stack/agents/dewey/work/chat-02/BRIEF-r1-314da8b0.md

CHAT-02 brief: read-only histories and Age (#1507, row 5)

Author: Dewey, 2026-09-26. Draft for Sage, then Filbert. No source edits. Plan row: docs/plans/2026-09-13_webui-session-chat.md line 214. Depends on CHAT-01 (28d4e98a) and uses the CHAT-01C companion (b023841c).

1. Is there a second defect in the return path?

Short answer: no second defect in the return path shows up for repository Pi seats. The answers reach the transcript, the board reads them, and the Console shows them. What Jason hit on 09-13 is best explained by the product gap: the Console offered one "Last assistant text" field, clipped at 240 characters, in an inspector, with no reply thread and no pending signal. 42c08d52 fixed the pending display. The clip and the missing thread are CHAT-02's job.

I found one real defect on the send side, not the return side (§1.3).

1.1 Evidence so far

1. Transcripts. I scanned every .pi/state/*/sessions file (darkwing, dewey, filbert, researcher) for board sends. The script is evidence/sends.mjs and the output is evidence/board-sends-repo-pi.jsonl. There are 30 sends and 29 have a final answer in the same file. Latency is 1.5 s minimum, 14.6 s median and 1019 s maximum. In 16 of the 29 answers the text exceeds 240 characters, so the Console showed them clipped.
2. The unanswered send. Filbert, 2026-09-12T16:33:17Z. The seat started working on it: it read files until 16:35:44Z, the last entry is an aborted assistant turn, and a new filbert session began at 16:36:26Z. The seat was relaunched mid-turn. That lost the answer, but the loss is in the seat, not in the return path.
3. Jason's report window. At 2026-09-13T00:49:59Z he asked dewey "What are the Gate E criteria?". The answer landed at 00:50:07Z with 365 characters, clipped to 240 in the Console. He wrote the report at 00:56:51Z. The report says the sessions "are all cards within the project dashboard and not available as independent chat interfaces". That describes the missing thread, not a missing answer.
4. The 09-13 Console, replayed. I extracted ea00ec66 into a scratch directory and ran the return-flow test with the pending and Age assertions removed. It passed: the new answer arrived through the 10 s poll, once, in the open inspector. That page did not drop answers.
5. Jason's live send today. Researcher, user entry at 2026-09-26T20:10:57.429Z, answer "pong" at 20:11:06.173Z (8.7 s). At 20:15:18Z /api/board showed the researcher row as idle with lastAssistantText: "pong", lastActivity 20:11:06.173Z, the same sessionFile, and a live registration (evidence/board-researcher-*.json). The live WebUI (pid 1266267, port 7330) serves app.js byte-identical to HEAD (d1a51646…), so the 42c08d52 pending notice is live. The WebUI server code has not changed since that process started on 09-13 16:19 CDT. The board (pid 3977979, port 7331) started today at 15:03 CDT.

1.2 What is still unverified

• What Jason's screen showed. For today's send I have the transcript and the board JSON, not the Console DOM or a screenshot. The one missing fact is whether "pong" appeared in the open inspector without a manual refresh.
• The 09-13 processes. The board keeps no reply receipts and the 09-13 board and WebUI processes are gone. I cannot prove what those processes served at 00:50Z. I can only show that the same code, replayed, works.
• Fleet seats. I did not scan ~/.mosaic fleet transcripts. The rule for this phase is to leave them alone, and CHAT-02 does not cover fleet catalogues.
• Clipping in a live send. "pong" is 4 characters, so today's send does not exercise the 240-character clip.

1.3 Send-side defect found in passing

Today's user entry starts with /: /[dragon-lin:control-board -> dragon-lin:researcher] ping. agent-send.sh adds the header, so the / was already sitting in the Pi editor when the paste arrived. send-message.sh pastes into whatever is in the composer (paste-buffer -p, then Enter). It does not clear it first. Here Pi treated the result as plain text. If the composer had held a real command prefix, a board reply could have become a slash command. This is the unmediated-ingress hazard that CHAT-01 assigns to CHAT-03I (B3). tools/tmux/** is excluded from the WebUI plan, so I record it and do not propose a fix here. Sage decides whether to note it on #1507 for the CHAT-03I charter.

1.4 Evidence that settles it

One question for Jason settles the live case: after today's ping, did "pong" appear in the Console inspector without pressing Refresh?

• If yes, there is no second defect. CHAT-02 proceeds as a product-gap fix.
• If no, /api/board had the answer, so the defect is in the Console or the WebUI proxy. Then capture the Console DOM and the network log for /api/board during one send.
• If /api/board had lacked it, the defect would be in the scanner (wrong newest file). If the transcript had lacked it, the defect would be in the transport or the seat. Neither happened today.

To make the check complete for the clip, one more live send can ask for an answer over 240 characters. Every step above then repeats, and the Console should show the text ending in "…". That proves the clip is what cut the 09-13 answer short.

2. Scope

CHAT-02 per the plan: both harness catalogues, safe full branch history with pagination and cursors, separate timestamps, relative Age from last activity, and fixtures for malformed, truncated, replaced and touched logs, cross-project and symlink denial, and no log writes. Read-only: opening a conversation never resumes, forks, launches or controls anything.

2.1 Backend: packages/conversation/**

The owner is not yet assigned. The plan proposes Darkwing; Sage now assigns.

• Catalogue. Approved source roots only: the board's repository specs (<repo>/.pi/state/<seat>/sessions) and live seat registrations. There is no global scan and no browser-supplied path. Each conversation gets an opaque ID mapped server-side to a file. Catalogue creation time, engine launch time and last activity are separate nullable fields.
• No-write Pi parser. Never SessionManager.open, which can migrate old files (CHAT-00 line 67). Open read-only. Record size, mtime and inode before and after each read, and the fixture asserts they are unchanged.
• Branches. Build the id/parentId tree and select a leaf explicitly. The default leaf is the last appended entry, which the implementer must confirm against the pinned Pi session docs. Other leaves are read-only branches. Compaction entries render as markers. Pi get_entries order is not a branch transcript (CHAT-00 line 66), so the parser does not use it.
• Pages. CHAT-01 limits: at most 100 parts, 8 MiB serialized UTF-8, 64 blocks per part and 262144 characters per string. Oversize content splits into continuation parts and is never clipped. The cursor binds actor, purpose, conversation, branch, snapshot, source epoch and expiry. A replaced file (new inode or a shorter length) is a new source epoch, so old cursors refuse and the client keeps its view with a reconcile marker. It never silently switches files.
• Damaged input. A malformed line becomes an unavailable part with its position, and reading continues. A truncated trailing line counts as incomplete, not as an error.
• Denials. A symlink anywhere under the root, a path that resolves outside the root, a parent-session reference, or a conversation from another project all refuse, with fixtures.

2.2 Console: packages/webui/** (Dewey)

This is the smallest piece that answers the 09-13 complaint: a read-only conversation view per session, opened from the card, table or inspector. It renders the selected branch in full, unclipped: user text, assistant text, tool calls and results collapsed, thinking hidden by default, and Markdown as untrusted text with no active HTML. Age stays as 42c08d52 shipped it. Reply keeps the existing board path unchanged. The view gets the new answer through polling, as the inspector does now. Streaming belongs to CHAT-03.

The full chat UI (sidebar, composer, queue, approvals, uploads) stays in CHAT-05.

2.3 Return-flow regression (required by the plan)

Extend packages/webui/tests/return-flow.test.mjs, or add a sibling test. Send from the conversation view. Then the seat appends user, toolCall, toolResult and a final answer longer than 240 characters. Assert that the whole answer appears once, unclipped, in the same open view, with no manual refresh, and that the draft and caret survive. Add a delayed-result variant where the toolResult lands after a poll. This covers the Pi engine only; see D2.

3. Decisions needed from Sage

• D1: deferred items. CHAT-01 line 342 defers "the actual execution/writer-claim record" to CHAT-02. CHAT-01C line 217 defers R3-1, reconciling dispatched but unconsumed input, to "CHAT-02 adapter evidence". A read-only reader needs neither. I recommend moving both to CHAT-03, which owns binding and the single writer, and recording that on #1507.
• D2: Claude catalogue. B1 is open: Claude's persisted branch format and leaf selection are unverified (CHAT-00 line 66). I recommend shipping Pi in CHAT-02, with the Claude catalogue refusing unsupported-harness under a fixture, and adding Claude once B1 evidence exists. Rocko is the only Claude seat. The plan asks for both harnesses, so this is a scope change for Sage to accept or refuse.
• D3: endpoint home. The WebUI knows only the board URL, and the board owns discovery and the approved roots. I recommend that packages/conversation stay a library with no server, and that the board add two read-only routes: catalogue and page. That is a board integration (packages/control-board/src/serve.mjs, scan.mjs), which the plan says needs approval and Darkwing's coordination. The alternative is for the WebUI to read seat registrations itself, which duplicates discovery.
• D4: backend author for packages/conversation/**, and the sequencing: backend first, then the Console against its fixtures.
• D5: Jason's gate. CHAT-02..08 were held pending Jason. This brief does not lift that hold. Building needs Sage's go under whatever Jason has ruled.

4. Acceptance

• The new packages/conversation tests cover every fixture in §2.1, including a before/after hash, mtime and inode check proving no log writes.
• The return-flow regression from §2.3 passes on the served WebUI.
• These still pass: node docs/plans/chat-00/check.mjs, chat-01/check.mjs, chat-01c/check.mjs, and the control-board, webui and seat suites.
• Browser evidence: conversation view at 320 and 1440 in both themes, with a long answer, a tool call, a malformed-line marker and a stale-cursor reconcile marker. No horizontal overflow at 320.
• Filbert approves the exact candidate hashes. Live check: one board send with a long answer, visible in full in the view.

5. Not in scope

Live adapters, control, streaming, queues, uploads, approvals, fleet seats, Claude history until B1, any change to tools/tmux/**, roles/** or session logs, and the / paste hazard (CHAT-03I).