feat(skills): add reply-router

skills/reply-router/REPORT.md

@@ -0,0 +1,22 @@
+# Reply Router Bounty Report
+
+## What was built
+
+- Added `skills/reply-router` as a public runx skill with a default `route_reply` agent-task runner.
+- Added typed inputs for inbound reply content, original send receipt, suppression policy, recipient stream coordinates, optimistic concurrency, and idempotency.
+- Added typed outputs for `classification`, `suppression_result`, `routing_decision`, and `escalation_lane`.
+- Added inline harness coverage for sealed unsubscribe suppression, sealed interested routing, and ambiguous/unsealed stop behavior.
+
+## Behavioral boundaries
+
+- Unsubscribe replies emit a ready-to-append suppression packet with the recipient aggregate, expected version, idempotency key, and `reply.unsubscribe_suppressed` event payload.
+- Interested replies emit a bounded `runx.reply.routing.v1` packet naming a later governed `send-as` run.
+- The skill never sends a message or mutates a provider directly.
+- Ambiguous replies or unsealed original send receipts stop with `needs_agent`.
+
+## Verification
+
+- `runx-cli 0.6.14` is available through `npx.cmd -y @runxhq/cli@0.6.14`.
+- `runx skill inspect ./skills/reply-router route_reply --json` passes with status `ok`.
+- Local Windows harness and registry publish are blocked by the native sealed receipt store error `os error 87` while syncing the receipt directory.
+- The blocker is environmental to this Windows host; the skill package itself parses and inspects successfully.

skills/reply-router/SKILL.md

@@ -0,0 +1,166 @@
+---
+name: reply-router
+description: Classify inbound replies against a suppression policy, emit durable suppression packets, and produce bounded routing decisions without sending.
+runx:
+  category: ops
+---
+
+# Reply Router
+
+Reply Router handles inbound replies to previously governed outbound sends. It
+protects the compliance boundary: an unsubscribe reply becomes a ready-to-append
+recipient-keyed suppression packet, while non-unsubscribe replies produce only a
+typed routing decision for a separate governed `send-as` run.
+
+## Quality Profile
+
+- Purpose: prevent missed suppressions and keep reply follow-up routing separate
+  from the act of sending.
+- Audience: operators running outreach, lifecycle, support, or sales workflows
+  that need a replayable reply decision before any next send.
+- Artifact contract: emit `classification`, optionally emit a
+  `reply.unsubscribe_suppressed` suppression result packet with CAS inputs, and
+  optionally emit a `runx.reply.routing.v1` routing decision that names a
+  downstream governed `send-as` run.
+- Evidence bar: every suppression or route must cite reply text, the sealed
+  original send receipt, the suppression policy signal, and the prior recipient
+  projection.
+- Stop conditions: unsealed original receipt, ambiguous reply intent, missing
+  suppression policy, policy confidence below threshold, missing recipient
+  aggregate, unreadable prior suppression state, or any attempt to send from this
+  skill.
+
+## Runner
+
+`route_reply` is the default agent-task runner.
+
+It performs one bounded classification and packet-authoring act:
+
+1. Classify the reply using only the supplied reply, sealed original send
+   receipt, policy, and recipient stream coordinates.
+2. For unsubscribe intent, emit a `runx.data.operation_result.v1` packet with
+   `operation: append_event`, `expected_version`, `idempotency_key`, and the
+   `reply.unsubscribe_suppressed` event payload.
+3. For follow-up intent, emit a bounded `runx.reply.routing.v1` packet naming a
+   downstream governed `send-as` run.
+4. For ambiguity or unsealed receipts, stop with `needs_agent`.
+
+The runner never sends a follow-up. For interested, objection, out-of-office, or
+wrong-person replies, the output is a routing decision naming a separate
+governed `send-as` run that a downstream driver or operator must issue.
+
+## Inputs
+
+- `inbound_reply` (required): object with `content`, `received_from`, and
+  `received_at`.
+- `original_send_receipt` (required): object with `send_plan`, `principal`,
+  `receipt_id`, `checksum`, and `sealed`.
+- `suppression_policy` (required): object with `unsubscribe_signals` and
+  `confidence_threshold`.
+- `data_source_ref` (required): logical data source for suppression events.
+- `store_id` (optional): deterministic local fixture store id for downstream
+  data-store replay.
+- `resource` (optional): suppression event resource, defaults to
+  `reply_suppressions`.
+- `aggregate_id` (required): recipient aggregate id, for example
+  `recipient:buyer@example.com`.
+- `expected_version` (required): stream version read before append.
+- `idempotency_key` (required): stable key derived from recipient, original send,
+  and classification.
+
+## Output Schema
+
+For unsubscribe replies:
+
+```json
+{
+  "classification": {
+    "type": "unsubscribe",
+    "confidence": 0.98,
+    "evidence": ["Reply text contains unsubscribe."]
+  },
+  "suppression_result": {
+    "aggregate_id": "recipient:buyer@example.com",
+    "idempotency_key": "recipient:buyer@example.com:seq-001:unsubscribe:v1",
+    "before_version": 0,
+    "after_version": 1
+  },
+  "routing_decision": null
+}
+```
+
+For routed replies:
+
+```json
+{
+  "classification": {
+    "type": "objection",
+    "confidence": 0.86,
+    "evidence": ["Reply asks about pricing."]
+  },
+  "routing_decision": {
+    "schema": "runx.reply.routing.v1",
+    "classification": "objection",
+    "send_target": "send-as:reply-followup",
+    "principal": "runx:principal:sales-ops"
+  },
+  "suppression_result": null
+}
+```
+
+## Suppression Semantics
+
+An unsubscribe classification must meet all of these conditions:
+
+- The original send receipt is sealed.
+- The reply text contains unsubscribe intent named by
+  `suppression_policy.unsubscribe_signals`.
+- The confidence is at or above `suppression_policy.confidence_threshold`.
+- The suppression result names `append_event` with `aggregate_id` equal to the
+  recipient entity, `expected_version`, and an idempotency key derived from
+  recipient plus original send plus decision.
+- No routing send target is emitted alongside the suppression.
+
+The suppression packet is the compliance block a downstream data-store lane
+commits and the next send preflight reads. Receipt history is evidence, not the
+state store.
+
+## Edge Cases And Stop Conditions
+
+- `needs_receipt`: `original_send_receipt.sealed` is false or the receipt id is
+  missing.
+- `needs_policy`: suppression signals or threshold are missing.
+- `ambiguous_reply`: the reply does not clearly match a suppression or routing
+  class.
+- `below_threshold`: matched signals exist but confidence is below the policy
+  threshold.
+- `needs_projection`: prior suppression state cannot be read.
+- `conflict`: the suppression stream version no longer matches
+  `expected_version`.
+- `send_attempted`: any attempt to send from this skill is invalid; route by
+  naming a downstream governed run instead.
+
+## Harness
+
+The inline harness covers:
+
+- `sealed_unsubscribe_suppression`: a sealed unsubscribe classification emits a
+  recipient suppression event packet with CAS coordinates.
+- `sealed_interested_route`: a sealed interested reply emits a bounded routing
+  packet and no suppression result.
+- `stop_ambiguous_or_unsealed`: an ambiguous reply with an unsealed original
+  receipt stops at `needs_agent` and performs no suppression write.
+
+## Invocation
+
+```bash
+runx skill reply-router route_reply \
+  --input-json inbound_reply='{"content":"unsubscribe me","received_from":"buyer@example.com","received_at":"2026-06-30T10:00:00Z"}' \
+  --input-json original_send_receipt='{"send_plan":{"recipient":"buyer@example.com","sequence_id":"seq-001"},"principal":"runx:principal:sales-ops","receipt_id":"runx:receipt:sha256:sealed","checksum":"sha256:...","sealed":true}' \
+  --input-json suppression_policy='{"unsubscribe_signals":["unsubscribe","opt out"],"confidence_threshold":0.8}' \
+  -i data_source_ref=local://reply-router/demo \
+  -i aggregate_id=recipient:buyer@example.com \
+  --input-json expected_version=0 \
+  -i idempotency_key=recipient:buyer@example.com:seq-001:unsubscribe:v1 \
+  --json
+```