fix(reminders): surface the autonomy hold instead of failing silently (#1806)

[dolho]

What

Makes the autonomy hold on self-reminders visible. The gate is not changed — a paused agent still must not wake itself. Only the silence is fixed.

Why

get_active_reminders filters ao.autonomy_enabled = 1, so a reminder on a paused agent is never armed. No job is created, so there is no skip to log — and the API reported a healthy-looking row:

status: pending | fire_at: 2026-07-27T10:43:51.215559Z | attempts: None | error: None

Clean A/B on a fresh 0.8.5 instance, two agents, identical 60s reminder set in the same second, only autonomy_enabled differing:

10:43:53Z  scout(autonomy ON)=fired    corbin(autonomy OFF)=pending
10:44:08Z  scout(autonomy ON)=fired    corbin(autonomy OFF)=pending

Worse than the sibling schedules case (#1796) on every observability axis: schedules at least log skipped: … autonomy is disabled, and scheduling.md documents the dependency. Reminders had neither.

Why not reject at create

Considered and rejected. Chat is not autonomy-gated, so an agent can be mid-conversation with a user who says "remind me in an hour". A 409 would fail a legitimate request, and queue-until-unpaused is genuinely useful — the reminder does fire past-due once autonomy returns (verified: held reminder fired ~1 reconcile cycle after enabling). The problem was never the holding; it was not saying so.

Changes

Layer	Change
models.py	autonomy_hold on ReminderSummary (inherited by Reminder), derived at read time — no column, no dual-track migration. Defaults False.
routers/reminders.py	_autonomy_hold() applied to create, list, cancel. Terminal reminders never flagged and skip the lookup. List resolves autonomy once per page (all rows share one agent — a per-row resolve would be a pointless N+1). Fail-open on a read error.
src/scheduler	log a count of held reminders per reconcile pass, only when non-zero, best-effort. The only place that can report them, since they never enter the job store.
MCP set_reminder	explicit warning sentence when held — not just the boolean, since the caller is an LLM about to tell a user "done" — plus the dependency in the tool description (Invariant #13).
docs/user-docs/automation/agent-reminders.md	documents the dependency, mirroring scheduling.md.

Verification

Unit — tests/unit/test_1806_reminder_autonomy_hold.py, 11 cases, all passing: held for pending/firing, not held when autonomy is on, never flagged for fired/cancelled/failed (asserting the lookup is not even attempted), fail-open on a read error, precomputed autonomy short-circuits the per-row lookup, the shared response shaper carries the flag, and both models default to False.

Live, patched into a running 0.8.5 instance (its backend source is bind-mounted):

CREATE on paused agent   → status: pending | autonomy_hold: True
CREATE on active agent   → status: pending | autonomy_hold: False
LIST on paused agent     → fired   | autonomy_hold: False      # terminal, correctly not flagged
                           pending | autonomy_hold: True

Then toggling autonomy on, without touching the row:

pending | autonomy_hold: False   <- flipped, proving it is derived live, not frozen at create

The instance was restored to stock afterwards and verified clean.

MCP server compiles (tsc via the image build, exit 0).

Not verified

The scheduler log line is unit-and-inspection level only — the scheduler image is not bind-mounted on my test stack, so I did not rebuild it to watch the line appear. The query is the exact complement of get_active_reminders' predicate and is wrapped best-effort.

Follow-up worth its own issue

A reminder held for days fires immediately on unpause, and several would fire at once. "Remind me in an hour" arriving three days late may be worse than never arriving. Deliberately out of scope here — flagging it rather than widening this fix before the release.

Related to #1806

🤖 Generated with Claude Code

Fixes #1806

[vybe]

Validated via /validate-pr: derived autonomy_hold flag (read-time, no column — correctly avoids a dual-track migration), all three surfaces in sync (router+models, MCP tool with an explicit LLM-facing HELD warning, scheduler once-per-pass held count). db.get_autonomy_enabled verified present (db/agent_settings/autonomy.py:53). Gate behavior unchanged; fail-open on settings-read failure; terminal rows skip the lookup; list path avoids N+1. 7 named regression tests + user-docs update. Security scan clean. Added 'Fixes #1806' for status auto-promotion.