feat: ai schema repair and session ps

.github/workflows/test.yml

@@ -27,7 +27,22 @@ jobs:
       - uses: actions/setup-go@v5
         with:
           go-version: "1.26.x"
-      - run: go test ./...
+      # Persist the embedded-postgres binary so most runs skip the download
+      # entirely. The version is pinned by the embedded-postgres library (V16).
+      - name: Cache embedded-postgres binary
+        uses: actions/cache@v4
+        with:
+          path: ~/.embedded-postgres-go
+          key: embedded-postgres-${{ runner.os }}-16.9.0
+      # The DB integration tests each start their own embedded postgres. Run
+      # packages serially (-p 1): concurrent StartEmbedded calls otherwise race
+      # the shared cold-cache Maven download (throttled -> "no version found
+      # matching 16.9.0") and collide on ports (FreePort closes the listener
+      # before postgres binds), which fails StartEmbedded with the server never
+      # becoming ready.
+      - run: go test -p 1 ./...
+        env:
+          CAPTAIN_DB_EMBEDDED_TEST: "1"
 
   build:
     runs-on: ubuntu-latest

.github/workflows/update-llm-model-registry.yml

@@ -0,0 +1,72 @@
+name: Update LLM Model Registry
+
+on:
+  schedule:
+    - cron: "0 3 * * 1"
+  workflow_dispatch:
+
+permissions:
+  contents: write
+  pull-requests: write
+
+concurrency:
+  group: update-llm-model-registry
+  cancel-in-progress: false
+
+jobs:
+  update:
+    runs-on: ubuntu-latest
+    env:
+      PR_BRANCH: chore/update-llm-model-registry
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+          ref: ${{ github.event.repository.default_branch }}
+
+      - name: Install Go
+        uses: actions/setup-go@v5
+        with:
+          go-version: "1.26.x"
+
+      - name: Generate model registry
+        run: make models-update
+
+      - name: Check for registry changes
+        id: changes
+        run: |
+          if git diff --quiet -- pkg/api/registry/models.json; then
+            echo "changed=false" >> "$GITHUB_OUTPUT"
+          else
+            git diff --stat -- pkg/api/registry/models.json
+            echo "changed=true" >> "$GITHUB_OUTPUT"
+          fi
+
+      - name: Commit and push registry update
+        if: steps.changes.outputs.changed == 'true'
+        run: |
+          git config user.name "github-actions[bot]"
+          git config user.email "github-actions[bot]@users.noreply.github.com"
+          git checkout -B "$PR_BRANCH"
+          git add pkg/api/registry/models.json
+          git commit -m "chore: update llm model registry"
+          git push --force-with-lease origin "$PR_BRANCH"
+
+      - name: Create pull request
+        if: steps.changes.outputs.changed == 'true'
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          if gh pr view "$PR_BRANCH" --repo "${{ github.repository }}" >/dev/null 2>&1; then
+            echo "Pull request already exists for $PR_BRANCH"
+            exit 0
+          fi
+
+          gh pr create \
+            --repo "${{ github.repository }}" \
+            --base "${{ github.event.repository.default_branch }}" \
+            --head "$PR_BRANCH" \
+            --title "chore: update llm model registry" \
+            --body "Automated update from models.dev using \`make generate-llm-model-registry\`."

.gitignore

@@ -13,6 +13,7 @@ pkg/cli/webapp/node_modules/
 pkg/cli/webapp/*.tsbuildinfo
 # goreleaser output at repo root.
 /dist/
+dist/
 # The built webapp under pkg/cli/webapp/dist is gitignored — it's a large
 # minified bundle whose vite build needs a local clicky-ui link: dependency
 # that isn't available in CI. A global `dist/` ignore excludes the directory,
@@ -23,6 +24,8 @@ pkg/cli/webapp/*.tsbuildinfo
 !pkg/cli/webapp/dist/
 pkg/cli/webapp/dist/*
 !pkg/cli/webapp/dist/.gitkeep
-!pkg/cli/webapp/dist/index.html
 .grite/
 hack/
+.ok/
+.okignore
+.ginkgo/

AGENTS.md

@@ -0,0 +1,14 @@
+# captain — agent notes
+
+Shared ways of working, the gavel todo workflow, and global skills come from the root ~/.agents/AGENTS.md.
+
+## Memory
+- [Commons-db migrations (HCL + SQL)](.agents/memory/commons-db-migrations.md) — HCL for tables, post-HCL SQL for views/triggers/deferred constraints; verify with a seeded PostgreSQL smoke, not dry apply.
+- [AI model fallback, recommendations & whoami catalogs](.agents/memory/ai-model-catalog-fallback.md) — IsFallbackEligible/ErrNoAPIKey classification in pkg/ai, per-family model retention in model_lists.go, keyless cmux backends in whoami.
+- [AI schema shaping, generation config & runtime logging](.agents/memory/ai-schema-generation-logging.md) — SchemaJSONForBackend provider transforms vs local validation, EffortConfig in generation_config.go, agent:model[:effort] log identity.
+- [Commit grouping for mixed Captain diffs](.agents/memory/commit-grouping.md) — group by semantic product slice not directory tree, tests travel with features, every file accounted for exactly once.
+- [Session viewer, Codex/Claude parsing & model normalization](.agents/memory/sessions-viewer-metadata.md) — SessionBrowser/SessionInspector seams, shared Codex extraction, synthetic event tools, unified Title/InitialPrompt derivation.
+- [History discovery, performance & bash categorization](.agents/memory/history-discovery-performance.md) — case-insensitive session-dir matching in both finders, prune Claude/Codex discovery early, pkg/bash priority-based classification.
+- [Prompt CLI runs: JSON contract, context.dir, run tables & Codex app-server](.agents/memory/prompt-cli-runs.md) — AIPromptResult full-request JSON, context.dir normalization into cmd.Dir, -M comparison tables, app-server JSON-RPC seams.
+- [Prompt workbench, serve UI, catalog & backend modes](.agents/memory/prompt-workbench-serve-ui.md) — captain serve hosts the workbench, catalog-backed family-first selectors, cmux-style backend modes land across four seams, rebuild linked clicky-ui dist.
+- [Lint cleanup: gavel lint sweeps & betterleaks fixes](.agents/memory/lint-cleanup.md) — snapshot gavel lint baseline, split by ownership slices, betterleaks comment hits fixed by neutral wording.

Makefile

@@ -1,7 +1,7 @@
 # Makefile stub - forwards to Taskfile
 # Install task: https://taskfile.dev/installation/
 
-.PHONY: build lint test install fmt tidy clean all
+.PHONY: build lint test install fmt tidy clean all models-update
 
 build:
 	@task build
@@ -26,3 +26,6 @@ clean:
 
 all:
 	@task all
+
+models-update:
+	@task models:update

README.md

@@ -17,6 +17,46 @@ It provides tools to:
 - run a web UI for launching and chatting with AI agents
 - configure default backend, model, and AI safety toggles
 
+```xml h=185px w=358px preview
+<div style="font-family:system-ui,sans-serif;padding:20px;color:var(--foreground)">
+  <label for="amt" style="font-size:14px;font-weight:600">Monthly budget</label>
+  <div id="out" style="font-size:30px;font-weight:700;color:var(--chart-1);margin:6px 0">$2,500</div>
+  <input id="amt" type="range" min="500" max="10000" step="100" value="2500"
+    style="width:100%;accent-color:var(--primary)" />
+  <p style="font-size:13px;color:var(--muted-foreground)">Drag to adjust — the figure updates live.</p>
+  <script>
+    var amt = document.getElementById('amt');
+    var out = document.getElementById('out');
+    amt.addEventListener('input', function () {
+      out.textContent = '$' + Number(amt.value).toLocaleString();
+    });
+  </script>
+</div>
+```
+
+```html preview
+<div style="font-family:system-ui,sans-serif;padding:20px">
+  <div id="cards" style="display:flex;gap:14px;flex-wrap:wrap"></div>
+  <script>
+    var stats = [
+      ['Active users', '12,480', '+8.2% MoM', 'var(--chart-2)'],
+      ['Revenue', '$48.2k', '+3.1% MoM', 'var(--chart-1)'],
+      ['Churn', '2.4%', '-0.5% MoM', 'var(--chart-5)']
+    ];
+    document.getElementById('cards').innerHTML = stats.map(function (s) {
+      return '<div style="flex:1;min-width:150px;padding:16px;background:var(--card);' +
+        'color:var(--card-foreground);border:1px solid var(--border);' +
+        'border-radius:var(--radius)">' +
+        '<div style="font-size:13px;color:var(--muted-foreground)">' + s[0] + '</div>' +
+        '<div style="font-size:26px;font-weight:700;margin-top:4px">' + s[1] + '</div>' +
+        '<div style="font-size:12px;font-weight:600;margin-top:4px;color:' + s[3] + '">' +
+        s[2] + '</div>' +
+        '</div>';
+    }).join('');
+  </script>
+</div>
+```
+
 ## What it does
 
 From the codebase, Captain is organized around a few core capabilities:
@@ -93,7 +133,7 @@ Supported backends are inferred from code and dependencies, including:
 ### 6. Adapter status and configuration
 
 - `whoami` — lists every AI adapter (API providers and CLI agents), how each is authenticated, whether its binary is installed, and the models it exposes
-- `configure` — interactive wizard to set default backend, model, reasoning effort, budget, timeout, and feature toggles (caching, MCP, hooks, skills, user/project settings, memory) saved to `~/.captain.yaml`
+- `configure` — configure per-provider agent, model, and effort defaults in `~/.captain.yaml`, or validate and save direct-provider API tokens in `~/.config/captain/vault`
 
 ### 7. Web UI and MCP server
 
@@ -459,9 +499,18 @@ Lists every AI adapter (API providers and CLI agents: `anthropic`, `openai`, `ge
 
 ```bash
 captain configure
+captain configure openai
+captain configure openai --test
+captain configure openai --agent codex-agent --model gpt-5.6-sol --effort high --active
 ```
 
-Interactive wizard that writes `~/.captain.yaml` with defaults for backend, model, reasoning effort, budget, timeout, and feature toggles (caching, MCP, hooks, skills, user/project settings, memory). These defaults apply to `captain ai prompt`, `captain ai agent`, `captain ai test`, and other AI commands.
+The providerless interactive wizard writes `~/.captain.yaml` with defaults for backend, model, reasoning effort, budget, timeout, and feature toggles (caching, MCP, hooks, skills, user/project settings, memory).
+
+With a provider and any of `--agent`, `--model`, `--effort`, or `--active`, `configure` saves provider-specific runtime defaults in `~/.captain.yaml`. `--effort default` clears an explicit effort so the selected model chooses its own default. The active provider supplies the agent and model for completely flagless runs. Explicit command flags still win, and each fallback model independently inherits defaults from its own provider.
+
+Passing `anthropic`, `openai`, `gemini`, or `deepseek` securely prompts for an API token, validates it with the provider's model-list endpoint, and saves it to `~/.config/captain/vault` only when validation succeeds. `--test` validates the currently effective token without writing. Automation may use `--token`, but interactive input is preferred because command-line arguments can be retained in shell history or process listings. Explicit runtime keys take precedence over the vault, and the vault takes precedence over environment variables.
+
+The `/whoami` page in `captain serve` exposes the same token operations plus agent, model, effort, and active-provider controls. Stored tokens are never displayed; only a masked identifier and credential source are returned.
 
 ### Serve
 
@@ -474,7 +523,7 @@ task www:dev
 task www:build
 ```
 
-Starts an HTTP API and embedded web UI. The UI launches `captain ai agent` operations and opens follow-up chat windows that resume the returned session. `--dev` starts the Vite dev server from `pkg/cli/webapp` and proxies `/api` back to the Go process. Use `task www:dev` for the local Go-backed Vite proxy with the browser opened, and `task www:build` to rebuild the embedded web UI assets.
+Starts an HTTP API and embedded web UI. The UI launches `captain ai agent` operations and opens follow-up chat windows that resume the returned session. `--dev` starts the Vite dev server from `pkg/cli/webapp`, binds the Go API to a random free port, and proxies `/api` to it. Pass `--port` to use a specific API port in development. Use `task www:dev` for the local Go-backed Vite proxy with the browser opened, and `task www:build` to rebuild the embedded web UI assets.
 
 ### MCP server
 
@@ -610,4 +659,4 @@ Start the web UI:
 - It is both an analysis tool and an execution/control tool.
 - The container/sandbox functionality is a major part of the project, not a side feature.
 - Many commands assume the presence of Claude local state under the user’s Claude config/projects directories.
-- Configuration is persisted to `~/.captain.yaml` via `captain configure`.
+- Runtime defaults are persisted to `~/.captain.yaml`; validated provider tokens are persisted separately to the private `~/.config/captain/vault` file.

Taskfile.yaml

@@ -82,6 +82,11 @@ tasks:
     cmds:
       - go mod tidy
 
+  models:update:
+    desc: Regenerate the embedded LLM model registry from models.dev + local patches
+    cmds:
+      - go run ./pkg/ai/internal/gen-model-registry --patches pkg/ai/internal/gen-model-registry/patches.json --output pkg/api/registry/models.json
+
   clean:
     desc: Remove build artifacts
     cmds: