Harden Wave 2 picker and compaction review follow-ups

CHANGELOG.md

@@ -37,6 +37,20 @@ Format loosely follows [Keep a Changelog](https://keepachangelog.com/). Versions
   mouse back for native terminal selection; Alt+C remains the keyboard copy
   path for whole messages, tool outputs, and diffs.
 
+### Session
+
+- **Superseded read stubs.** When compaction keeps more than one successful
+  `read_file` of the same path (or same path+offset+limit range), older results
+  become a one-line stub and the newest stays whole. Errors stay verbatim.
+  Dedup only considers turns that survive compaction, so a summarized re-read
+  cannot hollow a kept older body.
+
+### Tooling
+
+- **Vendored patch ledger.** `bin/vendor-patch-diff` and
+  `vendor/intx-inference/PATCHES.md` site markers prove local patches against
+  upstream without a manual re-sync checklist.
+
 ## [0.2.95] - 2026-08-09
 
 Tool-only auto-pause that no longer stops healthy work, resume the last session

docs/TUI.md

@@ -424,6 +424,9 @@ narrows the list in place (printable keys claimed by the filter row, same
 pattern as the command palette); Enter selects. Escape closes the picker.
 The row matching the session's live active model gets a `(current)` suffix.
 Alt+F on a model row still toggles favorite when a favorite hook is wired.
+While type-to-filter is active, bare `j`/`k` type into the filter rather than
+moving the highlight — use arrow keys (or the filtered list's navigation) to
+move.
 
 Onboarding (the standalone provider-setup screen, `provider-setup.ts`) and
 the satellite pickers used for session resume and session-mode selection
@@ -663,11 +666,6 @@ asserted as fact:
   tiny-terminal, sub-24-row path) has a corresponding test that pins the
   exact row counts, or whether some of that path is only exercised
   indirectly.
-- Whether the `(current)` marking on a provider *group* row
-  (`withGroupMark` in `openModels`, `product-host.ts`) is reachable and
-  correct in every case where the active model's provider itself has no
-  favorites/recents entry — the code path exists but was not traced through
-  a live picker session.
 - Full coverage of which chords are guaranteed deliverable on every terminal
   emulator Corbits Code targets (Shift+Enter and Alt+letter reporting depend
   on kitty-protocol negotiation the harness cannot test — see Test-harness

src/session/compactor.ts

@@ -207,7 +207,7 @@ export type CompactorConfig = {
 };
 
 const DEFAULT_COMPACTOR_CONFIG: CompactorConfig = {
-  keepRecentTurns: 5,
+  keepRecentTurns: 6,
   summaryMaxChars: 2000,
   maxAnchorTurns: 8,
 };
@@ -231,11 +231,17 @@ const ANCHOR_SCORE_THRESHOLD = 5;
 // Tool names whose results are path-keyed for re-read dedup during compaction.
 const READ_TOOLS = new Set(["read_file"]);
 
-// Call-id index for stub rendering (name + path). Not path-keyed — that is
-// buildPathToReads below.
+// Call-id index for stub rendering (name + path). Dedup keys live on `readKey`.
 type ToolCallInfo = {
   name: string;
+  /** Display path for stubs (always the raw path arg when present). */
   pathArg?: string;
+  /**
+   * Dedup identity for re-read stubbing. Full-file reads share the path alone;
+   * ranged reads (offset/limit) get a distinct key so chunked reads of the same
+   * file do not hollow each other.
+   */
+  readKey?: string;
 };
 
 type PathRead = {
@@ -245,7 +251,20 @@ type PathRead = {
   isError: boolean;
 };
 
-function pathArgFromArguments(raw: unknown): string | undefined {
+function scalarArg(value: unknown): string {
+  if (typeof value === "number" && Number.isFinite(value)) return String(value);
+  if (typeof value === "string") return value;
+  return "";
+}
+
+/**
+ * Extract path + re-read identity from a tool_call's arguments.
+ * Identity is path alone for full-file reads; path+offset+limit when either
+ * range arg is present so partial reads don't supersede each other.
+ */
+function readIdentityFromArguments(
+  raw: unknown,
+): { path: string; readKey: string } | undefined {
   let args: unknown = raw ?? {};
   if (typeof args === "string") {
     try {
@@ -255,8 +274,16 @@ function pathArgFromArguments(raw: unknown): string | undefined {
     }
   }
   if (args === null || typeof args !== "object" || Array.isArray(args)) return undefined;
-  const path = (args as Record<string, unknown>)["path"];
-  return typeof path === "string" && path.length > 0 ? path : undefined;
+  const rec = args as Record<string, unknown>;
+  const path = rec["path"];
+  if (typeof path !== "string" || path.length === 0) return undefined;
+  const offsetPart = scalarArg(rec["offset"]);
+  const limitPart = scalarArg(rec["limit"]);
+  const readKey =
+    offsetPart === "" && limitPart === ""
+      ? path
+      : `${path}\0${offsetPart}\0${limitPart}`;
+  return { path, readKey };
 }
 
 // callId → tool name/path for readable stubs. Inverse of path-to-reads.
@@ -266,18 +293,26 @@ function buildCallIndex(turns: readonly ConversationTurn[]): Map<string, ToolCal
     for (const block of turn.content) {
       if (block.type !== "tool_call") continue;
       const info: ToolCallInfo = { name: block.name };
-      const path = pathArgFromArguments(block.arguments);
-      if (path !== undefined) info.pathArg = path;
+      const identity = readIdentityFromArguments(block.arguments);
+      if (identity !== undefined) {
+        info.pathArg = identity.path;
+        info.readKey = identity.readKey;
+      }
       index.set(block.id, info);
     }
   }
   return index;
 }
 
 /**
- * Path → every read_file result that targeted it, in session order.
- * Groups repeated reads so older successful results can be stubbed when a
- * later read of the same path survives compaction.
+ * Read-identity → every read_file result that matched it, in session order.
+ * Groups repeated full-file (or same-range) reads so older successful results
+ * can be stubbed when a later identical read survives compaction.
+ *
+ * Callers must pass only turns that survive compaction (anchors + recent).
+ * Computing supersession over the full transcript would hollow a kept older
+ * read when the newer re-read was summarized away — leaving the model with a
+ * stub and no full body.
  */
 function buildPathToReads(
   turns: readonly ConversationTurn[],
@@ -289,14 +324,14 @@ function buildPathToReads(
     for (const block of turn.content) {
       if (block.type !== "tool_result") continue;
       const info = callIndex.get(block.callId);
-      if (info === undefined || !READ_TOOLS.has(info.name) || info.pathArg === undefined) continue;
+      if (info === undefined || !READ_TOOLS.has(info.name) || info.readKey === undefined) continue;
       const entry: PathRead = {
         callId: block.callId,
         order: order++,
         isError: block.isError === true,
       };
-      const list = pathToReads.get(info.pathArg);
-      if (list === undefined) pathToReads.set(info.pathArg, [entry]);
+      const list = pathToReads.get(info.readKey);
+      if (list === undefined) pathToReads.set(info.readKey, [entry]);
       else list.push(entry);
     }
   }
@@ -305,8 +340,9 @@ function buildPathToReads(
 
 /**
  * Call ids of successful read_file results that are superseded by a later
- * successful read of the same path. Error results never appear here — they
- * stay verbatim so the model still sees the failure.
+ * successful read of the same identity (path, or path+offset+limit). Error
+ * results never appear here — they stay verbatim so the model still sees the
+ * failure.
  */
 function supersededReadCallIds(pathToReads: ReadonlyMap<string, PathRead[]>): Set<string> {
   const superseded = new Set<string>();
@@ -494,7 +530,7 @@ export function createPruningCompactor(
 
   return {
     name: "pruning-compactor",
-    version: "1.3.0",
+    version: "1.3.1",
     async apply(
       turns: ConversationTurn[],
       _ctx: StrategyContext,
@@ -520,12 +556,9 @@ export function createPruningCompactor(
         };
       }
 
-      // callId → name/path for stubs; path → ordered reads for re-read dedup.
-      // Only older successful reads of a path re-read later are stubbed — not a
-      // blanket strip of every kept tool_result (see CL-5595 / CL-4374).
+      // callId → name/path for stubs. Built over the full transcript so a kept
+      // result can still name its path even when its call turn was summarized.
       const callIndex = buildCallIndex(aged.turns);
-      const pathToReads = buildPathToReads(aged.turns, callIndex);
-      const supersededReads = supersededReadCallIds(pathToReads);
 
       const keepCount = Math.min(cfg.keepRecentTurns, aged.turns.length - 1);
       const keepFrom = aged.turns.length - keepCount;
@@ -568,6 +601,12 @@ export function createPruningCompactor(
       const anchorTurns = sortedAnchorIndices.map((i) => olderTurns[i]!);
       const summarizedTurns = olderTurns.filter((_, i) => !anchorIndices.has(i));
 
+      // Path-dedup only among turns that survive. Supersession over the full
+      // transcript would hollow a kept older read when the newer re-read is only
+      // in the summary (CL-4374 review follow-up).
+      const pathToReads = buildPathToReads([...anchorTurns, ...recentTurns], callIndex);
+      const supersededReads = supersededReadCallIds(pathToReads);
+
       const summary = cfg.summarize !== undefined
         ? await cfg.summarize(summarizedTurns)
         : buildTurnSummary(summarizedTurns, cfg.summaryMaxChars, anchorTurns.length);
@@ -584,11 +623,11 @@ export function createPruningCompactor(
       };
 
       // Anchors and recent turns stay contentful except for path-dedup: when the
-      // same file was read successfully more than once, older results become a
-      // one-line stub and the newest stays whole. Error results are never
-      // stubbed. SummarizedTurns lose content wholesale via the summary above.
-      // Anchors are already image-aged (outside the recent window). Recent turns
-      // keep live base64 so a just-pasted screenshot still reaches the model.
+      // same file was read successfully more than once among kept turns, older
+      // results become a one-line stub and the newest stays whole. Error results
+      // are never stubbed. SummarizedTurns lose content wholesale via the summary
+      // above. Anchors are already image-aged (outside the recent window). Recent
+      // turns keep live base64 so a just-pasted screenshot still reaches the model.
       const process = (t: ConversationTurn): ConversationTurn =>
         stubSupersededReads(t, supersededReads, callIndex);
       const output = coalesceAdjacentTextTurns([

src/tui/product-host.test.ts

@@ -333,9 +333,9 @@ describe("mountProductHost", () => {
   })
 })
 
-describe("provider-first model picker", () => {
-  // Mirrors the bug-report shape: several providers, one (codex) with three
-  // accounts, plus a favorite so the top level has a reachable-without-descending pick.
+describe("flat type-to-filter model picker", () => {
+  // Several providers, one (codex) with three accounts, plus a favorite so the
+  // top of the flat list has a reachable pick without typing.
   const providers = {
     "codex/abk-labs": { models: ["gpt-5.5", "gpt-5.6-sol"] },
     "codex/dirtroad": { models: ["gpt-5.5", "gpt-5.6-sol"] },
@@ -524,6 +524,67 @@ describe("provider-first model picker", () => {
       harness.destroy()
     }
   })
+
+  test("Enter on a no-matches filter does not apply a model", async () => {
+    const { harness, host, selected } = await mountPicker()
+    try {
+      host.openModels?.()
+      await harness.renderOnce()
+      for (const ch of "zzzz-no-such-model") {
+        harness.pressKey(ch)
+      }
+      await harness.renderOnce()
+      expect(host.shell.overlayItems).toEqual(["(no matches)"])
+      acceptOverlaySelection(host.shell)
+      expect(selected).toEqual([])
+    } finally {
+      host.dispose()
+      harness.destroy()
+    }
+  })
+
+  test("filtered accept uses the filtered row id, not the unfiltered catalog index", async () => {
+    // Catalog order puts favorites/recents first; after filtering to "grok",
+    // index 0 is the grok row — accepting must still apply the grok id, never
+    // the catalog's index-0 favorite.
+    const { harness, host, selected } = await mountPicker()
+    try {
+      host.openModels?.()
+      await harness.renderOnce()
+      for (const ch of "grok") {
+        harness.pressKey(ch)
+      }
+      await harness.renderOnce()
+      // Accept whatever is focused after filter (should be the sole match).
+      acceptOverlaySelection(host.shell)
+      expect(selected).toEqual(["xai/thegreataxios:grok-4.5"])
+    } finally {
+      host.dispose()
+      harness.destroy()
+    }
+  })
+
+  test("Alt+F on the no-matches sentinel does not toggle a favorite", async () => {
+    const favorites: string[] = []
+    const { harness, host } = await mountPicker({
+      onFavoriteToggle: (id) => favorites.push(id),
+    })
+    try {
+      host.openModels?.()
+      await harness.renderOnce()
+      for (const ch of "zzzz-no-such-model") {
+        harness.pressKey(ch)
+      }
+      await harness.renderOnce()
+      expect(host.shell.overlayItems).toEqual(["(no matches)"])
+      harness.pressKey("f", { meta: true })
+      await harness.renderOnce()
+      expect(favorites).toEqual([])
+    } finally {
+      host.dispose()
+      harness.destroy()
+    }
+  })
 })
 
 describe("mount failure", () => {

src/tui/product-host.ts

@@ -86,13 +86,9 @@ export type ProductHostDeliver = (
 ) => void
 
 /**
- * `section` groups rows for the provider-first picker: "recent" and
- * "favorites" stay flat at the top (already single models, reachable without
- * descending); "provider" rows are grouped into one top-level entry per
- * provider (or per account, since each configured provider entry is already
- * account-scoped — `codex/abk-labs`, `codex/dirtroad`); "unconnected" stays
- * flat as a "connect →" row. Omitted (from a caller not using
- * buildModelsFirstCatalog) falls back to one flat list, unwrapped.
+ * `section` tags catalog rows for grouping/ordering in `buildModelsFirstCatalog`
+ * (recent and favorites first, then provider models, then unconnected connect
+ * rows). The live picker is flat + type-to-filter — it does not nest by section.
  */
 export type ProductHostModelOption = {
   readonly id: string
@@ -131,7 +127,7 @@ export type ProductHostConfig = {
    * `models`/`describeModel` via `setModels` and reopens the picker.
    */
   readonly onConnectProvider?: (providerName: string) => void
-  /** `f` on a focused model/provider row; absent rows (connect →) are skipped by the caller. */
+  /** Alt+F on a focused model row; connect rows are skipped by the caller. Bare `f` is claimed by type-to-filter. */
   readonly onFavoriteToggle?: (itemId: string) => void
   /** Command palette catalog (registry-backed). */
   readonly commands?: readonly PaletteCommand[]
@@ -519,8 +515,11 @@ export async function mountProductHost(
         // Flat list: type to narrow rather than drill into a provider pane.
         typeToFilter: true,
         onAccept: (sel) => {
-          const id = sel.id ?? items[sel.index]?.id
-          if (!id) return
+          // Prefer the stable id from the (possibly filtered) row. Do not fall
+          // back to `items[sel.index]` — that index is into the filtered list,
+          // not the unfiltered catalog, so it would pick the wrong model.
+          const id = sel.id
+          if (id === undefined || id.length === 0) return
           const providerName = id.startsWith("connect:") ? id.slice("connect:".length) : null
           if (providerName !== null) {
             onConnect?.(providerName)
@@ -535,16 +534,15 @@ export async function mountProductHost(
                 // Alt+F, never bare f — type-to-filter claims printable keys.
                 const name = typeof key.name === "string" ? key.name.toLowerCase() : ""
                 if (name !== "f" || key.ctrl || !(key.meta || key.option)) return false
-                if (itemId.startsWith("connect:")) return false
+                // Empty id is the "(no matches)" filter sentinel — not a model.
+                if (itemId.length === 0 || itemId.startsWith("connect:")) return false
                 onFavoriteToggle(itemId)
                 return true
               },
             }
           : {}),
       })
     }
-    ;(shell as AppShell & { __openModels?: () => void }).__openModels =
-      openModels
   }
   const setModels = (
     models: readonly ProductHostModelOption[],

src/tui/shell.ts

@@ -1,8 +1,8 @@
 /**
  * OpenTUI app shell — sticky transcript, prompt chrome, inset overlay.
  *
- * Wave 3 product skin on the Wave 2 platform. Functional wrappers around
- * @opentui/core class renderables. Not wired to production CLI; Ink remains production.
+ * Functional wrappers around @opentui/core class renderables. This is the
+ * production interactive CLI surface (Ink is no longer the live path).
  */
 
 import { homedir } from "node:os"