…esearch#5689)

Salvages the core fix from PR NousResearch#5673 (egerev) onto current main.

The chatgpt.com/backend-api/codex endpoint streams valid output items
via response.output_item.done events, but the OpenAI SDK's
get_final_response() returns an empty output list. This caused every
Codex response to be rejected as invalid.

Fix: collect output_item.done events during streaming and backfill
response.output when get_final_response() returns empty. Falls back
to synthesizing from text deltas when no done events were received.

Also moves the synthesis logic from the validation loop (too late, from
NousResearch#5681) into _run_codex_stream() (before the response leaves the
streaming function), and simplifies the validation to just log
diagnostics since recovery now happens upstream.

Co-authored-by: Egor <egerev@users.noreply.github.com>

…esearch#5689)

Salvages the core fix from PR NousResearch#5673 (egerev) onto current main.

The chatgpt.com/backend-api/codex endpoint streams valid output items
via response.output_item.done events, but the OpenAI SDK's
get_final_response() returns an empty output list. This caused every
Codex response to be rejected as invalid.

Fix: collect output_item.done events during streaming and backfill
response.output when get_final_response() returns empty. Falls back
to synthesizing from text deltas when no done events were received.

Also moves the synthesis logic from the validation loop (too late, from
NousResearch#5681) into _run_codex_stream() (before the response leaves the
streaming function), and simplifies the validation to just log
diagnostics since recovery now happens upstream.

Co-authored-by: Egor <egerev@users.noreply.github.com>

…tomizations

Brings in 30+ commits of upstream Hermes changes (including the codex
output[] backfill fix from NousResearch#5689 / commit 0e336b0) AND closes the
loop on the branch-safe update flow that left this branch silently
behind upstream for 11+ days.

Symptom that triggered this work: every gateway turn was failing with
"Invalid API response (attempt 1/3): response.output is empty" on
gpt-5.4 via openai-codex.  The fix landed upstream 2026-04-06; without
this merge it never reached the customizations branch because
``hermes update`` only updated origin/main and switched back to
blaize-customizations without merging main into it.

== Conflict resolution highlights ==

run_agent.py: kept main's _touch_activity(desc) API + main's codex
backfill in _run_codex_stream; preserved HEAD's _reasoning_deltas_fired
reset and the public touch_activity() wrapper for delegate_tool /
gateway/run.py callers (now delegates to _touch_activity for
description sync).  Guarded the cached-agent touch_activity() reset
with hasattr() so test mocks don't break.

hermes_cli/config.py: bumped _config_version 18 → 19 and added HEAD's
progress-aware-timeout migration as a new 18 → 19 step (idempotent
via "if 'timeout' not in config" guard, so users on either v12 or v18
land in a correct state).

hermes_cli/main.py: kept HEAD's branch-safety guards
(should_restore_original_branch, should_auto_restart_gateway) and
swapped in main's improved multi-profile gateway restart logic
(supports_systemd_services, find_gateway_pids, retry-on-die).

gateway/run.py: kept HEAD's per-channel overrides + two-threshold
progress-aware timeout monitor (CLAUDE.md documents this as the
intentional design); added main's _notify_long_running periodic
"Still working" notifications and main's service_tier /
request_overrides plumbing on cached agent reuse.

tools/cronjob_tools.py: restored both 'reason' (HEAD) and 'script'
(main) schema entries that the auto-merger had collided.  Restored
both timeout_seconds (HEAD) and script (main) function args.

hermes_cli/commands.py: kept HEAD's priority_skills reordering for
Telegram menus while taking main's _collect_gateway_skill_entries
refactor (priority_skills now applied as a post-processing step on
the helper's output).  Kept both new CommandDefs (restart-gateway
from HEAD, debug from main).

cron/scheduler.py: took main's inactivity-based timeout structure but
restored HEAD's per-job timeout_seconds lookup
(job.get("timeout_seconds")) so per-job overrides still work.

gateway/platforms/telegram.py: kept HEAD's _menu_config_mtime AND
main's _model_picker_state, _approval_state, plus all of main's
new helper methods.

== Update flow fix (prevents future drift) ==

hermes_cli/main.py cmd_update: after restoring the working tree to
the customizations branch, run ``git merge --no-edit origin/main``
into the customizations branch so it actually catches up to main.
On clean merge, log success and proceed.  On conflict, ``git merge
--abort`` so the working tree stays clean, surface the conflict to
the user, and force should_auto_restart_gateway = False.  Applies
in both the up-to-date-already path and the new-commits-pulled path.

Adds two regression tests in TestUpdateMergesMainIntoCustomizations:
  - test_clean_merge_runs_after_branch_restore: verifies the merge
    is invoked when on the customizations branch
  - test_conflict_aborts_merge_and_blocks_auto_restart: verifies
    merge --abort runs on conflict and launchd_restart is skipped

== Test fixes ==

tests/hermes_cli/test_config.py: bumped expected config version 18 → 19
tests/tools/test_browser_camofox_state.py: same bump

== Pre-existing upstream test failures (unrelated to this merge) ==

Verified failing on clean origin/main:
  - test_wsl_with_systemd: macOS lacks systemctl
  - test_concurrent_inserts_settle_at_cap: ~70s slow concurrent test
  - test_file_staleness::test_warning_when_file_modified_externally
  - test_file_staleness::test_patch_warns_on_stale_file
    (macOS treats /var/folders as a sensitive system path)
  - test_transcription::test_explicit_local_no_cloud_fallback
  - test_transcription::test_local_nothing_available

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…esearch#5689)

Salvages the core fix from PR NousResearch#5673 (egerev) onto current main.

The chatgpt.com/backend-api/codex endpoint streams valid output items
via response.output_item.done events, but the OpenAI SDK's
get_final_response() returns an empty output list. This caused every
Codex response to be rejected as invalid.

Fix: collect output_item.done events during streaming and backfill
response.output when get_final_response() returns empty. Falls back
to synthesizing from text deltas when no done events were received.

Also moves the synthesis logic from the validation loop (too late, from
NousResearch#5681) into _run_codex_stream() (before the response leaves the
streaming function), and simplifies the validation to just log
diagnostics since recovery now happens upstream.

Co-authored-by: Egor <egerev@users.noreply.github.com>

…esearch#5689)

Salvages the core fix from PR NousResearch#5673 (egerev) onto current main.

The chatgpt.com/backend-api/codex endpoint streams valid output items
via response.output_item.done events, but the OpenAI SDK's
get_final_response() returns an empty output list. This caused every
Codex response to be rejected as invalid.

Fix: collect output_item.done events during streaming and backfill
response.output when get_final_response() returns empty. Falls back
to synthesizing from text deltas when no done events were received.

Also moves the synthesis logic from the validation loop (too late, from
NousResearch#5681) into _run_codex_stream() (before the response leaves the
streaming function), and simplifies the validation to just log
diagnostics since recovery now happens upstream.

Co-authored-by: Egor <egerev@users.noreply.github.com>

…esearch#5689)

Salvages the core fix from PR NousResearch#5673 (egerev) onto current main.

The chatgpt.com/backend-api/codex endpoint streams valid output items
via response.output_item.done events, but the OpenAI SDK's
get_final_response() returns an empty output list. This caused every
Codex response to be rejected as invalid.

Fix: collect output_item.done events during streaming and backfill
response.output when get_final_response() returns empty. Falls back
to synthesizing from text deltas when no done events were received.

Also moves the synthesis logic from the validation loop (too late, from
NousResearch#5681) into _run_codex_stream() (before the response leaves the
streaming function), and simplifies the validation to just log
diagnostics since recovery now happens upstream.

Co-authored-by: Egor <egerev@users.noreply.github.com>

…esearch#5689)

Salvages the core fix from PR NousResearch#5673 (egerev) onto current main.

The chatgpt.com/backend-api/codex endpoint streams valid output items
via response.output_item.done events, but the OpenAI SDK's
get_final_response() returns an empty output list. This caused every
Codex response to be rejected as invalid.

Fix: collect output_item.done events during streaming and backfill
response.output when get_final_response() returns empty. Falls back
to synthesizing from text deltas when no done events were received.

Also moves the synthesis logic from the validation loop (too late, from
NousResearch#5681) into _run_codex_stream() (before the response leaves the
streaming function), and simplifies the validation to just log
diagnostics since recovery now happens upstream.

Co-authored-by: Egor <egerev@users.noreply.github.com>