feat(issue-24): startup dialogs come from the agent adapter and are handled while waiting for the TUI

handle_startup_dialogs was a claude-only string ladder called only when
AGENT=claude, and wait_for_tui_ready does not treat a dialog screen as ready.
So agy, grok and codex sat on their directory-trust prompt for the full
readiness timeout, the session was recorded as tui-ready-timeout, and
onboarding injection was disabled (create_session.sh:313-314) - a human had to
accept each dialog by hand.

The signatures and accept keys move to the adapters as `startup_dialogs`
(pattern, keys, kind), published through the existing `facts` eval contract as
MAM_STARTUP_DIALOGS. claude's ladder moves over verbatim; agy takes Enter, grok
y, codex Enter, and codex's update notice takes 2 (Skip), which blocks a start
just as hard while changing nothing. kind is `trust` or `notice`:
--no-auto-trust leaves trust dialogs to a human and still answers notices,
since trusting a directory makes codex load its project-local config, hooks and
exec policy.

Re-sending is gated by time (MAM_DIALOG_RESEND_S, default 5s), not by frames.
#24's comment reports that `--source visible` keeps showing an accepted dialog
for a while, and _pane_capture hardcodes that source, so "still on screen" is
not evidence the keys were missed; acceptance is judged by the ready tokens
alone.

Two existing tests asserted the old shape and are updated, not weakened:
b19's "Escape branch stays" now looks for it in claude's adapter, and c1b's
"Enter once per iteration" becomes "never twice for one frame", which the
cooldown satisfies strictly more tightly.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

tests/test_b19_headless_reconcile_fixes.py

@@ -282,13 +282,19 @@ if _pane_dialog_open dummy; then echo "DIALOG_OPEN"; else echo "DIALOG_CLOSED";
 
 
 def test_mam_dialog_tokens_exclude_yes_try_it():
-    """N-1 / F-1: token list must not contain Yes, try it; Escape branch stays."""
+    """N-1 / F-1: token list must not contain Yes, try it; Escape branch stays.
+
+    Issue #24 moved the per-agent ladder out of lib.sh into the agent adapter,
+    so the Escape branch is now claude's own `startup_dialogs` entry. The two
+    token-list assertions are unchanged; only where the branch lives moved.
+    """
     content = open(_LIB_SH, encoding="utf-8").read()
     modal_line = next(l for l in content.splitlines() if l.startswith("_MAM_MODAL_TOKENS="))
     hint_line = next(l for l in content.splitlines() if l.startswith("_MAM_HINT_TOKENS="))
     assert "Yes, try it" not in modal_line
     assert "Yes, try it" not in hint_line
-    assert "grep -q 'Yes, try it'" in content
+    adapter = os.path.join(os.path.dirname(_LIB_SH), "lib_py", "agents", "adapters", "claude.py")
+    assert "('Yes, try it', ['Escape'], 'notice')" in open(adapter, encoding="utf-8").read()
 
 
 def test_wait_for_tui_ready_succeeds_on_fullscreen_tip():

tests/test_c1_tui_readiness.py

@@ -55,7 +55,16 @@ if _pane_dialog_open dummy; then echo "DIALOG_OPEN"; else echo "DIALOG_CLOSED";
 
 
 def test_c1b_hint_press_enter_sends_enter():
-    """C-1b (B-3): 'Press Enter to continue' alone must return dialog CLOSED, and wait_for_tui_ready must send Enter exactly once per iteration (30 total across 30 iterations, not 60)."""
+    """C-1b (B-3): 'Press Enter to continue' alone must return dialog CLOSED,
+    and wait_for_tui_ready must never send Enter twice for one frame.
+
+    Issue #24 replaced "once per iteration" with a time cooldown
+    (MAM_DIALOG_RESEND_S): a `--source visible` snapshot keeps showing an
+    accepted dialog for a while, so re-sending on every poll types into a TUI
+    that already moved on. This test stubs `sleep`, so the wall clock barely
+    advances and one cooldown window covers the whole run - hence 1, not 30.
+    The guard the case was written for (never 2 per iteration) is unchanged
+    and strictly stronger now."""
     screen = """Some startup banner
 Press Enter to continue
 """
@@ -73,7 +82,7 @@ export MAM_STRONG_READY_TOKENS='NonExistentStrong'
 export MAM_WEAK_READY_TOKENS=''
 export MAM_INPUT_PLACEHOLDER=''
 rc=0
-wait_for_tui_ready dummy-sess custom || rc=$?
+wait_for_tui_ready dummy-sess claude || rc=$?
 echo "RC=$rc"
 echo "KEYS_COUNT=$(grep -c "send-keys -t dummy-sess Enter" "$KEYS" || echo 0)"
 rm -f "$KEYS"
@@ -81,7 +90,9 @@ rm -f "$KEYS"
     res = _run_lib_helpers(script)
     assert res.returncode == 0, res.stderr + res.stdout
     assert "DIALOG_CLOSED" in res.stdout
-    assert "KEYS_COUNT=30" in res.stdout
+    count = int(res.stdout.split("KEYS_COUNT=")[1].split()[0])
+    assert count >= 1, "the dialog was never answered: " + res.stdout
+    assert count <= 30, "one frame must never be answered twice: " + res.stdout
 
 
 def test_c1c_transition_to_upsell_modal_sends_escape_not_enter():

tests/test_issue23_24_startup_path.py

@@ -302,3 +302,145 @@ def test_s13_the_loop_watcher_applies_the_same_rule(mam_sandbox):
 
     watcher.wait_once("agt", session="iso-x", run=_run)
     assert seen["env"] is not None and "HERDR_SESSION" not in ("",)
+
+
+# --------------------------------------------------------------------------
+# item 2 — start-up dialogs come from the adapter (#24)
+#
+# MOCKED, and this is the limit of it: the pane text is a fixture string and
+# send-keys is a shell stub, so these fix OUR contract - which screen signature
+# gets which keys, once. They do NOT show that the keys drive a real agy, grok,
+# codex or opencode TUI; no such TUI can be started in this environment. That
+# check is a live herdr run after the loop.
+# --------------------------------------------------------------------------
+
+def _run_dialogs(tmp_path, agent, pane_text, timeout=1, env=None, prelude=""):
+    lib = tmp_path / ".agents" / "skills" / "lib.sh"
+    pane_file = tmp_path / "pane.txt"
+    pane_file.write_text(pane_text)
+    keys_log = tmp_path / "sent_keys.txt"
+    if keys_log.exists():
+        keys_log.unlink()
+    script = '''
+source "%s"
+_pane_tail() { cat "%s"; }
+_sks_herdr() { printf '%%s\\n' "$*" >> "%s"; }
+%s
+handle_startup_dialogs test-sess %d %s
+echo "rc=$?"
+''' % (lib, pane_file, keys_log, prelude, timeout, agent)
+    run_env = dict(os.environ)
+    run_env.pop("MAM_AUTO_TRUST", None)
+    run_env.update(env or {})
+    res = subprocess.run(["bash", "-c", script], capture_output=True, text=True,
+                         cwd=str(tmp_path), env=run_env)
+    sent = keys_log.read_text().splitlines() if keys_log.exists() else []
+    return res, sent
+
+
+AGENT_DIALOGS = [
+    ("agy", "Do you trust the contents of this project?\n > Yes, I trust this folder",
+     ["send-keys -t test-sess Enter"]),
+    ("grok", "Do you trust the contents of this directory?",
+     ["send-keys -t test-sess y"]),
+    ("codex", "Do you trust the contents of this directory?\n > 1. Yes, continue",
+     ["send-keys -t test-sess Enter"]),
+    ("codex", "Update available\n 1. Update now\n 2. Skip\n 3. Skip until next version",
+     ["send-keys -t test-sess 2"]),
+    ("claude", "Do you trust the files in this folder?\n Yes, I trust this folder",
+     ["send-keys -t test-sess Enter"]),
+    ("claude", "Try the new fullscreen renderer?\n Yes, try it",
+     ["send-keys -t test-sess Escape"]),
+]
+
+
+@pytest.mark.parametrize("agent,pane,expected", AGENT_DIALOGS)
+def test_s4_each_agents_dialog_signature_gets_its_own_keys(
+        mam_sandbox, agent, pane, expected):
+    """S4 (stubbed pane + stubbed send-keys): the adapter's signature is
+    matched and exactly its keys are sent."""
+    res, sent = _run_dialogs(mam_sandbox, agent, pane)
+    assert sent == expected, (res.stdout, res.stderr, sent)
+
+
+def test_s4b_every_adapter_publishes_the_fact(mam_sandbox):
+    """S4b (real `lib_py.agents facts`): MAM_STARTUP_DIALOGS is part of the
+    facts contract for every agent, so the shell never has to know an agent
+    name. An agent with no known blocking modal publishes an empty list."""
+    repo = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+    skills = os.path.join(repo, ".agents", "skills")
+    seen = {}
+    for agent in ("claude", "agy", "grok", "codex", "opencode", "hermes"):
+        out = subprocess.run(
+            [os.path.join(repo, ".venv", "bin", "python"), "-m", "lib_py.agents",
+             "facts", agent],
+            capture_output=True, text=True, cwd=repo,
+            env=dict(os.environ, PYTHONPATH=skills)).stdout
+        line = [l for l in out.splitlines() if l.startswith("MAM_STARTUP_DIALOGS=")]
+        assert line, "no MAM_STARTUP_DIALOGS for %s" % agent
+        seen[agent] = json.loads(line[0].split("=", 1)[1].strip("'"))
+    for agent in ("claude", "agy", "grok", "codex"):
+        assert seen[agent], "%s must publish its start-up dialogs" % agent
+        for pattern, keys, kind in seen[agent]:
+            assert kind in ("trust", "notice"), (agent, kind)
+            assert keys, (agent, pattern)
+    assert seen["opencode"] == [] and seen["hermes"] == []
+
+
+def test_s5_an_accepted_dialog_still_on_screen_is_not_resent(mam_sandbox):
+    """S5 (stubbed pane + stubbed send-keys): #24's comment reports that
+    `--source visible` keeps showing the dialog for a while after it was
+    accepted, and _pane_capture hardcodes that source. The pane here never
+    changes, so a frame-count rule would send the keys again every poll; the
+    time cooldown must keep it at one."""
+    res, sent = _run_dialogs(
+        mam_sandbox, "grok", "Do you trust the contents of this directory?",
+        timeout=3, env={"MAM_DIALOG_RESEND_S": "60"})
+    assert sent == ["send-keys -t test-sess y"], (res.stdout, res.stderr, sent)
+
+
+def test_s5b_the_cooldown_expiring_allows_one_more_try(mam_sandbox):
+    """S5b: the cooldown is a rate limit, not a one-shot - a dialog that really
+    did not take the keys is retried once the window passes."""
+    res, sent = _run_dialogs(
+        mam_sandbox, "grok", "Do you trust the contents of this directory?",
+        timeout=4, env={"MAM_DIALOG_RESEND_S": "1"})
+    assert len(sent) >= 2, (res.stdout, res.stderr, sent)
+    assert set(sent) == {"send-keys -t test-sess y"}, sent
+
+
+def test_s5c_readiness_is_judged_by_the_ready_tokens_not_by_the_dialog_leaving(
+        mam_sandbox):
+    """S5c: a ready screen returns 0 without sending anything."""
+    res, sent = _run_dialogs(mam_sandbox, "grok", "Grok ready\n>>> ", timeout=3)
+    assert sent == [], sent
+    assert "rc=0" in res.stdout, res.stdout + res.stderr
+
+
+def test_s6_no_auto_trust_leaves_trust_dialogs_alone_but_answers_notices(
+        mam_sandbox):
+    """S6 (stubbed): --no-auto-trust. codex trust is left to a human with a
+    warning; the update notice still gets answered, since skipping it changes
+    nothing and it blocks the session just as hard."""
+    res, sent = _run_dialogs(
+        mam_sandbox, "codex", "Do you trust the contents of this directory?\n > 1. Yes, continue",
+        timeout=1, env={"MAM_AUTO_TRUST": "0"})
+    assert sent == [], sent
+    assert "no-auto-trust" in res.stderr, res.stderr
+
+    res, sent = _run_dialogs(
+        mam_sandbox, "codex", "Update available\n 1. Update now\n 3. Skip until next version",
+        timeout=1, env={"MAM_AUTO_TRUST": "0"})
+    assert sent == ["send-keys -t test-sess 2"], (res.stderr, sent)
+
+
+def test_s6b_create_session_no_longer_gates_dialogs_on_claude(mam_sandbox):
+    """S6b (source read): the #24 root cause was the `AGENT = claude` gate at
+    the create_session call site, so the gate's absence is the regression."""
+    repo = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+    src = open(os.path.join(repo, ".agents", "skills", "multi-agent-mux-create",
+                            "scripts", "create_session.sh")).read()
+    call = [l for l in src.splitlines() if "handle_startup_dialogs" in l]
+    assert call, "create_session no longer answers start-up dialogs at all"
+    assert all('"$AGENT"' in l for l in call), call
+    assert 'if [ "$AGENT" = "claude" ] && [ -z "$LAST_VISIBLE_STATUS_OVERRIDE" ]' not in src