test(agent): harden hostname anchoring against gethostname failures

[beplee]

What does this PR do?

Addresses @lorencato23 review hardening points on PR #102192 (stop synchronized sessions from misidentifying the active host).

Changes

1. **Fail closed when raises ** — the previous code silently skipped the hostname check when failed, which meant a local stored prompt could be treated as verified when it shouldn't be. Now we reject the stored prompt.

2. Fail closed when returns empty string — same conservative behavior: reject the stored prompt rather than silently treating it as verified.

3. Regression tests for failure modes — added tests in both and covering:

   • raising
   • returning empty string
   • Legacy local prompt without hostname (already existed, now also tested with failure)

4. Document hostname as host anchoring, not authentication — added comments explaining that hostname is a useful human-readable signal but not a globally unique gateway identity (two cloned VMs can share a hostname).

5. Same-hostname/different-profile test — makes the remaining limitation visible and prevents future overclaiming of the guarantee.

Type of Change

• Bug fix (non-breaking change that fixes an issue)

How to Test

scripts/run_tests.sh tests/run_agent/test_compression_persistence.py -k 'machine_hostname or legacy_local_prompt or gethostname or same_hostname'
scripts/run_tests.sh tests/agent/test_prompt_builder.py -k 'identifies_live_machine or omits_hostname'

Related

• Follow-up to fix(agent): stop synchronized sessions from misidentifying the active host #102192
• Review comments by @lorencato23

[Enough1122]

AI code review — automated review for reference; please use your judgment.

PR #102436 — harden hostname anchoring against gethostname failures

Verdict: Looks good. Thoughtful host-anchoring for DB-synced prompts with honest scoping (anchoring, not auth — stated in the comment) and fail-closed behavior on unreadable hostnames.

What the change does

• agent/prompt_builder.py:96-101 — local-backend prompts gain a Machine hostname: line (omitted when empty/raising).
• agent/conversation_loop.py:20-64 — _stored_prompt_matches_runtime rejects cross-machine reuse (hostname mismatch), one-time-rejects legacy hostname-less prompts to upgrade them, fails closed on OSError/empty; read window widened to idx-3..idx+4 since host fields now precede the anchor.
• contributors/emails/bepleecanton@gmail.com — salvage attribution for fix(agent): stop synchronized sessions from misidentifying the active host #102192.
• Tests: hostname emit/omit matrix, cross-machine stale, legacy rebuild, fail-closed cases, same-host-different-profile still valid.

Non-blocking

1. conversation_loop.py:63 — casefold() comparison is right for Windows (ATLAS vs atlas), but hostnames that legitimately differ in case cosmetics across dual-boot/synced profiles would force a rebuild per switch. Cheap (one rebuild) — noting only.
2. Hostname is a weak anchor (cloned VMs, DHCP defaults like localhost, containers sharing a hostname): two different machines with the same hostname and cwd will still share a prompt. The code comment acknowledges this and tracks a stable-identity follow-up — agree with that framing; just making sure the follow-up is actually filed/tracked somewhere.
3. conversation_loop.py:31 — widening the scan to 3 lines before the anchor slightly enlarges the spoof surface (a Host:-looking line in user-controlled prompt text just above the anchor could satisfy the read). First-occurrence anchoring on Hermes' own block mitigates it; and stored prompts are local-DB content, not remote input. Acceptable — noting the tradeoff of the window change.

Tests: The same-host/different-profile test pins the anchoring-vs-auth boundary explicitly. Good.