fix(core): preserve speaker names in prior dialogue context

[RemilioNubilio]

Summary

• preserves connector-provided sender names when RECENT_MESSAGES history is rendered as structured prior_message:user blocks
• keeps the existing provider-text dedupe behavior, so full # Conversation Messages provider text still stays out of Stage 1 prompts
• adds a regression test for the Discord failure shape where a later turn asks about a named participant from earlier chat context

Why

A live Discord exchange showed Remilio/Nubilio losing speaker attribution after earlier chat context was converted into anonymous prior_message:user blocks. The model saw prior text, but not who said it, so it failed follow-up questions like "look in the chat" / "her and botdick" even though the relevant messages were present.

This patch is connector-agnostic: it reads the existing metadata.sender / entity-name fields already written by connectors and prefixes prior dialogue content with the sender name. Reply references already use this shape.

Validation

• Synced branch to latest origin/develop (fa83ddee6f) via fast-forward before patching
• bun vitest run src/__tests__/message-runtime-stage1.test.ts src/features/basic-capabilities/providers/recentMessages.test.ts src/runtime/__tests__/planner-loop.test.ts from packages/core: 94 tests passed
• bun run typecheck from packages/core: passed
• bun run build from packages/core: passed
• Baseline after sync before patch:
  • core focused suite: 119 tests passed
  • orchestrator focused suite: 120 tests passed
• Live Discord smoke after restart from local source:
  • setup message: 1507674411552084059
  • prompt: 1507674422939750474
  • bot reply: 1507674432225939507
  • trajectory: tj-224f2528525703
  • model: gpt-oss-120b
  • tool calls: 0, failures: 0
  • result: answered the shebotdick/botdick compatibility question using the recent channel context instead of claiming missing information

Greptile Summary

This PR fixes speaker attribution in prior-dialogue context by prefixing each prior_message:user block with the sender's name drawn from metadata.sender / entity-name fields, and adds a core.contextual_identity_lookup_requires_recall evaluator that reroutes short, locally-scoped "who is X?" questions through memory/messaging rather than letting the simple shortcut guess.

• Speaker attribution (priorDialogueSpeakerName + priorDialogueContent): reads connector-supplied sender metadata and prepends \"name: text\" to prior-dialogue segments, preserving existing dedup behaviour for provider text.
• Identity-lookup evaluator (extractContextualIdentityLookupSubject + isShortContextualLookupSubject): detects single-turn questions about short, local-looking handles and injects an identity_lookup_policy context slice, routing the planner toward recall before answering.
• Tests: three new Stage 1 test cases cover the Discord failure shape, the routing trigger, and a negative case for public-entity questions.

Confidence Score: 4/5

Safe to merge for the speaker-attribution fix; the identity-lookup evaluator has a case-sensitivity inconsistency that causes it to silently skip capitalized chat handles.

The speaker-attribution change is narrow and well-tested. The identity-lookup evaluator's isShortContextualLookupSubject tests the third looksLikeLocalName condition against the original-case subject instead of the lowercased normalized, so a Discord display name like Shebotdick (capital S) would not trigger recall routing while the lowercase spelling would, quietly reverting to the guessing path the feature is meant to prevent.

packages/core/src/services/message.ts — specifically the isShortContextualLookupSubject function and its looksLikeLocalName regex.

Important Files Changed

Filename	Overview
packages/core/src/services/message.ts	Adds speaker-name attribution to prior-dialogue context blocks and a new core.contextual_identity_lookup_requires_recall evaluator; contains a case-sensitivity inconsistency in isShortContextualLookupSubject where the last looksLikeLocalName regex tests the original-case subject instead of normalized, causing capitalized chat handles to bypass recall routing.
packages/core/src/tests/message-runtime-stage1.test.ts	Adds three new test cases: speaker-name preservation in prior dialogue blocks, contextual-identity routing through recall, and a negative case for public-entity questions; tests are well-scoped and directly exercise the changed code paths.

_{Reviews (2): Last reviewed commit: "fix(core): route local identity lookups ..." | Re-trigger Greptile}

[RemilioNubilio]

Additional trajectory evidence from the live Discord investigation:

Before

Source channel: 1481030966565797888

Failing exchange:

• user prompt 1507667963766116494: whats the compatibility between her and botdick
• bot reply 1507667972901044325: said it did not have enough information about who her and botdick were
• trajectory tj-0acd867623935c

Follow-up:

• user prompt 1507668050311254127: look in the chat bozo
• bot reply 1507668083551244350: claimed it scanned recent chat but found no mention
• trajectory tj-0b1dd2e55d91ad

What the trajectories showed: the relevant text was present in the Stage 1 prompt, but structured history had been reduced to anonymous prior_message:user blocks, so the model saw message contents without reliable speaker identity.

After

Smoke channel: 1490836448755060920

Live verification after restarting from local source:

• setup message 1507674411552084059: botdick, meet shebotdick...
• prompt 1507674422939750474: what is the compatibility between her and botdick?
• bot reply 1507674432225939507: correctly answered that shebotdick and botdick complement each other
• trajectory tj-224f2528525703
• model: gpt-oss-120b
• tool calls: 0
• tool failures: 0

Note: the live after smoke had the setup message in the connector-provided recent channel context, which already includes names. The new unit regression directly covers the lower-level prior_message:user path from the failing trajectories and asserts it now renders botdick: ... / 1gig: ... while still suppressing full provider text duplication.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: 067b59bf-f33e-4c14-86c0-bb0bfc3878a0

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[RemilioNubilio]

Follow-up pushed in b2d62be0cd after another live Discord failure exposed a related routing gap.

What changed:

• Added a built-in response-handler evaluator for short local/chat identity lookups such as who is queuebot?.
• If Stage 1 tries to answer these through the simple shortcut, the evaluator clears the ungrounded reply and routes through the planner with general context plus a MESSAGE action hint when available.
• The guard is generic and intentionally narrow: local-looking short subjects only. A regression keeps ordinary public questions like who is obama? on the direct reply path.
• Added planner context-slice rendering so evaluator-added grounding policy is visible to Stage 2.

Evidence:

• Before: Discord prompt 1507683002614812734, bot reply 1507683017735278716, trajectory tj-41867f0b8ac1e7. Stage 1 used gpt-oss-120b, plannerIterations=0, toolCallsExecuted=0, and guessed that shebotdick was the same bot/persona as botdick. The original defining Discord message 1507664178456825866 was outside the bounded Stage 1 input.
• Second failure found while inspecting latest trajectories: tj-6054febf8f622e showed the first version cleared the bad reply but selected unavailable memory/messaging contexts in the live guest role, producing an empty send. Fixed by always including general as the portable planner context.
• After: test-channel prompt 1507693238914387978, bot reply 1507693264726130698, trajectory tj-66c28d4517bbf8. Route is now messageHandler -> toolSearch -> planner, plannerIterations=1, toolCallsExecuted=0, toolCallFailures=0, and the trajectory contains core.contextual_identity_lookup_requires_recall, identity_lookup_policy, and MESSAGE in the action surface.

Validation:

• bunx @biomejs/biome check --write src/services/message.ts src/__tests__/message-runtime-stage1.test.ts
• bun vitest run src/__tests__/message-runtime-stage1.test.ts src/features/basic-capabilities/providers/recentMessages.test.ts src/runtime/__tests__/response-handler-evaluators.test.ts — 66 tests passed
• bun run typecheck from packages/core
• bun run build from packages/core
• Branch rechecked against upstream: origin/develop...HEAD is 0 behind / 2 ahead.