Keep historical outcomes out of live activity and document rollout gates

docs/delivery.md

@@ -229,3 +229,30 @@ company. Remaining boundaries include OS-enforced tool isolation, a real media
 capture/publication pilot, provider billing receipts, domain-specific quality and
 business-impact evaluations, and distributed-host ownership. These are reported
 as unproven, not inferred from the presence of modules or dashboard cards.
+
+## Deployment Gates
+
+A merged PR is not automatically a deployed runtime. `bin/run_autopull.sh` follows
+the immutable commit in `runtime/deploy-approved-sha`, not the newest `main`.
+Only advance this pin for an explicitly approved release after testing and checking
+active workers. Updating the checkout alone is temporary: the next autopull will
+restore the approved commit. Keep the guard, dispatcher-only setting and existing
+schedules intact during a dashboard rollout.
+
+Read-only monitoring can run independently of that execution release. Export a
+reviewed commit with `git archive` into a private persistent release directory,
+install its pinned dependencies into that directory's `.venv`, and set the user
+service's `WorkingDirectory` and `ExecStart` to that release. Set
+`Environment=ORCH_ROOT=%h/agent-os` so observations still read the real runtime.
+Do not point a lasting service at a temporary development worktree. This starts
+only `orchestrator.dashboard.server`, not the coordinator or queue. A stale
+coordinator alert is expected until the execution release is actually enabled.
+
+The GitHub identity used by the runtime needs Project read/write access in
+addition to repository access. Reading issues and posting comments can work while
+Project updates fail. Inspect scopes with `gh auth status` under the runtime
+identity and check whether `GH_TOKEN` or `GITHUB_TOKEN` overrides the stored login;
+never print token values. For an OAuth login, the account owner can grant the
+missing access with `gh auth refresh --hostname github.com --scopes project`.
+That requires an account authorization, not a code retry or GitHub Actions job.
+Pending updates stay in the outbox and retry after access is restored.

docs/worklog/2026-09-16-delivery-operations.md

@@ -0,0 +1,33 @@
+# Delivery Operations Rollout, 2026-09-16
+
+Agent-OS PR358 adds persistent intent/program ownership, bounded execution,
+independent acceptance, durable notifications and the live Proof dashboard.
+Proof PRs 1-3 add operational observations separate from benchmark results and
+keep historical reconciliation out of live delivery metrics and activity charts.
+
+The deployment check found two operational gates, not worker failures:
+
+- The host's immutable approved execution pin still selects `dc4559703af49126d2403132601992fff12dd4b1`.
+  An initial fast-forward was reverted by the intentional autopull guard. Approval
+  to advance the pin was requested; it was not silently overwritten or bypassed.
+- The active GitHub OAuth login has repository access but not Project access.
+  No environment token override was present. GitHub Project projection remains
+  queued, rather than being reported as delivered.
+
+The private read-only dashboard is installed as a user service on localhost:8765
+from a separate immutable release directory with its own virtual environment.
+No new cron, public listener, execution-mode change or worker was added. Monitor
+health and execution health are separate: an inactive coordinator raises an alert.
+
+Issue #355 was independently reconciled against merged PR356 and its actual README
+diff. Its Telegram completion notice was acknowledged. It remains closed and is
+marked historical, not counted as a new task or a near-instant delivery. The live
+check exposed a historical-event chart inconsistency, corrected in Proof PR3 and
+covered by both Proof and Agent-OS HTTP regression tests.
+
+Validation of the foundation: 794 Agent-OS tests, 43 Proof tests, passing remote
+Agent-OS CI, desktop/mobile browser checks, safe rendering of hostile titles,
+outage/stale-snapshot behavior, and installed-wheel static assets. The follow-up
+adds two Proof regression cases and an Agent-OS HTTP integration case. No claims
+of real non-code publication, complete provider billing, OS sandboxing, or measured
+business impact are inferred from those checks. See [delivery limits](../delivery.md).

requirements.txt

@@ -1,4 +1,4 @@
 PyYAML>=6.0
 pytest>=8.0
 duckduckgo-search>=7.0
-proof @ git+https://github.com/kai-linux/proof.git@1daeef95bd820cf1f0e93d4daf6f3211cf092446
+proof @ git+https://github.com/kai-linux/proof.git@466558a3ac9410a93dd3824850c6424fa3fce5bc

tests/test_delivery_dashboard.py

@@ -1,4 +1,5 @@
 import http.client
+import json
 import sys
 from pathlib import Path
 from threading import Thread
@@ -67,6 +68,25 @@ def test_real_http_dashboard_and_live_snapshot(server):
     assert b"proof.operations.v1" in body
 
 
+def test_historical_reconciliation_is_not_new_delivery_activity(server, tmp_path):
+    store = DeliveryStore(store_path({"root_dir": str(tmp_path)}))
+    goal = store.upsert(
+        "historical-issue",
+        "Previously delivered",
+        "Historical request",
+        metadata={"historical_import": True},
+    )
+    store.record_evidence(goal["id"], 1, "human_acceptance", True, "test", {})
+    assert store.verify(goal["id"])
+    status, _, body = request(server, "/api/delivery")
+    assert status == 200
+    snapshot = json.loads(body)
+    assert snapshot["goals"][0]["verified"] is True
+    assert snapshot["metrics"]["historical_imports"] == 1
+    assert snapshot["metrics"]["verified_delivery"]["denominator"] == 0
+    assert snapshot["timeline"] == []
+
+
 def test_dns_rebinding_host_rejected(server):
     status, _, _ = request(server, "/api/delivery", host="evil.example")
     assert status == 403