feat(agents): show a parked child as parked, not "waiting for input" - #5942
Conversation
#5921 recorded the park/ask distinction (`parked_at_turn_end` on the checkpoint and the worker record) but deliberately left every UI projection alone. The result is the complaint that remains on #5906: a child parked because its parent's turn ended sits in the Agents panel, the sidebar and Agent Details wearing the same "waiting for input" label as a child that actually asked a question — with no way to tell them apart, and a "Pending question" line quoting a resume recipe nobody asked for. The runtime is why the surfaces got it wrong: parking hands the child a `needs_input` note phrased as a question ("Resume this parked child with ..."), and every projection keys off `needs_input`. So the flag has to outrank it. - `AgentCurrentActivityStatus::Parked` joins the existing status vocabulary (no second one). It is derived in exactly one place — `subagent_routing::subagent_is_parked`, reading the checkpoint flag — ahead of `needs_input`, and made sticky across the `Interrupted` terminal projection the same way `Waiting` already was. - The work-surface rows, the sidebar and Agent Details render it as `parked`, with a one-line recovery in place of the fake question: `resume_from` to continue, `cancel` to dismiss — the verbs the `agent` tool actually exposes, quoted. - Parked lands in `WorkBucket::Ready`, not `Attention`: it sorts below live and answerable work and stops inflating the `blocked` chip. - `RosterState::Parked` gives the receipts roster its own state and glyph, and parked husks sink below live and waiting rows there too. The wire `state` string gains `parked`. - `AgentStatusParked` / `AgentStatusParkedRecovery` ship translated in all 15 complete packs; a new parity test also pins `resume_from` and `cancel` as untranslatable tokens, so no pack can tell an operator to type a verb that does not exist. Agent Details additionally suppresses the parked child's `needs_input` note as a pending question — it asks the operator nothing. Verified (rebased onto origin/main, which now carries #5921) cargo fmt --all -- --check clean cargo check -p codewhale-tui clean RUST_MIN_STACK=16777216 cargo test -p codewhale-tui --lib \ -- tui::work_surface tui::agent_details tui::sidebar \ tui::agent_roster localization test result: ok. 227 passed; 0 failed; 0 ignored (216 before) RUST_MIN_STACK=16777216 cargo test -p codewhale-tui --lib test result: ok. 11733 passed; 0 failed; 13 ignored Eleven new tests. The five UI guards were each shown failing on their exact assertion with `subagent_is_parked` reverted to `false`. Local-test evidence only: no provider call, no deploy, and no manual reproduction of the original incident in a running TUI. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0188XYyJaw9Mh9uSrqQBoqhm Signed-off-by: CodeWhale Bot <bot@codewhale.net>
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_bf594a95-9393-40bb-976b-a494ca8671e2) |
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
|
Claude finished @Hmbown's task in 5m 13s —— View job Review complete
|
There was a problem hiding this comment.
🟡 Changes recommended
Agent Details may incorrectly render a redundant/confusing Current: line for Parked status unless Parked is excluded like Waiting.
Once you've addressed the issues Copilot identified, you can request another Copilot review.
Pull request overview
This PR updates the TUI’s agent-status derivation and rendering so that turn-end–parked child agents are displayed as parked (distinct from truly waiting for input) across the work surface, sidebar, agent details, and receipts roster, using a single derived status and fully translated copy.
Changes:
- Introduces
AgentCurrentActivityStatus::Parkedand derives it centrally fromcheckpoint.parked_at_turn_end, outrankingneeds_input. - Updates work-surface rows, sidebar rows, agent details, and the agent roster to render parked (including recovery guidance using
resume_from/cancel) and to adjust sorting/bucketing so parked husks don’t inflate “blocked”. - Extends localization and protocol/parity wiring to support the new
parkedvocabulary and wire state.
File summaries
| File | Description |
|---|---|
| crates/tui/src/tui/work_surface/model.rs | Uses localized label for Parked and buckets it into Ready to avoid “blocked” inflation. |
| crates/tui/src/tui/work_surface/mod.rs | Adds work-surface tests covering parked vs real waiting behavior, sorting, and narrow-row rendering. |
| crates/tui/src/tui/subagent_routing.rs | Centralizes parked detection + recovery detail and ensures parked outranks needs_input in derived activity. |
| crates/tui/src/tui/sidebar.rs | Localizes parked status word in sidebar rows and adds coverage tests. |
| crates/tui/src/tui/app.rs | Adds AgentCurrentActivityStatus::Parked and documents derivation constraints. |
| crates/tui/src/tui/agent_roster.rs | Adds RosterState::Parked, glyph, and sorting rule to sink parked rows. |
| crates/tui/src/tui/agent_roster/tests.rs | Adds roster tests validating parked state mapping, sorting, and redispatch clearing. |
| crates/tui/src/tui/agent_details.rs | Localizes parked state and suppresses parked “Pending question” projection. |
| crates/tui/src/localization.rs | Adds message IDs and parity tests ensuring parked copy is translated and preserves tool tokens. |
| crates/tui/src/core/protocol_parity.rs | Adds parked to roster-state string mapping for protocol parity. |
| crates/tui/locales/en.json | Adds English strings for parked status and recovery line. |
| crates/tui/locales/de.json | Adds German parked strings. |
| crates/tui/locales/ca.json | Adds Catalan parked strings. |
| crates/tui/locales/es-419.json | Adds LATAM Spanish parked strings. |
| crates/tui/locales/fr.json | Adds French parked strings. |
| crates/tui/locales/pt-BR.json | Adds Brazilian Portuguese parked strings. |
| crates/tui/locales/ru.json | Adds Russian parked strings. |
| crates/tui/locales/uk.json | Adds Ukrainian parked strings. |
| crates/tui/locales/vi.json | Adds Vietnamese parked strings. |
| crates/tui/locales/hi.json | Adds Hindi parked strings. |
| crates/tui/locales/id.json | Adds Indonesian parked strings. |
| crates/tui/locales/ja.json | Adds Japanese parked strings. |
| crates/tui/locales/ko.json | Adds Korean parked strings. |
| crates/tui/locales/zh-Hans.json | Adds Simplified Chinese parked strings. |
| crates/tui/locales/zh-Hant.json | Adds Traditional Chinese parked strings. |
| crates/protocol/src/event_msg.rs | Updates roster state documentation to include parked. |
Review details
- Files reviewed: 26/26 changed files
- Comments generated: 1
- Review effort level: Lite
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
| | AgentCurrentActivityStatus::Interrupted | ||
| | AgentCurrentActivityStatus::Waiting | ||
| ) | ||
| { | ||
| let mut current = vec![activity_status_label(activity.status).to_string()]; | ||
| let mut current = vec![activity_status_label(activity.status, app.ui_locale).into_owned()]; |
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
| pub(crate) fn subagent_is_parked(agent: &SubAgentResult) -> bool { | ||
| agent | ||
| .checkpoint | ||
| .as_ref() | ||
| .is_some_and(|checkpoint| checkpoint.parked_at_turn_end) |
There was a problem hiding this comment.
🟡 Parked agents still say waiting
The Agents modal still labels a parked child “Waiting for you” through WhaleState::for_subagent. This contradicts its parked status and falsely implies user input can resume it.
Prompt for agents
Extend the shared parked-state derivation to the /subagents Agents modal. WhaleState::for_subagent in crates/tui/src/tui/whales.rs currently checks needs_input first, so a parked checkpoint still becomes WhaleState::Waiting and renders the localized “Waiting for you” badge. Make that modal consume the typed parked distinction without reintroducing string inference or a parallel lifecycle system. Add focused coverage pairing a parked child with a genuinely asking child.
Was this helpful? React with 👍 or 👎 to provide feedback.
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: ab9d13b0b3
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| pub(crate) fn subagent_is_parked(agent: &SubAgentResult) -> bool { | ||
| agent | ||
| .checkpoint | ||
| .as_ref() | ||
| .is_some_and(|checkpoint| checkpoint.parked_at_turn_end) |
There was a problem hiding this comment.
After agent(action="cancel") is used on a parked child, terminalize_settled_worker_on_cancel changes the worker record to Cancelled but leaves both the checkpoint's parked_at_turn_end flag and the record flag set. Because this helper treats that historical flag as unconditional, the next AgentList reconciliation changes the child back to Parked; RosterState::from_record likewise reports it as parked and nonterminal. The advertised “cancel to dismiss” action therefore leaves every new projection showing the recovery prompt and can keep parent settled counts wrong. Clear the parked flags on cancellation or require a nonterminal worker state when deriving Parked.
AGENTS.md reference: crates/tui/AGENTS.md:L16-L17
Useful? React with 👍 / 👎.
| /// actually answer. It is its own state here because the recovery is | ||
| /// different: nobody will answer it, and it is continued through | ||
| /// `resume_from` (a *new* agent) or dismissed with `cancel`. | ||
| Parked, |
There was a problem hiding this comment.
A parked child has stopped executing, but adding this state without including it in the elapsed-time latch means agent_is_terminal returns false as soon as the Parked activity takes precedence. Consequently, freeze_terminal_agent_elapsed never records its duration even though the manager recomputes SubAgentResult::duration_ms from started_at.elapsed() on every snapshot, so the parked work row's elapsed time continues ticking until archival; the durable roster has the same problem because its WaitingForUser record has no completed_at_ms. Treat Parked as stopped for timing purposes without dropping its unfinished recovery detail.
AGENTS.md reference: crates/tui/AGENTS.md:L18-L19
Useful? React with 👍 / 👎.
| // answerable (#5906). They are the one class of row the operator is not | ||
| // meant to scan past to find real work, and the sort is stable so the | ||
| // history order survives inside each group. | ||
| rows.sort_by_key(|row| row.state == RosterState::Parked); |
There was a problem hiding this comment.
This boolean sort moves parked rows behind every nonparked row, including later completed, failed, and canceled receipts. For concurrent children where an older long-running child parks after a newer child has already completed, the historical roster therefore changes from creation order to showing the newer completion first, even though only live and answerable rows need to rank above parked work. Use a rank that promotes live/waiting rows while retaining creation order among parked and terminal receipts.
Useful? React with 👍 / 👎.
| let parked = subagent_is_parked(agent); | ||
| // Parked outranks `needs_input`: the park note *is* phrased as a | ||
| // question, and reading it as one is exactly the bug (#5906). | ||
| let mut structured_status = if parked { | ||
| AgentCurrentActivityStatus::Parked |
There was a problem hiding this comment.
A turn-end parked child still has SubAgentStatus::Interrupted, and reconcile_terminal_subagent_card_retention runs before this new checkpoint-based derivation. On the first later reconciliation more than 45 seconds after parking, that retention pass therefore removes the child from subagent_cache as a terminal card, causing the work-surface Agents view, sidebar projection, and Agent Details entry to lose the newly advertised resume_from/cancel recovery even though Parked is classified as unfinished and actionable. Exempt parked checkpoints from terminal-card expiry until they are resumed or dismissed.
Useful? React with 👍 / 👎.
| "WhaleStateBlocked": "Blocked", | ||
| "WhaleStateOffline": "Offline", | ||
| "AgentStatusParked": "parked", | ||
| "AgentStatusParkedRecovery": "nobody is waiting on it; resume_from to continue, or cancel to dismiss", |
There was a problem hiding this comment.
The recovery strings now embed the resume_from argument and cancel action directly in every locale pack, while the scoped TUI contract requires commands and key names to be composed in code. The added parity test merely pins fifteen duplicated literals, so a tool-schema rename or a partial locale correction can leave user instructions referring to an obsolete action. Use named placeholders in the translated prose and interpolate the canonical tool tokens in parked_recovery_detail.
AGENTS.md reference: crates/tui/AGENTS.md:L25-L26
Useful? React with 👍 / 👎.
![codewhale-agent[bot]](https://avatars.githubusercontent.com/u/101357273?s=60&v=4){kind=small}
There was a problem hiding this comment.
Codewhale review
PR introduces a centralized Parked status derived from checkpoint.parked_at_turn_end and propagates it through the work surface, sidebar, agent roster, and agent details, with localized copy. The approach is mostly clean and test coverage is strong, but the Agent Details current-activity guard does not appear to include the new variant, and the roster sort key may over-sink parked rows.
Findings
- [ERROR] Agent Details current-activity guard does not include Parked (
crates/tui/src/tui/agent_details.rs:254)
project_agent_details still guards the block that renders theState: ...line with amatches!over Queued..Waiting. A parked child hasAgentCurrentActivityStatus::Parked, so it will not enter this branch andState: parkedwill not render. The new test in this file should fail if this path is exercised, so the guard must be updated. - [WARNING] Roster sort puts parked rows below terminal states too (
crates/tui/src/tui/agent_roster.rs:153)
build_agent_roster sorts withrow.state == RosterState::Parked, a bool key, so parked rows sink below every non-parked row, including Done/Failed/Cancelled. The comment and PR description describe sinking below live and answerable rows only, and there is no test covering terminal rows. - [INFO] Localization parity test only uses substring checks for tool tokens (
crates/tui/src/localization.rs:4855)
Assertingrecovery.contains("cancel")does not provecancelis an intact tool token; for example,cancelarorcancelledwould pass but are not the verb the agent tool exposes. Current packs ship the literal token, but the parity gate should be stricter.
Suggestions
-
crates/tui/src/tui/agent_details.rs:254— Add Parked to the guard so parked children render the status line.| AgentCurrentActivityStatus::Parked | AgentCurrentActivityStatus::Waiting
Assessment
Good centralized derivation and broad test coverage, but the Agent Details guard likely needs to add Parked before the feature works on that surface. The roster ordering and the token parity test could also be tightened.
Advisory review by Codewhale (codewhale review --pr 5942 --post, head ab9d13b0b32d891611dd563be7c44837ff2add3d). Line-specific findings are also posted as inline review comments; mechanical fixes arrive as committable suggestions you can apply from the Files tab. CODEOWNERS approval still governs merge.
| @@ -253,8 +253,8 @@ | |||
| | AgentCurrentActivityStatus::Interrupted | |||
| | AgentCurrentActivityStatus::Waiting | |||
There was a problem hiding this comment.
[ERROR] Agent Details current-activity guard does not include Parked
project_agent_details still guards the block that renders the State: ... line with a matches! over Queued..Waiting. A parked child has AgentCurrentActivityStatus::Parked, so it will not enter this branch and State: parked will not render. The new test in this file should fail if this path is exercised, so the guard must be updated.
| // ...except parked husks, which sink below everything still live or | ||
| // answerable (#5906). They are the one class of row the operator is not | ||
| // meant to scan past to find real work, and the sort is stable so the | ||
| // history order survives inside each group. |
There was a problem hiding this comment.
[WARNING] Roster sort puts parked rows below terminal states too
build_agent_roster sorts with row.state == RosterState::Parked, a bool key, so parked rows sink below every non-parked row, including Done/Failed/Cancelled. The comment and PR description describe sinking below live and answerable rows only, and there is no test covering terminal rows.
| "{} lost the resume verb the agent tool exposes: {recovery}", | ||
| locale.tag() | ||
| ); | ||
| assert!( |
There was a problem hiding this comment.
[INFO] Localization parity test only uses substring checks for tool tokens
Asserting recovery.contains("cancel") does not prove cancel is an intact tool token; for example, cancelar or cancelled would pass but are not the verb the agent tool exposes. Current packs ship the literal token, but the parity gate should be stricter.
| @@ -253,8 +253,8 @@ | |||
| | AgentCurrentActivityStatus::Interrupted | |||
| | AgentCurrentActivityStatus::Waiting | |||
There was a problem hiding this comment.
Add Parked to the guard so parked children render the status line.
| | AgentCurrentActivityStatus::Waiting | |
| | AgentCurrentActivityStatus::Parked | |
| | AgentCurrentActivityStatus::Waiting |
No-Issue: display half of #5906 (the claim half closed with #5921); the founder's "2 sub agents waiting for input" report.
Refs #5906
Builds on the recording work that landed in #5921 (now on
main), whichadded
parked_at_turn_endtoSubAgentCheckpointandAgentWorkerRecordbut deliberately left every UI projection of
WaitingForUseralone. Thisis the display half.
The complaint that remained
A child parked because its parent's turn ended shows up in the Agents
panel, the sidebar and Agent Details as "waiting for input" — the same
label as a child that actually asked the operator a question. There is no
way to tell them apart, and Agent Details prints a
Pending question:line quoting a resume recipe nobody asked for.
The runtime is why the surfaces got it wrong. Parking hands the child a
needs_inputnote phrased as a question:Every projection keys off
needs_input, so the parked husk reads asanswerable. The flag has to outrank it.
What changed
One derivation, one status vocabulary.
AgentCurrentActivityStatusgains
Parked; it is derived in exactly one place —subagent_routing::subagent_is_parked, reading the checkpoint flag —ahead of
needs_input, and made sticky across theInterruptedterminal projection the same way
Waitingalready was. No surface sniffsa reason string, and no second status enum was introduced.
Three surfaces say
parked. The work-surface / Agents rows, thesidebar and Agent Details render the state, with a one-line recovery
standing in for the fake question:
resume_fromto continue,canceltodismiss — the verbs the
agenttool actually exposes, quoted exactly.Agent Details also suppresses the parked child's
needs_inputnote as apending question; it asks the operator nothing.
Parked sorts below live work and stops inflating the chip. Parked
lands in
WorkBucket::Ready, notAttention, so it ranks under bothActive and Attention and the work heading's
blockedcount covers onlychildren a person is genuinely holding up.
RosterState::Parkedgives thereceipts roster its own state and glyph (
◌, hollow-but-unfinished), andparked husks sink below live and waiting rows there too despite the rail's
otherwise-chronological order. The wire
statestring gainsparked.Translated, not fallen back.
AgentStatusParkedandAgentStatusParkedRecoveryship in all 15 complete packs. A new paritytest additionally pins
resume_fromandcancelas untranslatabletokens in every pack, so no translation can tell an operator to type a
verb that does not exist.
Verified
Eleven new tests, each built from a child constructed exactly as the
turn-end parking projection builds one (
Interrupted+WaitingForUserneeds_inputnote + a checkpoint with the flag),paired against a child that genuinely asked:
sidebar: parked row saysparked, the asking row still sayswaiting; the parked word follows the UI locale (asserted againsttr(De, AgentStatusParked), not a literal).work_surface: the parked row's detail leads withparked, namesresume_fromandcancel, and never replays the parking note; theasking row keeps
waiting for inputand its real question; theblockedchip counts one, not two; the parked row sorts below both thelive and the asking row; and the status word survives the narrow-row
ladder at widths 96 / 72 / 56, where the existing degradation test
already pins
running.agent_details: parked body saysState: parked, carries both recoveryverbs, and has no
Pending questionline; a child that really askedstill reports
State: waitingwith its question.agent_roster: a parked record isRosterState::Parkedwith a distinctglyph and is not terminal; parked husks sort below live and waiting rows
despite being oldest; clearing the flag on re-dispatch returns the row
to
Running.localization: every complete pack translates both keys away fromEnglish and keeps
resume_from/cancelintact.The five UI guards were each shown failing on their exact assertion with
subagent_is_parkedreverted tofalse.Not verified / not done
reproduction of the original incident in a running TUI.
whales.rsis untouched.WhaleState::for_subagentstill maps aparked child to
Waitingvianeeds_input. That is the ambient whaleglyph vocabulary, not the Agents/sidebar/details rows this issue is
about; folding it in would have widened the change past the surfaces
named in the complaint.
Readybucket is a reuse, not a new one. A parked husk is not"queued", but
Readyis the existing rank that sits below Active andAttention while staying actionable, and adding a fifth bucket would have
rippled through the heading copy and every bucket match in the module
for no behavioral gain.
resume_fromis not offered as a keybinding. The recovery linenames the two tool actions that exist; wiring a row action to invoke
them is separate work.
🤖 Generated with Claude Code
https://claude.ai/code/session_0188XYyJaw9Mh9uSrqQBoqhm
Note
Low Risk
TUI presentation and sorting only; no changes to auth, persistence, or agent execution semantics beyond how parked checkpoints are displayed.
Overview
Fixes #5906 by treating turn-end–parked child agents as their own state instead of “waiting for input”, which previously matched real operator questions because parking still sets a question-shaped
needs_inputnote.Derivation and activity model.
subagent_is_parkedreadscheckpoint.parked_at_turn_endinsubagent_routingand maps toAgentCurrentActivityStatus::Parkedbeforeneeds_input; that status stays sticky acrossInterruptedterminal projection likeWaiting. Parked rows show localized recovery copy (resume_from/cancel) instead of the fake question, and Agent Details no longer surfaces a Pending question for parked husks.Surfaces and sorting. Work surface, sidebar, agent details, and the agent roster all label parked (with
RosterState::Parked, glyph◌, wirestate: "parked"). Parked work lands inWorkBucket::Readyso it does not inflate the blocked chip or sort above live/answerable agents; the roster sinks parked rows below live and waiting rows.i18n.
AgentStatusParkedandAgentStatusParkedRecoveryship in all locale packs, with a parity test that keepsresume_fromandcancelas literal tool tokens.Reviewed by Cursor Bugbot for commit ab9d13b. Bugbot is set up for automated code reviews on this repo. Configure here.