Distinguish namesake agents and selected recipients

[loganj]

🤖

Summary

Keep unique agent names clean, and distinguish agents that share a displayed name. Distinct exact keys with the same name show a short qualifier such as Honey · 2abc, including across different profiles.

Rebased onto main after #140, #141, and #147 merged. This is no longer a stacked PR.

Behavior

• Compare trimmed displayed names, preserving case. Count distinct public keys, not duplicate records or shared profile links.
• Derive qualifiers from public keys and lengthen them when short suffixes collide. Live directory name changes update shared naming surfaces.
• Keep qualifiers presentation-only. The merged Show consistent agent names across conversations #141 picker/completion fix separates display labels from serialized source names; message bodies and exact notification keys remain unchanged.
• Preserve Show selected mentions as inline identity chips #147’s inline recipient chips, not the removed avatar strip. Selected chips remain bound to their original source names and exact keys across directory renames. Their existing compact key labels are independent of directory qualifiers.
• Do not change discovery, import, profile grouping, agent startup, or the production composer/CSS.

Validation

At acf51510e2090db8b32458190108563f8d045a45:

• 82 focused tests passed across directory, identity-name service, composer, and real picker/completion → composer → received-message coverage.
• All three mentions journeys passed in Chromium and WebKit: 6 passes.
• Push hooks passed 888 tests across 61 files and design checks.
• One browser case added, none removed relative to main. It checks visible inline qualifiers and viewport fit at 390px touch width, live directory collisions, removal/reselection, and exact notification keys. Browser layout/touch behavior cannot be established by DOM tests alone. Colocated tests cover source text and routing stability. No separate deliberate mutation/fail-then-pass run was performed for this integration.
• Hosted CI: run 35864807289 passed on its first attempt: JavaScript, Rust/tool integration, Chromium/WebKit journeys, browser measurements, and CI required. DCO, Semgrep, and zizmor also passed. Windows native validation is skipped by workflow policy. No live credential import or agent launch was exercised.

Review

The prior qualified-mention round-trip finding is addressed by the merged #141 source/display separation. The existing changes-requested decision has not yet been cleared. Human review coordination remains with the PR owner. The separate WebKit flake investigation is outside this PR.

Completion feedback repair

At 52dadd9bf905879f002faccc2839a7af970383a1, admission includes source names alongside display labels. This restores completed-name dismissal without new conditional branches or changes to qualified filtering/multi-word completion.

One additional browser case (two total added relative to main; none removed) exercises the real completion producer and native Enter/Tab handling. It failed in both engines before the repair and passed afterward. Both complete affected browser files passed (10 tests), query/round-trip unit coverage passed (9 tests), and push hooks passed 899 tests. New CI: https://github.com/block/buzz-app/actions/runs/35867401918 (passed on the first attempt; required CI, security, and DCO green). Both addressed review threads are resolved; review approval remains outstanding.

[wesbillman]

Carl, an automated reviewer, commenting via Wes’s GitHub account.

Changes requested: one P2 qualified-mention round-trip defect

Reviewed head b95faf299ce52afe2c5a6cf43e453933042088c8 against stacked base 65f700cdc7a6cd47de18df3f962343e431a08732 (feat/agents-directory).

The new qualifier-bearing display name is also serialized by the existing mention picker/composer into message text. The received-message renderer still recognizes only the raw name prefix and then adds the qualified name again. The resulting timeline text repeats the suffix. See the inline finding for the narrow reproduction and exit criteria.

Validation: a review-only React diagnostic used the real name provider, MentionPicker, MessageComposer send boundary, and MessageMarkdown renderer with synthetic session data. Selecting one of two Honey identities yielded outgoing @Honey · rcaj with exactly the intended recipient key, then rendered Honey · rcaj · rcaj. A separate direct renderer probe reproduced the same defect. Both tests characterize the defect, not an acceptance pass. No production edits or broad local CI duplication.

Applicable hosted JavaScript, Rust/tool integration, Chromium/WebKit journeys, browser measurements, security and DCO checks pass; Windows native validation is skipped. No live native import/start or live relay acceptance was performed. The previously reported #140 reconnect defect is inherited and remains tracked there, not duplicated as a new #142 finding.

The recipient/UI lane independently confirmed the finding and found no additional routing defect. I completed the collision/source/cache trace; the outstanding delegated algorithm lane was explicitly cancelled before publication.

[wesbillman]

Carl, an automated reviewer, commenting via Wes’s GitHub account.

Changes requested: one P2 completion-admission regression

Reviewed head acf51510e2090db8b32458190108563f8d045a45 against base 60ed428b016ea67cfadbff03c56b812990e170a9.

The previous doubled-suffix finding is fixed. The new blocker is that presentation qualifiers keep completion open after a fully typed name plus a space; Enter then selects a namesake instead of sending the intended plain text. See the inline finding for reproduction and the narrow fix.

I independently reproduced this through the real fixture/composer in both Chromium and WebKit. Without directory qualifiers, @Honey dismisses suggestions and Enter publishes no p tags. With two qualified Honey identities, suggestions remain; the first Enter selects a key and the next Enter publishes a p tag for it. A separate real-service/SSR probe verified the old round-trip fix: raw @Honey, one exact key, one displayed qualifier.

All three delegated lanes returned; directory/collision, notification and geometry review found no additional blocker. Production composer/CSS remain unchanged by this rebased diff. No production edits or broad local CI rerun; browser probes used synthetic identities, not live credentials or agent launches.

Hosted snapshot: JavaScript, Rust/tool integration, measurements, all four browser shards, security and DCO passed; the aggregate CI required job was still running. Windows native validation was skipped. Those checks do not cover this qualified-name keyboard path.

Exit criterion: restore completed-name-plus-space dismissal with qualified names and cover keyboard acceptance versus plain-text submission without changing exact-key routing or multi-word completion.

[wesbillman]

Carl, an automated reviewer, commenting via Wes’s GitHub account.

Rereview: no remaining blockers

Reviewed 52dadd9bf905879f002faccc2839a7af970383a1 against base 60ed428b016ea67cfadbff03c56b812990e170a9, focusing on the delta from reviewed head acf51510e2090db8b32458190108563f8d045a45.

The completed-name dismissal finding is fixed: admission includes source names while qualified filtering stays unchanged. Deliberate Enter/Tab acceptance still selects an exact key and inserts the raw name; typing the complete name plus a space dismisses suggestions and sends plain text without recipient tags. The earlier doubled-suffix fix is unchanged.

Validation: traced the real producer → completion key handling → composer send boundary and ran a focused real-helper probe. Inspected hosted Chromium and WebKit report artifacts from run 35867401918: both complete relevant browser files passed, 10/10 tests with no retries, including the new keyboard regression. The optional independent follow-up lane did not return and was explicitly cancelled before publication; this verdict is based on my completed review and verified test evidence. All applicable hosted checks are green; Windows native validation is skipped. No broad local suite duplication or live relay/native launch validation.

Both prior finding threads are resolved. No additional code changes requested. This is a review comment, not formal approval; Wes has not requested approval.