Fleet/agents: first-class sub-agent and workflow management in the TUI (agents bottom view: live list, per-agent status/elapsed/tokens, focus, message, stop, history over time)

## Combined Core execution: C01

[Core execution plan]() owns order and scope. This issue owns C01's current status, path claims, blockers and evidence. Other linked implementation issues retain their own delivery history.

**Issue mapping:** [SHA-6236 / GitHub #5479]()
owns the retained agent/workflow register, status/elapsed/usage, focus, message,
stop and history requirements. Credit the current repair; qualify the remaining
controls and runtime-backed history through C09 rather than create a second
roster authority.

**Owner: Root integration. Dependencies: C00 and resumed GUI authorization.**
The `b6fad79373` build/install is complete; retain the linked binary identities. Preserve the
user's running sessions and the required separately matched Desktop runtime.
Verify that the Agents register retains completed/parked workers beyond the
compact cache's 45-second expiry, including while another worker runs, without
duplicate IDs or cross-conversation rows. Verify fresh state wins, waiting for
permission/input remains distinct from parked, and checkpoint continuation
preserves intended profile/model, ownership and idempotence.

The same-session 14:21 and 14:22:57 screenshots showed three workers, a
successor, then “no agents have run this session.” Persisted records retained
four workers. Source inspection found the register consuming the expiring
compact cache despite a retained roster. The fix is now committed, locally tested and installed; the screenshot's exact transition and duplicate live
execution were not proven. Retained fallback indentation still needs a
truthful disposition in the final presentation check.

**Completion evidence required:** build/install hashes plus bounded final-package acceptance
after GUI permission resumes; correct history and checkpoint/successor behavior
with no fabricated lifecycle state. See roster diagnosis (retained Ops evidence: `../agent-roster-20260909/REPORT.md`)
for source evidence, whose earlier “not applied” status is superseded above.

---

## Original issue and contributor history

## Ask (owner, 2026-08-17)

Bring Codewhale's sub-agent / workflow management up to the level a modern agent harness has natively: a persistent, glanceable **agents rail** in the TUI that shows *every* agent that ran during the session over time — not just the currently running one — with what each is doing, how long it has been going, tokens used, and direct controls.

Reference behavior (what the owner is comparing against): a footer/status strip line per agent — `○ 3m 29s · ↓ 96.3k tokens` — with the parent listed first (`● main`), workflows shown as one row with `5/6 agents done`, and keys to jump into an agent (`←`) or manage it (`↓`).

## What exists

* `agent` tool (12-field schema, [#5458]()), Fleet roster / worker records, `/jobs`, the Work bar rail (`work_bar_subagents_pty` tests), `⌥V` output view, `task_status`, sub-agent transcripts stored as handles, `/goal` `/workflow` `/auto` orchestration (#5439 says they are shipped but buried).

## Spec (proposed; break into slices)

1. **Agents rail** (always available, collapsible): one row per agent spawned this session, newest last, retained after completion (until session end or explicit clear). Row = status glyph (running / waiting / done / failed / cancelled), name or type, the agent's current step or last tool in one line, elapsed, tokens (↓ input/↑ output), cost if priced. Workflows/fleets aggregate to one row with `n/m done` and expand.
2. **Keys**: jump to an agent's transcript (`←`), manage (`↓` → message / interrupt / cancel / open details / copy id), back to main. Discoverable in `/help` and the fn+F1 keys overlay.
3. **Truth**: every number comes from runtime receipts (worker records / usage events), never estimated; a finished agent keeps its final numbers.
4. **Time**: rows show wall time and, when finished, duration; the rail keeps history so "multiple sub agents over a period of time" is visible, not just the live one.
5. **/agents** command lists the same data in the transcript for headless/exec parity; `exec` stream-json gains an `agent_summary` event.
6. Docs: docs/SUBAGENTS.md gets the rail; #5439's discoverability points here.

Out of scope for the first slice: new orchestration semantics, remote/cloud agents, redesigning the Work bar.

No-Issue linkage: this *is* the issue. Refs #5439, #5442, #5123, #5324.

Start with the retained-roster diagnosis in ../agent-roster-20260909/REPORT.md and the existing work_bar_subagents_pty tests; confirm how the Agents register and compact cache represent completed workers. Then trace the /agents entry point and exec stream-json behavior, with docs/SUBAGENTS.md covering the documentation slice. Done means runtime-backed history and controls preserve distinct states and identities, with bounded acceptance evidence after GUI permission resumes.

Written by the indexing model from the issue text.