[codex] Preserve participant roles in knowledge synthesis

[vsumner]

Summary

• Updates the cortex knowledge synthesis prompt to preserve concise participant/user role facts when they affect future work.
• Keeps the exclusion for the agent's own identity, recent activity, status, and ephemeral data.
• Marks R3 fixed in the working-memory triage tracker and aligns the implementation plan prompt boundary.

Testing

• cargo test -p spacebot prompts::engine::tests::knowledge_synthesis_prompt_preserves_participant_roles -- --test-threads=1
• cargo test -p spacebot prompts::engine::tests -- --test-threads=1
• cargo fmt --all
• just preflight
• just gate-pr

Notes

• This is stacked on fix(memory): make cortex synthesis non-blocking #570 and closes working-memory triage item R3.

Note

AI-Generated Summary: This PR updates the cortex knowledge synthesis prompt to properly preserve stable participant role facts that influence future decision-making and context assembly. The change refines what gets stored in the agent's knowledge synthesis layer—excluding the agent's own identity and ephemeral data while retaining user role information like "the CEO" that provides behavioral context for future interactions. The implementation includes a targeted test verifying that participant roles are preserved while role explanations and full user profiles are appropriately excluded.

_{Written by Tembo for commit 9447a87. This will update automatically on new commits.}

[coderabbitai]

Walkthrough

The PR updates the knowledge synthesis prompt to explicitly preserve stable participant or user role facts (e.g., "the user is the CEO") while forbidding restatement of the agent's identity. Design documentation is updated to reflect this behavior change, and a new unit test validates the prompt output.

Changes

Cohort / File(s)	Summary
Design Documentation docs/design-docs/working-memory-implementation-plan.md, docs/design-docs/working-memory-triage.md	Updated design docs to reflect the new behavior of preserving participant/user role facts in knowledge synthesis and marked finding R3 as fixed.
Prompt Template Update prompts/en/cortex_knowledge_synthesis.md.j2	Modified inclusion criteria to retain stable participant/user role facts when they impact future work; forbade restatement of agent identity and clarified exclusion rules.
Test Coverage src/prompts/engine.rs	Added unit test knowledge_synthesis_prompt_preserves_participant_roles to verify the prompt includes participant role facts and the specified marker text.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~12 minutes

Suggested reviewers

• jamiepine

🚥 Pre-merge checks | ✅ 3

✅ Passed checks (3 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and concisely summarizes the main change: preserving participant roles in the knowledge synthesis prompt, which aligns with all four modified files.
Description check	✅ Passed	The description is directly related to the changeset, detailing the updates to knowledge synthesis prompt, testing approach, and marking R3 as fixed.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch codex/knowledge-synthesis-participant-roles

---

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.}

[coderabbitai]

🧹 Nitpick comments (2)

docs/design-docs/working-memory-implementation-plan.md (1)

666-681: Doc drift: inline prompt snippet diverges from the actual template.

The inline snippet here is a simplified, lossy paraphrase of prompts/en/cortex_knowledge_synthesis.md.j2 (missing the Include ONLY / NEVER Include section structure, the "Do not restate the agent's own identity or recent activity" sentence, and the "Until participant context owns them…" clarifier at the bottom). Since this document is the implementation plan being kept in lockstep with the prompt per R3, consider either trimming this block to a brief reference ("see prompts/en/cortex_knowledge_synthesis.md.j2") or mirroring the actual template content to avoid future drift.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@docs/design-docs/working-memory-implementation-plan.md` around lines 666 -
681, The inline prompt snippet in the "Synthesize the agent's long-term
knowledge..." block diverges from the actual template
prompts/en/cortex_knowledge_synthesis.md.j2 (it omits the "Include ONLY"/"NEVER
Include" sections, the sentence forbidding restating the agent's identity/recent
activity, and the final clarifier); update the snippet to exactly mirror
prompts/en/cortex_knowledge_synthesis.md.j2 or replace the block with a
single-line reference pointing to that template (e.g., "See
prompts/en/cortex_knowledge_synthesis.md.j2") so the implementation plan stays
in lockstep and avoids doc drift, ensuring the {{ max_words }} placeholder
remains intact.

prompts/en/cortex_knowledge_synthesis.md.j2 (1)

13-23: Minor tension between "NEVER Include" block and the participant-role carve-out.

The structure reads: "NEVER Include … Who the agent is, its role …" then after the bulleted exclusions a separate paragraph on line 23 re-permits participant/user role statements. That's logically consistent (agent role ≠ participant role), but a reader scanning the "NEVER Include" header may miss the carve-out, and the LLM may over-index on the prohibition. Consider either:

• Moving the line-23 clarifier into the Include ONLY section next to the existing participant-role bullet (line 8), or
• Scoping the exclusion bullet explicitly, e.g. "Who the agent is, the agent's role…" to make it clear the prohibition is about the agent's own identity, not participants'.

📝 Suggested wording tweak

-- Who the agent is, its role, or the company description (Layer 1: Identity)
+- Who the agent is, the agent's own role, or the company description (Layer 1: Identity)

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@prompts/en/cortex_knowledge_synthesis.md.j2` around lines 13 - 23, The "NEVER
Include" block is ambiguous about participant roles vs agent identity; update
the template so readers and the LLM can't misinterpret it by either moving the
existing participant-role clarifier (the sentence on line 23) into the "Include
ONLY" section next to the participant-role bullet, or explicitly scope the
exclusion bullet to the agent (change the bullet text to something like "Who the
agent is, the agent's role, or the company description") so the prohibition
clearly applies to the agent's identity and not participants.

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Nitpick comments:
In `@docs/design-docs/working-memory-implementation-plan.md`:
- Around line 666-681: The inline prompt snippet in the "Synthesize the agent's
long-term knowledge..." block diverges from the actual template
prompts/en/cortex_knowledge_synthesis.md.j2 (it omits the "Include ONLY"/"NEVER
Include" sections, the sentence forbidding restating the agent's identity/recent
activity, and the final clarifier); update the snippet to exactly mirror
prompts/en/cortex_knowledge_synthesis.md.j2 or replace the block with a
single-line reference pointing to that template (e.g., "See
prompts/en/cortex_knowledge_synthesis.md.j2") so the implementation plan stays
in lockstep and avoids doc drift, ensuring the {{ max_words }} placeholder
remains intact.

In `@prompts/en/cortex_knowledge_synthesis.md.j2`:
- Around line 13-23: The "NEVER Include" block is ambiguous about participant
roles vs agent identity; update the template so readers and the LLM can't
misinterpret it by either moving the existing participant-role clarifier (the
sentence on line 23) into the "Include ONLY" section next to the
participant-role bullet, or explicitly scope the exclusion bullet to the agent
(change the bullet text to something like "Who the agent is, the agent's role,
or the company description") so the prohibition clearly applies to the agent's
identity and not participants.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: 80245db4-2cbd-45e2-b244-f27a25300f72

📥 Commits

Reviewing files that changed from the base of the PR and between 0c55f54 and e20c5fb.

📒 Files selected for processing (4)

• docs/design-docs/working-memory-implementation-plan.md
• docs/design-docs/working-memory-triage.md
• prompts/en/cortex_knowledge_synthesis.md.j2
• src/prompts/engine.rs

[jamiepine]

LGTM. Good call keeping role facts here until participant context owns them.