carthagent 1.1.0-preview.12 → 1.1.0-preview.13

package/docs/release-notes-preview.13.md

@@ -0,0 +1,68 @@
+# 1.1.0-preview.13 experimental release
+
+Experimental preview. Install it explicitly with the experimental channel;
+stable 1.0.7 is unchanged and published preview.12 remains immutable.
+
+## What changes
+
+**Directory-first Projects rail.** The native left rail is organized around
+folders: a current-folder card with Browse, New session and model controls,
+distinct folder and group headings, and an active-session marker. A
+folder-browser dialog adds Back, Forward, Up, Home, direct paths and
+filtering, with async cancellation, exact path handling and preserved drafts.
+
+**Shared folder/model setup for new sessions.** Global and per-folder New
+session entries route through one setup flow. The session is created only on
+the first send, keeping the chosen folder, agent, model and applicable
+variant; unavailable selections are blocked instead of silently substituted.
+Pre-session `/model` is supported.
+
+**Isolated source-preview launcher.** `script/ctg/source-preview.sh` runs the
+checked-in development patch against its own profile for verification without
+touching the published runtime or its caches. The combined-v5 patch is a
+replacement patch: it contains all of combined-v4 plus these sources and
+their tests, and applies directly to the pinned upstream.
+
+The upstream revision, lockfile and detached runtime identity are unchanged
+from preview.12; the runtime is rebuilt because each runtime manifest is
+bound to its exact package version.
+
+**Ship and Voyages follow-up fixes.** Beam replay requests include the required
+nonempty reasoning placeholder when earlier assistant messages have no recorded
+reasoning. The delivery plan entry point stays visible in the code-mode catalog,
+and final review accepts single-assertion aliases and assertion arrays. Voyages
+and quest lists activate on mouse release so newly opened dialogs remain open;
+rail clicks select the intended quest and respect modal focus.
+Delivery tool guidance explicitly identifies JSON-text responses and requires
+relative file paths for promised outputs, avoiding prose-as-filename contracts.
+
+## Provenance
+
+| Input | Exact value |
+| --- | --- |
+| Package | `1.1.0-preview.13` |
+| Prebuild source | `e5de273811a11c2f8678b54605c0068ab8cfcf38` |
+| Prebuild workflow run | [41380550073](https://github.com/B-icy/Carthagent/actions/runs/41380550073) |
+| Full manifest SHA-256 | `5083dbe10b6f4be2550ab0abb04c9299197dacabe4b9d3556bd3b03fc9f6151a` |
+| Upstream revision | `843a6c514e606b436a7fe713422ace0a93ad3d61` |
+| CTG patch (combined-v5) SHA-256 | `ec4bb182308ed8b311513aad5c9c7d57dcd0db103f0882f3dc3517d535b4c975` |
+| Lock SHA-256 | `9d78379cd264584a495440a217b9d2ecab524c9a56f7f8b4b5bed3926daa330d` |
+| Runtime / compatibility | `0.0.0-ctg.2.0.32` / `2.0.32` |
+
+## Trying it
+
+```sh
+ctg update --channel experimental --check
+ctg update --channel experimental
+ctg --version   # 1.1.0-preview.13
+```
+
+Or install the exact version from npm or the exact tag:
+
+```sh
+npm install -g carthagent@1.1.0-preview.13
+npm install -g --install-links=true github:B-icy/Carthagent#v1.1.0-preview.13
+```
+
+`ctg update --channel stable` returns to stable 1.0.7 without `--force`. See
+the [upgrade guide](native-upgrade.md) for profile preservation.

package/docs/source-preview.md

@@ -0,0 +1,120 @@
+# Projects sidebar source preview
+
+This preview runs the redesigned Projects sidebar from OpenCode source plus the
+Carthagent plugin in this checkout. It is a development preview. Nothing here
+builds, publishes, installs, or changes the published `ctg` runtime.
+
+## What is pinned where
+
+| Purpose | Patch | Recipe | Hash reference |
+| --- | --- | --- | --- |
+| Published runtime (preview.13) | `docs/patches/opencode-v2.0.32-session-ui-combined-v5.patch` | `script/ctg/build-owned-runtime.sh`, publish workflow | `CTG_RUNTIME_PATCH_SHA256`, `lib/opencode/runtime-release.json` |
+| Source preview (development) | `docs/patches/opencode-v2.0.32-session-ui-combined-v5.patch` | `script/ctg/prepare-development-runtime.sh` | `CTG_DEVELOPMENT_RUNTIME_PATCH_SHA256` |
+
+Both patches apply to clean upstream `https://github.com/anomalyco/opencode` at
+`843a6c514e606b436a7fe713422ace0a93ad3d61`. v5 is a replacement patch: it
+contains everything in v4 plus the sidebar, folder browser and new-session
+setup sources and their tests under `packages/tui/test/sidebar`. Do not apply
+v4 first.
+
+## Setup
+
+1. `npm ci` in this checkout (the plugin runs from `plugins/carthagent`).
+2. Clone upstream into a new directory, check out the pinned commit, and run
+   `script/ctg/prepare-development-runtime.sh <checkout>`. The script refuses a
+   different commit, a dirty checkout or a patch with the wrong SHA-256, then
+   runs `git apply --check` and `git apply`. It does not install or compile.
+3. Run `bun install` with Bun 1.4.2 in the prepared checkout.
+
+## Launch
+
+```sh
+script/ctg/source-preview.sh --source DIR --profile DIR --bun PATH [--workspace DIR] [-- ARGS...]
+```
+
+- All three of `--source`, `--profile` and `--bun` are required. Relative
+  paths resolve against your current directory, so you can launch from
+  anywhere. Paths with spaces work when quoted.
+- The launcher checks that `packages/cli/src/index.ts` and `node_modules`
+  exist and that `bun --version` matches the checkout's `packageManager` pin.
+  It fails before touching the profile when a prerequisite is missing.
+- It runs Bun with `--no-env-file` from `<source>/packages/cli` (required for
+  JSX resolution) under `env -i`. Only these variables reach OpenCode: `HOME`,
+  `PATH` (Bun's directory, Node's directory if found, `/usr/bin`, `/bin`),
+  `TERM`, `COLORTERM`, `TMPDIR`, the four `XDG_*` directories,
+  `OPENCODE_CONFIG_DIR`, `OPENCODE_TEST_HOME`,
+  `OPENCODE_DISABLE_PROJECT_CONFIG=1`, `OPENCODE_DISABLE_AUTOUPDATE=1` and
+  `BUN_RUNTIME_TRANSPILER_CACHE_PATH=0`. Provider API-key variables are never
+  passed through.
+- First launch creates the profile (0700) with `config/opencode/opencode.json`
+  (the plugin path), `cli.json` (Carthage theme) and `themes/carthage.json`
+  (all 0600). Each file is created only if absent; existing configuration,
+  sessions and authentication are left untouched. If an existing
+  `opencode.json` does not list the plugin, the launcher prints a warning and
+  still leaves the file alone.
+- `--standalone` is always used, so the CLI starts its own private server on
+  a loopback ephemeral port (`--port 0`) that exits with the TUI. `--server`
+  is rejected. Without `--`, the workspace (default: current directory) is
+  opened. With `--`, the remaining arguments run another CLI command in the
+  same profile, for example `-- auth login`.
+
+## Use
+
+- **Folder card** (top of Projects): current folder, abbreviated path (click
+  for the full path), Browse, New session and the model action. Narrow layouts
+  show compact `⌕` search, `▤` Browse and `◇` model controls.
+- **Browse** (`/browse`, palette “Browse folders”): Filter and Path inputs
+  (Tab switches), Up/Down select, Enter opens, Alt+Left/Right/Up for
+  Back/Forward/Up, Alt+H for Home, Ctrl+S confirms, Escape cancels.
+- **New session** (`Ctrl+X N`, `/new`, palette “New session” or “New session
+  in folder…”, folder card, compact `+`, folder right-click menu): choose the
+  folder, then the model. Global entries start at the folder a plain new
+  session would use; project entries start at that project's folder. Nothing
+  is created until the first message is sent; cancelling leaves the current
+  session, draft and model unchanged. The new session keeps the agent you were
+  using and, when you keep the same model, its variant (including the default).
+  If the chosen agent or model is not available at the folder when you send,
+  the send is blocked with your text kept and no session created; pick another
+  agent (`Shift+Tab`) or model (`/model`, `Ctrl+X M`) and send again. There is
+  no automatic fallback to another model. `/model` also works before a session
+  exists. The model step inside setup has no integration shortcut; connect
+  integrations from the standalone model picker (`Ctrl+X M`).
+- **Exit**: `/exit` or palette “Exit the app”. The private server exits with
+  the TUI.
+- **Resume**: relaunch with the same `--profile` and select the session in
+  Projects. Automatic resume is not part of this preview.
+
+## Tests
+
+Run serially. Close the preview first.
+
+```sh
+# Product (from this checkout)
+node --test --test-concurrency=1 tests/source-preview.test.mjs tests/native-ship-plan.test.mjs tests/native-model-command.test.mjs
+npx eslint . && npm run typecheck
+
+# Native (from <source>/packages/tui, using the same isolated variables as the launcher)
+bun --no-env-file test --timeout 30000 test/navigation-rail.test.tsx \
+  test/component/session-tabs-rail.test.tsx test/context/session-tabs-model.test.ts \
+  test/ui/layout.test.ts test/sidebar
+```
+
+`tests/source-preview.test.mjs` uses a Bun test double and synthetic profiles,
+so it proves argument handling, isolation and config preservation, not a real
+launch. It runs the launcher with `/bin/sh` on Linux and macOS and is skipped
+on native Windows, where the POSIX launcher is not supported.
+
+## Limits
+
+- Source preview only. No native binary, package or release includes these
+  changes, and the published runtime remains v4.
+- The launcher is a POSIX `sh` script for Linux and macOS (GNU and BSD
+  utilities). Native Windows is not supported.
+- Real launches were verified on Linux with Bun 1.4.2. No cross-platform
+  launch matrix was run.
+- The native type checker (`tsgo`) and native build are not part of this
+  workflow; Bun tests compile the changed modules but are not a full static
+  type check.
+- Some native tests outside the focused list fail in the same way on the
+  pre-change sources (for example missing `ToastProvider` in
+  `session-tabs-status` harnesses).

package/lib/delivery-review.mjs

@@ -41,13 +41,23 @@ export async function captureDeliveryReview(state, cwd, directory, signal) {
   return { ...capture, diff: text.slice(0, 12000), truncated: text.length > 12000, unresolvedFindings: (state.reviewFindings || []).filter(f => !f.resolution), next: 'Read the full capture path if truncated; new text source is included. Inspect binary assets separately. Probe critical component connections and user outcomes, not only individual functions or artifact existence. Record coverage, findings and limitations; earlier findings require explicit resolutions with evidence, never silently move them to limitations.' };
 }
 
+/** Models commonly send `assertion` or a string array; both carry the same claim. */
+function coverageAssertions(item) {
+  const value = item.assertions ?? item.assertion;
+  if (Array.isArray(value)) return value.filter(v => typeof v === 'string' && v.trim()).join('; ');
+  return typeof value === 'string' ? value : '';
+}
+
 export function recordDeliveryReview(state, cwd, params) {
   const capture = state.reviewCapture;
   const hash = fingerprint(cwd, ['.']);
   if (!capture || capture.id !== params.captureId || capture.runId !== state.runId || capture.revision !== state.revision || capture.fingerprint !== hash) throw Error('Review capture missing or stale. Use delivery_review action=inspect after final source edits.');
   if (pendingRequirements(state, hash).length) throw Error('Run fresh checks for every requirement before recording review.');
   const nonempty = value => typeof value === 'string' && value.trim().length > 0;
-  if (!Array.isArray(params.coverage) || (state.plan.acceptance || []).some(a => !params.coverage.some(c => c.requirement === a.requirement && nonempty(c.assertions)))) throw Error('Review coverage must explain concrete assertions for every exact acceptance requirement, including required validators.');
+  if (!Array.isArray(params.coverage) || params.coverage.some(c => !c || typeof c !== 'object')) throw Error('coverage must be an array of {requirement, assertions, status?} objects.');
+  params = { ...params, coverage: params.coverage.map(c => ({ ...c, assertions: coverageAssertions(c) })) };
+  const uncovered = (state.plan.acceptance || []).filter(a => !params.coverage.some(c => c.requirement === a.requirement && nonempty(c.assertions))).map(a => a.requirement);
+  if (uncovered.length) throw Error(`Review coverage must explain concrete assertions for every exact acceptance requirement, including required validators. Missing: ${JSON.stringify(uncovered.slice(0, 10))}${uncovered.length > 10 ? ` (+${uncovered.length - 10} more)` : ''}. Each item: {requirement: <exact acceptance text>, assertions: "<what the checks assert>", status?: "demonstrated"|"unverified"}.`);
   if (params.coverage.some(c => c.status !== undefined && !['demonstrated', 'unverified'].includes(c.status))) throw Error('Coverage status must be demonstrated or unverified.');
   if (!Array.isArray(params.probes) || !params.probes.length || params.probes.some(p => !nonempty(p))) throw Error('Record boundary/metamorphic probes and observed results, not just a generic review claim.');
   if (!Array.isArray(params.findings) || params.findings.some(p => !nonempty(p))) throw Error('findings must be an array; unresolved findings require repairs or a blocked handoff.');

package/lib/opencode/runtime-release.json

@@ -1,36 +1,36 @@
 {
   "version": 1,
-  "releaseVersion": "1.1.0-preview.12",
+  "releaseVersion": "1.1.0-preview.13",
   "compatibilityVersion": "2.0.32",
   "runtimeVersion": "0.0.0-ctg.2.0.32",
   "sourceRevision": "843a6c514e606b436a7fe713422ace0a93ad3d61",
-  "patchSha256": "1f2309bb3a31dfb8e8a0827a6a1144a1f76a005abff58f6f48857639ec02478a",
+  "patchSha256": "ec4bb182308ed8b311513aad5c9c7d57dcd0db103f0882f3dc3517d535b4c975",
   "lockSha256": "9d78379cd264584a495440a217b9d2ecab524c9a56f7f8b4b5bed3926daa330d",
   "assets": {
     "linux-x64": {
       "name": "opencode-linux-x64.tar.gz",
-      "archiveSha256": "d4c202e0825fc77321ccd23d91a7baede4951a4eb4cee28f258c171e72f03852",
-      "executableSha256": "24bcfff03d32db163e0d40c4d0b212eede300a71a3dc26aa23810b9088d52280"
+      "archiveSha256": "d7c1869ace6dde4672f8ade4a8ba7826b15769a04b943571122d9541e52b62a7",
+      "executableSha256": "8a02fe556eb64def1507ff34a2cdd012dee636c0a6678f14d1fe57a020efa620"
     },
     "linux-arm64": {
       "name": "opencode-linux-arm64.tar.gz",
-      "archiveSha256": "c3a1e68a6fd7db5909d2d38639c5b1e8df7a5de64fc40f6cc758211a280ea058",
-      "executableSha256": "b73708103d62de570e629a0bf9c2bce9b1a4db0a87a79b1c86c2d2ebe998db7b"
+      "archiveSha256": "7e5f7ea55e54a752d2923296f6917f7b48a175e5ae39a8caa540d844d89e6d8b",
+      "executableSha256": "eefc048de0b3932148050844b0d574c112bf13aeae001c767fac463dbbd1292e"
     },
     "darwin-x64": {
       "name": "opencode-darwin-x64.tar.gz",
-      "archiveSha256": "d733d269bd06a652c858b62b24e5b78a629400fc687d7dc71a999bb03bfab1d5",
-      "executableSha256": "f2bb0e772744bc04b96368d32cc35582a6a1bca0be795897704c9395e5059f37"
+      "archiveSha256": "84be12c83c8cd9ef1028467fb53e94016e2b514013df9f62347a451f3c84044c",
+      "executableSha256": "4de7f310debebc00d955d8b4c349749b15bac69ef6998a051144d4e86b79b1a4"
     },
     "darwin-arm64": {
       "name": "opencode-darwin-arm64.tar.gz",
-      "archiveSha256": "e0a0ac4c7f5b4d14eab8844428d717a9a596c61694b318d0e065080de5a42fda",
-      "executableSha256": "9bf83ef1fd9e1af8028a23941a05195f8256847deae271c2b016014d5363605b"
+      "archiveSha256": "8e9be7bda9df7a1fe84f1344738435f748ee56598c4635c2be8fdfef1433414a",
+      "executableSha256": "0963189d87daa88616980f72f3470b69902ac2b2b3b62cadb9e0fa436245b960"
     },
     "windows-x64": {
       "name": "opencode-windows-x64.tar.gz",
-      "archiveSha256": "0318a74ba0b67ee5119b4289c0b78124c620676c8770e48c24d347513e96e2aa",
-      "executableSha256": "b057ce293492cff6dabf3939d12072798f53d8b4cf7ab944fd97ef0bb0879bb9"
+      "archiveSha256": "5ca306190c354b5e6f690103da8a9db9e90bcf5463214f50cd1dfafd10109d4c",
+      "executableSha256": "4249c083b86b9bc034550954c2bc08ffdf9c578cf63c90a0adfabd103cac1594"
     }
   }
 }

package/lib/opencode/runtime.mjs

@@ -13,11 +13,12 @@ export const OPENCODE_VERSION = '2.0.32';
 export const CTG_RUNTIME_VERSION = '0.0.0-ctg.2.0.32';
 export const CTG_RUNTIME_SOURCE = '843a6c514e606b436a7fe713422ace0a93ad3d61';
 export const CTG_RUNTIME_LOCK_SHA256 = '9d78379cd264584a495440a217b9d2ecab524c9a56f7f8b4b5bed3926daa330d';
-export const CTG_RUNTIME_PATCH_SHA256 = '1f2309bb3a31dfb8e8a0827a6a1144a1f76a005abff58f6f48857639ec02478a';
+export const CTG_RUNTIME_PATCH_SHA256 = 'ec4bb182308ed8b311513aad5c9c7d57dcd0db103f0882f3dc3517d535b4c975';
 // Update only this development pin after reviewing a new checked-in patch and
-// its build recipe. The published release pin above and release manifest stay
-// immutable until a separately verified runtime release is promoted.
-export const CTG_DEVELOPMENT_RUNTIME_PATCH_SHA256 = '1f2309bb3a31dfb8e8a0827a6a1144a1f76a005abff58f6f48857639ec02478a';
+// script/ctg/prepare-development-runtime.sh. The published release pin above,
+// its build recipe, and the release manifest stay immutable until a separately
+// verified runtime release is promoted.
+export const CTG_DEVELOPMENT_RUNTIME_PATCH_SHA256 = 'ec4bb182308ed8b311513aad5c9c7d57dcd0db103f0882f3dc3517d535b4c975';
 const DEFAULT_ROOT = path.join(homedir(), '.carthagent', 'opencode');
 const require = createRequire(import.meta.url);
 const PACKAGE_ROOT = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '../..');

package/lib/providers/experiential.mjs

@@ -64,11 +64,11 @@ const EXPERIENTIAL_MODEL_PROFILES = Object.freeze({
   'glm-5.3-flash': { name: 'GLM 5.3 Flash', reasoning: true, thinkingLevelMap: TIERED_REASONING_LEVELS, contextWindow: 1_048_576, maxTokens: 32_768, input: ['text', 'image'], compat: { supportsReasoningEffort: true, thinkingFormat: 'zai' } },
   'glm-5.3': { name: 'GLM 5.3', reasoning: true, thinkingLevelMap: TIERED_REASONING_LEVELS, contextWindow: 1_048_576, maxTokens: 32_768, input: ['text'], compat: { supportsReasoningEffort: true, thinkingFormat: 'zai' } },
   'kimi-k3': { name: 'Kimi K3', reasoning: true, thinkingLevelMap: TIERED_REASONING_LEVELS, contextWindow: 1_048_576, maxTokens: 32_768, input: ['text'], compat: { supportsDeveloperRole: false, supportsReasoningEffort: true, thinkingFormat: 'deepseek' } },
-  'beam-1.0-501b-a23b': { name: 'Beam 1.0 501B A23B', reasoning: true, thinkingLevelMap: BEAM_REASONING_LEVELS, contextWindow: 1_048_576, maxTokens: 32_768, input: ['text'], compat: { supportsReasoningEffort: true } },
+  'beam-1.0-501b-a23b': { name: 'Beam 1.0 501B A23B', reasoning: true, thinkingLevelMap: BEAM_REASONING_LEVELS, contextWindow: 1_048_576, maxTokens: 32_768, input: ['text'], compat: { supportsReasoningEffort: true, reasoningField: 'reasoning_content', requireReasoning: true, reasoningPlaceholder: ' ' } },
   'mercury-3': { name: 'Mercury 3', reasoning: true, thinkingLevelMap: MERCURY3_REASONING_LEVELS, contextWindow: 1_048_576, maxTokens: 131_072, input: ['text'], compat: { supportsReasoningEffort: true } },
 });
 
-const MODEL_METADATA_FIELDS = ['reasoning', 'thinkingLevelMap', 'contextWindow', 'maxTokens', 'input', 'output', 'compat'];
+const MODEL_METADATA_FIELDS = ['reasoning', 'thinkingLevelMap', 'contextWindow', 'maxTokens', 'input', 'output', 'compat', 'cost'];
 
 export const EXPERIENTIAL_PROVIDER_ID = PROVIDER_ID;
 export const EXPERIENTIAL_PROVIDER_NAME = CARTHAGENT_SHIP_PROVIDER_NAME;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "carthagent",
-  "version": "1.1.0-preview.12",
+  "version": "1.1.0-preview.13",
   "workspaces": [
     "plugins/carthagent"
   ],

package/plugins/carthagent/src/daily-source-rpc.ts

@@ -8,7 +8,7 @@ export interface DailySourceReport {
   freshness: 'current' | 'stale';
   goal: string;
   change: string;
-  checks: Array<{ id: string; before: 'not_run'; after: 'passed' | 'failed' | 'not_run'; revision: string; result: string }>;
+  checks: Array<{ id: string; before: 'not_run' | 'failed'; after: 'passed' | 'failed' | 'not_run'; revision: string; result: string }>;
   /** Model-authored review/resolution claims, not independent verification. */
   findings: Array<{ id: string; issue: string; evidence: string; resolution: string }>;
   /** At least one field or item was omitted; the UI must warn before selection. */
@@ -20,7 +20,7 @@ const fingerprint = { type: 'string', minLength: 64, maxLength: 64 } as const;
 const goal = { type: 'string', maxLength: 1024 } as const;
 const check = {
   type: 'object', properties: {
-    id, before: { type: 'string', enum: ['not_run'] },
+    id, before: { type: 'string', enum: ['not_run', 'failed'] },
     after: { type: 'string', enum: ['passed', 'failed', 'not_run'] },
     revision: fingerprint, result: goal,
   }, required: ['id', 'before', 'after', 'revision', 'result'], additionalProperties: false,

package/plugins/carthagent/src/server/daily-source-rpc.ts

@@ -57,6 +57,16 @@ function localJson(path: string, max = MAX_FILE): { value: unknown; raw: string
   } catch (error) { if (missing(error)) return null; throw invalid(); }
 }
 
+/** Receipts already hold an output tail; keep its end within the byte budget. */
+function tail(value: unknown, limit: number, cut: () => void) {
+  if (typeof value !== 'string' || Buffer.byteLength(value, 'utf8') <= limit) return value;
+  cut();
+  const bytes = Buffer.from(value, 'utf8');
+  const kept = bytes.subarray(bytes.length - (limit - 4)).toString('utf8').replace(/^\uFFFD+/, '');
+  const line = kept.indexOf('\n');
+  return `…\n${line >= 0 && line < kept.length - 1 ? kept.slice(line + 1) : kept}`;
+}
+
 function validateState(value: unknown, directory: string, runId: string) {
   if (!record(value) || value.runId !== runId || !Number.isSafeInteger(value.revision) || (value.revision as number) < 1 || !record(value.plan) || !Array.isArray(value.plan.checks) || value.plan.checks.length > 20000) throw invalid();
   if (value.status !== undefined && !['implementing', 'verifying', 'verified', 'blocked'].includes(value.status as string)) throw invalid();
@@ -76,6 +86,7 @@ function validateState(value: unknown, directory: string, runId: string) {
   }
   if (value.reviewEvidence !== undefined && !record(value.reviewEvidence)) throw invalid();
   if (value.handoff !== undefined && !record(value.handoff)) throw invalid();
+  if (value.baselines !== undefined && !record(value.baselines)) throw invalid();
   if (value.reviewFindings !== undefined && (!Array.isArray(value.reviewFindings) || value.reviewFindings.length > 20000)) throw invalid();
   const findings = (value.reviewFindings ?? []) as RecordValue[];
   const findingIDs = new Set<string>();
@@ -129,9 +140,9 @@ function sourceFor(directory: string, sessionID: string): DailySourceReport | nu
   const result: DailySourceReport = {
     runId: parts[2], revision: current, planRevision: state.revision as number,
     freshness: stale ? 'stale' : 'current', goal: text((state.plan as RecordValue).goal, 1024),
-    // Native delivery_finish stores review/launch/limitations, not a change summary.
-    // Do not manufacture a change claim from those fields or from the goal.
-    change: '',
+    // Only delivery_finish's explicit change summary for this snapshot; never
+    // derived from review/launch/limitations or the goal.
+    change: record(state.handoff) && state.handoff.fingerprint === current ? text(state.handoff.change, 3072) : '',
     checks: checks.slice(0, 12).map(check => {
       const receipt = evidence[check.id];
       const fresh = Boolean(isFresh(check));
@@ -139,7 +150,10 @@ function sourceFor(directory: string, sessionID: string): DailySourceReport | nu
       // Omit unbound failure output rather than making old/scope-error text look
       // like an execution result for the current revision in the outbound DTO.
       const output = receipt?.passed === false && !failed ? undefined : receipt?.outputTail;
-      return { id: check.id, before: 'not_run', after: fresh ? 'passed' : failed ? 'failed' : 'not_run', revision: current, result: text(output, 1024) };
+      // A first-in-run execution that failed on different source is a recorded before-failure.
+      const baseline = record(state.baselines) && record(state.baselines[check.id]) ? state.baselines[check.id] as RecordValue : undefined;
+      const before = baseline?.passed === false && baseline.checkDigest === checkDigest(check) && hash(baseline.fingerprint) && baseline.fingerprint !== current ? 'failed' as const : 'not_run' as const;
+      return { id: check.id, before, after: fresh ? 'passed' : failed ? 'failed' : 'not_run', revision: current, result: text(tail(output, 1024, () => { truncated = true; }), 1024) };
     }),
     findings: findings.slice(0, 5).map(finding => {
       const resolution = record(finding.resolution) ? finding.resolution : {};

package/plugins/carthagent/src/server/delivery.ts

@@ -43,7 +43,7 @@ const short = S()
 const steps = A({ anyOf: [short, O({ id: { type: "string", pattern: "^[a-z][a-z0-9_-]{0,39}$" }, title: short, kind: { type: "string", enum: ["work", "regression"] }, dependsOn: A(short, 40), checks: A(short, 24) }, ["id", "title"])] }, 40)
 const checks = A(O({ id: { type: "string", pattern: "^[a-z][a-z0-9_-]{0,39}$" }, kind: { type: "string", enum: ["test", "runtime", "static"] }, argv: A(short, 80), timeoutSeconds: { type: "integer", minimum: 1, maximum: 300 } }), 24)
 const acceptance = A(O({ requirement: short, checks: A(short, 24) }), 40)
-const planSchema = O({ goal: short, assumptions: A(short), artifacts: A(short, 40), outputs: A(short, 40), steps, verification: { type: "string", enum: ["required", "advisory", "none"] }, acceptance, checks, design: { type: "object", additionalProperties: true }, workflow: { type: "object", additionalProperties: true }, outcomes: A({ type: "object", additionalProperties: true }, 32) }, ["goal", "assumptions", "artifacts", "steps", "acceptance", "checks"])
+const planSchema = O({ goal: short, assumptions: A(short), artifacts: A(short, 40), outputs: A(S("Relative file path that must exist at completion, e.g. invoice.mjs. Not a prose deliverable or command result; put test commands in checks. Omit outputs when no files are promised."), 40), steps, verification: { type: "string", enum: ["required", "advisory", "none"] }, acceptance, checks, design: { type: "object", additionalProperties: true }, workflow: { type: "object", additionalProperties: true }, outcomes: A({ type: "object", additionalProperties: true }, 32) }, ["goal", "assumptions", "artifacts", "steps", "acceptance", "checks"])
 
 function sessionKey(sessionID: string) { return `session/${sessionID}` }
 
@@ -149,8 +149,9 @@ export async function registerDelivery(ctx: Context, options: DeliveryOptions =
       } finally { release() }
     })
   }
+  // The code-mode catalog drops the longest signatures first; pin the plan entry point so it is always listed.
   const add = (editor: Parameters<Parameters<Context["tool"]["transform"]>[0]>[0], name: string, description: string, input: Schema, execute: (input: DeliveryInput, toolContext: RegisteredToolContext) => Promise<{ content: string }>) => editor.add({
-    name, description, input: input as Tool.ValueSchema<DeliveryInput>, execute: async (rawValue, tool) => {
+    name, description: `${description} Returns text; in code mode use JSON.parse(result) for JSON responses before reading fields. Plain-text messages are not JSON.`, input: input as Tool.ValueSchema<DeliveryInput>, execute: async (rawValue, tool) => {
       const value = rawValue as DeliveryInput
       const { active } = await scope(tool.sessionID)
       if (!active || !isShipAgent(tool.agent)) throw Error("Carthagent delivery tools are disabled for this session; select Ship mode to use them.")
@@ -161,12 +162,12 @@ export async function registerDelivery(ctx: Context, options: DeliveryOptions =
         return { ...result, metadata: { deliveryPlan: planSnapshot(data.state) } }
       }
       catch (error) { throw new NativeToolError({ message: error instanceof Error ? error.message : String(error) }) }
-    }, options: { namespace: "carthagent", codemode: true },
+    }, options: { namespace: "carthagent", codemode: true, ...(name === "delivery_plan" ? { pinned: true } : {}) },
   })
 
   registrations.push(await ctx.tool.transform(editor => {
     editor.namespace?.({ name: "carthagent", description: "Evidence-driven CTG delivery workflow tools" })
-    add(editor, "delivery_plan", "Create the durable acceptance contract and plan.d2. Informational asks may use verification none/advisory; product changes require fresh checks.", planSchema, async (input, tool) => operate(tool.sessionID, async ({ cwd, data, options: opts }) => {
+    add(editor, "delivery_plan", "Create the durable acceptance contract and plan.d2. Outputs are relative file paths, not prose or test results. Informational asks may use verification none/advisory; product changes require fresh checks.", planSchema, async (input, tool) => operate(tool.sessionID, async ({ cwd, data, options: opts }) => {
       if (data.configError) throw Error(data.configError)
       const plan = validatePlan(bindRequiredChecks({ ...input, verification: input.verification ?? "required" }, data.required), cwd)
       validateRevision(data.state, plan)
@@ -277,6 +278,9 @@ export async function registerDelivery(ctx: Context, options: DeliveryOptions =
         }
         const passed = result.code === 0 && !result.timedOut && !result.cancelled && !result.outputLimit && before === after
         data.state.evidence[check.id] = { passed, ...identity, executedRevision: data.state.revision, fingerprint: after, code: result.code, timedOut: result.timedOut, cancelled: result.cancelled, logPath: result.logPath, durationMs: result.durationMs, changedDuringCheck: before !== after, outputTail: result.output.slice(-1200) }
+        // The first execution in a run is its baseline; later receipts replace evidence only.
+        const baselines = (data.state.baselines ??= {}) as Record<string, unknown>
+        if (!baselines[check.id]) baselines[check.id] = { passed, code: result.code, checkDigest: identity.checkDigest, fingerprint: after, at: new Date().toISOString() }
         results.push({ id: check.id, advisory: verification !== "required", ...data.state.evidence[check.id], outputTail: result.output })
         await persist(cwd, tool.sessionID, data)
       }
@@ -285,12 +289,12 @@ export async function registerDelivery(ctx: Context, options: DeliveryOptions =
       const hash = fingerprint(cwd, ["."]); const issues = completionIssues(data.state, cwd, hash)
       return text({ results, pendingChecks: pendingChecks(data.state, hash), completionIssues: issues, readyToFinish: Object.values(issues).every(items => !items.length), next: "Complete pending steps, review the evidence, then call delivery_finish." })
     }))
-    add(editor, "delivery_review", "Inspect final diff or record assertion coverage, boundary probes and findings. Required for structured plans; assertions are model-authored.", O({ action: { type: "string", enum: ["inspect", "record"] }, captureId: short, coverage: A({ type: "object", additionalProperties: true }, 64), resolutions: A({ type: "object", additionalProperties: true }, 40), outcomes: A({ type: "object", additionalProperties: true }, 32), probes: A(short, 40), findings: A(short, 40), limitations: A(short) }, ["action"]), async (input, tool) => operate(tool.sessionID, async ({ cwd, data }) => {
+    add(editor, "delivery_review", "Inspect final diff or record assertion coverage, boundary probes and findings. Required for structured plans; assertions are model-authored.", O({ action: { type: "string", enum: ["inspect", "record"] }, captureId: short, coverage: A({ type: "object", properties: { requirement: { type: "string", description: "Exact acceptance requirement text" }, assertions: { anyOf: [{ type: "string" }, A({ type: "string" }, 32)], description: "Concrete assertions the linked checks make" }, assertion: { type: "string", description: "Alias for a single assertion" }, status: { type: "string", enum: ["demonstrated", "unverified"] } }, required: ["requirement"], additionalProperties: true }, 64), resolutions: A({ type: "object", additionalProperties: true }, 40), outcomes: A({ type: "object", additionalProperties: true }, 32), probes: A(short, 40), findings: A(short, 40), limitations: A(short) }, ["action"]), async (input, tool) => operate(tool.sessionID, async ({ cwd, data }) => {
       if (!data.state) throw Error("Call delivery_plan first")
       const result = input.action === "inspect" ? await captureDeliveryReview(data.state, cwd, runDirectory(cwd, tool.sessionID, data.state.runId), tool.signal) : input.action === "record" ? recordDeliveryReview(data.state, cwd, input) : (() => { throw Error("Review action must be inspect or record") })()
       await persist(cwd, tool.sessionID, data); return text(result)
     }))
-    add(editor, "delivery_finish", "Finish delivery. Verified requires every required validator, fresh check, output, step, requirement and review gate; blocked requires specific limitations.", O({ status: { type: "string", enum: ["verified", "blocked"] }, review: short, launch: short, limitations: A(short) }), async (input, tool) => operate(tool.sessionID, async ({ cwd, data }) => {
+    add(editor, "delivery_finish", "Finish delivery. Verified requires every required validator, fresh check, output, step, requirement and review gate; blocked requires specific limitations.", O({ status: { type: "string", enum: ["verified", "blocked"] }, change: { ...short, description: "One-paragraph summary of what changed and why" }, review: short, launch: short, limitations: A(short) }), async (input, tool) => operate(tool.sessionID, async ({ cwd, data }) => {
       if (!data.state) throw Error("Call delivery_plan first")
       if (data.configError) throw Error(data.configError)
       const verification = verificationMode(data.state.plan); let hash: string | null = null

package/plugins/carthagent/src/server/providers.ts

@@ -98,7 +98,7 @@ function modelInfo(providerID: string, profile: CatalogModel, defaultBaseURL: st
 
 function mapCompatibility(compat: Record<string, unknown>): NonNullable<Model.Info['compatibility']> {
   const keys = [
-    'reasoningField', 'requireReasoning', 'maxTokensField', 'requireFinishReason',
+    'reasoningField', 'requireReasoning', 'reasoningPlaceholder', 'maxTokensField', 'requireFinishReason',
     'requireAssistantAfterTool', 'supportsPromptCacheKey', 'cacheFormat',
     'supportsThinkingBlockBinding', 'supportsEffortUpdates',
   ] as const;

package/plugins/carthagent/src/server/ship-guidance.ts

@@ -8,6 +8,8 @@ import { unframe } from "./framing.mjs"
 
 export const STANDARD_GUIDANCE = "Standard delivery: inspect relevant files (glob/read/code_nav); create delivery_plan with complete requirements, artifact roots, a few steps and real checks. A valid required plan enables implementation immediately: design components/scenario ceremony is optional. Revisions merge design/workflow fields, replace supplied arrays, preserve obligations and require fresh final evidence, not repeated approval. Mark implementation steps done independently of integrated check execution. Use delivery_check all, final delivery_review inspect/record, then delivery_finish with honest limitations. No fabricated evidence, scope reduction or unrequested dependency changes. Informational plans do not authorize implementation. For bounded environment versions use delivery_probe kind=node-version or npm-version; inspect other APIs through read."
 
+export const TOOL_SURFACE_GUIDANCE = "File edits and commands use the direct read/write/edit/bash tools. A code-mode search() catalog lists only namespaced tools.* helpers, so their absence there does not mean they are unavailable. If a tool is locked, its error names the required delivery step; take that step instead of searching for alternatives."
+
 const browserCheck = fileURLToPath(new URL("../../../../tools/browser-check.mjs", import.meta.url))
 const browserGuidance = existsSync(browserCheck)
   ? `For web/browser tasks, a self-contained browser check is available: ${JSON.stringify(browserCheck)}. Declare it as a runtime check, e.g. ["node",${JSON.stringify(browserCheck)},"--page","index.html","--assert","#app","--assert-count","#items:3","--click","#btn","--then-text","#out:done","--console-clean","--screenshot","artifacts/ui.png"]. It serves the workspace over HTTP, runs jsdom assertions with real inline-script execution (clicks, text, selectors, console-error detection), and captures a real screenshot with the bundled lightweight headless browser (auto-downloaded on first use).`
@@ -24,8 +26,8 @@ export function shipGuidance({ state, workflow, prompt = "", profiles = [], disc
   const task = originalTask(prompt)
   const mode = state && !["verified", "blocked"].includes(state.status) ? state.workflowMode : workflow
   const parts = mode === "standard"
-    ? ["Preserve every user requirement and promised output.", STANDARD_GUIDANCE]
-    : [DELIVERY_INVARIANTS, deliveryPhaseGuidance(state)]
+    ? ["Preserve every user requirement and promised output.", STANDARD_GUIDANCE, TOOL_SURFACE_GUIDANCE]
+    : [DELIVERY_INVARIANTS, deliveryPhaseGuidance(state), TOOL_SURFACE_GUIDANCE]
   if (/\b(test|typescript|route|backend|api)\b/i.test(task)) parts.push(TEST_INFRASTRUCTURE_GUIDANCE)
   if (profiles.some(profile => /web|browser/.test(profile.id)) || /\b(browser|web|frontend|screenshot)\b/i.test(task)) parts.push(browserGuidance)
   if (imageInput === false) parts.push("This session's model cannot process images: do not attempt to view, read, or reason over image files or screenshots — visual review is unavailable. Verify visual behavior through DOM structure, computed styles, and console evidence instead; screenshots remain useful as evidence artifacts for the user even though you cannot inspect them.")

package/plugins/carthagent/src/tui.tsx

@@ -17,6 +17,7 @@ import { VoyagesNavigation } from "./ui/voyages-navigation.js";
 import { nativeNavigation } from "./ui/native-navigation.js";
 import { provisionNativeThemes } from "./ui/theme-provision.mjs";
 import { setSessionAutoCompaction } from "./ui/session-preferences.mjs";
+import { runModelCommand } from "./ui/model-command.js";
 import { createNativeReviewConsumer, nativeReviewMode, sendNativeReviewKickoff, subscribeNativeReviewEvents } from "./ui/native-review.mjs";
 
 const THEMES = ["carthage", "tokyonight", "nebula", "ember", "forest", "mono", "obsidian", "midnight", "solarized-dark", "nord", "contrast-dark", "okabe-dark", "paper", "daylight", "solarized-light", "contrast-light", "okabe-light", "solarized", "okabe", "contrast", "system"];
@@ -45,8 +46,13 @@ export default Plugin.define({
         else reviewLoops.delete(sessionID);
         void updateReviewLoopState((state) => { state.sessions = [...reviewLoops]; });
       },
+      // No formal delivery (unavailable) or a verified one; blocked/unfinished runs are not offered.
+      eligible: async (sessionID: string, signal?: AbortSignal) => {
+        const delivery = await workspaceClient(context, context.data.session.get(sessionID)?.location).delivery({ sessionID }, signal ? AbortSignal.any([signal, AbortSignal.timeout(15_000)]) : AbortSignal.timeout(15_000));
+        return delivery.status === "unavailable" || delivery.status === "verified";
+      },
       offer: async ({ sessionID, writes, start }: { sessionID: string; writes: number; start: () => Promise<boolean> }) => {
-        const accepted = await context.ui.dialog.confirm({ title: `Review ${context.data.session.get(sessionID)?.title ?? sessionID}?`, message: `This successful run changed ${writes} file${writes === 1 ? "" : "s"}. Start the PR → fresh-context review → fixes workflow?`, label: { confirm: "Start review", cancel: "Not now" } });
+        const accepted = await context.ui.dialog.confirm({ title: `Review ${context.data.session.get(sessionID)?.title ?? sessionID}?`, message: `This run changed ${writes} file${writes === 1 ? "" : "s"}. Start the PR → fresh-context review → fixes workflow?`, label: { confirm: "Start review", cancel: "Not now" } });
         if (accepted) await start();
       },
       kickoff: async (reviewSessionID: string, signal?: AbortSignal) => sendNativeReviewKickoff({ client: context.client, sessionID: reviewSessionID, signal }),
@@ -175,19 +181,17 @@ export default Plugin.define({
           }
           return;
         }
-        case "model": {
-          await context.data.location.model.sync(location);
-          const models = context.data.location.model.list(location) ?? [];
-          const query = arg.toLowerCase();
-          const options = models.filter((model) => !query || `${nativeProviderName(model.providerID)} ${model.providerID}/${model.modelID} ${model.name}`.toLowerCase().includes(query))
-            .map((model) => ({ title: `${nativeProviderName(model.providerID)}/${model.modelID}`, value: model, description: model.name }));
-          const selected = await context.ui.dialog.select({ title: "Select provider model", placeholder: "Search provider/model…", options });
-          if (selected && sessionID) {
-            await context.client.session.switchModel({ sessionID, model: { providerID: selected.providerID, id: selected.id } });
-            toast(`Selected ${nativeProviderName(selected.providerID)}/${selected.modelID}.`, "success");
-          } else if (selected) toast("Open or create a session before selecting its model.", "warning");
+        case "model":
+          await runModelCommand({
+            sync: () => context.data.location.model.sync(location),
+            list: () => context.data.location.model.list(location) ?? [],
+            select: (options) => context.ui.dialog.select({ title: "Select provider model", placeholder: "Search provider/model…", options }),
+            switchModel: (target, model) => context.client.session.switchModel({ sessionID: target, model }),
+            openNativePicker: (query) => context.keymap.dispatch("model.list", query),
+            providerName: nativeProviderName,
+            toast,
+          }, { sessionID, arg });
           return;
-        }
         case "thinking": {
           if (!selectedModel) { toast("Choose a model first.", "warning"); return; }
           const variants = context.ui.model.variant.list();

package/plugins/carthagent/src/ui/daily-quests.tsx

@@ -1,8 +1,9 @@
 /** @jsxImportSource @opentui/solid */
-import { For, Show, createEffect, createSignal, onCleanup, onMount } from "solid-js";
+import { For, Show, createEffect, createSignal, on, onCleanup, onMount } from "solid-js";
 import { createHash, randomUUID } from "node:crypto";
 import type { ScrollBoxRenderable, SelectRenderable, TextareaRenderable } from "@opentui/core";
 import { usePlugin } from "@opencode/plugin/tui";
+import { clickSelect } from "./select-mouse.js";
 import { dailyQuestsClient, dailySourceClient } from "./client.js";
 import type { DailyEvidence, DailyQuestDetail, DailySnapshot, DailySubmission, DailyRecord } from "../daily-quests-rpc.js";
 import type { DailySourceReport } from "../daily-source-rpc.js";
@@ -192,6 +193,8 @@ export function DailyQuestsPage(props: { sessionID?: string; questID?: string })
     // A navigator link selects only an assignment in the authenticated slate.
     // It never composes evidence or opens a consent dialog.
     void controller.refresh().then(() => { if (props.questID) return controller.select(props.questID); });
+    // The router reuses this page when another rail quest is opened.
+    createEffect(on(() => props.questID, (questID) => { if (questID) void controller.select(questID); }, { defer: true }));
     onCleanup(() => { focus.dispose(); stop(); actions = undefined; controller.dispose(); closeOwnDialog(); });
   });
   return <box flexDirection="column" flexGrow={1} padding={1}>
@@ -227,6 +230,6 @@ export function DailyQuestsPage(props: { sessionID?: string; questID?: string })
       <textarea ref={editor} flexBasis={0} flexGrow={1} minHeight={2} initialValue={view().draft?.text ?? ""} focused={focus.editorEnabled()} onContentChange={() => syncDailyEditor(editor, controller)} keyBindings={[{ name: "p", ctrl: true, action: "submit" }]} onSubmit={preview} />
     </Show>
     <text flexShrink={0} fg={context.theme.text.faint}>↑/↓ Enter · PgUp/Dn · Ctrl+U/D details · Cancel: no send</text>
-    <select ref={(node) => { actions = node; focus.attachActions(node); }} flexBasis={0} flexGrow={1} flexShrink={0} minHeight={3} showScrollIndicator options={options()} focused={focus.actionsEnabled()} keyBindings={[{ name: "pageup", action: "move-up-fast" }, { name: "pagedown", action: "move-down-fast" }]} onSelect={(_index, option) => select(option?.value)} />
+    <select ref={(node) => { actions = node; focus.attachActions(node); }} onMouseUp={(event) => { if (focus.actionsEnabled()) clickSelect(actions, event); }} flexBasis={0} flexGrow={1} flexShrink={0} minHeight={3} showScrollIndicator options={options()} focused={focus.actionsEnabled()} keyBindings={[{ name: "pageup", action: "move-up-fast" }, { name: "pagedown", action: "move-down-fast" }]} onSelect={(_index, option) => select(option?.value)} />
   </box>;
 }

package/plugins/carthagent/src/ui/model-command.ts

@@ -0,0 +1,32 @@
+type CatalogModel = { providerID: string; modelID: string; id: string; name: string };
+
+export type ModelCommandHost = {
+  sync(): Promise<unknown>;
+  list(): readonly CatalogModel[];
+  select(options: Array<{ title: string; value: CatalogModel; description: string }>): Promise<CatalogModel | undefined>;
+  switchModel(sessionID: string, model: { providerID: string; id: string }): Promise<unknown>;
+  /** Opens the native configured-model picker (the `/models` command) with an optional search. */
+  openNativePicker(query: string): void;
+  providerName(providerID: string): string;
+  toast(message: string, variant: "success" | "warning"): void;
+};
+
+/**
+ * `/model [name]`. Without a session the choice belongs to the next session's draft, which only the native
+ * picker owns, so it is opened directly; inside a session the selection is committed to that session.
+ */
+export async function runModelCommand(host: ModelCommandHost, input: { sessionID?: string; arg: string }) {
+  if (!input.sessionID) {
+    host.openNativePicker(input.arg);
+    return;
+  }
+  await host.sync();
+  const query = input.arg.toLowerCase();
+  const options = host.list()
+    .filter((model) => !query || `${host.providerName(model.providerID)} ${model.providerID}/${model.modelID} ${model.name}`.toLowerCase().includes(query))
+    .map((model) => ({ title: `${host.providerName(model.providerID)}/${model.modelID}`, value: model, description: model.name }));
+  const selected = await host.select(options);
+  if (!selected) return;
+  await host.switchModel(input.sessionID, { providerID: selected.providerID, id: selected.id });
+  host.toast(`Selected ${host.providerName(selected.providerID)}/${selected.modelID}.`, "success");
+}

package/plugins/carthagent/src/ui/native-review.d.mts

@@ -10,6 +10,8 @@ export interface NativeReviewConsumerOptions {
   diff: (sessionID: string, signal?: AbortSignal) => Promise<unknown>;
   offer: (value: NativeReviewOffer) => Promise<unknown>;
   kickoff: (sessionID: string, signal?: AbortSignal) => Promise<unknown>;
+  /** False or a rejection suppresses the offer (e.g. blocked formal delivery). */
+  eligible?: (sessionID: string, signal?: AbortSignal) => Promise<boolean>;
   isLoopActive?: (sessionID: string) => boolean;
   setLoopActive?: (sessionID: string, active: boolean) => void;
   signal?: AbortSignal;