Shell-mode (`!cmd`) output is not displayed while the agent is Working

[probepark]

Summary

Shell-mode commands entered while the agent is Working (streaming) are not shown in the transcript. Typing !pwd mid-turn runs the command, but the execution block flashes and then disappears — no command header, no output, no exit status. The same !pwd typed while idle renders normally.

Repro

1. Start an interactive session and send any prompt so the agent is Working.
2. While the spinner is active, type !pwd and press Enter.
3. Observe: the shell block vanishes as soon as the command finishes; output is never shown for the rest of the turn.
4. Type !pwd while idle → renders correctly.

Differential: $ (python/eval) commands typed mid-turn stay visible. Only ! (bash) commands disappear.

Root cause (evidence)

CommandController.handleBashCommand renders into two different parents depending on streaming state — packages/coding-agent/src/modes/controllers/command-controller.ts:1144:

const isDeferred = this.ctx.session.isStreaming;
this.ctx.bashComponent = new BashExecutionComponent(command, this.ctx.ui, excludeFromContext);

if (isDeferred) {
    this.ctx.pendingMessagesContainer.addChild(this.ctx.bashComponent);  // :1149
    this.ctx.pendingBashComponents.push(this.ctx.bashComponent);         // :1150
} else {
    this.ctx.chatContainer.addChild(this.ctx.bashComponent);             // :1152
}

Three defects on the deferred branch:

1. Completion detaches the component without ever re-parenting it — command-controller.ts:1180:

const bashComponent = this.ctx.bashComponent;
if (isDeferred && bashComponent && this.ctx.pendingBashComponents.includes(bashComponent)) {
    this.ctx.pendingMessagesContainer.detachChild(bashComponent);   // :1182 — removed from screen
}
this.ctx.bashComponent = undefined;

The component is detached from pendingMessagesContainer and never added to chatContainer. It stays in pendingBashComponents (a plain array, not a rendered container), so it has no parent and is not drawn. For a fast command like pwd this happens within milliseconds of submit, which is why nothing is visible.

handlePythonCommand (command-controller.ts:1189) has no equivalent detach block — which matches the observed asymmetry between ! and $.

2. The only flush runs on the next non-streaming submit — packages/coding-agent/src/modes/controllers/input-controller.ts:949:

if (this.ctx.session.isStreaming) {
    ...
    this.ctx.updatePendingMessagesDisplay();
    return;                                        // :944 — flush is never reached while busy
}

// Normal message submission
// First, move any pending bash components to chat
this.ctx.flushPendingBashComponents();             // :949

flushPendingBashComponents() (packages/coding-agent/src/modes/utils/ui-helpers.ts:1241) is the only place that moves the parked components into the chat transcript, and it is only reachable from the non-streaming submit path. So the output stays invisible until the user sends another prompt after the turn ends.

3. Any pending-bar refresh during the turn disposes the live component. updatePendingMessagesDisplay() starts with a clear() — ui-helpers.ts:981:

updatePendingMessagesDisplay(): void {
    this.ctx.pendingMessagesContainer.clear();     // :982

and Container.clear() disposes children — packages/tui/src/tui.ts:504:

clear(): void {
    for (const child of this.children) child.dispose?.();

updatePendingMessagesDisplay() has ~25 callsites and fires on ordinary mid-turn events, e.g. packages/coding-agent/src/modes/controllers/event-controller.ts:406 and :441 (queued/dequeued user messages). So a still-streaming ! command can be torn down mid-flight, stopping its loader and dropping buffered output. flushPendingBashComponents explicitly documents that these components must be detached rather than disposed (ui-helpers.ts:1242-1244), but clear() violates that.

Additionally, renderInitialMessages() (ui-helpers.ts:903-904) clears both pendingMessagesContainer and pendingBashComponents, permanently dropping anything parked there if a transcript rebuild happens before a flush.

Expected

A !command submitted while the agent is Working should render its header, streaming output, and exit status continuously, and should land in the chat transcript when it completes — same as the idle path.

Suggested fix

• On completion of a deferred bash command, move the component into the chat transcript (detachChild + addChatChild) and remove it from pendingBashComponents, instead of just detaching it (command-controller.ts:1180-1183).
• Keep in-flight execution components out of the container that updatePendingMessagesDisplay() clears, or have that method rebuild only the queued-message chips and re-attach live execution components (ui-helpers.ts:981-982).
• Add regression coverage for ! and $ submitted mid-turn asserting the component is parented and visible during streaming and after completion.

Environment

• Repo: gajae-code @ b6198e748
• Affected: interactive TUI mode, bash (!, !!) shell input

[Yeachan-Heo]

Batch triage — ADMITTED / ACTIVE CONTRIBUTOR OWNER PATH

Canonical baseline: dev@732856b3ccb3fade6e9fbc17908a4fbca5a7682f. The defect remains observable in current source: streaming-time Bash results are queued at agent-session.ts:15551-15555, terminal agent_end publication at agent-session.ts:2549-2564 does not flush them durably, and transcript rebuild clears the pending Bash display at ui-helpers.ts:898-905. This is the reported flash/disappear class, not a duplicate of #3674/#3680 viewport/scrollback work.

Contributor ledger

• fix(coding-agent): retain deferred shell output #3652 — sj0618, head 2bbb01605708e4237c160375804b44dfca26dcc7, base 66475d50377073cd2e028da0a0377964babbaf45; closed stale/conflicting. Current checks: 15 success / 6 skipped.
• fix(coding-agent): retain deferred shell output #3660 — sj0618, head d564f7ef86392698dee4ecad182110a0e4d93878, base bd5aea72e7d336fca481bfd49ef0f88de3f34dd7; closed stale/conflicting. Current checks: 12 success / 3 failed / 6 skipped.
• fix(coding-agent): retain deferred shell output #3697 — sole live mutation lane, primary author sj0618 with probepark co-author/review contribution. Exact head 5be7c7b49b90249ef70acd467822730c77944eea addresses the prior late-rejection, uncertain-append, concurrent-publication, FIFO, rebuild-identity, and settlement findings.

Current owner path

#3697 is 1 commit ahead and 10 behind current dev; merge base is a8fa63a599e9fa3d0189e48bfce778a36353ccdd. GitHub's current synthetic merge is 7fe493771631fd52ea2c33f3e3767f3684135b1a with parents 732856b3... and 5be7c7b4..., but that merge ref has zero check runs. The head itself has 22 checks (16 success / 6 skipped) from the older base and still carries CHANGES_REQUESTED.

Resume @sj0618 on #3697: rebase onto exact current dev, preserve sj0618 primary credit and probepark co-authorship, rerun affected durability/rebuild/failure-recovery checks, then obtain fresh maintainer exact-head review. No duplicate PR is admitted.

—
Signed: Yeachan-Heo · GPT-heavy hostile batch triage · 2026-08-05

[Yeachan-Heo]

Hostile re-review — glm-deepseek lane · ADMITTED / EXACT-HEAD VERIFIED / MERGE_READY PENDING EXACT-HEAD REVIEW SIGN-OFF

Continuing the prior GPT-heavy triage at canonical baseline dev@732856b3.... I verified the prior triage's claims against current source, the full PR diff, and the review lineage, then ran the focused tests at the exact contributor head.

Correction to the prior triage (important)

The prior comment states head 5be7c7b49b90249ef70acd467822730c77944eea "addresses the prior late-rejection, uncertain-append, concurrent-publication, FIFO, rebuild-identity, and settlement findings" as if settled. That assertion was not independently verified by the prior lane, and there is no signed exact-head review at 5be7c7b4. The last signed review (2026-08-03T17:13:01Z, Yeachan-Heo) was at 9f350aa0 with verdict MERGE_BLOCKED (2 code findings + 4 missing regressions); 9f350aa0 is now orphaned because the branch was rebased. I performed the exact-head verification the prior lane assumed.

Exact-head verification — 5be7c7b4 (single commit on a8fa63a5, MERGEABLE/CLEAN)

I fetched the contributor head into a throwaway worktree and verified the patch against every finding in the 9f350aa0 MERGE_BLOCKED verdict:

1. Late-rejection race (HIGH, found at 7f2380c8). Resolved. command-controller.ts now promotes the completed deferred component on both success and rejection paths — the recorded flag is removed and the condition is isDeferred && !isStreaming && pendingIndex !== -1. No stranded-in-pending outcome.
2. Uncertain-append entry-id loss (HIGH). Resolved. #appendBashMessageDurably retains pending.uncertainEntryId and reconciles via exact entry id (sessionManager.getBranch().find(id === uncertainEntryId && role === bashExecution)) before retry, with a bounded 2-attempt loop.
3. Concurrent publication divergence (MEDIUM). Resolved. #awaitAgentEndPublication serializes concurrent waiters through one shared publication promise; #retryDeferredAgentEndPublication checks identity (failedPublication !== current) before retrying. All three prior direct #agentEndPublicationPromise await sites (lines 6053, 6242, 8592, 8867/8870) are migrated.
4. Missing regressions. All four added: committed-append/recovery-rejection (reconciles a committed deferred shell append after recovery rejects), concurrent waiters (shares terminal retry settlement between concurrent public waiters and emits once), partial dequeue (dequeues only durable deferred shell records after a partial failure), and lease settlement (covered in the concurrent-waiters case with leaseCloseCount === 1, claimProducer called 1×). Plus durable-before-publication and rebuild-identity regression cases.

Tests run at exact head (local, read-only worktree)

bun test packages/coding-agent/test/modes/controllers/bash-command.test.ts \
          packages/coding-agent/test/agent-session-user-shortcut-hooks.test.ts
→ 13 pass, 0 fail, 71 expect() calls

bun --cwd=packages/coding-agent run check
→ biome: 2508 files clean; tsc --noEmit: clean

GitHub CI at this head (run 30868433973, after the rebase): 16 success / 6 skipped / 0 failed. The synthetic merge 7fe4937 has zero check runs; it is stale from the old base and is not authoritative.

Contributor ledger — confirmed

• fix(coding-agent): retain deferred shell output #3652 (sj0618) closed stale/conflicting. fix(coding-agent): retain deferred shell output #3660 (sj0618) closed stale/conflicting. fix(coding-agent): retain deferred shell output #3697 is the sole live lane, primary author sj0618, probepark co-author/review contribution. Cross-repo branch (sj0618/gajae-code); sj0618 is sole push owner.
• Issue author: probepark. Not a duplicate of [Bug] [TUI] Beginning of streaming assistant responses can be lost from native scrollback and replaced by stale tool output #3674/Repair TUI ask viewport and scrollback accessibility #3680 (closed viewport/scrollback work).

Disposition

MERGE_READY pending an exact-head review sign-off at 5be7c7b4. The prior lane's CHANGES_REQUESTED state on this PR is pinned to orphaned heads and does not reflect the applied state. Resume @sj0618: no further code change is required; request fresh exact-head maintainer review at 5be7c7b4, then merge. No duplicate PR admitted.

—
Signed: Yeachan-Heo · glm-deepseek hostile re-review · 2026-08-05

[Yeachan-Heo]

Closing during the emergency maintenance freeze. This item is not in the retained critical or maintainer-owned set. Do not open a replacement issue unless a maintainer explicitly directs it.

—
[repo owner's gaebal-gajae (clawdbot) 🦞]