Skip to content

Show who sent every message in every thread - #112

Merged
nullStack65 merged 3 commits into
release/native-0.1.10from
feat/message-sender-labels
Nov 19, 2026
Merged

nullStack65 merged 3 commits into
release/native-0.1.10from
feat/message-sender-labels

Conversation

@nullStack65

@nullStack65 nullStack65 commented Nov 19, 2026 •

Copy link
Copy Markdown
Owner

What this does

Every message in every thread now says who sent it:

  1. A person: "Kameron · " or "Jeff · ". The server stamps this from the signed-in client session (sentBy), and strips any client-sent sender fields, so a client cannot spoof it.
  2. Another thread or agent: named and linked. A relayed or pasted message keeps its real sender. A relay is never relabelled as a person, even when it carries a forged person stamp.
  3. The thread's own model reply: provider and model, for example "Claude · Opus 5.5".
  4. System and automated events: labelled as system or automation.
  5. Old messages: a sender appears only where stored data proves it. Otherwise the message shows "Unknown sender" (or "Unknown person" for a signed-in client that has no founder assigned). Nothing is guessed.

Filtering:

  • A "From" picker in in-thread find (Ctrl/Cmd+F) covers Kameron, Jeff, people, agents, model and system.
  • Cross-thread thread search takes a sender filter for Kameron, Jeff, people and model replies. Agents and system messages are not in the search index, so search can't filter on them; only the in-thread filter covers those.

Setup: Settings → Sharing & access gets a per-client "who uses this" picker (Unknown person / Kameron / Jeff). A client cannot name its own person, so another trusted client assigns it; bots are never assigned. A device keeps its assignment when its session renews (same device key, or the local desktop app).

PR #110 (standing bot threads) supplied the persona and avatar work reused here.

Notes for review

  • Migration 097_AuthSessionFounder adds a nullable founder column to auth_sessions. The open placement-schedules branch also claims number 097 (097_ScheduledTaskOccurrences). Whichever lands second must renumber.
  • Mobile shows the person sender line on user bubbles. The model sender line is web-only for now.

Status

START: scope = sender labels, filters and server stamping in nullStack65/t3code, based on release/native-0.1.10 at 7af284f. Checkpoints follow as comments.

Known gaps (from review): two older client command paths still attribute messages to an agent; sender search covers the newest 20,000 messages (10,000 per thread); model labels use current provider and model names; trusted clients can assign each other.

Deploy: fleet

Model: claude-opus-5-5
Task: message sender labels in every T3 thread

🤖 Generated with Claude Code

Each message now names its sender: Kameron or Jeff (stamped by the server
from the signed-in client session, so a client cannot spoof it), another
thread or agent (named and linked, and a relay keeps its real sender), the
thread's own model reply (provider and model), or system/automated events.
Old messages show a sender only where stored data proves it; otherwise
"Unknown sender".

Adds a "From" filter to in-thread find and a sender filter to cross-thread
thread search (people and model replies, which the search index covers).
Connections settings can assign a paired client to Kameron or Jeff.

Model: claude-opus-5-5
Task: message sender labels in every T3 thread
Agent: Claude Code

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@nullStack65

Copy link
Copy Markdown
Owner Author

CHECKPOINT 1: committed and pushed head dd5d7ed, rebased on release/native-0.1.10 at 7af284f.

  • Typecheck is clean for contracts, server, client-runtime, web and mobile (targeted, per package).
  • Tests pass: ThreadSearch.test.ts 33/33 (includes "filters messages by who sent them, keeping a relay's real sender"), messageSender.test.ts 5/5 (includes a relay with a forged founder stamp staying an agent), MessagesTimeline.logic.test.ts 122/122 (sender grouping and the in-thread filter).
  • One independent review (Claude Fable 5.2) has started. Next is the headless browser pass with synthetic data.

Model: claude-opus-5-5
Task: show who sent every message (PR 112)
Agent: Claude Code

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@nullStack65

Copy link
Copy Markdown
Owner Author

CHECKPOINT 2: head 6dfe668, pushed.

  1. Headless browser pass on a local dev server with synthetic data (project "demo-harbor", host name masked as demo-host):
    • Before (base 7af284f): Kameron's message bubble has no sender; model replies read "T3 Code".
    • After: "Kameron · Work laptop" and "Jeff · Work laptop" above each person's bubble. The Kameron message stays Kameron after the client is reassigned to Jeff, because the sender is stamped when the message is sent. Model replies read "Replay (dev) · Showcase" (provider · model).
    • Settings → Sharing & access: each authorized client gets a "Kameron / Jeff / Not a founder" picker, saved on the server session.
    • Find in thread → From = Kameron leaves only Kameron's message.
  2. Fix found by that pass (this commit): the new "From" picker squeezed the find text box to about 40px. The picker is now narrower and the box keeps a minimum width (now about 74px in a 280px bar).
  3. Screenshots: before-timeline, after-timeline, after-settings and after-filter were captured. They could not be attached here as GitHub image uploads because the headless browser has no GitHub sign-in. They were not committed (repo rule). They are shown in the agent thread that owns this PR.
  4. Pending: independent review (one round), ARC CI "Voice + founder inbox (Linux ARC)".

- A client can no longer name its own person; another trusted client
  assigns it. Bots, expired and revoked sessions are refused.
- A device keeps its person across session renewal (same proof key, or
  the local desktop bootstrap).
- Edits to queued messages and answers to runtime requests are stamped
  with whoever made them.
- A stamped person's message cannot be disguised as an automation.
- Mobile only groups consecutive messages from the same sender.
- Unsent messages no longer read "Unknown sender"; unstamped old
  messages read "Unknown sender" rather than "You".
- Sender search excludes scheduled-task messages from "People".

Model: claude-opus-5-5
Task: message sender labels (nullStack65/t3code PR 112)
Agent: Claude Code

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@nullStack65

Copy link
Copy Markdown
Owner Author

CHECKPOINT 3: head 78bb88b, pushed. This answers the one independent review round (Fable).

Fixed:

  1. Self-naming: a client can no longer assign its own person, including through another session on the same device key. Another trusted client assigns it in Settings → Sharing & access. Bots, expired and revoked sessions are refused. New test: "lets a client name another client's person but never its own".
  2. Restart and renewal: a device keeps its person across new sessions when the server can tell it is the same device (same proof key, or the local desktop app's bootstrap).
  3. Queued edits: editing a queued message re-stamps it with whoever made the edit.
  4. Approval feedback and answers to agent questions are stamped with the person who answered.
  5. Disguise: a message carrying a person's stamp cannot pass as an automation. Covered by a new unit test.
  6. Mobile: consecutive messages are grouped only when they come from the same sender.
  7. Unsent messages no longer flash "Unknown sender". Unstamped old messages are headed "Unknown sender" instead of "You".
  8. Sender search: "People" no longer picks up scheduled-task messages.

Known gaps, not fixed:

  1. Two older client command paths still create messages attributed to an agent.
  2. Search by sender looks only at the newest 20,000 messages (10,000 per thread), so very large histories can miss older matches.
  3. Model labels use the provider and model names as they are now. They are not a snapshot from when the reply was sent.
  4. Policy: trusted clients name each other, so two trusted clients could still assign each other.
  5. Old messages without any stamp stay under "People" (shown as "Unknown sender"). They are never guessed.

Checks run locally (all pass):

  • Typecheck: contracts, client-runtime, server, web and mobile.
  • Tests:
    • messageSender: 5
    • mobile pending-thread-feed: 10
    • web MessagesTimeline.logic: 122
    • server ThreadSearch, SessionStore, EnvironmentAuth and auth http: 85 plus the new one

@nullStack65
nullStack65 marked this pull request as ready for review November 19, 2026 22:33
@nullStack65
nullStack65 merged commit 49ab827 into release/native-0.1.10 Nov 19, 2026
5 of 6 checks passed
@nullStack65

Copy link
Copy Markdown
Owner Author

RESULT: complete for the servers; the two desktop apps are not yet verified (staged updates).

  1. What each message now shows:
    • Messages from people show who sent them and from which client, for example "Kameron · Work laptop" or "Jeff · Phone". The server stamps this from the signed-in session when the message arrives, so a client cannot set it. A client also cannot name its own person; another trusted client assigns it.
    • A message relayed by another thread or agent shows that thread's name, linked to it. Even a forged person stamp cannot make it read as a founder's message.
    • Model replies show the provider and model, for example "Claude · Opus 5.5".
    • Automations read "Automation" and system notices read "T3 Code".
    • Find in thread has a "From" filter (Kameron, Jeff, people, agents, model, system). Cross-thread search can filter by sender too.
  2. Old messages: none were changed. Messages without a stored stamp read "Unknown sender". A client with no person assigned reads "Unknown person". Nothing is guessed.
  3. PR: Show who sent every message in every thread #112, squash-merged as 49ab827 at head 78bb88b. All checks passed, including ARC CI on the merged commit.
  4. Running it, by readback (fleet drift check, 23:36 UTC, through the normal rollout path):
    • Both fleet T3 servers are live on 49ab827 (descriptor buildCommit).
    • The two desktop apps use staged updates. One could not be reached. The other still runs an earlier build, with this one not yet staged.
  5. Not verified:
    • The two desktop apps.
    • The new build's UI on the live servers. It was checked in a headless browser on a local dev server with synthetic data.
    • Screenshots are not attached here, because the headless browser has no GitHub sign-in.
    • Known gaps from review:
      • Two older client command paths still attribute messages to an agent.
      • Sender search covers only the newest 20,000 messages (10,000 per thread).
      • Model labels use current names.
      • Trusted clients can assign each other.
      • The mobile app does not show the model line yet.
  6. Founder action: in Settings → Sharing & access, assign each client to Kameron or Jeff. Do it from a different client than the one being assigned. Until then, new messages read "Unknown person".

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant