Delivery-failure retry loop: wake the agent to fix and resend undelivered messages

[dimavrem22]

Mirrors inkbox-ai/hermes-agent-plugin#48 for the Claude Code bridge — the same delivery-failure retry loop, adapted to this repo's per-contact session architecture.

What

An outbound message can die two ways, and until now the bridge only learned about one:

• Rejected at send time — the server's outbound content policy 422s the send (markdown artifacts, emoji overload…), the recipient has opted out, the address is bad, or the body is too long. The error comes back on the send call and was surfaced as a one-shot recovery turn in sessions.py.
• Failed after acceptance — the carrier flags it (e.g. error 40002) or the mail server bounces it, reported later via text.delivery_failed / imessage.delivery_failed / message.bounced / message.failed webhooks. The bridge already woke the session for these, but with no cap and no shared budget.

Both surfaces now feed one delivery-failure loop: the agent is woken in the affected contact's session (a run_consult side-turn — it acts via its Inkbox tools rather than replying on the channel that just failed) with the exact error (rule/code/detail) plus its own undelivered body, and instructed to fix and resend.

Guardrails (feature behavior)

• Hard cap: 3 total sends per logical reply. Failures 1 and 2 wake the agent; failure 3 goes quiet with a loud log line. The budget is shared across both failure surfaces (keyed by conversation + recipient, merged by max).
• Budget resets on a fresh inbound, a delivered receipt, or a 30-minute TTL.
• Webhook replays are deduped per failed message.
• Transient (5xx) send failures are excluded — a bare resend clears those.
• The synchronous surface routes through ContactSession._deliver_reply via a new on_send_rejected gateway callback, replacing its old one-shot recovery turn.
• Subscriptions now include message.bounced / message.failed for mail and the outbound delivery lifecycle for text/iMessage (so delivered-reset and delivery_failed-wake fire for real); the reconciler auto-patches the event set on startup.

Testing

• 19 unit tests (tests/test_delivery_failure_retry.py): cap semantics, budget sharing, resets, replay dedup, group recipient rows, opt-out and bounce paths, all three channels.
• Two live tests (real-model leg, tests/live/test_sms.py): test_sms_retry_after_carrier_delivery_failure (forged, org-key-signed text.delivery_failed webhook → real follow-up SMS) and test_sms_retry_after_internal_spam_block (emoji-overload bait → asserts the block reached the agent via wake-up or inline tool rejection; a compliant follow-up is best-effort since the ask is inherently unsatisfiable).
• Live-suite conversation-health reset guardrail (tests/live/conftest.py): reads the opener's window and only resets when it's running low, so the shared 10DLC pool's duplicate_body / unanswered_limit rules don't false-red the suite. Driver SMS bodies are diversified with a unique ref across the SMS, voice, and cross-channel suites.

Raises the inkbox SDK floor to 0.4.20 (ships blocked_spam_filter) across package metadata, the setup wizard, and doctor. Bumps the plugin 0.1.2 → 0.1.3.

Notes vs the hermes PR

• The hermes PR also suppressed a leaked 🔍 progress glyph in outbound messages; that mechanism (_is_hermes_admin_notice / tool-progress-glyph list) does not exist in this bridge, so there's nothing to mirror here.
• The wake-up is a run_consult capture turn (this bridge's existing delivery-failure design) rather than an auto-replying event turn, so the prompt instructs the agent to send via its Inkbox tools; the cap now bounds what was previously a deliberately-uncapped-but-non-auto-replying path.