-level overflow path converges. Two new unit tests in `CodableRoundTripTests`: - `test_audioLevel_nanIsClampedToZero` — encodes `.audioLevel(rms: .nan)`, decodes, asserts `rms == 0.0` and `rms.isFinite`. - `test_audioLevel_infinityIsClampedToZero` — companion for ±∞ so a future refactor cannot narrow the guard to NaN-only. `swift test --package-path packages/Bus --filter CodableRoundTripTests` passes (30 tests, +2 from this commit). Closes #47 Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>	2026-05-14 09:50:50 -05:00
james.maes	31f0a8289a	docs(handoff): multi-runner shipped + v0.2 wave plan (#192 ) Captures the multi-runner expansion (Grogu-jarvis, -2, -3 live), auto-Release flow proof (v0.1.0-dev.64), Superpowers process established, and Memory test flake disposition. Lists v0.2 Wave 1 candidates (#19, #45, #46, #37, #178, #47) for next session pickup. Sibling to docs/HANDOFF-2026-05-13-v0.1-shipped.md. Refs #180 #188	2026-05-14 09:48:26 -05:00
james.maesandClaude Opus 4.7	439e1ee2f5	ci: build hygiene — webview cleanup + OpenSSF Scorecard (#46 , #178 ) (#194 ) * ci(build-webview): sweep stale hashed chunks before rsync (#46) Vite emits content-hashed asset names (e.g. `index-CXPb0I5z.js`), so the active bundle filename rotates on every source/dependency edit. The existing rsync runs without `--delete` (to preserve `bus-harness.html` and `.gitignore`), so old chunks accumulated in `App/Resources/webview/assets/` — observed 9 stale `index-.js` / `.js.map` pairs locally before this fix. Add an explicit `rm -f` of the asset glob (JS + JS map + CSS + CSS map) before the rsync drops the fresh bundle. The preserved files (`bus-harness.html`, `index.html`, `.gitignore`) live one level up in `$DEST`, not in `$DEST/assets`, so the rm cannot reach them. Verified locally: `bash scripts/build-webview.sh` now leaves exactly one `index-.js` + `.js.map` in `assets/` (matching the reference in `index.html`). Refs #46 Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * ci(scorecard): add OpenSSF Scorecard workflow + README badge (#178) Adds `.github/workflows/scorecard.yml` running `ossf/scorecard-action@v2.4.0` on: - Weekly cron (Mondays 06:00 UTC) for posture drift detection - Push to develop for prompt regression visibility - workflow_dispatch for ad-hoc runs Permissions follow the canonical Scorecard set (`id-token: write`, `contents: read`, `security-events: write`) with top-level `read-all` and job-scoped writes. Results are published to the OpenSSF public dataset (powers the README badge URL) and uploaded as SARIF to code-scanning. The workflow consumes no untrusted user inputs in `run:` blocks, so it is not exposed to workflow-injection patterns documented in github.blog/security/.../how-to-catch-github-actions-workflow-injections. README.md gains the standard Scorecard badge next to the existing license/platform/swift badges. Refs #178 Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>	2026-05-14 09:29:48 -05:00
james.maesandClaude Opus 4.7	dced4cb9a3	docs(superpowers): v0.2 Wave 1 sequencing + execution spec (#193 ) * docs(handoff): multi-runner shipped + v0.2 wave plan Captures the multi-runner expansion (Grogu-jarvis, -2, -3 live), auto-Release flow proof (v0.1.0-dev.64), Superpowers process established, and Memory test flake disposition. Lists v0.2 Wave 1 candidates (#19, #45, #46, #37, #178, #47) for next session pickup. Sibling to docs/HANDOFF-2026-05-13-v0.1-shipped.md. Refs #180 #188 * docs(superpowers): v0.2 Wave 1 sequencing + execution spec Three-PR plan for #19, #37, #45, #46, #47, #178 grouped by surface area: doc drift, build hygiene, small code fixes. Landing order is build hygiene first (lowest blast radius), doc drift second, code fixes last. Memory-flake disposition documented. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>	2026-05-14 09:08:24 -05:00
james.maes	00f1075ea5	fix(agent): bundle assistant text with toolUse + document cache hint dormancy (#185 ) Two related AgentCore tactical fixes from the 2026-05-12 audit. #24 — Assistant text blocks alongside tool_use were dropped from message history. Opus 4.7 commonly narrates ("Let me check the time for you...") before calling a tool. On the next round-trip the appended assistant message contained only the .toolUse block, so the model had no record of the narration — a user follow-up like "what did you say before checking the time?" would deny it. Fix: introduce `assistantTextSinceLastFlush` parallel to `assistantTextSoFar` (the latter stays cumulative for D-02 vision routing). On every toolUseRequested, flush the per-flush buffer into a .text(...) block bundled into the same assistant message as the .toolUse, trim whitespace-only text to avoid Anthropic 400s on empty text blocks, and reset the flush buffer so subsequent tool_uses in the same round get only the text emitted between them. D-02 untouched. #23 — CacheHints documentation flagged that the eligibility threshold (4096 chars ≈ 1024 tokens) is structurally false on healthy boot (preamble ~700 chars, empty initial context). Added a paragraph explicitly documenting the dormancy, the intentional cost-acceptance rationale, and the natural recovery once M-4 (v0.8) lands searchFacts-driven hydration. No code change. Tests: 3 new tests in AssistantTextWithToolUseTests cover text+toolUse bundling, toolUse-alone (no spurious empty text block), and whitespace-only-text rejection. Full AgentCore suite: 248 tests, 0 failures, 1 skipped. Closes #23, #24	2026-05-14 06:53:35 -05:00
james.maes	774e60a020	ci(actions): create GitHub Release directly in publish-tag job (#187 ) The default GITHUB_TOKEN cannot trigger downstream workflows (GitHub prevents the chain to avoid infinite loops). publish-tag's `git push origin "$TAG"` therefore never fires release.yml's `push: tags:` listener — observed on v0.1.0-dev.50, where the tag landed but the Release object was missing until manually created. Fold the Release creation into the publish-tag job using `gh release create` with GH_TOKEN=secrets.GITHUB_TOKEN. Direct API calls work fine with the default token; only the push→workflow chain is blocked. - Idempotent: skips if release already exists. - Notes: auto-generated from `git log $prev..$TAG` plus a Verification block linking back to the CI run. - Prerelease flag respects steps.version.outputs.prerelease (true for dev/rc tags, false for vX.Y.Z final tags on main). release.yml stays untouched as a fallback for manually-pushed tags (humans pushing tags directly CAN trigger workflows). Closes #186	2026-05-13 18:28:07 -05:00
james.maes	a26f49fdbd	docs(superpowers): multi-runner Grogu spec + implementation plan (#190 ) * docs(superpowers): brainstorm/spec for multi-runner expansion on Grogu Brainstorm-phase design for scaling Grogu from 1 to 3 self-hosted CI runners. Covers components, precondition (PR #189 must land first), sequential registration sequence, acceptance validation, failure modes + mitigations, rollback, and explicit out-of-scope items. Acceptance bar (minimum): green develop CI with >=3 SPM jobs in flight simultaneously. Warm cache + runner-health monitoring are explicit follow-ups. Path follows the superpowers:brainstorming skill convention: docs/superpowers/specs/YYYY-MM-DD-<topic>-design.md. Refs #180 #182 #186 #188 * docs(superpowers): implementation plan for multi-runner Grogu expansion Bite-sized task breakdown for the multi-runner spec at docs/superpowers/specs/2026-05-13-multi-runner-grogu-design.md. Five tasks, each task several 2-5 minute steps with concrete commands and expected output: 1. Verify PR #189 (age-filter) is on develop — precondition gate 2. Register Grogu-jarvis-2 (verify pre-stage, token, config, install service, verify online) 3. Register Grogu-jarvis-3 (same sequence) 4. Acceptance — rerun PR #185 CI, observe >=3 SPM jobs concurrent 5. Update CLAUDE.md runner inventory + open PR Self-review pass: every spec component covered, no placeholders, type-name consistency verified across tasks. Refs #180 #188	2026-05-13 17:58:29 -05:00
james.maes	4cdb3560e9	ci(actions): age-filter zombie reaper (>= 1h) for multi-runner safety (#189 ) Grogu will host multiple self-hosted runners simultaneously (planned: 3 runners). Each runner's job-start cleanup step previously killed ANY xctest/swift-test/swift-driver — meaning sibling runners would nuke each other's in-flight tests. Real CI jobs complete in <10 min. Zombies from interactive Xcode use are typically >hours old. A 1-hour age cutoff cleanly separates the two populations. awk filter: ps -o etime formats are `MM:SS` (<1h), `HH:MM:SS` (>=1h), or `DD-HH:MM:SS` (days). Match the latter two by testing for `-` (days prefix) or three `:`-separated groups. Single-runner behavior unchanged. Validated against synthetic ps output covering all three etime format cases. Closes #188	2026-05-13 17:33:03 -05:00
james.maes	81ce64c3f7	ci(actions): reap orphan xctest/swift-test on self-hosted before each job (#183 ) Grogu is a shared workstation, not an ephemeral VM. Interactive Xcode test runs, aborted swift test sessions, or debugger crashes can leave orphan xctest/swift-test/swift-driver processes around for hours, holding Mach services + sockets + ports that subsequent CI jobs block on when mounting their own test infrastructure. Concrete incident: develop run 25824607025 hung in SPM AgentCore for 34 min (0% CPU, sleeping state) while 6 unrelated xctest zombies aged 22-24h were alive from prior interactive use. The runner only executes one job at a time, so any such process at job start belongs to no live job and is safe to kill. Add a first step that runs SIGTERM then SIGKILL on any xctest/swift-test/swift-driver process. Gated to runner.environment == 'self-hosted' since github-hosted VMs start clean. Closes #182	2026-05-13 16:37:27 -05:00
james.maes	5fc2d3fabf	docs: reconcile Ollama model baseline to actual production names (#184 ) The codebase moved to qwen3.6:latest (memory extractor) and qwen2.5-coder:32b-instruct-q8_0 (agent fallback) in commit `8d0a959`, but CLAUDE.md and the MemoryRegressionCorpusTests doc-comments still cited the vanilla qwen2.5-coder:32b string. Both updated to match what ModelID.swift, MemoryExtractor.swift, and BootHealth.swift actually use today. No behavior change — documentation alignment only. Closes #58	2026-05-13 16:25:00 -05:00