diff --git a/.agents/skills/changeset-versioning/SKILL.md b/.agents/skills/changeset-versioning/SKILL.md
index 4645a6a7..4f955185 100644
--- a/.agents/skills/changeset-versioning/SKILL.md
+++ b/.agents/skills/changeset-versioning/SKILL.md
@@ -111,10 +111,40 @@ Changelogs are auto-generated by `@changesets/changelog-github` and include:
 
 The changelog is written to `CHANGELOG.md` during the version step.
 
+## What a publish triggers downstream
+
+A real `@openrouter/agent` publish fans out to three repos via
+`repository_dispatch` (event type `openrouter-agent-published`). All of these run
+*after* npm already has the package, so none of them can fail the release:
+
+| Hop | Target | Effect |
+| --- | --- | --- |
+| B | `openrouter-web` | Bumps the pinned `@openrouter/agent` used by server tools |
+| C | `python-agent` | Opens a PR porting the release delta into the Python port |
+| C | `go-agent` | Opens a PR porting the release delta into the Go port |
+
+The port repos treat this repo as their **reference spec**: they run
+[Upstreamer](https://github.com/mountgram/upstreamer) against the release tag and
+gate the generated port on a mechanical verifier plus a behavioral parity eval
+before advancing their sync state. Their contracts live at
+`.upstreamer/upstreamer.md` in each repo.
+
+Hop C is dispatched with the release tag (e.g. `@openrouter/agent@0.8.0`), not a
+branch, so a port reproduces the exact published tree. If that tag is not on
+origin — the manual publish path pushes tags best-effort — the dispatch falls
+back to the publishing run's commit SHA, which is the same tree. If a dispatch
+fails it logs a warning rather than failing the release: recover by running the
+port repo's **Upstreamer Port** workflow manually with that ref, or wait for its
+weekly cron.
+
+Practical consequence: a breaking change to the `callModel` surface will produce
+port PRs in two other repos on release. If the ports need a coordinated change,
+sequence it the same way `@openrouter/sdk` coordination works.
+
 ## Configuration
 
 - `.changeset/config.json` — Changesets configuration
-- `.github/workflows/publish.yaml` — Release workflow (`push: main` + `workflow_dispatch`)
+- `.github/workflows/publish.yaml` — Release workflow (`push: main` + `workflow_dispatch`); also carries the HOP B/C downstream dispatches
 - `.github/workflows/ci.yaml` — PR validation (lint, typecheck, test)
 
 ## Common Commands
diff --git a/.github/workflows/publish.yaml b/.github/workflows/publish.yaml
index 92a35b68..3f5329f5 100644
--- a/.github/workflows/publish.yaml
+++ b/.github/workflows/publish.yaml
@@ -137,6 +137,20 @@ jobs:
         if: github.event_name == 'workflow_dispatch' && inputs.mode == 'publish' && !inputs.dry-run
         run: pnpm exec changeset publish --no-git-checks
 
+      # `changeset publish` only creates release tags locally; on the push path
+      # changesets/action pushes them, but nothing does here. Push them so the
+      # HOP C dispatch below names a ref the port repos can actually resolve.
+      #
+      # Best-effort: the packages are already on npm by the time this runs, so a
+      # rejected push (tag protection ruleset, or a re-run where the tag exists
+      # on origin at a different commit) must not turn a successful publish red
+      # — and must not skip the HOP B/C dispatch steps below. HOP C independently
+      # verifies the tag is on the remote and falls back to the commit SHA.
+      - name: Push release tags (manual publish)
+        if: github.event_name == 'workflow_dispatch' && inputs.mode == 'publish' && !inputs.dry-run
+        continue-on-error: true
+        run: git push origin --tags
+
       # `changeset publish` has no native --dry-run. Fall back to pnpm's
       # recursive dry-run, which simulates publishing every workspace package
       # rather than only the ones changesets would pick. Output set may be
@@ -195,3 +209,74 @@ jobs:
             -F "client_payload[version]=${{ steps.published.outputs.version }}" \
             -F "client_payload[source_run_url]=${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}"
           echo "Dispatched openrouter-agent-published (version ${{ steps.published.outputs.version }}) to openrouter-web"
+
+      # HOP C trigger: tell the Python and Go ports of @openrouter/agent that a
+      # new version shipped, so each opens a PR porting the delta. Those repos run
+      # Upstreamer against this repo as their reference spec (see their
+      # .upstreamer/upstreamer.md) and gate the result on a mechanical verifier
+      # plus a parity eval before advancing their sync state.
+      #
+      # Deliberately release-triggered rather than a cron over this repo's main:
+      # ports track published versions, so the delta they see always lands on a
+      # release boundary instead of a mid-flight commit.
+      #
+      # `ref` is the release tag changesets created, so a port reproduces the exact
+      # published tree rather than whatever main has drifted to since. If that tag
+      # never reached origin, the step falls back to this run's commit SHA — same
+      # tree, still resolvable.
+      #
+      # This step must not fail the release: the packages are already on npm by
+      # now, so a red job here would misreport a successful publish. A failed
+      # dispatch is a warning — re-run it from the port repo's own
+      # "Upstreamer Port" workflow (workflow_dispatch), or let its weekly cron
+      # pick the change up.
+      #
+      # `!cancelled()` keeps this independent of the HOP B step above: a failed
+      # monorepo dispatch should not also stop the ports from being told.
+      - name: Dispatch port repos
+        if: ${{ !cancelled() && !inputs.dry-run && steps.published.outputs.version != '' }}
+        continue-on-error: true
+        env:
+          # Same cross-repo PAT as HOP B; additionally needs contents:write on
+          # OpenRouterTeam/python-agent and OpenRouterTeam/go-agent.
+          GH_TOKEN: ${{ secrets.GH_TOKEN }}
+          VERSION: ${{ steps.published.outputs.version }}
+          SOURCE_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
+        run: |
+          set -euo pipefail
+          TAG="@openrouter/agent@${VERSION}"
+          FAILED=""
+
+          # The tag is only usable as a ref if it actually reached origin. On the
+          # push path changesets/action pushes it; on the manual publish path the
+          # push is best-effort and may have been rejected or skipped. Dispatching
+          # an unresolvable ref would make the ports fail on checkout rather than
+          # degrade, so fall back to this run's commit SHA — which points at the
+          # same published tree.
+          if git ls-remote --exit-code --tags origin "refs/tags/${TAG}" >/dev/null 2>&1; then
+            REF="$TAG"
+          else
+            REF="${{ github.sha }}"
+            echo "::warning::${TAG} is not on origin; dispatching ref=${REF} instead."
+          fi
+
+          for REPO in python-agent go-agent; do
+            # -f (--raw-field) everywhere: -F treats values starting with `@`
+            # (like TAG) as filenames to read, which would error out before the
+            # request is even sent.
+            if gh api "repos/OpenRouterTeam/${REPO}/dispatches" \
+                 -f event_type=openrouter-agent-published \
+                 -f "client_payload[version]=${VERSION}" \
+                 -f "client_payload[ref]=${REF}" \
+                 -f "client_payload[source_run_url]=${SOURCE_RUN_URL}"; then
+              echo "Dispatched openrouter-agent-published (ref ${REF}) to ${REPO}"
+            else
+              echo "::warning::Failed to dispatch to OpenRouterTeam/${REPO}. Trigger its Upstreamer Port workflow manually with ref=${REF}, or wait for its weekly cron."
+              FAILED="$FAILED ${REPO}"
+            fi
+          done
+
+          # Report, but never fail the release — npm is already updated.
+          if [ -n "$FAILED" ]; then
+            echo "Port dispatch incomplete:$FAILED"
+          fi