hermes: add read-only Gitea PR comment filter, break the reply loop

The gitea-events subscription woke the agent on every delivery. That is an
unbounded loop as soon as she is given a prompt that tells her to answer on
the PR: her answer is itself a pull_request_comment, which wakes her again.

Adds a Hermes route script that drops the deliveries that must never reach an
LLM call: luna's own comments (the loop guard), "deleted" actions (the body
is still in the payload, so acting on one means acting on a request that was
explicitly withdrawn), non-pull-request comments, empty bodies, and edits
that did not actually change the body — a label or attachment change fires
"edited" too. Everything else passes through unchanged.

Mounted READ-ONLY from the nix store rather than written into hermesHome.
Hermes resolves route scripts under ~/.hermes/scripts, which here is inside
/opt/data — HERMES_WRITE_SAFE_ROOT — so a filter written there would be a
loop guard sitting in the writable root of the agent it constrains. Deleting
it fails closed (Hermes treats a missing script as "ignore"), but rewriting
it to always-allow would silently restore the loop. Read-only from the store
makes that impossible and keeps the guard in git.

The script also normalises changes.body.from to always exist. Gitea omits
`changes` entirely on created events, and Hermes replaces the prompt payload
with whatever JSON the script emits, so guaranteeing the key here means a
prompt referencing {changes.body.from} renders empty instead of leaving an
unfilled placeholder.

Note the stdout contract (gateway/platforms/webhook.py): only exactly
"[SILENT]", empty output, or a nonzero exit drop a delivery. Any OTHER text
on stdout lets it through and is attached as script_output — so a stray
debug print would silently defeat the filter. All diagnostics go to stderr,
and gitea-pr-comment-filter-test.py asserts that discipline along with each
drop rule (25 cases). Run it after any edit: the fail-closed behaviour means
a syntax error produces silence, not an error.

--events is still unset; event selection remains runtime-tunable policy.
The filter covers only what must not be.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S94o42aQ8VkBmEWvDem5xa

hosts/mars/gitea-pr-comment-filter-test.py

@@ -0,0 +1,101 @@
+"""Contract test for gitea-pr-comment-filter.py.
+
+Hermes treats "[SILENT]"/empty/nonzero-exit as ignore, a JSON object as a
+payload replacement, and ANY OTHER stdout text as allow-with-script_output.
+So each case asserts on the exact stdout discipline, not just the decision.
+"""
+import json, subprocess, sys, pathlib
+
+SCRIPT = str(pathlib.Path(__file__).with_name("gitea-pr-comment-filter.py"))
+
+def payload(action="created", author="darman", body="please fix the typo",
+            previous=None, is_pull=True, cid=42, number=7):
+    p = {"action": action, "is_pull": is_pull,
+         "comment": {"id": cid, "body": body, "user": {"login": author},
+                     "html_url": "https://git.mgaction.town/darman/homelab/pulls/7#issuecomment-42"},
+         "issue": {"number": number, "title": "some PR"},
+         "repository": {"full_name": "darman/homelab"},
+         "sender": {"login": author}}
+    if previous is not None:
+        p["changes"] = {"body": {"from": previous}}
+    return p
+
+def run(p):
+    r = subprocess.run([sys.executable, SCRIPT], input=json.dumps(p),
+                       capture_output=True, text=True)
+    return r.returncode, r.stdout, r.stderr
+
+def classify(rc, out):
+    """Replicate Hermes's own interpretation of the script result."""
+    if rc != 0 or out.strip() == "" or out.strip() == "[SILENT]":
+        return "IGNORED"
+    try:
+        v = json.loads(out)
+        return "ALLOWED" if isinstance(v, dict) else "ALLOWED(script_output)"
+    except ValueError:
+        return "ALLOWED(script_output)"
+
+fails = []
+def check(name, p, expect):
+    rc, out, err = run(p)
+    got = classify(rc, out)
+    ok = got == expect
+    print(f"{'PASS' if ok else 'FAIL'}  {name:<52} {got}")
+    if not ok:
+        fails.append(name); print(f"      expected {expect}; stdout={out!r} stderr={err.strip()!r}")
+    return out
+
+# --- the loop guard, the whole reason this exists ---
+check("luna's own comment is dropped (LOOP GUARD)", payload(author="luna"), "IGNORED")
+check("luna in different case is dropped",          payload(author="LUNA"), "IGNORED")
+
+# --- action handling ---
+check("created by human is allowed",                payload(), "ALLOWED")
+check("deleted is dropped",                         payload(action="deleted"), "IGNORED")
+check("edited with changed body is allowed",
+      payload(action="edited", body="new text", previous="old text"), "ALLOWED")
+check("edited with unchanged body is dropped",
+      payload(action="edited", body="same", previous="same"), "IGNORED")
+check("unknown action is dropped",                  payload(action="reopened"), "IGNORED")
+
+# --- misc guards ---
+check("issue comment (is_pull=false) is dropped",   payload(is_pull=False), "IGNORED")
+check("empty body is dropped",                      payload(body="   "), "IGNORED")
+check("missing comment object is dropped",          {"action": "created"}, "IGNORED")
+check("malformed payload is dropped",               "not-a-dict", "IGNORED")
+
+# --- normalisation: the prompt's {changes.body.from} must always resolve ---
+out = check("created event still allowed", payload(), "ALLOWED")
+norm = json.loads(out)
+c1 = norm.get("changes", {}).get("body", {}).get("from")
+print(f"{'PASS' if c1 == '' else 'FAIL'}  {'created: changes.body.from normalised to empty':<52} {c1!r}")
+if c1 != "": fails.append("normalise-created")
+
+out = check("edited event still allowed", payload(action="edited", body="new", previous="old"), "ALLOWED")
+c2 = json.loads(out).get("changes", {}).get("body", {}).get("from")
+print(f"{'PASS' if c2 == 'old' else 'FAIL'}  {'edited: changes.body.from preserved':<52} {c2!r}")
+if c2 != "old": fails.append("normalise-edited")
+
+# --- payload passthrough: prompt paths must survive the transform ---
+norm = json.loads(run(payload())[1])
+for path in [("comment","id"), ("comment","body"), ("comment","user","login"),
+             ("comment","html_url"), ("issue","number"), ("issue","title"),
+             ("repository","full_name"), ("action",)]:
+    cur, ok = norm, True
+    for k in path:
+        if isinstance(cur, dict) and k in cur: cur = cur[k]
+        else: ok = False; break
+    label = ".".join(path)
+    print(f"{'PASS' if ok else 'FAIL'}  {'prompt path survives: {' + label + '}':<52} {cur if ok else 'MISSING'}")
+    if not ok: fails.append(f"path-{label}")
+
+# --- stdout discipline: an ignore must emit EXACTLY [SILENT] ---
+rc, out, err = run(payload(author="luna"))
+print(f"{'PASS' if out == chr(91)+'SILENT'+chr(93)+chr(10) else 'FAIL'}  {'ignore emits exactly [SILENT] on stdout':<52} {out!r}")
+if out != "[SILENT]\n": fails.append("silent-exact")
+print(f"{'PASS' if err.strip() else 'FAIL'}  {'ignore explains itself on stderr':<52} {err.strip()[:40]!r}")
+if not err.strip(): fails.append("stderr-reason")
+
+print()
+print("ALL PASSED" if not fails else "FAILURES: " + ", ".join(fails))
+sys.exit(1 if fails else 0)

hosts/mars/gitea-pr-comment-filter.py

@@ -0,0 +1,107 @@
+#!/usr/bin/env python3
+"""Hermes webhook filter for Gitea pull_request_comment deliveries.
+
+Contract (gateway/platforms/webhook.py): the payload arrives on stdin as JSON.
+STDOUT IS A PROTOCOL CHANNEL, not a log:
+
+  - exactly "[SILENT]"  -> delivery ignored, no agent run, no tokens spent
+  - a JSON object       -> REPLACES the payload used by the prompt template
+  - any other text      -> delivery is ALLOWED THROUGH and the text is attached
+                           as script_output
+
+That last case is why every diagnostic here goes to stderr. A stray print()
+would not drop an event, it would let one through.
+
+Empty stdout, a nonzero exit, a missing script, or a timeout also count as
+"ignored", so this script fails CLOSED: if it breaks, nothing reaches the
+agent rather than everything. That is the right direction for a loop guard,
+but it does mean a syntax error silently disables the whole integration --
+run the test file next to this one after editing.
+
+Two jobs:
+
+1. Filter. Drop the deliveries that must never wake the agent -- above all
+   luna's own comments, which would otherwise loop forever: the prompt tells
+   her to reply on the PR, and her reply is itself a pull_request_comment.
+2. Normalise. Guarantee changes.body.from always exists, so the prompt's
+   {changes.body.from} renders as empty rather than as an unfilled
+   placeholder on "created" events, where Gitea omits `changes` entirely.
+"""
+import json
+import sys
+
+# Comment authors whose comments must never wake the agent. luna is the agent
+# herself (loop guard). Add "ci-bot" here if CI ever starts commenting on PRs
+# and you do not want her reacting to build output.
+IGNORED_AUTHORS = {"luna"}
+
+# Gitea's HookIssueCommentAction values are created / edited / deleted.
+# "deleted" is dropped: the payload still carries the comment body, so letting
+# it through would have her act on a request that was explicitly withdrawn.
+ALLOWED_ACTIONS = {"created", "edited"}
+
+
+def ignore(reason: str) -> None:
+    print(f"gitea-pr-comment-filter: ignoring delivery: {reason}", file=sys.stderr)
+    print("[SILENT]")
+    raise SystemExit(0)
+
+
+def main() -> None:
+    try:
+        payload = json.loads(sys.stdin.read())
+    except (ValueError, OSError) as exc:
+        ignore(f"unparseable payload: {exc}")
+
+    if not isinstance(payload, dict):
+        ignore("payload is not a JSON object")
+
+    comment = payload.get("comment") or {}
+    issue = payload.get("issue") or {}
+    action = (payload.get("action") or "").strip().lower()
+    author = ((comment.get("user") or {}).get("login") or "").strip()
+
+    if action not in ALLOWED_ACTIONS:
+        ignore(f"action={action or '<missing>'}")
+
+    if author.lower() in IGNORED_AUTHORS:
+        ignore(f"author={author} is the agent itself (loop guard)")
+
+    # Belt and braces: the route already filters to pull_request_comment, but
+    # if that filter is ever loosened this keeps issue comments out. Only
+    # enforced when the key is actually present.
+    if "is_pull" in payload and not payload.get("is_pull"):
+        ignore("not a pull request comment (is_pull=false)")
+
+    body = (comment.get("body") or "").strip()
+    if not body:
+        ignore("empty comment body")
+
+    # Gitea omits `changes` on created events and populates changes.body.from
+    # with the pre-edit text on edits. Normalise it to a plain string so the
+    # prompt template always resolves, and drop no-op edits (a label or
+    # attachment change can fire "edited" without touching the body).
+    changes = payload.get("changes") or {}
+    previous = ((changes.get("body") or {}).get("from") or "") if isinstance(changes, dict) else ""
+    if action == "edited":
+        if previous.strip() == body:
+            ignore("edited but comment body is unchanged")
+        if not previous.strip():
+            print(
+                "gitea-pr-comment-filter: edited delivery carries no previous body; "
+                "passing through so the agent can reconcile from the PR thread",
+                file=sys.stderr,
+            )
+
+    payload["changes"] = {"body": {"from": previous}}
+
+    print(
+        "gitea-pr-comment-filter: allowing comment id=%s action=%s author=%s pr=%s"
+        % (comment.get("id"), action, author, issue.get("number")),
+        file=sys.stderr,
+    )
+    json.dump(payload, sys.stdout)
+
+
+if __name__ == "__main__":
+    main()

hosts/mars/hermes-agent.nix

@@ -90,6 +90,19 @@ let
   # is hers to make, anywhere inside HERMES_WRITE_SAFE_ROOT=/opt/data.
   giteaHost = "git.mgaction.town";
 
+  # luna's webhook filter, mounted READ-ONLY below. It lives in the nix store
+  # rather than being written into hermesHome because hermesHome IS
+  # HERMES_WRITE_SAFE_ROOT: a filter dropped there is a loop guard sitting
+  # inside the writable root of the agent it constrains, and she could edit
+  # it back out. Deleting it would fail closed (Hermes treats a missing
+  # script as "ignore"), but rewriting it to always-allow would silently
+  # restore the reply loop. Read-only from the store makes that impossible
+  # and keeps the guard versioned in git — same reasoning as the git/tea
+  # binaries mounted below.
+  prCommentFilter = pkgs.writeText "gitea-pr-comment-filter.py" (
+    builtins.readFile ./gitea-pr-comment-filter.py
+  );
+
   # hermesHome as the CONTAINER sees it (the bind mount below). Anything
   # written host-side that gets READ back inside the container must use this
   # prefix, not hermesHome — see the credential.helper below, which was
@@ -147,6 +160,11 @@ in
     script = ''
       mkdir -p ${hermesHome}
       mkdir -p ${dropboxDir}
+      # Parent for the read-only filter bind-mounted at
+      # /opt/data/scripts/gitea-pr-comment-filter.py. /opt/data is itself a
+      # bind mount of hermesHome, so this directory has to exist HOST-side
+      # before podman can mount a file inside it.
+      mkdir -p ${hermesHome}/scripts
 
       export HOME=${hermesHome}
       export GIT_CONFIG_GLOBAL=${hermesHome}/.gitconfig
@@ -187,6 +205,12 @@ in
       chown ${hermesUid}:${hermesGid} \
         ${hermesHome}/.gitconfig \
         ${hermesHome}/.git-credentials
+      # Same cont-init caveat as the files above: the directory is created
+      # here as root, and Hermes reads its scripts as uid ${hermesUid}. The
+      # mounted filter itself is world-readable 0444 from the store, so only
+      # the directory needs handing over.
+      chown ${hermesUid}:${hermesGid} ${hermesHome}/scripts
+
       if [ -d ${hermesHome}/.config ]; then
         chown ${hermesUid}:${hermesGid} ${hermesHome}/.config
       fi
@@ -216,6 +240,10 @@ in
       # is read-only content-addressed build output, not a source of
       # secrets, so mounting the whole thing read-only costs nothing beyond
       # the two specific binaries actually being reachable.
+      # Read-only: see prCommentFilter above. Hermes resolves route scripts
+      # under ~/.hermes/scripts, which is /opt/data/scripts in here.
+      "${prCommentFilter}:/opt/data/scripts/gitea-pr-comment-filter.py:ro"
+
       "/nix/store:/nix/store:ro"
       "${pkgs.git}/bin/git:/usr/local/bin/git:ro"
       "${pkgs.tea}/bin/tea:/usr/local/bin/tea:ro"

services/dev/gitea-hermes-webhook-relay.nix

@@ -75,6 +75,14 @@ in
   # That only works because the relay supplies X-GitHub-Event — see the
   # header comment above.
   #
+  # --script does the selection that MUST NOT be retunable at runtime.
+  # hosts/mars/gitea-pr-comment-filter.py drops luna's own comments before
+  # any LLM call, which is what stops the reply loop: the prompt tells her to
+  # answer on the PR, and her answer is itself a pull_request_comment. It is
+  # bind-mounted read-only from the nix store (see hosts/mars/hermes-agent.nix)
+  # so the agent cannot edit its own guard out. Hermes resolves the name
+  # relative to ~/.hermes/scripts, hence the bare filename here.
+  #
   # The secret is read from the CONTAINER's environment ($GITEA_HERMES_
   # WEBHOOK_SECRET, injected via sops.templates."hermes-agent.env"), which is
   # why hosts/mars/secrets.nix restarts podman-hermes-agent BEFORE this unit
@@ -108,6 +116,7 @@ in
         hermes webhook subscribe gitea-events \
           --secret "$GITEA_HERMES_WEBHOOK_SECRET" \
           --description "Forward authenticated Gitea events to L.U.N.A." \
+          --script gitea-pr-comment-filter.py \
           --deliver telegram --deliver-chat-id "15151223"
       '
     '';