Coalesce same-tool transcript rows and stream live shell output

CHANGELOG.md

@@ -15,10 +15,18 @@ parallel copies under `docs/` or `scripts/notes/`. At cut time: rename
 
 ### Added
 
+### Added
+
 - Plan and counsel workers require substance in Findings (files/paths,
   acceptance criteria, non-goals, risks, ordered steps). Four headings
   with stub Findings salvage as `incomplete-report`, not an attachable
   plan. Implement and review envelope completeness is unchanged.
+- Consecutive same-tool transcript calls collapse into one row with a count
+  chip (`· ×N`). Settled lanes use a past-tense head (`Grepped ×3 · "corbits"`).
+  `spawn_agent` stays one row per dispatch; `manage_tasks` paints no row.
+- Pending `run_shell` rows stream up to three live output lines from a
+  bounded 8 KiB feed, then a last-three preview and a non-zero `exit N` at
+  settle.
 
 ### Fixed
 
@@ -31,7 +39,6 @@ parallel copies under `docs/` or `scripts/notes/`. At cut time: rename
   ask is dropped. Do not poll `list_agents`.
 
 
-
 ## [0.3.21] - 2026-09-11
 
 ### Added

docs/TUI.md

@@ -63,7 +63,7 @@ down the left edge (`userBubbleLines` in `src/tui/stream.ts`). Each bubble
 keeps one empty bar row above and below its text so the operator's voice
 stays easy to find while scrolling through denser assistant and tool rows —
 the pad is part of the bubble itself, not an extra turn-boundary gap, and
-assistant/tool rows are unchanged.
+assistant rows are unchanged; tool rows are lanes (see Tool lanes below).
 
 Parent live reasoning paints through the existing thinking row — never a
 third mid-turn stream lane. While `inference.thinking.delta` arrives,
@@ -75,6 +75,63 @@ same one row per turn (`reasoning-fold`); `inference.text.delta` grows the
 open assistant streaming row in place. Worker spawn_agent-row thinking is a
 separate path and is unchanged by this preview.
 
+## Tool lanes
+
+Transcript tool rows group by tool, not by sentence. A call for a tool the
+previous row already represents folds onto that row instead of opening a new
+one (`src/tui/tool-rows.ts`); the row narrates the newest call's subject and,
+while calls are in flight, carries a dim count chip (`· ×N`) after the
+subject. `spawn_agent` never folds — each dispatch is its own live anchor —
+and `manage_tasks` paints no transcript row at all.
+
+When the lane settles (its last outstanding answer landed), the head rewrites
+to a past-tense count over the latest subject — `Grepped ×3 · "corbits"` —
+with the count taken over calls, never over payload items (a lane of three
+greps says `×3` even if the payloads returned forty matches in total; nothing
+here can substantiate a payload total). Each call's own answer stays behind
+the expand arrow. For an edit lane the per-call `+n/-n path` addenda remain
+in the expanded body, so folding never buries which file each call touched.
+Single-call rows (count <= 1) render exactly as they did before lanes.
+
+The past tense is a map keyed by raw tool name (`src/tui/tool-formatter.ts`:
+`grep` → `Grepped`, `read_file` → `Read`, `write_file` → `Wrote`,
+`edit_file` → `Edited`, `run_shell` → `Ran`, `list_dir` → `Listed`,
+`search_files` → `Searched`, …); an unknown tool falls back to its display
+name.
+
+Because a lane's row identity moves to the newest call, a lane also carries
+the call ids it absorbed (`memberIds`, newest appended). A
+result resolves its lane when its call id is the row's own id **or** one of
+its members — this is what pairs a resumed transcript's parallel batch
+(call, call, result, result) correctly. An id matching nothing still answers
+nothing: it is appended as its own row, never folded onto the newest
+same-name lane.
+
+Alt+C on a lane copies the most recent call's full output (see the Alt+C
+bullet under Clipboard and mouse).
+
+### Shell output lanes
+
+A pending `run_shell` row shows the command head and elapsed clock as today,
+plus up to three dim tail lines of the command's live output and — when the
+feed window holds more than three lines — the same dim `⋯ +N lines` elision
+marker as the settle preview, `N` counting lines within the live window (the
+feed keeps only the most recent 8 KiB). The output
+travels through a polled bounded feed (`src/session/shell-output-feed.ts`),
+not a reactor event: the plugin appends chunks (capped at an 8 KiB tail per
+call, emitting at most once per 100 ms plus a final flush at settle), and
+the product host's sticky poll reads the feed snapshot every 200 ms and
+repaints the pending row frame-coalesced. Nothing from the feed is
+persisted. When the feed is not wired (tests, the demo shell), the row
+renders exactly as before — silent degradation.
+
+At settle the live tail is replaced by a preview of the full output: its
+last three lines plus a dim `⋯ +N lines` elision marker when more were
+produced (the marker carries the count; the "N lines" stat is not painted
+on shell rows). A non-zero exit adds an `exit N` stat; a zero exit adds
+none. The existing expand idiom (Alt+E / click / the row arrow) reveals the
+full output, hiding the preview; the idiom and the arrow are unchanged.
+
 The prompt box's border carries the metadata that would otherwise cost a
 titlebar row: the model label sits right-aligned in the top rule as
 `profile · model · effort` (empty segments omitted), and a
@@ -249,7 +306,10 @@ clocks.
 `runtime-bridge` paints each `spawn_agent` call as a transcript stream row for
 **spawn / final / fail anchors**. While the agents strip is sticky, sticky-poll
 `syncAgentProgress` rewrites are gated off so the transcript is not a dual live
-rail. Ordinary in-flight tool rows keep their own elapsed clock
+rail. A `spawn_agent` call never folds into a tool lane: each dispatch keeps
+its own transcript row for its whole lifetime, because the row is the live
+progress anchor, not just a call record. Ordinary in-flight tool rows keep
+their own elapsed clock
 (`syncToolElapsed`) without the current-tool suffix.
 
 ### Unprompted fleet reports
@@ -725,7 +785,9 @@ running its own selection. Two chords cover remaining copy needs:
   (`enterCopyMode`) that resolves through the system clipboard port
   (`src/tui/system-clipboard.ts` — a native helper binary per
   platform, `pbcopy`/`clip`/`wl-copy`/`xclip`/`xsel`, falling back to an OSC
-  52 escape sequence when no helper is available, e.g. over SSH).
+  52 escape sequence when no helper is available, e.g. over SSH). On a
+  coalesced tool lane the copy resolves to the most recent call's full
+  output; a single-call row copies its own output, exactly as before.
 
 Arrow keys never scroll anything — inside the prompt they are caret motion
 or, at the buffer's edges, prompt-history recall; inside an open overlay's
@@ -787,6 +849,10 @@ terminal. It cannot observe:
 
 - **Real paint.** Tests assert on the shell's in-memory row/rect state, not
   on what a terminal emulator actually draws to a screen buffer.
+- **The live shell tail's wall clock.** The feed itself is pure and the
+  cadence is injectable, so the live tail is headless-testable by driving a
+  fake feed and a fake clock through the same sync path the sticky poll
+  uses; what the harness cannot see is real-time emission timing.
 - **Modifier reporting.** Whether a real terminal can report Shift+Enter,
   Alt+letter, or similar modifier combinations depends on the terminal
   negotiating the kitty keyboard protocol (or an equivalent) with the actual

src/agent/posix-tool-plugins.ts

@@ -30,6 +30,7 @@ import {
 import type { PermissionGate } from "../permission/gate.js";
 import { createWorktreeRootsProvider } from "../permission/worktree-roots.js";
 import type { CompactionArchive } from "../session/compaction-archive.js";
+import type { ShellOutputFeedMap } from "../session/shell-output-feed.js";
 
 export interface CorePosixToolPluginsArgs {
   cwd: string;
@@ -47,6 +48,9 @@ export interface CorePosixToolPluginsArgs {
   // Live getter for the background-shell registry (run_shell background:true).
   // Omitted makes background runs fail closed in shell-guard.
   getBackgroundShellRegistry?: () => BackgroundShellRegistry | undefined;
+  // Live getter for the per-call bounded shell-output feeds the transcript
+  // polls for a running command's live tail. Omitted leaves the tail unwired.
+  getShellOutputFeeds?: () => ShellOutputFeedMap | undefined;
   /** Primary-only evidence archive; workers omit this getter. */
   getEvidenceArchive?: () => CompactionArchive | undefined;
 }
@@ -86,6 +90,7 @@ export function buildCorePosixToolPlugins(
     getContextDir,
     shellEnv,
     getBackgroundShellRegistry,
+    getShellOutputFeeds,
     getEvidenceArchive,
   } = args;
   // Pre-gate sandboxes honor yolo mode so outside-workspace path tools and shell
@@ -119,6 +124,7 @@ export function buildCorePosixToolPlugins(
       ...(getBackgroundShellRegistry !== undefined
         ? { getBackgroundShellRegistry }
         : {}),
+      ...(getShellOutputFeeds !== undefined ? { getShellOutputFeeds } : {}),
     }),
     ...(getEvidenceArchive !== undefined
       ? [evidenceArchiveSearchPlugin(getEvidenceArchive)]

src/agent/tools.ts

@@ -90,6 +90,7 @@ import {
   createBackgroundShellRegistry,
   type BackgroundShellExit,
 } from "../shell/background-shell.js";
+import { createShellOutputFeedMap } from "../session/shell-output-feed.js";
 import { createListDirTool } from "../util/list-dir.js";
 import {
   createExaMCPWebFetchTool,
@@ -289,6 +290,9 @@ export interface AgentToolset {
     callbacks: MCPConnectCallbacks,
     signal?: AbortSignal,
   ) => Promise<void>;
+  // Per-call bounded live-output tails of foreground shells, polled by the
+  // transcript for each pending run_shell row's live lines.
+  shellOutputFeed: ReturnType<typeof createShellOutputFeedMap>;
   // Connect one newly persisted server through the same lifecycle as startup MCP.
   connectMCPServer: (
     config: MCPServerConfig,
@@ -360,6 +364,10 @@ export async function createAgentToolset(
       : {}),
   });
   const shellCollect = createShellCollectTool(backgroundShells);
+  // Per-call bounded live-output tails of foreground shells, polled by the TUI
+  // for each pending run_shell row's live lines. Workers get a map too; nothing
+  // reads it unless a transcript polls it (silent degradation).
+  const shellOutputFeed = createShellOutputFeedMap();
   const sessionBlobReader =
     getBlobReader !== undefined
       ? createLazyBlobReader(getBlobReader)
@@ -447,6 +455,7 @@ export async function createAgentToolset(
       ...(getEvidenceArchive !== undefined ? { getEvidenceArchive } : {}),
       ...(shellEnv !== undefined ? { shellEnv } : {}),
       getBackgroundShellRegistry: () => backgroundShells,
+      getShellOutputFeeds: () => shellOutputFeed,
     }),
   });
 
@@ -1179,6 +1188,7 @@ export async function createAgentToolset(
   return {
     dynamicRunner,
     connectMCP,
+    shellOutputFeed,
     connectMCPServer: publicConnectMCPServer,
     disconnectMCPServer: publicDisconnectMCPServer,
     hasMCPServer: (name) =>

src/plugins/shell-guard-plugin.test.ts

@@ -10,9 +10,12 @@ import { spawnSync, type ChildProcess } from "node:child_process";
 import { randomUUID } from "node:crypto";
 
 import { createBackgroundShellRegistry } from "../shell/background-shell.js";
+import { createShellOutputFeed } from "../session/shell-output-feed.js";
+
 import {
   BoundedShellOutput,
   MAX_SHELL_OUTPUT_BYTES,
+  SHELL_FEED_EMIT_MS,
   advertiseShellGuardTimeout,
   resolveShellTimeoutMs,
   reapLiveChildren,
@@ -37,6 +40,35 @@ describe("runGuardedShell", () => {
     expect(output).toContain("hello");
   });
 
+  test("a rate-limited second write reaches the feed before the process exits", async () => {
+    const feed = createShellOutputFeed();
+    let finished = false;
+    const running = runGuardedShell(
+      { command: "echo first; sleep 0.02; echo second; sleep 0.4" },
+      neverAbort(),
+      undefined,
+      undefined,
+      (text) => {
+        feed.append(text);
+      },
+    ).then((result) => {
+      finished = true;
+      return result;
+    });
+    const deadline = Date.now() + SHELL_FEED_EMIT_MS + 80;
+    while (
+      !feed.snapshot().includes("second") &&
+      Date.now() < deadline &&
+      !finished
+    ) {
+      await Bun.sleep(10);
+    }
+    expect(finished).toBe(false);
+    expect(feed.snapshot()).toContain("second");
+    const result = await running;
+    expect(result.exitCode).toBe(0);
+  });
+
   test("omitted timeout does not arm a timer", async () => {
     const start = Date.now();
     const { exitCode, timedOut, output } = await runGuardedShell(
@@ -309,6 +341,66 @@ describe("background run_shell (shellGuardPlugin)", () => {
     expect(registry.runningCount()).toBe(0);
     registry.disposeAll("test done");
   });
+
+  test("an unwired shell-output feed spawns fine and paints no tail", async () => {
+    const handler = defined(
+      shellGuardPlugin(process.cwd(), undefined, undefined, {}).middleware,
+    )(fallback);
+    const result = await handler(
+      { id: "fg2", name: "run_shell", arguments: { command: "echo hi" } },
+      neverAbort(),
+    );
+    expect(result.isError).toBeUndefined();
+    expect(String(result.content)).toContain("hi");
+  });
+
+  test("a wired feed receives the output tail at cadence with a final flush", async () => {
+    const feed = createShellOutputFeed();
+    let emits = 0;
+    const handler = defined(
+      shellGuardPlugin(process.cwd(), undefined, undefined, {
+        getShellOutputFeeds: () => {
+          const wrapped = {
+            append: (text: string) => {
+              emits += 1;
+              feed.append(text);
+            },
+            snapshot: () => feed.snapshot(),
+            clear: () => feed.clear(),
+          };
+          return {
+            forCall: () => wrapped,
+            get: () => wrapped,
+            drop: () => undefined,
+          };
+        },
+      }).middleware,
+    )(fallback);
+    const result = await handler(
+      {
+        id: "fg3",
+        name: "run_shell",
+        arguments: {
+          // Fifteen lines ~10 ms apart: far more chunk arrivals than one
+          // cadence window per 100 ms can allow. Without the Date.now() gate
+          // in emitPendingOutput every arrival emits (~16 emissions) and this
+          // ceiling fails — the assertion is what pins the cadence.
+          command:
+            "i=1; while [ $i -le 15 ]; do echo line$i; sleep 0.01; i=$((i+1)); done",
+        },
+      },
+      neverAbort(),
+    );
+    expect(result.isError).toBeUndefined();
+    // The final flush lands the tail (including the last line) in the feed.
+    expect(feed.snapshot()).toContain("line1");
+    expect(feed.snapshot()).toContain("line15");
+    // At most one emit per 100 ms of wall time (~300 ms with the pwd probe
+    // trailer), plus the final flush. Still far below the ~16 arrivals, so a
+    // broken cadence gate cannot pass.
+    const elapsedMs = 350;
+    expect(emits).toBeLessThanOrEqual(Math.ceil(elapsedMs / 100) + 1);
+  });
 });
 
 describe("advertiseShellGuardTimeout", () => {