diff --git a/docs/prds/0020-start-and-attach-from-dashboard.md b/docs/prds/0020-start-and-attach-from-dashboard.md
new file mode 100644
index 0000000..251cde0
--- /dev/null
+++ b/docs/prds/0020-start-and-attach-from-dashboard.md
@@ -0,0 +1,336 @@
+# PRD 0020: Start and attach to agents from inside the dashboard
+
+- **Status:** Draft
+- **Author:** didericis
+- **Created:** 2026-05-26
+
+## Summary
+
+Today the dashboard is read-only: it surfaces pending proposals
+and active agents (PRD 0019) but can't *start* an agent or
+*re-enter* one. The operator's path is split — they launch
+agents from one terminal (`./cli.py start <name>`), and watch
+them from another (`./cli.py dashboard`).
+
+This PRD collapses that split. The dashboard becomes the
+operator's single surface: pressing a key opens an agent picker,
+selecting one runs the existing prepare → preflight → launch
+flow inside a curses-friendly variant, and on yield drops to a
+full-screen `docker exec -it … claude` session (the "handoff"
+shape from `docs/research/claude-code-pane-in-dashboard.md`).
+When the operator exits claude, the dashboard re-renders with
+the now-running bottle visible in the agents pane.
+
+Crucially, the bottle's lifetime is owned by the *dashboard
+process*, not by the individual claude session. Exit claude →
+back to dashboard, bottle still running. Start another agent →
+two bottles up at once. Quit the dashboard → all dashboard-
+launched bottles tear down.
+
+## Problem
+
+Two real frictions today:
+
+1. **Two terminals for one workflow.** The dashboard is the
+   right shape to *watch* agents — proposals queue, status
+   updates, operator-edit verbs — but it's the wrong shape to
+   *start* them. Today you open a second terminal for that. In
+   parallel use (3–5 bottles), the operator has 5+ terminals
+   open and the dashboard's "active agents" pane is hopelessly
+   behind reality because they just spawned three in a row.
+
+2. **`./cli.py start` ties the bottle to a single claude
+   session.** The start command's `ExitStack` brings the bottle
+   up, runs claude, and tears down on Ctrl-D — fine for a one-
+   shot session, wrong for "let me bounce in and out of this
+   bottle a few times while triaging proposals." Today the only
+   way to re-enter a bottle after exiting claude is to start a
+   fresh one and lose all in-bottle state.
+
+The dashboard already discovers active bottles, scopes
+operator-edit verbs to a selected agent (PRD 0019), and
+captures full-merged logs per bottle (PRD 0018). It already
+*wants* to be the primary surface. This PRD finishes that.
+
+## Goals / Success Criteria
+
+1. From inside `./cli.py dashboard`, pressing `n` (new) opens
+   an agent picker listing every agent defined in the manifest.
+   Selecting one runs `prepare → preflight → launch`.
+2. The preflight Y/N summary renders cleanly — either as a
+   curses modal or via `curses.endwin() → text-mode prompt
+   → restore`, matching the existing editor-flow pattern.
+3. On launch success, the dashboard performs a handoff (option
+   1 from the research doc): `curses.endwin()` → `docker exec
+   -it claude-bottle-<slug> claude --dangerously-skip-permissions`
+   → on exit, `stdscr.refresh()` and re-render with the new
+   bottle in the agents pane.
+4. The bottle's lifetime is owned by the dashboard process, NOT
+   by any single claude session. Exiting claude (Ctrl-D, `/exit`)
+   returns to the dashboard with the bottle still running. The
+   operator can start more agents and re-enter previous ones.
+5. Pressing Enter on a selected row in the agents pane re-
+   attaches to that agent's bottle via the same handoff — drops
+   to full-screen claude, returns on exit.
+6. Pressing `x` (or similar — keybinding decided in design)
+   on a selected agent stops just that bottle (compose down +
+   state cleanup) without quitting the dashboard.
+7. Quitting the dashboard (`q`) tears down every bottle the
+   dashboard started, unless something has explicitly preserved
+   the state (capability-block, crash). Matches today's
+   start.py teardown semantics.
+
+## Non-goals
+
+- **A pane that hosts the claude TUI alongside proposals.** The
+  embedded-emulator option from the research doc is out of
+  scope. The handoff (option 1) is the v1; option 2 is a
+  separate PRD if and when handoff is observably insufficient.
+- **Adopting bottles started by an out-of-dashboard `./cli.py
+  start` invocation.** Those have their own ExitStack-owner and
+  the dashboard treats them as read-only-watch (already does
+  today). Re-attach only applies to bottles the *current
+  dashboard process* started.
+- **Persisting a "bottle pool" across dashboard runs.** When
+  the dashboard quits, its bottles go. Resume across dashboard
+  invocations is `./cli.py resume <identity>`, which is
+  unchanged.
+- **Multi-window UI.** Single curses window, two existing
+  panes (proposals + agents); the agent picker is a modal, not
+  a third pane.
+- **Removing `./cli.py start`.** Stays as the script-friendly /
+  legacy entry point. The dashboard is the new default.
+
+## Scope
+
+### In scope
+
+- Manifest-driven agent picker (curses modal): list view with
+  j/k navigation + Enter to confirm, Esc to abort.
+- Preflight rendering inside the dashboard's curses surface
+  (modal or drop-and-resume — picked in design).
+- A new `_dashboard_start_flow` that wraps prepare + preflight
+  + launch and returns a `DockerBottle` handle the dashboard
+  retains alongside its `pending` and `agents` lists.
+- A `bottles: dict[slug, DockerBottle]` map on the main loop
+  that owns every dashboard-launched handle. ExitStack tears
+  them all down on dashboard exit.
+- `Enter` on an agents-pane row → re-attach handoff (docker
+  exec -it claude into the existing container).
+- `x` (or similar) on an agents-pane row → explicit per-bottle
+  stop without quitting.
+- `q` (existing quit key) → tear down all dashboard-launched
+  bottles before returning.
+
+### Out of scope
+
+- Changes to `./cli.py start` itself. It keeps its current
+  shape; the dashboard reuses its internal pieces (backend.
+  prepare / backend.launch) without reaching through the CLI
+  layer.
+- Changes to `backend.launch`'s context-manager contract; the
+  dashboard's bottle map just holds the context-manager-yielded
+  Bottle and calls `__exit__` on quit / explicit stop.
+- New manifest fields. The picker reads what's already there.
+- Adopting non-dashboard bottles into the dashboard's owned set.
+
+## Proposed design
+
+### Bottle ownership
+
+Today's flow:
+
+```
+./cli.py start agent
+  └─ with backend.launch(plan) as bottle:        ← bottle alive while inside `with`
+       bottle.exec_claude([...], tty=True)       ← blocks until claude exits
+     # context exits → compose down → state cleanup
+```
+
+The proposed dashboard-owned flow:
+
+```
+./cli.py dashboard
+  └─ stack = ExitStack()
+     bottles: dict[str, DockerBottle] = {}
+
+     # operator presses `n`, picks agent
+     ctx = backend.launch(plan)
+     bottle = stack.enter_context(ctx)           ← bottle stays alive
+     bottles[plan.slug] = bottle
+
+     # operator interacts via:
+     curses.endwin()
+     bottle.exec_claude([...], tty=True)         ← blocks; returns on Ctrl-D
+     stdscr.refresh()
+     # bottle is STILL ALIVE here — only the claude process exited
+
+     # ... operator does other things, eventually `q`:
+     stack.close()                               ← tears down every bottle
+```
+
+The shift is one line of code semantically but the change in
+operator experience is real: bottles outlive any single claude
+session.
+
+### Agent picker
+
+Pressing `n` opens a centered modal listing every agent name
+from `spec.manifest.agents`. j/k navigates; Enter selects; Esc
+aborts. Width is the longest name + bottle name + a column for
+"already running?" so the operator can see at a glance whether
+picking an agent starts a fresh one (different slug suffix) or
+not.
+
+```
+┌─ start agent ───────────────────────────┐
+│   implementer       dev      (running)  │
+│ > researcher        dev                 │
+│   triage-bot        sandbox             │
+└─ Enter: start  Esc: cancel ─────────────┘
+```
+
+Starting an agent that already has a running bottle is allowed
+— each `start` mints a fresh slug — but the picker surfaces the
+already-running state so the operator doesn't accidentally
+double-launch.
+
+### Preflight Y/N
+
+Two viable shapes:
+
+**Modal** — render the preflight summary lines (`agent / env /
+skills / bottle / git gate / egress`) in a centered curses
+modal with `[y/N]` at the bottom. Capture the next keypress.
+
+**Drop-and-resume** — `curses.endwin()`, print the preflight to
+stderr, read y/N from stdin, restore curses. Matches the
+editor-flow + handoff pattern; lower implementation cost.
+
+Lean toward **modal** for the y/N because it doesn't flash the
+terminal between dashboard frames. Drop-and-resume is acceptable
+if modal proves fiddly.
+
+### Re-attach (Enter on agent)
+
+Same handoff pattern the new-agent flow uses. The dashboard
+already holds the `DockerBottle` for any slug it started —
+`bottle.exec_claude([...], tty=True)` does the right `docker
+exec -it claude …` and returns on session exit. Re-attach is
+"already-running" + the same exec call; the agent picker isn't
+involved.
+
+For agents the dashboard didn't start (read-only watch), Enter
+is a no-op with a status hint ("dashboard didn't start this
+bottle; resume with `./cli.py resume <identity>` outside the
+dashboard"). PRD-0019's selection model already differentiates
+focus; this layer just gates the action.
+
+### Explicit per-bottle stop
+
+`x` on a selected dashboard-owned agent invokes
+`stack.pop_callback`-style targeted teardown: take that bottle
+out of the map, call its `close()` to tear down compose + state,
+update the agents pane on the next refresh. Bottles the
+dashboard didn't start (`x` on a read-only-watch row) → no-op
+with a status hint.
+
+### Dashboard quit
+
+`q` (existing) calls `stack.close()` before exit; every
+dashboard-launched bottle goes through its normal teardown
+(`compose down` + state settle). Preserve markers (capability-
+block, crash) still keep state across teardown. The dashboard
+process itself returns 0.
+
+If the operator wants to keep bottles alive past dashboard
+exit, the existing path is unchanged: launch them via
+`./cli.py start` in a separate terminal. That ownership stays
+out-of-band.
+
+## Implementation chunks
+
+Sized for one PR each.
+
+1. **Refactor `_launch_bottle` so the launch + exec_claude
+   pieces are separable.** Today's `cli/start.py` runs both
+   inside one function. Extract `prepare_with_preflight(spec,
+   *, render_preflight, prompt_yes)` and `attach_claude(bottle,
+   *, remote_control)`. The CLI's existing one-shot use binds
+   them as before; the dashboard binds them with curses-aware
+   render + prompt callables. No behavior change.
+2. **Agent picker modal + new-agent flow.** New key `n` opens
+   the picker; `prepare_with_preflight` runs against the
+   selected agent; on Y, `backend.launch(plan)` enters the
+   dashboard's ExitStack; handoff invokes `attach_claude`.
+3. **Re-attach via Enter on owned agents-pane row.** Looks up
+   the slug in the dashboard's `bottles` map; if present →
+   handoff; else → status-line hint pointing at `./cli.py
+   resume`.
+4. **Explicit per-bottle stop (`x` keybinding).** Pop the
+   bottle's `close` callback off the stack, call it, refresh.
+5. **Quit-cleanup (`q`).** Hook `stack.close()` into the
+   normal return path. Document the "exiting dashboard tears
+   down every bottle it started" contract in `dashboard.py`'s
+   module docstring.
+
+## Open questions
+
+1. **Modal vs. drop-and-resume for preflight Y/N.** Both work;
+   modal is nicer if the curses geometry handling is
+   straightforward. Pick during chunk 2 by prototyping the
+   modal in ~30 lines and seeing if it looks right.
+
+2. **Agent picker: text-filter typing?** v1 is j/k navigation
+   only. If the manifest has 20+ agents the picker gets noisy;
+   add fzf-style filter input later if needed.
+
+3. **What happens if `attach_claude` exits because the
+   container died** (not a clean claude exit — e.g., OOM,
+   panic)? Today's `_settle_state` marks the bottle preserved
+   for non-zero exit codes. The dashboard's re-render needs to
+   notice the bottle is gone (compose down or container-not-
+   running state) and surface a status line. Probably:
+   transcript snapshot + mark preserved + remove from
+   `bottles` map + status line "claude session for [slug]
+   ended with exit N; preserved for resume".
+
+4. **Double-start of the same agent.** Allowed by design — slugs
+   are unique per launch — but the picker should make it clear
+   this is a "start a SECOND bottle" decision, not a "re-enter
+   the first." Probably handled by showing the running-count in
+   the picker row.
+
+5. **Should `q` confirm before tearing down N running
+   bottles?** A 5-bottle dashboard with 5 in-flight sessions
+   loses non-trivial state on accidental `q`. Probably yes:
+   curses modal "quit and tear down N bottles? [y/N]". Skip
+   confirmation when there are zero owned bottles.
+
+6. **Race between handoff and 1s refresh tick.** While the
+   dashboard's `stdscr.timeout` is set, a key press fires the
+   handoff and the dashboard sits in `docker exec` for minutes.
+   `discover_active_agents` / `discover_pending` don't poll
+   during that window, which is fine — the moment we
+   `stdscr.refresh()` after exec returns, the next loop iter
+   runs discovery and the panes reflect reality. Worth calling
+   out in the design but no special handling needed.
+
+7. **Multi-bottle resource use.** Five bottles up means five
+   compose projects: 5×(agent + pipelock + egress optional +
+   git-gate optional + supervise optional) containers, plus 5×2
+   networks. On a 16-GiB host this is fine; on something
+   smaller the operator might want a soft cap or a warning.
+   Out of v1; flag for follow-up if it bites.
+
+## References
+
+- PRD 0018 — compose-per-instance lifecycle (the `backend.
+  launch` context-manager contract this PRD layers against)
+- PRD 0019 — active-agents pane + selection model (the
+  agents-pane row the re-attach + stop verbs hook into)
+- `docs/research/claude-code-pane-in-dashboard.md` — option 1
+  (handoff) is what `attach_claude` implements here; options 2
+  / 3 are out of scope for this PRD
+- `claude_bottle/cli/start.py:_launch_bottle` — the function
+  chunk 1 extracts the prepare + attach pieces out of