feat(webapp): facts the session may have changed, suggested by the local model for the GM to use (#1280) #1325

Merged

claude-bot merged 11 commits from feat/1280-fact-suggestions into main 2026-11-03 16:27:20 +00:00

Conversation 2 Commits 11 Files changed 36 +4238 -15

claude-bot commented 2026-11-03 15:01:05 +00:00

Collaborator

Copy link

Closes #1280 (D323–D326; the drawing is lane 6j, also in this PR).

What it does

• After a session is summarised, the worker asks the campaign's LLM one schema-constrained question per page that appeared: did its Status, Allegiances, Ruler, Leader or Owner change?
  • Status is Alive, Dead, Missing or Unknown, on NPC, PC, Creature and Deity pages. Allegiances is on NPC pages, Ruler on Settlement and Country, Leader on Organization, and Owner on Business.
  • The question covers only the transcript lines that mention the page, plus one line either side.
• Every change must quote the transcript. The quote is matched after normalising, within one line or across two adjacent lines. The speaker and the time come from the matched line, never from the model. A change is dropped and counted when its quote isn't found, its value is unchanged, its fact isn't allowed for the type, or its Status isn't one of the four. Counts go on the run (summarisation_runs.fact_suggestions).
• Suggestions are GMs-only until used. "Facts that may have changed" in Suggested wiki updates offers Use it, Change it…, Not this one, and a receipt with Undo. The page editor shows a waiting suggestion under its fact. A used fact keeps its own "Who sees it"; a new one is for everyone at the table (D324). It's written through the normal edit path, so history records it.
• The queueing hook comes after summarisation, in a try, so it can't fail or slow that path. Without an LLM configured, a session is skipped.

Privacy (three security-review findings, all fixed with tests)

• Retention: when a transcript is destroyed, its pending suggestions go, and decided ones lose their quote and speaker.
• Trash: while a transcript is binned, no read shows its suggestions.
• Erasure: each suggestion records the seq of its quoted line or lines, so erasing a member's lines removes the suggestions quoting them, however the scrub matched the speaker. A mixed erasure is tested.
• Reprocess: replacing the lines renumbers them, so a reprocess forgets the session's suggestions first, and the next run suggests again.
• Deleting a session, page or campaign cascades.

Server: migration f1280a0b1c2d (revises e1279a0b1c2d) adds fact_suggestions and summarisation_runs.fact_suggestions; it has a real downgrade, and Alembic has a single head. New routes: GET /api/campaigns/{id}/wiki/fact-suggestions, plus …/{sid}/use, /undo and /dismiss, all for GMs only. The article read (GM only) and the GM session read carry the waiting suggestions, and the attention count includes them. The bot is unaffected.

After the deploy (D325, on the owner's say): run this once over past sessions, newest first. It's safe to repeat.

celery -A app.tasks.celery_app call app.tasks.fact_tasks.suggest_fact_changes_backfill

Checked

• Backend, on the dev host: every migration test, tests/wiki, tests/export, attention, transcript bin, member erasure, enqueue-off-loop, sessions, audio tasks and retention gave 3,056 passed and 0 failed. ruff format and check are clean; the version and changelog checks are OK.
• Frontend: src/pages/wiki and src/wiki, 43 files and 625 tests passed. eslint has 0 errors.
• Not yet: the full backend suite (CI) and the render check, which is starting.

New wording goes on the Wording Review page before merge.

🤖 Generated with Claude Code

Closes #1280 (D323–D326; the drawing is lane 6j, also in this PR). **What it does** - After a session is summarised, the worker asks the campaign's LLM one schema-constrained question per page that appeared: did its Status, Allegiances, Ruler, Leader or Owner change? - Status is Alive, Dead, Missing or Unknown, on NPC, PC, Creature and Deity pages. Allegiances is on NPC pages, Ruler on Settlement and Country, Leader on Organization, and Owner on Business. - The question covers only the transcript lines that mention the page, plus one line either side. - **Every change must quote the transcript.** The quote is matched after normalising, within one line or across two adjacent lines. The speaker and the time come from the matched line, never from the model. A change is dropped and counted when its quote isn't found, its value is unchanged, its fact isn't allowed for the type, or its Status isn't one of the four. Counts go on the run (`summarisation_runs.fact_suggestions`). - Suggestions are **GMs-only until used.** "Facts that may have changed" in Suggested wiki updates offers Use it, Change it…, Not this one, and a receipt with Undo. The page editor shows a waiting suggestion under its fact. A used fact keeps its own "Who sees it"; a new one is for everyone at the table (D324). It's written through the normal edit path, so history records it. - The queueing hook comes after summarisation, in a `try`, so it can't fail or slow that path. Without an LLM configured, a session is skipped. **Privacy (three security-review findings, all fixed with tests)** - **Retention:** when a transcript is destroyed, its pending suggestions go, and decided ones lose their quote and speaker. - **Trash:** while a transcript is binned, no read shows its suggestions. - **Erasure:** each suggestion records the `seq` of its quoted line or lines, so erasing a member's lines removes the suggestions quoting them, however the scrub matched the speaker. A mixed erasure is tested. - **Reprocess:** replacing the lines renumbers them, so a reprocess forgets the session's suggestions first, and the next run suggests again. - Deleting a session, page or campaign cascades. **Server:** migration `f1280a0b1c2d` (revises `e1279a0b1c2d`) adds `fact_suggestions` and `summarisation_runs.fact_suggestions`; it has a real downgrade, and Alembic has a single head. New routes: `GET /api/campaigns/{id}/wiki/fact-suggestions`, plus `…/{sid}/use`, `/undo` and `/dismiss`, all for GMs only. The article read (GM only) and the GM session read carry the waiting suggestions, and the attention count includes them. The bot is unaffected. **After the deploy (D325, on the owner's say):** run this once over past sessions, newest first. It's safe to repeat. ``` celery -A app.tasks.celery_app call app.tasks.fact_tasks.suggest_fact_changes_backfill ``` **Checked** - **Backend,** on the dev host: every migration test, `tests/wiki`, `tests/export`, attention, transcript bin, member erasure, enqueue-off-loop, sessions, audio tasks and retention gave 3,056 passed and 0 failed. ruff format and check are clean; the version and changelog checks are OK. - **Frontend:** `src/pages/wiki` and `src/wiki`, 43 files and 625 tests passed. eslint has 0 errors. - **Not yet:** the full backend suite (CI) and the render check, which is starting. New wording goes on the Wording Review page before merge. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

claude-bot added 8 commits 2026-11-03 15:01:05 +00:00

docs(design): lane 6j, facts the session may have changed (#1280) ... fd7d4eacaf

Suggested wiki updates gains "Facts that may have changed": one card
per page, old -> new with the quote from the recording, and Use it,
Change it... or Not this one. The editor shows a pending suggestion
under its fact. Status, Allegiances, Ruler, Leader and Owner first.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

docs(design): D323–D326, facts the session may have changed (#1280) ... ae62f6fd89

The owner's answers to Q214–Q217.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

wip: fact suggestions server (#1280) 9d7b7a7529

feat(webapp): facts the session may have changed, in the queue and the editor (#1280) ... e9e4a46db1

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

wip: fact suggestions tests and transcript lifecycle (#1280) 1151ad3e4d

feat(webapp): facts the session may have changed, the server (#1280) ... efefb22d85

fact_suggestions table and run counts, the suggest_fact_changes task and
its backfill, GM-only routes, the article and session reads, the attention
count, and the transcript's lifecycle (retention, bin, erasure).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Merge remote-tracking branch 'origin/feat/1280-fe' into feat/1280-fact-suggestions 483a011706

fix(webapp): a reprocess forgets the session's fact suggestions before renumbering its lines (#1280) ... 4352bfffd9

A suggestion links its quote to transcript lines by seq. A reprocess
replaces the lines and numbers them afresh, so the old links named the
wrong lines and a later erasure could miss the quote. Pending ones go
and decided ones lose their quote; the next run suggests from the new
lines. Found by the security review.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

claude-bot commented 2026-11-03 15:22:25 +00:00

Author

Collaborator

Copy link

Render check of 4352bfff, on the seeded journeys stack built from this branch. The migration reached f1280a0b1c2d. Suggestions were inserted directly, since that stack has no LLM. Light and dark, 1280, 820 (queue only) and 390, 44 shots.

• Every page returned HTTP 200, with no axe violations (WCAG 2.2 AA), no console errors from the pages, and no sideways scroll at 390.
• Queue: "Facts that may have changed · 5", one card per page and one row per change, each with its quote and who said it when.
  • Change it… gives a Select for Status and a text field otherwise.
  • Use it turns the row into "Status is now Dead." with Undo (POST /use 200), and Undo brings it back (200).
  • Not this one removes the row (POST /dismiss 200).
• A player: GET …/wiki/fact-suggestions returns 403. The article read carries no fact_suggestions, and the page shows Status "Alive", nothing suggested.
• Editor: a pending suggestion shows under Status. Use it writes "Dead", focus lands on the value with a visible ring, and a later Save of another fact is a 200 with no version conflict.
• Targets: 34 px tall or more.

Fixed after the check (latest commit): a long new value left "— →" alone on a line; on a phone, Undo stretched across the receipt; and a quote-less row drew empty "“”". There are two new tests.

Seen, as designed:

• The session line counts only the latest played sessions, as it already did; older sessions' facts still show in the group.
• The editor shows the newest suggestion for a fact, and an older one for the same fact is in the queue.

**Render check** of 4352bfff, on the seeded journeys stack built from this branch. The migration reached `f1280a0b1c2d`. Suggestions were inserted directly, since that stack has no LLM. Light and dark, 1280, 820 (queue only) and 390, 44 shots. - Every page returned HTTP 200, with no axe violations (WCAG 2.2 AA), no console errors from the pages, and no sideways scroll at 390. - **Queue:** "Facts that may have changed · 5", one card per page and one row per change, each with its quote and who said it when. - Change it… gives a Select for Status and a text field otherwise. - Use it turns the row into "Status is now Dead." with Undo (POST `/use` 200), and Undo brings it back (200). - Not this one removes the row (POST `/dismiss` 200). - **A player:** `GET …/wiki/fact-suggestions` returns 403. The article read carries no `fact_suggestions`, and the page shows Status "Alive", nothing suggested. - **Editor:** a pending suggestion shows under Status. Use it writes "Dead", focus lands on the value with a visible ring, and a later Save of another fact is a 200 with no version conflict. - **Targets:** 34 px tall or more. **Fixed after the check (latest commit):** a long new value left "— →" alone on a line; on a phone, Undo stretched across the receipt; and a quote-less row drew empty "“”". There are two new tests. **Seen, as designed:** - The session line counts only the latest played sessions, as it already did; older sessions' facts still show in the group. - The editor shows the newest suggestion for a fact, and an older one for the same fact is in the queue.

rbrooks added 1 commit 2026-11-03 15:22:27 +00:00

fix(webapp): the facts queue after its render check (#1280) ... f9fbcb47f7

- The arrow travels with a long new value instead of sitting alone.
- On a phone, Undo sits beside the receipt rather than across it.
- A quote blanked with its transcript draws no empty quotation marks.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

claude-bot scheduled this pull request to auto merge when all checks succeed 2026-11-03 15:32:13 +00:00

rbrooks added 2 commits 2026-11-03 15:49:59 +00:00

fix(webapp): the facts backfill picks sessions by their lines and the bin, never the transcript text (#1280) ... 1a008cd715

CI's transcript-bin guard caught a raw .transcript read in
sessions_to_backfill. It now selects sessions with transcript lines
and no binned transcript; the task already refused a binned one.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

test: list the fact suggestions' uses of the transcript rows (#1280) ... 08da4836e8

The suggestion task reads rows only once readable_transcript is not
None; the backfill's EXISTS is filtered on the bin; the rest are type
annotations.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

claude-bot commented 2026-11-03 15:50:08 +00:00

Author

Collaborator

Copy link

CI: tests/test_transcript_bin_readers.py failed. That file is the transcript-bin guard (#680): a binned transcript must never be served, searched or sent to a model.

• 1a008cd7: sessions_to_backfill read Session.transcript to find processed sessions. It now picks sessions with transcript lines (an EXISTS) whose transcript isn't binned, and never reads the text. A new test covers that a binned session is skipped and a restored one comes back.
• 08da4836: the guard's sibling check lists every use of the transcript rows. The suggestion task reads rows only once readable_transcript(session) is not None, as it already did. The backfill's EXISTS is filtered on the bin, and the other four uses are type annotations.

Checked: the fact-suggestion and transcript-bin tests passed on the dev host, apart from the listing fixed in 08da4836, which I checked with the guard's own collector. ruff is clean. The PR is re-scheduled to merge when CI passes.

**CI:** `tests/test_transcript_bin_readers.py` failed. That file is the transcript-bin guard (#680): a binned transcript must never be served, searched or sent to a model. - **1a008cd7:** `sessions_to_backfill` read `Session.transcript` to find processed sessions. It now picks sessions with transcript lines (an EXISTS) whose transcript isn't binned, and never reads the text. A new test covers that a binned session is skipped and a restored one comes back. - **08da4836:** the guard's sibling check lists every use of the transcript rows. The suggestion task reads rows only once `readable_transcript(session)` is not None, as it already did. The backfill's EXISTS is filtered on the bin, and the other four uses are type annotations. Checked: the fact-suggestion and transcript-bin tests passed on the dev host, apart from the listing fixed in 08da4836, which I checked with the guard's own collector. ruff is clean. The PR is re-scheduled to merge when CI passes.

claude-bot merged commit 7420a31f32 into main 2026-11-03 16:27:20 +00:00

claude-bot referenced this pull request from a commit 2026-11-03 16:27:21 +00:00

Merge pull request 'feat(webapp): facts the session may have changed (#1280)' (#1325)

claude-bot referenced this pull request 2026-11-03 16:28:04 +00:00

Wiki facts suggested by the local model after a session (Status first), for GM approval #1280

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels   

backend

Issue related to the backend

  

bot

Issue related to the bot

  

bug

Something is not working

  

duplicate

This issue or pull request already exists

  

enhancement

New feature

  

frontend

Issue related to the frontend

  

help wanted

Need some help

  

invalid

Something is wrong

  

question

More information is needed

  

recording

Issue related to the recording system

  

security

Security hardening / vulnerability work

  

wontfix

This won't be fixed

No labels

backend

bot

bug

duplicate

enhancement

frontend

help wanted

invalid

question

recording

security

wontfix

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

2 participants

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

rbrooks/Quest-Board!1325

No description provided.