Skip to content

feat(people): let agents read a person's notes instead of the system prompt - #2158

Merged
zhangfand merged 9 commits into
claude/project-thread-v81aa3from
claude/project-thread-v81aa3-read-person
Nov 14, 2026
Merged

zhangfand merged 9 commits into
claude/project-thread-v81aa3from
claude/project-thread-v81aa3-read-person

Conversation

@zhangfand

Copy link
Copy Markdown
Contributor

Requested by zhangfan · project thread

Why

The inbox app copied the notes of whoever sent the message that opened a chat's session into the system prompt. In a group, the agent then applied one member's notes to everyone. A session the guardian opened had none, and one reopened by a scheduled wake-up had none either. Now main reads a person when it meets someone it does not recognize, and sentinel reads the sender's own notes on each message it triages.

Sketch

 system prompt (inbox)
   ## Guardian
-  ## Sender info            <- whoever opened the session
 main
+  core:read_person { name } -> name, bond level, accounts, notes
 sentinel prompt, each message
+  <sender>…the sender's own notes…</sender>
   Triage this message from Ana (other): …

Invariants

  • No session's system prompt carries a sender's notes. The agent reads a person's notes on purpose, by name.
  • Sentinel, which answers untrusted senders and holds no tools, sees only the notes of the sender of the message it triages.
  • A visitor's chat never reads a person's notes.

Scope

Merge danger

Two-way door, no schema change. Blast radius: main's and sentinel's prompts in chats from connections. Main loses the notes from its system prompt and reads them with the tool instead.

Evidence

  • core:read_person answers every person by that name, ignoring case, with accounts and notes, answers no one for an unknown name, and refuses a visitor's chat.
  • An untrusted message: sentinel's prompt carries that sender's <sender> block, and no agent's system prompt names the sender.
  • pnpm dev:all reaches Rome started with core:read_person registered.

🤖 Generated with Claude Code

…prompt

The inbox app put the notes of whoever opened a chat's session into the
system prompt, which in a group described one sender for every message.
Main now reads a person with core:read_person when it meets someone it
does not recognize. Sentinel, which holds no tools, reads the sender's
own notes on each message it triages. The system prompt keeps only the
guardian's profile.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@zhangfand zhangfand self-assigned this Nov 14, 2026

@zoolsher zoolsher left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Design & Alternatives review: 💬 COMMENT

Problem: When a connection talks to Rome, the inbox puts the notes of whoever opened the session into the system prompt. In a group, the agent then applies one member's notes to everyone. Sessions opened by the guardian or by a wake-up get no notes at all. The cause is that the system-prompt suffix is fixed when the session opens, but the sender changes from message to message.

Goals: no sender's notes in any system prompt; each sender's notes reach the agent when that sender speaks; sentinel only ever sees the notes of the sender it is triaging; visitor chats never read notes. Non-goals: giving sentinel tools; schema changes; changing how sender labels work (#2157).

Approaches considered (all grounded in the code):

  • A. This PR (mixed): sentinel gets a <sender> block on every message (pushed). Main and other agents get core:read_person(name) plus one line in main.yaml telling the model to call it for "someone you do not recognize" (pulled).
  • B. Push for the current sender everywhere, pull only for others: reuse buildSenderInfo and prepend the same <sender> block in handleTrustedMessage (it currently sends just prompt: text). Add "sender" to packages/core/src/core/sticky-headers.ts, so withoutRepeatedHeaders (agent-session.ts:3727) only re-sends it when the speaker changes. concierge_owner already uses this pattern for "the person it serves". Keep core:read_person for people who are only mentioned, or who appear in caught-up group history.
  • C. Fix it at the root in core: have SenderLabels.readUnseen (chats/sender-labels.ts), which already resolves each message's Person through byAccount, attach notes for each distinct linked sender. This covers every channel, not just inbox. But it puts file reads in the prompt-facts path and grows every catch-up batch.
  • D. Do nothing / keep notes for private threads only: this still leaves a group-of-one session opened by a wake-up or by the guardian without notes, so it doesn't solve the stated problem.
Approach Added complexity Correctness/completeness Blast radius Reversibility Maintenance Effort
A. PR (push to sentinel, pull for main) Low: 1 action + prompt line Groups fixed; whether the replied-to sender's notes get applied depends on the model choosing to call the tool Main's and sentinel's prompts in connection chats Two-way (new action is additive, no schema change) Low Done
B. Push current sender to all agents + sticky header, keep tool for others Low: ~5 lines + one STICKY_HEADERS entry Every reply deterministically sees the speaker's notes; others on demand Same as A Two-way Low; same buildSenderInfo for sentinel and main Small
C. Notes in SenderLabels (core) Medium: file I/O in prompt facts, dedup per session Most complete, every channel All channel sessions Two-way Medium Medium
D. Status quo / private-only suffix None Groups and wake-up sessions still wrong none n/a n/a none

None of these is a one-way door. core:read_person is an agent-facing action name, cheap to change.

Recommendation: The direction is right: notes come off the session-fixed system prompt, sentinel gets the push it needs, and the name-based tool is a reasonable way for main to look up mentioned people. The weak spot is the person main is replying to. Before this PR, a 1:1 DM reliably carried that person's notes, which can include the guardian's own instructions for handling them. Now they only arrive if the model judges the sender as "someone you do not recognize", and a labelled sender like Ana (Ana Lopez, inner circle) will often look recognized. Option B closes this gap with the existing sticky-header mechanism and keeps the PR's invariants.

Minor, outside the design question: the guard comment in read-person.ts says "a request with no session, never reads them", but if (actor && …) allows a missing actor, and the first test relies on that. Either the comment or the code should change.

Verdict: COMMENT — Taking notes out of the session-fixed system prompt is right, but for the person main is actually answering, it should get their notes pushed onto the turn (as sentinel does), not only when the model decides to call a tool. The choice is reversible, so this is a suggestion, not a blocker.

1 finding(s) posted as inline comments below.

Severity Category File Title
P2 design packages/core/agents/main.yaml Main no longer reliably sees the notes of the person it is replying to

Automated review by RomeOS Code Review · commit 8af493b · Full review history

When a Rome platform capability fails or is missing in a way only the platform can fix, report it with `system:send_feedback`.
Use `system:send_message` to proactively send messages through any registered channel.
Always check memory for relevant context before responding.
Each message names who said it. When a message comes from, or mentions, someone you do not recognize, read who they are with `core:read_person` before you answer.

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[P2] design — Main no longer reliably sees the notes of the person it is replying to

Before this PR, a connection's 1:1 DM always carried that sender's notes. Those notes can include the guardian's instructions for handling the person. Now main only reads them if the model decides the sender is "someone you do not recognize", and a sender labelled Ana (Ana Lopez, inner circle) will usually look recognized. Suggestion: in handleTrustedMessage, prepend the same <sender>${buildSenderInfo(...)}</sender> block that sentinel gets, and add "sender" to STICKY_HEADERS (packages/core/src/core/sticky-headers.ts), so it is only re-sent when the speaker changes, the way concierge_owner already works. Keep core:read_person for people who are only mentioned or who appear in caught-up history.

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Not changed, on purpose. Taking the person's notes out of the system prompt and having main read them with core:read_person when it meets someone it does not recognize is the design the owner chose, so the runtime no longer decides which notes main reads. A 1:1 DM still names the sender on every message (who="Ana Lopez, inner circle"), so main knows who it is talking to and can read their notes in one call. I've raised the DM case with the owner as a design question; if they want the notes back for DMs it will be a follow-up.

@zoolsher zoolsher left a comment •

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔁 This review has been superseded. See the latest review.

Comment thread packages/core/src/people/read-person.ts Outdated
// The notes are the guardian's own. A visitor's chat, or a request with
// no session, never reads them.
const actor = actionExecutionContext.getStore()?.actor;
if (actor && actor.kind !== "guardian") {

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[P2] security — Visitor guard lets through turns with no actor, contradicting the comment and the PR invariant

The comment says "A visitor's chat, or a request with no session, never reads them", but if (actor && actor.kind !== "guardian") lets any turn with no actor through. The first test (read("ana lopez") with no context) relies on that. Per rome-tool-ports.ts, a turn no person started (a visitor's scheduled task run, a followed event, a resumed turn) has no actor. So a visitor-owned chat can still read the guardian's notes, which breaks the PR's invariant. core:set_agent_name already covers this case: it checks ownerOfChat(channelContext.threadId) for webchat threads, with tests for "visitor-owned chat without an actor". Do the same here (inbox turns, which have no webchat owner, can still be allowed). Otherwise, at least fix the comment so it does not claim protection the code lacks.

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Fixed in 3a369b8. An actorless turn in a webchat chat is now checked against the chat's owner, and anything but the guardian's chat is refused.

@zhangfand zhangfand left a comment •

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔁 This review has been superseded. See the latest review.

Comment thread packages/core/src/people/read-person.ts Outdated
// The notes are the guardian's own. A visitor's chat, or a request with
// no session, never reads them.
const actor = actionExecutionContext.getStore()?.actor;
if (actor && actor.kind !== "guardian") {

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Fixed in 3a369b8. Same change: core:read_person asks Concierge who owns the chat and refuses a visitor's chat, so a delegated task run for a visitor cannot read notes.

prompt: `Triage this message from ${displayName} (${bondLevel}):\n\n${text}`,
// The sender's own notes go on each message: the sentinel's session is
// the chat's, where many people speak, and it reads no one else's.
prompt: `<sender>\n${senderInfo}\n</sender>\n\nTriage this message from ${displayName} (${bondLevel}):\n\n${text}`,

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Fixed in 3a369b8. Sentinel's session is now sentinel:<channel>:<thread>:<sender>, so each sender's notes stay in their own session.

Comment thread packages/core/src/people/read-person.ts Outdated
id,
...(delegate ? { delegate } : {}),
})),
notes: person.profilePath ? await readNotes(person.profilePath) : null,

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Fixed in 3a369b8. personProfilePath returns the stored path when it is under memory/relationship/, else memory/relationship/<id>.md, and both core:read_person and the people resource use it.

zhangfand and others added 2 commits November 14, 2026 05:42
# Conflicts:
#	docs/domains/people.md
`core:read_person` refuses a visitor-owned chat as well as a visitor actor.
Sentinel's session is kept per sender, so one sender's notes never reach
another's triage. A person with no stored profile path reads its notes at
the conventional path under memory/relationship/.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Comment thread packages/core/src/people/read-person.ts Outdated
// The notes are the guardian's own. A visitor's chat, or a request with
// no session, never reads them.
const actor = actionExecutionContext.getStore()?.actor;
if (actor && actor.kind !== "guardian") {

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Fixed in 3a369b8. Same change: core:read_person asks Concierge who owns the chat and refuses a visitor's chat, so a delegated task run for a visitor cannot read notes.

prompt: `Triage this message from ${displayName} (${bondLevel}):\n\n${text}`,
// The sender's own notes go on each message: the sentinel's session is
// the chat's, where many people speak, and it reads no one else's.
prompt: `<sender>\n${senderInfo}\n</sender>\n\nTriage this message from ${displayName} (${bondLevel}):\n\n${text}`,

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Fixed in 3a369b8. Sentinel's session is now sentinel:<channel>:<thread>:<sender>, so each sender's notes stay in their own session.

Comment thread packages/core/src/people/read-person.ts Outdated
id,
...(delegate ? { delegate } : {}),
})),
notes: person.profilePath ? await readNotes(person.profilePath) : null,

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Fixed in 3a369b8. personProfilePath returns the stored path when it is under memory/relationship/, else memory/relationship/<id>.md, and both core:read_person and the people resource use it.

@zoolsher zoolsher left a comment •

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔁 This review has been superseded. See the latest review.

@zhangfand zhangfand left a comment

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Code Review: ✅ APPROVE

The PR replaces session-wide sender notes with an explicit person-reading action and per-message sentinel context, including sender-isolated sessions and visitor access checks. The changes fit the existing architecture. Focused tests could not run because dependencies are not installed (rstest was unavailable).

Verdict: APPROVE — No actionable bugs or security regressions were found in the reviewed changes.

No issues found. Code looks good!


Automated review by RomeOS Code Review · commit 3a369b8

@zoolsher zoolsher left a comment •

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔁 This review has been superseded. See the latest review.

* stranger. Sentinel reads it with each message it triages. The agents that
* answer read a person's notes themselves, with `core:read_person`.
*/
export function buildSenderInfo(

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[P2] correctness — Sentinel's <sender> block misses notes stored under the person-id file

buildSenderInfo returns "Sender profile markdown unavailable" whenever person.profilePath is null. createPerson never sets that field, and BONDS.md has agents write notes at memory/relationship/<id>.md. So for every person created on the People page, sentinel never sees their notes, which defeats the per-message block this PR adds. The new Casey assertion locks in that fallback. This is the same gap fixed for read_person. The inbox can't import personProfilePath, but the fix is one line: const path = ctx.person.profilePath ?? \memory/relationship/${ctx.person.id}.md`, then read that path. Add a test with a null profilePath` and an existing id-named notes file.

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Fixed in 78f5bd1. buildSenderInfo now falls back to memory/relationship/<id>.md when no path is stored, the same place core:read_person reads. A new test covers it.

Comment thread packages/core/src/people/read-person.ts Outdated
name: READ_PERSON_ACTION,
type: "system",
description:
"Read who a person is: their bond level, their accounts, and the guardian's notes on them. Call it when a message comes from, or mentions, someone you do not recognize. Pass the person's name as a message's sender label shows it in parentheses, such as `Ana Lopez` from `Ana (Ana Lopez, inner circle)`. A sender marked `not linked` is nobody the guardian tracks and has no notes.",

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[P3] code-quality — Tool description cites a sender-label format Rome never renders

The description says to take the name from parentheses, as in Ana (Ana Lopez, inner circle). Rome actually renders who="Ana Lopez, inner circle" on <current_message>, Ana [Ana Lopez, inner circle] in <conversation_context>, and agent acting for Ana Lopez, … for delegates (sender-labels.ts). Since matching is exact apart from case, a model that passes Ana Lopez, inner circle or agent acting for Ana Lopez gets no one back. Describe the real format: the person's name is the text before the comma in who, after agent acting for when present.

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Fixed in 78f5bd1. The description now cites the labels Rome renders: who="Ana Lopez, inner circle" and Ana [Ana Lopez, inner circle].

@zhangfand zhangfand left a comment •

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔁 This review has been superseded. See the latest review.

zhangfand and others added 2 commits November 14, 2026 06:30
Sentinel's `<sender>` block now falls back to `memory/relationship/<id>.md`
when a person has no stored profile path, as `core:read_person` does. The
`core:read_person` description now cites the sender label Rome renders.

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

@zoolsher zoolsher left a comment •

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔁 This review has been superseded. See the latest review.

@zhangfand zhangfand left a comment •

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔁 This review has been superseded. See the latest review.

};
if (context?.actor && context.actor.kind !== "guardian") return refused;
const thread = context?.channelContext;
if (thread?.channel === "webchat") {

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Fixed in 29c95ca, by failing closed. core:read_person now refuses a turn with neither an actor nor a chat, so a summoned agent cannot read notes unless its turn shows it is the guardian's. The trade-off is that a summoned agent can't use the tool even when a guardian chat summoned it. Carrying the caller's chat through system:summon would fix that, but it changes summon for every action, so I've left it for a follow-up. A new test covers the actorless, chatless turn.

zhangfand and others added 2 commits November 14, 2026 06:41
…he guardian's

A turn with neither an actor nor a chat, such as an agent summoned from a
visitor's task, is now refused by `core:read_person`.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
};
if (context?.actor && context.actor.kind !== "guardian") return refused;
const thread = context?.channelContext;
if (thread?.channel === "webchat") {

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Fixed in 29c95ca, by failing closed. core:read_person now refuses a turn with neither an actor nor a chat, so a summoned agent cannot read notes unless its turn shows it is the guardian's. The trade-off is that a summoned agent can't use the tool even when a guardian chat summoned it. Carrying the caller's chat through system:summon would fix that, but it changes summon for every action, so I've left it for a follow-up. A new test covers the actorless, chatless turn.

@zoolsher zoolsher left a comment •

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔁 This review has been superseded. See the latest review.

@zhangfand zhangfand left a comment •

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔁 This review has been superseded. See the latest review.

@zhangfand
zhangfand added this pull request to stack #2167 November 14, 2026 06:03
# Conflicts:
#	rome_apps/inbox/src/actions/message-handler/index.test.ts

@zoolsher zoolsher left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Code Review: ✅ APPROVE

Goal: The inbox system-prompt suffix is fixed when a session opens, so it no longer carries a sender's notes. Main and the other * agents now read a person when they need to, with the new core:read_person action. Sentinel gets the current sender's notes in a <sender> block on each message, and each sender now has their own sentinel session (sentinel:<channel>:<thread>:<sender>).

Checked beyond the diff:

  • read-person.ts access check. It matches the existing core:set_agent_name pattern in agent-name.ts:
    • A turn whose actor is not the guardian is refused.
    • A webchat turn is allowed only when the chat's owner is the guardian.
    • A turn with neither an actor nor a chat is refused.
    • A turn on a connection, which has a non-webchat channelContext, is allowed. That is the intended path from handleTrustedMessage.
    • The tests cover all of these cases.
  • Wiring in index.ts. peopleRepo (line 451) and concierge (line 1208) both exist before the action is registered. The action is also in COORDINATION_ACTIONS, so it runs in the main process with the action context set.
  • personProfilePath extraction. people/resource.ts gives the same result as before: the stored file name if there is one, else <id>.md, checked against readdir. The profilePath value is written only by internal code, so the startsWith check is not reachable by user input.
  • New sentinel session key. The new key still parses as kind: "sentinel" in agent-session-key.ts. Nothing reads the watched threadId, sessionKeyChannel returns undefined for a sentinel as it did before, and nothing else looks up sentinel sessions by prefix. Sentinel has no tools, so it never creates subagent or fork key suffixes that walkToRoot would have to parse. Old per-chat sentinel sessions are simply no longer used. This is harmless.
  • buildSenderInfo fallback. It now reads memory/relationship/<id>.md when no path is stored, as core:read_person does, and a test covers it.
  • Earlier threads. The visitor and summon access gaps, the id-file fallback and the tool-description format are all fixed. The DM-notes design question has the author's reasoned reply and is with the owner, so I don't raise it again.

Design: The pull approach (fetch notes when needed) for answering agents, plus the push approach (notes included automatically) for the tool-less sentinel, is a reasonable and minimal fix for the root cause: a session-wide suffix can't describe each sender. I found nothing over-complicated in it.

Remaining (optional): a stale comment that an earlier review raised and that is still there, and buildMessageContext inputs that are now unused.

Tests: I did not run the tests; I reviewed them by reading.

Verdict: APPROVE — I found no bugs or security regressions at head. The two remaining findings are optional P3 cleanups.

2 finding(s) posted as inline comments below.

Severity Category File Title
P3 code-quality rome_apps/inbox/src/actions/message-handler/index.ts Stale comment still says sentinel's session is the chat's
P3 code-quality rome_apps/inbox/src/actions/message-handler/index.ts buildMessageContext no longer uses most of its inputs

Automated review by RomeOS Code Review · commit 1c48424 · Full review history

agentName: "sentinel",
prompt: `Triage this message from ${displayName} (${bondLevel}):\n\n${text}`,
channelThreadKey: `sentinel:${channelThreadKey}`,
// The sender's own notes go on each message: the sentinel's session is

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[P3] code-quality — Stale comment still says sentinel's session is the chat's

An earlier review raised this, and it has no reply yet. The comment says "the sentinel's session is the chat's, where many people speak", but four lines below, channelThreadKey now gives each sender their own session, and the next comment says so. A later reader could wrongly conclude that senders share a sentinel session.

Suggested change
// The sender's own notes go on each message: the sentinel's session is
// The sender's own notes go on each message, and sentinel reads no one
// else's.

// info) belongs in this system-prompt suffix — both stay in the cacheable
// system prefix. The per-thread, dynamic thread + project framing is rendered
// separately as a `<thread_context>` block on the session's FIRST user message
// Only the STABLE Guardian profile belongs in this system-prompt suffix, so

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[P3] code-quality — buildMessageContext no longer uses most of its inputs

buildMessageContext now only reads GUARDIAN.md through resolveProfilePath. The person, bondLevel, strangerPersonId, displayName and other fields that the call at line 770 builds are unused, and MessageContext exists only to type them. Optional cleanup: narrow the parameter to { resolveProfilePath } and give buildSenderInfo its own small type. Then MessageContext and the 12-field call go away.

@zhangfand zhangfand left a comment

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Code Review: ✅ APPROVE

Moves person notes out of inbox system prompts into an explicit read action and per-message sentinel context, with sender-isolated sentinel sessions and visitor authorization checks. Reviewed the changed files and surrounding call paths. Tests could not run because rstest is not installed in the clone.

Verdict: APPROVE — No actionable bugs or regressions were found in the reviewed changes.

No issues found. Code looks good!


Automated review by RomeOS Code Review · commit 1c48424

@zhangfand
zhangfand merged commit f709881 into main Nov 14, 2026
13 checks passed
@zhangfand
zhangfand deleted the claude/project-thread-v81aa3-read-person branch November 14, 2026 06:16
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.

2 participants