fix(tui): complete sub-agent detail panel state coverage + artifact handles (#4094)

[Hmbown]

Fixes #4094

Completes the four items the reopen comment listed as still-open after PR #4182 (which shipped the live-activity trail, tool-call name/status, final summary/handoff, and bounded histories). Single-file change to crates/tui/src/tui/sidebar.rs (the detail-panel renderer #4182 added): three test additions, two small width-safety fixes they exposed, and one bounded artifact-handle feature.

The four remaining items

1. Many-tool-call state coverage + tests. A worker that has fired hundreds of tool calls now provably renders correctly — non-empty, carrying the live tool-call trail (name + compact status) and the step count, width-bounded, plus an inspect-more handle. Distinct from #4182's existing many-agents bound test.

• Test: subagent_expanded_detail_renders_many_tool_calls_without_overflow — asserts the dossier line carries 247 step(s) and grep_files, no blank lines, and (rendered narrow) no line overflows the column.

2. Narrow-width readability + tests. Every rendered line — header, role-mix, label, dossier, handle — is now clamped to the column. This exposed and fixed two real overflows:

• the dossier line's format!(" {}", truncate(.., width-2)) overflowed by the 2-space indent at content_width < 3 → new indented_detail_line() clamps indent + body together;
• the two-tone N running / M header was never truncated and spilled past narrow columns → it now collapses to a single truncated span when it doesn't fit (status clipped, never spilled).
• Test: subagent_detail_readable_and_bounded_across_narrow_widths — asserts no line overflows across widths [1,2,3,5,8,12,16,20,24,32,48] and that the status verb stays visible at a usable-narrow width (24).

3. Explicit running/done/failed state matrix + tests.

• Test: subagent_status_matrix_renders_marker_verb_and_style — for running/done/failed/canceled/interrupted, asserts the label renders the right marker ([~]/[✓]/[!]/[-]), the right state color on the span, and that the dossier surfaces the status verb.

4. Artifact/receipt handle for full output (vs. inline dump). Instead of dumping the full sub-agent transcript inline (the freeze/emptiness risk this issue tracks), the panel renders a bounded preview plus a copyable handle to the fuller artifact: ▸ full output · handle_read agent:<id>/full_transcript. This reuses the existing, documented sub-agent transcript var-handle — the same agent:<id> / full_transcript handle that subagent_session_projection materializes and that default_subagent_artifacts documents as the inspect path — so it adds no new artifact infrastructure. The handle is gated (subagent_output_handle) so an empty transcript is never advertised, guarded by max_rows, and truncated so it never overflows.

• Tests: subagent_completed_worker_surfaces_output_handle_not_inline_dump (completed worker shows the transcript handle and still shows the bounded summary; handle line is non-clickable), subagent_output_handle_gated_on_inspectable_output (fresh zero-step worker → no handle; running-with-steps and terminal workers → handle).

Scope of item 4 (transparent boundary)

This is a bounded feature per the issue's feasibility gate — a copyable handle reference the user runs via the existing handle_read tool, reusing existing plumbing. It intentionally does not add a new artifact system. Optional future enhancements (not required by the AC, flagged for the maintainer): making the handle line click-to-open the transcript in-TUI, and surfacing richer receipt kinds (patch / test_result / report) beyond the transcript handle.

Notes

• The handle is the single deliberate place a raw agent id is surfaced in the detail panel — a functional, copyable handle on its own dedicated line — while the dossier line stays id-free per v0.8.58: Hide internal numeric/id slop from normal UI — stable labels for turns, agents, steps #3030.
• Guard v0.9.0 guard: preserve verified active modules and DEEPSEEK env names #4172 respected: no verified-active modules or DEEPSEEK* env names renamed/removed.
• Bounded/lock-free render path (pure over projected rows, no aggregation on the UI thread) keeps the expand/click path from blocking the render loop, per the freeze reopen comment.

Verification

• cargo fmt --all --check → clean
• cargo clippy --workspace --all-features --locked -D warnings (gate allow-list) → clean
• cargo test --workspace --locked → 5996 passed; the only 2 failures are the known pre-existing unrelated ones (git_repo_root_reports_attempted_paths_when_no_repo_found, skill_hotbar_action_activates_skill_through_dollar_alias), confirmed failing on pristine origin/main with this change stashed
• cargo build --release -p codewhale-tui → ok

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[claude]

Claude finished @Hmbown's task in 3m 32s —— View job

---

Review: PR #4197 — sub-agent detail panel state coverage + artifact handles

• Gather context (diff, CLAUDE.md/AGENTS.md guidance)
• Understand the request (code review of sidebar.rs changes)
• Review correctness (panics, off-by-one, exhaustiveness)
• Review provider/model/route safety invariant (v0.8.65 EPIC v0.8.65 EPIC: Separate provider facts, model facts, offerings, and route resolution #2608)
• Review reuse/simplification/dead code
• Review test coverage
• Review security concerns
• Post final review

Summary

Single-file, well-scoped diff to crates/tui/src/tui/sidebar.rs. The narrow-width overflow fixes (indented_detail_line, the header-collapse guard) and the status-matrix/many-tool-call tests are solid, targeted fixes for real bugs the reopen comment called out — good width-safety discipline (content_width.max(1) used consistently, no unwrap/expect added, no panics found).

Bug: advertised output handle doesn't exist for running workers (posted inline at sidebar.rs:2766)

subagent_output_handle (sidebar.rs:2766) returns Some("agent:{id}/full_transcript") whenever steps_taken > 0, including for running workers — this is exactly what the PR's own subagent_expanded_detail_renders_many_tool_calls_without_overflow test exercises (a running row, steps_taken: 247, asserting the handle line appears).

But agent:<id>/full_transcript is only ever populated by insert_subagent_full_transcript_handle (crates/tui/src/tools/subagent/mod.rs:4551), which is called exclusively from terminal branches (Cancelled/Interrupted/BudgetExhausted/end-of-run) — never while a worker is still running. The codebase's own subagent_session_projection (mod.rs:3174-3192) makes this explicit: it only reuses the full_transcript handle if snapshot.status != SubAgentStatus::Running, and inserts a differently-named "transcript" handle otherwise.

So for a running worker, calling handle_read on the panel's advertised handle fails with "handle_read: no payload found for handle agent:<id>/full_transcript" (crates/tui/src/tools/handle.rs:281-286). That's worse than the "empty transcript" case the PR says it guards against — the handle is nonexistent, not empty, for the running-worker case that's actually most common while a worker is expanded and active. The new unit tests can't catch this because they only assert against the SidebarAgentRow struct in isolation, never against the real handle store.

Suggested fix: gate the handle on sidebar_agent_status_is_terminal(row.status) only, or point running workers at the live agent:<id>/transcript name to match what subagent_session_projection actually inserts mid-run.

Minor notes (not blocking)

• agent_status_marker/sidebar_agent_status_is_terminal are pre-existing and unchanged by this diff — not in scope, but worth noting sidebar_agent_status_is_terminal checks for a literal "budget" status string while subagent_status_name actually emits "budget_exhausted" (pre-existing, unrelated to this PR).
• No provider/model-prefix inference issues found — this diff doesn't touch model/provider selection.
• No security concerns (no new input parsing, no secrets, no shell/exec paths touched).

[claude]

Bug: the advertised handle doesn't exist yet for running workers.

subagent_output_handle returns Some("agent:{id}/full_transcript") whenever steps_taken > 0, even while the worker is still running (see the subagent_expanded_detail_renders_many_tool_calls_without_overflow test, which exercises exactly this: a running row with steps_taken: 247).

But the backing var-handle agent:<id>/full_transcript is only ever inserted by insert_subagent_full_transcript_handle (crates/tui/src/tools/subagent/mod.rs:4551, called only from the Cancelled/Interrupted/BudgetExhausted/end-of-loop terminal branches — mod.rs:5038,5168,5233,5331,5609). The codebase's own subagent_session_projection (mod.rs:3174-3192) is explicit about this: it only looks up the full_transcript handle if snapshot.status != SubAgentStatus::Running; otherwise it inserts a differently-named "transcript" handle instead.

So for a running worker, handle_read on the advertised agent:<id>/full_transcript reference will fail with "handle_read: no payload found for handle agent:<id>/full_transcript" (crates/tui/src/tools/handle.rs:281-286). That directly contradicts the PR's stated guarantee ("gated ... so an empty transcript is never advertised") — it's not empty, it's nonexistent. The new unit tests only assert against the SidebarAgentRow struct in isolation, so they can't catch this mismatch against the real handle store.

Suggested fix: only advertise the handle once the worker is terminal (sidebar_agent_status_is_terminal(row.status)), or point running workers at the live agent:<id>/transcript name instead of full_transcript to match what subagent_session_projection actually inserts while running.

[Fix this →](https://claude.ai/code?q=Fix%20subagent_output_handle%20in%20crates%2Ftui%2Fsrc%2Ftui%2Fsidebar.rs%20(around%20line%202766%2C%20branch%20codex%2Fv0868-fix-4094)%3A%20it%20currently%20returns%20Some(%22agent%3A%7Bid%7D%2Ffull_transcript%22)%20whenever%20steps_taken%20%3E%200%2C%20including%20for%20running%20workers.%20But%20the%20agent%3A%3Cid%3E%2Ffull_transcript%20var-handle%20is%20only%20inserted%20by%20insert_subagent_full_transcript_handle%20in%20crates%2Ftui%2Fsrc%2Ftools%2Fsubagent%2Fmod.rs%20at%20terminal%20states%20(Cancelled%2FInterrupted%2FBudgetExhausted%2Fend-of-loop)%2C%20never%20while%20running%20(see%20subagent_session_projection%20around%20mod.rs%3A3174%2C%20which%20explicitly%20branches%20on%20status%20%21%3D%20Running%20before%20looking%20up%20full_transcript%2C%20falling%20back%20to%20a%20differently-named%20%22transcript%22%20handle%20otherwise).%20Fix%20subagent_output_handle%20to%20only%20advertise%20the%20full_transcript%20handle%20when%20sidebar_agent_status_is_terminal(row.status)%20is%20true%2C%20and%20update%2Fadd%20tests%20accordingly%20(the%20many-tool-call%20running-worker%20test%20should%20no%20longer%20expect%20a%20full_transcript%20handle%20line%2C%20or%20should%20expect%20no%20handle%20at%20all%20while%20running).