Session page: the "limited setup" note is alarming for something expected #1187

New issue

Closed

opened 2026-10-14 03:46:49 +00:00 by claude-bot · 0 comments

claude-bot commented

2026-10-14 03:46:49 +00:00

Collaborator

Today (ReadView.jsx, session_service.summary_limits, titles from capability_service): a lock icon, "Summarised with a limited setup", then the titles of every capability gap recorded on the latest summarisation run, joined: "A typical session is summarised in 3 passes, not one. An over-long prompt loses its beginning rather than failing. An admin can see the detail under Bot settings." Shown to every member.

Session 14's run recorded two gaps: llm_window_needs_chunking (40,960-token window, 3 passes; the run actually used 4 windows, 76/86 beats validated, summary complete) and llm_front_truncates (severity low: the endpoint reports prompt tokens, so truncation would be detected and logged; none was logged).

Change

Say it calmly and say it is expected. Replace the heading and the joined titles on the session page with one short line in the page's normal voice (no lock icon: it reads as "restricted"), e.g. heading "Summarised in parts", body "Expected with this AI setup. An admin can see why under Bot settings." Final wording on the owner's Wording Review page.
Only what applied to this summary. The session page shows a gap only when it affected this summary: chunking when the run really used more than one window. A latent risk that did not happen (front-truncation when nothing was detected, the provider guessed from the URL, and so on) stays on Admin → Bot settings, not on every session.
A real problem still says so. If this run detected truncation (or another gap actually bit), the page keeps a plain warning for that case, with what to do. Find whether the run records a detected truncation; if it does not, record it.
Admin → Bot settings keeps the full detail unchanged.

Not in scope: changing the context window or the provider; who sees the note (unchanged).

Owner, 13 Oct 2026, reviewing Session 14 on prod: the note on the session page reads as an error, and it is not one. "It's alarming for GMs, I think it should be shortened and made clear that this is expected." Today (`ReadView.jsx`, `session_service.summary_limits`, titles from `capability_service`): a lock icon, "Summarised with a limited setup", then the titles of every capability gap recorded on the latest summarisation run, joined: "A typical session is summarised in 3 passes, not one. An over-long prompt loses its beginning rather than failing. An admin can see the detail under Bot settings." Shown to every member. Session 14's run recorded two gaps: `llm_window_needs_chunking` (40,960-token window, 3 passes; the run actually used 4 windows, 76/86 beats validated, summary complete) and `llm_front_truncates` (severity low: the endpoint reports prompt tokens, so truncation would be detected and logged; none was logged). ## Change 1. **Say it calmly and say it is expected.** Replace the heading and the joined titles on the session page with one short line in the page's normal voice (no lock icon: it reads as "restricted"), e.g. heading "Summarised in parts", body "Expected with this AI setup. An admin can see why under Bot settings." Final wording on the owner's Wording Review page. 2. **Only what applied to this summary.** The session page shows a gap only when it affected this summary: chunking when the run really used more than one window. A latent risk that did not happen (front-truncation when nothing was detected, the provider guessed from the URL, and so on) stays on Admin → Bot settings, not on every session. 3. **A real problem still says so.** If this run detected truncation (or another gap actually bit), the page keeps a plain warning for that case, with what to do. Find whether the run records a detected truncation; if it does not, record it. 4. Admin → Bot settings keeps the full detail unchanged. Not in scope: changing the context window or the provider; who sees the note (unchanged).