fix: use documented aux model for update summaries

[Jordan-SkyLF]

Thinking Path

• The category-refinement portion of this PR shipped in v0.51.62 via stage-355.
• This branch is now rebased onto current origin/master so the PR carries only the remaining aux-model routing follow-up.
• The reviewer concern was that auxiliary.update_summary would be a WebUI-only magic key, not a documented Hermes Agent auxiliary task.

What Changed

• Update-summary generation now asks for the documented auxiliary.compression text-model slot first.
• The existing main-model fallback is preserved if auxiliary resolution or generation fails.
• Added a route comment explaining why summary generation maps to compression/summarization instead of inventing a new task name.
• Updated focused regression coverage to assert the route uses "compression" and does not reintroduce "update_summary" in api/routes.py.
• Updated the changelog with the standalone aux-routing follow-up.

Why It Matters

• Users can override summary generation through an existing documented auxiliary slot instead of guessing an undiscoverable WebUI-specific config key.
• The PR stays self-contained to hermes-webui and does not require a coordinated Hermes Agent registry/setup change.
• Existing installs without an auxiliary compression model still fall back to the configured main model.

Before / After

• No visual before/after evidence is needed for this follow-up: the diff only changes model-routing selection and regression coverage.
• The update-summary UI/category behavior and screenshots from the earlier scope have already shipped in v0.51.62.

Verification

• Rebased onto current origin/master at 18297f3aff71d11ac5951e064245ebd86e1c87aa; latest release checked: v0.51.62.
• python -m pytest tests/test_update_banner_fixes.py::TestUpdateSummaryRouteModelSelection -q → 1 passed in 2.39s
• python -m pytest tests/test_update_banner_fixes.py::TestUpdateSummaryRouteModelSelection tests/test_update_banner_fixes.py::TestWhatsNewSummaryToggle -q → 15 passed in 2.49s
• python -m pytest tests/test_update_banner_fixes.py -q → 60 passed in 3.74s
• python -m py_compile api/routes.py api/updates.py
• node --check static/ui.js && node --check static/boot.js && node --check static/panels.js
• git diff --check origin/master...HEAD

Risks / Follow-ups

• Generated summary quality still depends on the configured auxiliary.compression model, or the main model when fallback is used.
• If maintainers later want a dedicated update_summary aux task, that should be registered/documented in Hermes Agent rather than introduced only in WebUI.

Model Used

• OpenAI GPT-5.5 via Codex CLI, with local terminal/file tooling for GitHub PR refresh, rebase, tests, and verification.

[nesquena-hermes]

Heads-up: the earlier 3-PR portion of this PR (commit afbcc9a6 — "fix: wrap update banner on mobile" + 62eb703d — "fix: avoid duplicate update summary bullets") shipped in v0.51.61 as part of stage-354 (PR #2239, merged 2026-05-14 07:39Z).

You pushed a new commit c73d63e8 shortly after the merge — that one is not in master yet. It adds substantive new functionality (_split_summary_category + _unique_summary_bullets helpers + scrollable summary panel for longer summaries) that's separate from what shipped. Since the GitHub auto-close compares head SHA to merge SHA and they don't match, this PR stayed OPEN.

A few options:

1. Rebase against master and let this PR carry only the new c73d63e8 changes (the scrollable-summary + category-split helpers). The two earlier commits will fold into the rebase as already-applied.
2. Open a fresh PR for just the new functionality from c73d63e8 and close this one. Cleaner if the new work has its own thinking-path / before-after evidence.
3. Close this PR as-is if you intended c73d63e8 to be a separate concern that's still cooking.

Whichever you prefer — the shipped portion is good, no rework needed on what's already in master. Just need to figure out the right handle for the additional summary-category work.

Thanks again for the cleanup work — three improvements in one PR is a lot to land smoothly!

— maintainer

[nesquena-hermes]

Re-read the PR after the force-push — the two earlier commits (afbcc9a6 wrap-banner + 62eb703d dedup) are correctly out of the diff (they shipped via stage-354 / PR #2239), and what remains is the focused third improvement: categorized Notice: / Worth knowing: summary bullets + an auxiliary update_summary model routing layer + a scrollable summary panel. Net diff is now 158 +, 51 - across 5 substantive files. Notes from reading the actual code.

Categorized bullet parsing

api/updates.py:_split_summary_category() is the parser doing the work:

match = re.match(r'^\s*(?:[-*•]+|\d+[.)])?\s*(notice|what you(?:ll|\'ll| will) notice|user(?:s)? will notice|worth knowing|worth|note)\s*:\s*(.+)$', raw, re.I)

This is permissive — accepts Notice:, What you'll notice:, Worth:, Note:, Users will notice: — which is right for what LLMs realistically emit. The pairing with _categorized_summary_bullets_from_text() then partitions into notice/worth and dedups. The "fall back to the legacy single-bucket parser if zero notice bullets are recovered" path at _format_update_summary_sections() is the right safety net for when the model ignores the prefix instruction:

notice_items, worth_items = _categorized_summary_bullets_from_text(summary_text)
if not notice_items:
    notice_items = _summary_bullets_from_text(summary_text, fallback_items=_fallback_update_bullets(details))

The behavior change from the old code — which sliced the first 3 as notice and remainder as worth (bullets[:3] / bullets[3:]) — is meaningful: the LLM now controls the split via its prefix, and there's no longer a hard worth_items[:2] cap. The test test_update_summary_keeps_all_categorized_notice_and_worth_bullets pins exactly this (5 notice + 2 worth all preserved), so the cap removal is intentional and not a regression. Good.

Auxiliary update_summary routing — but this task isn't a known aux task

api/routes.py:5358-5370 adds:

from agent.auxiliary_client import get_text_auxiliary_client

aux_client, aux_model = get_text_auxiliary_client(
    "update_summary",
    main_runtime=main_runtime,
)
if aux_client is not None and aux_model:
    response = aux_client.chat.completions.create(
        model=aux_model,
        messages=messages,
    )
    return str(response.choices[0].message.content or "").strip()

This calls get_text_auxiliary_client("update_summary", ...) at ~/.hermes/hermes-agent/agent/auxiliary_client.py:3199-3221. That function resolves via _resolve_task_provider_model(task) at line 3800, which reads config["auxiliary"]["update_summary"] from hermes_cli.config.load_config().

Two things to verify:

1. update_summary is not in the documented list of known aux tasks. call_llm()'s docstring at auxiliary_client.py:4108-4111 enumerates "compression", "vision", "web_extract", "session_search", "skills_hub", "mcp", "title_generation". And hermes_cli/main.py:2046 lists the title_generation slug for the setup wizard. WebUI is effectively defining a new auxiliary task here without coordinating with the agent's known-task registry. That works at runtime — _resolve_task_provider_model() falls through to "auto" if no auxiliary.update_summary config block exists — but it means a user has to know to add an unlisted task name to their config.yaml to override the model. Worth either (a) documenting the new task name in the changelog/CONTRIBUTING and adding it to hermes_cli/main.py:2046 and hermes_cli/config.py:868-region defaults, or (b) re-using auxiliary.web_extract or another text-aux task with a route comment explaining why.

2. The fallback path includes building a full AIAgent instance just for one chat turn. The pre-fix code also built AIAgent; the only behavior change is the try-aux-first added on top. The try/except Exception around the aux call at api/routes.py:5358-5371 is broad but appropriate — anything unexpected (provider down, model 404, network error, missing config) falls through to the main-model AIAgent path. Good defensive shape.

Scrollable summary panel CSS contract

static/style.css adds one line that the test pins literally:

#updateSummaryPanel{max-height:min(34vh,260px);overflow:auto;overscroll-behavior:contain;scrollbar-gutter:stable;scrollbar-width:thin;scrollbar-color:var(--accent) transparent;}

The scrollbar-gutter:stable + overscroll-behavior:contain choice is correct for a panel inside an update banner — avoids layout-shift jitter when the scrollbar appears, and prevents the body from scrolling when the user reaches the bottom of the summary list. Worth keeping overscroll-behavior since some users will read summaries on small touchscreens where momentum bleed-through is jarring.

Verdict

The force-push correctly removed the already-merged hunks, and what remains is genuinely a follow-up (refines what stage-354 shipped). Strong recommend on item (1) above: either register update_summary as a known aux task in the agent's wizard/config defaults, or rename to reuse auxiliary.web_extract / auxiliary.compression. Otherwise users who want to override the model won't know the magic key exists. Everything else looks ready to land.

— maintainer

[Jordan-SkyLF]

Addressed the aux-routing note in 97b68f0: update summaries now reuse the documented auxiliary.compression text task instead of introducing a WebUI-only auxiliary.update_summary key.

Also refreshed the changelog, focused regression test, and PR body verification. CI is green on 3.11 / 3.12 / 3.13.

[nesquena-hermes]

Heads-up: the rebase + category-refinement portion of this PR (commit a291ffd) shipped in v0.51.62 as part of stage-355 (PR #2263, merged 2026-05-14 15:43Z).

You also pushed a new commit 97b68f0 ("fix: use documented aux model for update summaries") shortly after — that one is not in master and is a substantively new concern (which aux model the update summaries use, vs the category dedup work that just shipped).

Since the merge SHA didn't match the head SHA, GitHub kept this PR open. Stage-355 also includes 10 other PRs that landed since this branch's base, so the diff vs master is now ~30 files behind.

A few options:

1. Rebase this branch against current master and let the PR carry only the new 97b68f0 commit. The earlier two commits will be already-applied on rebase.
2. Close this PR and open a fresh one for the aux-model concern alone — likely cleaner since the new work has a different theme than the category dedup it was riding on.
3. Close this PR as-is if 97b68f0 was just a quick follow-up you'd rather table.

Whichever you pick — a291ffd's changes are good in master, no rework needed there. The aux-model question can land separately on a fresh branch off current master if that's easier.

Thanks for the iteration speed — 3 distinct improvements rolled across two release windows is a lot to land cleanly!

— maintainer

[Jordan-SkyLF]

Rebased this PR onto current master and narrowed it to the standalone aux-model follow-up in 339e625: update summaries now reuse documented auxiliary.compression with main-model fallback, and the already-shipped category/UI work is out of the diff.

Refreshed the title/body/changelog and reran verification. CI is green on 3.11 / 3.12 / 3.13.