Merge remote-tracking branch 'origin/main' into tmp/merge-pr-70

# Conflicts:
#	UPSTREAM_DIVERGENCE.md

Author: tyulyukov

AGENTS.md

@@ -168,6 +168,20 @@ The same applies to `py-*` vs `pt-*`/`pb-*`.
 
 Use these as implementation references when designing protocol handling, UX flows, and operational safeguards.
 
+## Upstream Merge: Divergence Ledger
+
+Before starting any upstream sync, read [`UPSTREAM_DIVERGENCE.md`](./UPSTREAM_DIVERGENCE.md). It is the audit trail of which upstream PRs are already ported (under possibly-different MarCode SHAs), which are already behaviorally equivalent and must NOT be re-ported, which are intentionally skipped, and what is genuinely pending. Its sister doc [`FEATURES.md`](./FEATURES.md) lists MarCode-exclusive features that must be preserved through any merge.
+
+The ledger's own preamble describes the exact `git cherry origin/main upstream/main` workflow for computing real pending work. Always subtract the "Already equivalent" and "Intentionally skipped" sets before scoping a cycle — re-porting an already-equivalent PR creates pointless conflicts on FEATURES.md-protected files.
+
+**After a cycle:**
+
+1. Add new rows to the "Ported in the current cycle" section (upstream PR number + new MarCode SHA + conflict-resolution notes, especially any FEATURES.md-protected substitutions like `AnalyticsServiceNoopLive` in place of upstream `AnalyticsService.layerTest`).
+2. Move finished rows out of "Pending real work".
+3. Advance "Baseline after cycle" to the last merge SHA of the cycle.
+4. Record any data-migration smoke results (real DB rows migrated, legacy shapes surviving / none).
+5. Run `bun run fmt` — CI enforces `oxfmt --check` on markdown, so unformatted table column widths will fail the Format step.
+
 ## Upstream Merge: Migration Ordering
 
 MarCode has its own database migrations that were added at specific IDs. When merging upstream changes that introduce NEW migrations, **never renumber existing MarCode migrations** — existing users already have them applied at their original IDs. The Effect SQL migrator tracks migrations by numeric ID in the `effect_sql_migrations` table; renumbering causes it to skip the new upstream tables (thinking those IDs are done) and attempt to re-create existing tables at the new IDs.

UPSTREAM_DIVERGENCE.md

@@ -37,7 +37,7 @@ Sections are ordered by action:
 
 ## Ported in the current cycle
 
-**Cycle:** 2026-04-24 · Baseline before cycle: `7c430aece` · Baseline after cycle: `9f6411d17`
+**Cycle:** 2026-04-24 · Baseline before cycle: `7c430aece` · Baseline after cycle: `ececcdcb1`
 
 ### Direct-to-main (no PR, user-approved)
 
@@ -82,13 +82,19 @@ Sections are ordered by action:
 - `toast.tsx` — took upstream's version wholesale (492 → 719 lines). Upstream already bundled `CopyErrorButton` at line 93, so no manual re-integration needed.
 - `Sidebar.tsx` — preserved MarCode's "Delete anyway" warning toast flow (action button → deferred close → `api.dialogs.confirm` with thread-count messaging → `removeProject({ force: true })` → inline error toast). Wrapped all toast calls via the new `stackedThreadToast(...)` helper for layout consistency.
 
+### PR #68 — cycle ledger bootstrap
+
+Introduced this doc. No upstream PR; meta-work that codifies the "Already equivalent" / "Intentionally skipped" sets so future cycles don't re-derive them.
+
+New SHA: `ee646a654`.
+
 ### PR #69 — sidebar row timestamp (part 1 of upstream #1996)
 
 | Upstream                                               | Subject                                                  | New SHA     |
 | ------------------------------------------------------ | -------------------------------------------------------- | ----------- |
 | [#1996](https://github.com/pingdotgg/t3code/pull/1996) | Use latest user message time for thread timestamps (row) | `524e93afd` |
 
-Narrow behavioral port: `Sidebar.tsx:715` thread-row label now falls back `thread.latestUserMessageAt ?? thread.updatedAt ?? thread.createdAt`. Matches the helper already in `CommandPalette.logic.ts:161`.
+Narrow behavioral port: `Sidebar.tsx:715` thread-row label now falls back `thread.latestUserMessageAt ?? thread.updatedAt ?? thread.createdAt`. Matches the helper already in `CommandPalette.logic.ts:161`. Preserves MarCode's structural-sharing event handling ([FEATURES.md §"Incremental Event Handling"](./FEATURES.md#incremental-event-handling--structural-sharing)).
 
 ### PR #70 — shell-stream authority for sidebar summary (part 2 of upstream #1996)
 
@@ -100,6 +106,27 @@ Narrow behavioral port: `Sidebar.tsx:715` thread-row label now falls back `threa
 
 What we deliberately did NOT port from upstream #1996: structural reorganization of `store.ts` (commentary, `ensureThreadRegistered` extraction, `retainThreadScopedRecord` changes). MarCode's Incremental Event Handling & Structural Sharing (FEATURES.md §1) already covers the functional behavior; upstream's cosmetic reshuffle is net risk for no gain.
 
+### PR #71 — provider model selection option arrays (upstream #2246)
+
+| Upstream                                               | Subject                                             | New SHA                                                                                                                                      |
+| ------------------------------------------------------ | --------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------- |
+| [#2246](https://github.com/pingdotgg/t3code/pull/2246) | Refactor provider model selections to option arrays | `d776a9d12` + `62fd6ea4b` + `b7f903e7e` + `e2601893d` + `5f75e700d` + `c1bce4fb7` + `9bb54da84` (A→E + formatter + test-alignment follow-up) |
+
+**Why ported:** forward compatibility with upstream's `provider-instance-registry` branch (`ceddb40 / 7a466f / 8a82b53 / 84b0d74`), which extends the new option-array shape. Any future sync that touches provider identity assumes #2246 is in.
+
+**Key deviations from upstream:**
+
+- Migration id renumbered `026 → 030` because MarCode already uses `26_AuthSessionLastConnectedAt`. Current head after port: `030_CanonicalizeModelSelectionOptions`.
+- `RoutingTextGeneration.ts` — kept MarCode's Claude→Codex fallback branching intact (FEATURES.md §"Claude-Powered Text Generation"); only the option access was retrofitted with `getModelSelectionStringOptionValue` / `getModelSelectionBooleanOptionValue` helpers.
+- `ClaudeTextGeneration.ts` — preserved MarCode's fork-exclusive progressive generation code path.
+- `composerProviderRegistry.tsx` (deleted upstream) — local additions hand-ported into the new `composerProviderState.tsx` rather than deleted.
+- `ProviderModelPicker.browser.tsx` + `composerDraftStore.ts` — near-total hand-rebuild since upstream ±399 LoC collides with MarCode's +397 LoC delta.
+- All new test layers routed through `AnalyticsServiceNoopLive` (no upstream `AnalyticsService.layerTest`).
+
+**Post-merge hotfix** (on main, direct commits): `composerDraftStore.ts` gained `normalizeModelSelectionByProviderMap` to coerce legacy object-shape `options` blobs in pre-existing v5 localStorage (migration 030 only touches SQLite — persisted drafts in the browser store needed a separate runtime normalizer). Regression guard in `composerDraftStore.test.ts` under `describe("composerDraftStore legacy modelSelection options migration")`.
+
+**Phase 4 smoke** (real dev DB, 2026-04-24): migration 030 applied cleanly — 155 `projection_threads` rows canonicalized, 0 legacy-shape survivors across `projection_threads`, `projection_projects`, and `orchestration_events` (`project.{created,meta-updated}`, `thread.{created,meta-updated,turn-start-requested}`).
+
 ---
 
 ## Already equivalent under a different SHA
@@ -160,9 +187,7 @@ MarCode ships semver alphas (`1.0.0-alpha.*`), not nightly builds. Adopting nigh
 
 ## Pending real work
 
-| Upstream                                               | Subject                                             | Risk | Notes                                                                                                                                                                                                                                                                                                                                                                                                                          |
-| ------------------------------------------------------ | --------------------------------------------------- | ---- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
-| [#2246](https://github.com/pingdotgg/t3code/pull/2246) | Refactor provider model selections to option arrays | High | Adds migration `026_CanonicalizeModelSelectionOptions` (renumber to `030` — MarCode's `26_AuthSessionLastConnectedAt` already owns that id). Touches every `*TextGeneration.ts` (collides with FEATURES.md §"Claude-Powered Text Generation") and the MarCode provider instance registry work (`0e71e3023`, `42b428826`, `4da47be23`). 67 files, +3403/-3828. Plan: `/Users/tyulyukov/.claude/plans/cached-napping-sundae.md`. |
+_None as of 2026-04-24._ Both previously-listed rows (#1996 and #2246) landed in this cycle — #1996 across PRs #69 + #70, #2246 via PR #71. Re-run the `git cherry origin/main upstream/main` workflow at the top of this doc when starting a new cycle to populate this section.
 
 ---
 

apps/server/src/git/Layers/ClaudeTextGeneration.test.ts

@@ -202,10 +202,10 @@ it.layer(ClaudeTextGenerationTestLayer)("ClaudeTextGenerationLive", (it) => {
           modelSelection: {
             provider: "claudeAgent",
             model: "claude-haiku-4-5",
-            options: {
-              thinking: false,
-              effort: "high",
-            },
+            options: [
+              { id: "thinking", value: false },
+              { id: "effort", value: "high" },
+            ],
           },
         });
 
@@ -265,10 +265,10 @@ it.layer(ClaudeTextGenerationTestLayer)("ClaudeTextGenerationLive", (it) => {
           modelSelection: {
             provider: "claudeAgent",
             model: "claude-opus-4-6",
-            options: {
-              effort: "max",
-              fastMode: true,
-            },
+            options: [
+              { id: "effort", value: "max" },
+              { id: "fastMode", value: true },
+            ],
           },
         });
 

apps/server/src/git/Layers/ClaudeTextGeneration.ts

@@ -29,10 +29,17 @@ import {
   sanitizeThreadTitle,
   toJsonSchemaObject,
 } from "../Utils.ts";
-import { normalizeClaudeModelOptionsWithCapabilities } from "@marcode/shared/model";
-import { resolveClaudeApiModelId } from "../../provider/Layers/ClaudeProvider.ts";
+import {
+  getModelSelectionStringOptionValue,
+  getProviderOptionDescriptors,
+} from "@marcode/shared/model";
+import {
+  getClaudeModelCapabilities,
+  normalizeClaudeCliEffort,
+  resolveClaudeApiModelId,
+  resolveClaudeEffort,
+} from "../../provider/Layers/ClaudeProvider.ts";
 import { ServerSettingsService } from "../../serverSettings.ts";
-import { getClaudeModelCapabilities } from "../../provider/Layers/ClaudeProvider.ts";
 
 const CLAUDE_TIMEOUT_MS = 180_000;
 
@@ -85,15 +92,24 @@ const makeClaudeTextGeneration = Effect.gen(function* () {
     modelSelection: ClaudeModelSelection;
   }): Effect.fn.Return<S["Type"], TextGenerationError, S["DecodingServices"]> {
     const jsonSchemaStr = JSON.stringify(toJsonSchemaObject(outputSchemaJson));
-    const normalizedOptions = normalizeClaudeModelOptionsWithCapabilities(
-      getClaudeModelCapabilities(modelSelection.model),
-      modelSelection.options,
-    );
+    const caps = getClaudeModelCapabilities(modelSelection.model);
+    const descriptors = getProviderOptionDescriptors({
+      caps,
+      selections: modelSelection.options,
+    });
+    const findDescriptor = (id: string) => descriptors.find((descriptor) => descriptor.id === id);
+    const rawEffortSelection = getModelSelectionStringOptionValue(modelSelection, "effort");
+    const resolvedEffort = resolveClaudeEffort(caps, rawEffortSelection);
+    const cliEffort = normalizeClaudeCliEffort(resolvedEffort);
+    const thinkingDescriptor = findDescriptor("thinking");
+    const fastModeDescriptor = findDescriptor("fastMode");
+    const thinking =
+      thinkingDescriptor?.type === "boolean" ? thinkingDescriptor.currentValue : undefined;
+    const fastMode =
+      fastModeDescriptor?.type === "boolean" ? fastModeDescriptor.currentValue : undefined;
     const settings = {
-      ...(typeof normalizedOptions?.thinking === "boolean"
-        ? { alwaysThinkingEnabled: normalizedOptions.thinking }
-        : {}),
-      ...(normalizedOptions?.fastMode ? { fastMode: true } : {}),
+      ...(typeof thinking === "boolean" ? { alwaysThinkingEnabled: thinking } : {}),
+      ...(fastMode ? { fastMode: true } : {}),
     };
 
     const claudeSettings = yield* Effect.map(
@@ -112,7 +128,7 @@ const makeClaudeTextGeneration = Effect.gen(function* () {
           jsonSchemaStr,
           "--model",
           resolveClaudeApiModelId(modelSelection),
-          ...(normalizedOptions?.effort ? ["--effort", normalizedOptions.effort] : []),
+          ...(cliEffort ? ["--effort", cliEffort] : []),
           ...(Object.keys(settings).length > 0 ? ["--settings", JSON.stringify(settings)] : []),
           "--tools",
           "",

apps/server/src/git/Layers/CodexTextGeneration.test.ts

@@ -247,10 +247,10 @@ it.layer(CodexTextGenerationTestLayer)("CodexTextGenerationLive", (it) => {
             modelSelection: {
               provider: "codex",
               model: "gpt-5.4",
-              options: {
-                reasoningEffort: "xhigh",
-                fastMode: true,
-              },
+              options: [
+                { id: "reasoningEffort", value: "xhigh" },
+                { id: "fastMode", value: true },
+              ],
             },
           });
         }),

apps/server/src/git/Layers/CodexTextGeneration.ts

@@ -5,6 +5,10 @@ import { ChildProcess, ChildProcessSpawner } from "effect/unstable/process";
 
 import { CodexModelSelection } from "@marcode/contracts";
 import { sanitizeBranchFragment, sanitizeFeatureBranchName } from "@marcode/shared/git";
+import {
+  getModelSelectionBooleanOptionValue,
+  getModelSelectionStringOptionValue,
+} from "@marcode/shared/model";
 
 import { resolveAttachmentPath } from "../../attachmentStore.ts";
 import { ServerConfig } from "../../config.ts";
@@ -156,7 +160,9 @@ const makeCodexTextGeneration = Effect.gen(function* () {
 
     const runCodexCommand = Effect.fn("runCodexJson.runCodexCommand")(function* () {
       const reasoningEffort =
-        modelSelection.options?.reasoningEffort ?? CODEX_GIT_TEXT_GENERATION_REASONING_EFFORT;
+        getModelSelectionStringOptionValue(modelSelection, "reasoningEffort") ??
+        CODEX_GIT_TEXT_GENERATION_REASONING_EFFORT;
+      const fastMode = getModelSelectionBooleanOptionValue(modelSelection, "fastMode") === true;
       const command = ChildProcess.make(
         codexSettings?.binaryPath || "codex",
         [
@@ -169,7 +175,7 @@ const makeCodexTextGeneration = Effect.gen(function* () {
           modelSelection.model,
           "--config",
           `model_reasoning_effort="${reasoningEffort}"`,
-          ...(modelSelection.options?.fastMode ? ["--config", `service_tier="fast"`] : []),
+          ...(fastMode ? ["--config", `service_tier="fast"`] : []),
           "--output-schema",
           schemaPath,
           "--output-last-message",

apps/server/src/git/Layers/CursorTextGeneration.test.ts

@@ -137,11 +137,11 @@ it.layer(CursorTextGenerationTestLayer)("CursorTextGenerationLive", (it) => {
           modelSelection: {
             provider: "cursor",
             model: "gpt-5.4",
-            options: {
-              reasoning: "xhigh",
-              fastMode: true,
-              contextWindow: "1m",
-            },
+            options: [
+              { id: "reasoning", value: "xhigh" },
+              { id: "fastMode", value: true },
+              { id: "contextWindow", value: "1m" },
+            ],
           },
         });
 

apps/server/src/git/Layers/CursorTextGeneration.ts

@@ -108,7 +108,7 @@ const makeCursorTextGeneration = Effect.gen(function* () {
         yield* applyCursorAcpModelSelection({
           runtime,
           model: modelSelection.model,
-          modelOptions: modelSelection.options,
+          selections: modelSelection.options,
           mapError: ({ cause, configId, step }) =>
             mapCursorAcpError(
               operation,

apps/server/src/git/Layers/GitCore.test.ts

@@ -2285,6 +2285,80 @@ it.layer(TestLayer)("git integration", (it) => {
       }),
     );
 
+    describe("readWorkingTreeDiff", () => {
+      it.effect("includes unstaged modifications to tracked files", () =>
+        Effect.gen(function* () {
+          const tmp = yield* makeTmpDir();
+          yield* initRepoWithCommit(tmp);
+          const core = yield* GitCore;
+
+          yield* writeTextFile(path.join(tmp, "README.md"), "# test\nunstaged change\n");
+
+          const result = yield* core.readWorkingTreeDiff(tmp);
+
+          expect(result.truncated).toBe(false);
+          expect(result.diff).toContain("diff --git a/README.md b/README.md");
+          expect(result.diff).toContain("+unstaged change");
+        }),
+      );
+
+      it.effect("includes untracked files as new-file diffs", () =>
+        Effect.gen(function* () {
+          const tmp = yield* makeTmpDir();
+          yield* initRepoWithCommit(tmp);
+          const core = yield* GitCore;
+
+          yield* writeTextFile(path.join(tmp, "new-file.txt"), "brand new line\n");
+
+          const result = yield* core.readWorkingTreeDiff(tmp);
+
+          expect(result.truncated).toBe(false);
+          expect(result.diff).toContain("new-file.txt");
+          expect(result.diff).toContain("new file mode");
+          expect(result.diff).toContain("+brand new line");
+        }),
+      );
+
+      it.effect("includes both tracked-unstaged changes and untracked files in one diff", () =>
+        Effect.gen(function* () {
+          const tmp = yield* makeTmpDir();
+          yield* initRepoWithCommit(tmp);
+          const core = yield* GitCore;
+
+          yield* writeTextFile(path.join(tmp, "README.md"), "# test\nmodified line\n");
+          yield* writeTextFile(path.join(tmp, "fresh.txt"), "untracked content\n");
+
+          const result = yield* core.readWorkingTreeDiff(tmp);
+
+          expect(result.diff).toContain("README.md");
+          expect(result.diff).toContain("+modified line");
+          expect(result.diff).toContain("fresh.txt");
+          expect(result.diff).toContain("+untracked content");
+        }),
+      );
+
+      it.effect("respects .gitignore when enumerating untracked files", () =>
+        Effect.gen(function* () {
+          const tmp = yield* makeTmpDir();
+          yield* initRepoWithCommit(tmp);
+          const core = yield* GitCore;
+
+          yield* writeTextFile(path.join(tmp, ".gitignore"), "ignored.txt\n");
+          yield* git(tmp, ["add", ".gitignore"]);
+          yield* git(tmp, ["commit", "-m", "add gitignore"]);
+          yield* writeTextFile(path.join(tmp, "ignored.txt"), "should be hidden\n");
+          yield* writeTextFile(path.join(tmp, "visible.txt"), "should appear\n");
+
+          const result = yield* core.readWorkingTreeDiff(tmp);
+
+          expect(result.diff).toContain("visible.txt");
+          expect(result.diff).toContain("+should appear");
+          expect(result.diff).not.toContain("ignored.txt");
+          expect(result.diff).not.toContain("should be hidden");
+        }),
+      );
+    });
+
     it.effect("falls back to empty remote branch data when remote lookups fail", () =>
       Effect.gen(function* () {
         const tmp = yield* makeTmpDir();