fix(deriver): prevent cross-subject self-observations

[Diaspar4u]

Summary

Fix deriver subject attribution without disabling peer self-observation or filtering stored messages.

• distinguish the message speaker from the fact's subject;
• allow target-specific derivation to retain facts supplied by any speaker;
• require the structured deriver output to classify whether each conclusion is a durable target fact;
• discard transient cross-subject speech/knowledge conclusions before persistence.

Motivation

With self-observation enabled, an assistant reply such as You play tennis on Tuesdays can produce an AI self-conclusion such as assistant knows the user plays tennis on Tuesdays, or worse, misattribute the tennis schedule directly to the assistant. Disabling self-observation prevents the corruption but also removes useful AI identity, capability, and preference learning.

This fixes the attribution boundary while preserving assistant messages, cross-peer observation, and genuine AI self-representation.

Implementation

The existing single deriver call now returns is_durable_target_fact for each explicit conclusion. Only conclusions classified as durable facts about the target peer are converted into persistent observations. The prompt includes speaker/subject rules and paired positive and negative examples.

This deliberately does not apply the SQL speaker filter proposed in #823. Interleaved messages remain available so another speaker can provide valid evidence about the target peer, matching the maintainer guidance on that PR.

End-to-end enqueue promotion for assistant-authored facts about the other participant is provided by the stacked follow-up #944.

Validation

• Added prompt, cache-prefix, and persistence-filter tests.
• Ruff and focused basedpyright checks pass.
• Repeated live model tests confirmed:
  • an echoed user tennis schedule is rejected from AI self-memory;
  • a genuine AI preference remains in AI self-memory;
  • when evaluated with the user as target, an assistant-authored user schedule is accepted as a durable user fact.

Fixes #817.
Supersedes the attribution approach in #823 while preserving its reported reproduction and maintainer feedback.

Summary by CodeRabbit

• New Features
  • Improved fact extraction to identify durable facts specifically about the target peer.
  • Added clearer handling for speaker attribution, pronouns, ambiguous subjects, and statements about others.
  • Excludes transient communication states and non-durable observations from results.
  • Returns no observations when no valid durable target-peer facts are found.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 6b57fa89-17ac-45b9-a827-c171127a5b86

📥 Commits

Reviewing files that changed from the base of the PR and between 1baed39 and a0e725f.

📒 Files selected for processing (9)

• src/deriver/prompts.py
• src/utils/representation.py
• tests/conftest.py
• tests/deriver/test_deriver_processing.py
• tests/deriver/test_prompts.py
• tests/deriver/test_representation_crud.py
• tests/integration/test_representation.py
• tests/integration/test_token_metrics.py
• tests/utils/test_length_finish_reason.py

🚧 Files skipped from review as they are similar to previous changes (8)

• src/deriver/prompts.py
• src/utils/representation.py
• tests/deriver/test_deriver_processing.py
• tests/utils/test_length_finish_reason.py
• tests/deriver/test_prompts.py
• tests/conftest.py
• tests/integration/test_token_metrics.py
• tests/integration/test_representation.py

---

Walkthrough

The deriver prompt now applies explicit speaker and subject attribution rules, marks durable target-peer facts, and excludes non-durable observations during representation conversion. Tests and fixtures adopt the new prompt-specific observation model and validate filtering behavior.

Changes

Durable target observations

Layer / File(s)	Summary
Prompt attribution and durability rules src/deriver/prompts.py, tests/deriver/test_prompts.py	The prompt distinguishes speakers from subjects, excludes ambiguous or transient observations, and defines durable target-fact labeling. Tests verify these rules and preserve the static prompt prefix.
Prompt representation and filtering src/utils/representation.py	Adds PromptExplicitObservation with is_durable_target_fact and converts only durable prompt observations into Representation.explicit.
Fixtures and conversion validation tests/conftest.py, tests/deriver/test_deriver_processing.py, tests/deriver/test_representation_crud.py, tests/integration/test_representation.py, tests/integration/test_token_metrics.py, tests/utils/test_length_finish_reason.py	Updates mocks, fixtures, integration inputs, and CRUD tests for the new observation model, including coverage for dropping non-durable facts.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Sequence Diagram(s)

sequenceDiagram
  participant Messages
  participant DeriverPrompt
  participant PromptRepresentation
  participant Representation
  Messages->>DeriverPrompt: provide speaker-labeled messages
  DeriverPrompt->>PromptRepresentation: emit target-peer observations with durability flags
  PromptRepresentation->>Representation: convert only durable target facts

Loading

Possibly related PRs

• plastic-labs/honcho#806: Both modify minimal_deriver_prompt and target-peer framing.
• plastic-labs/honcho#609: Both modify the prompt construction path in src/deriver/prompts.py.

Suggested reviewers: rajat-ahuja1997

Poem

I’m a rabbit guarding facts tonight,
Speakers and subjects now align right.
Durable truths pass through the gate,
Fleeting chatter has to wait.
Hop, hop—memories shine bright!

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly matches the main change: fixing deriver attribution to avoid cross-subject self-observations.
Linked Issues check	✅ Passed	The prompt, representation, and tests all address #817 by separating speakers from subjects and filtering non-durable or ambiguous observations.
Out of Scope Changes check	✅ Passed	The changes stay focused on deriver attribution, prompt wording, and related test updates with no obvious unrelated scope.
Docstring Coverage	✅ Passed	Docstring coverage is 80.00% which is sufficient. The required threshold is 80.00%.

✨ 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.}

[coderabbitai]

Actionable comments posted: 1

🧹 Nitpick comments (1)

src/deriver/prompts.py (1)

69-87: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low value

Wrap the changed prompt-source lines.

Lines 70, 72, 73, 77, and 87 exceed the repository’s 88-character limit. As per coding guidelines, Python code must keep lines within 88 characters to match Black-compatible formatting.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/deriver/prompts.py` around lines 69 - 87, Wrap the overlong prompt-source
lines in the changed prompt text, specifically the lines corresponding to 70,
72, 73, 77, and 87, so each remains within the repository’s 88-character limit
while preserving the exact prompt wording and behavior.

Source: Coding guidelines

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@src/deriver/prompts.py`:
- Around line 69-71: Make the prompt and structured-output schema consistently
accept durable facts about the target peer when directly supported by any
speaker’s message. In src/deriver/prompts.py lines 69-71, revise the
explicit-fact definition to use target-peer scope rather than speaker/user-only
scope; in lines 80-82, replace inferred or unanchored examples such as fixed
birthday, NYC residence, and education equivalence with directly supported
observations only. In src/utils/representation.py lines 130-133, update the
field description and inference guidance to match this any-speaker, target-peer
evidence contract.

---

Nitpick comments:
In `@src/deriver/prompts.py`:
- Around line 69-87: Wrap the overlong prompt-source lines in the changed prompt
text, specifically the lines corresponding to 70, 72, 73, 77, and 87, so each
remains within the repository’s 88-character limit while preserving the exact
prompt wording and behavior.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: f41a2a28-87c1-43da-9ba5-2dde31f0518b

📥 Commits

Reviewing files that changed from the base of the PR and between 3ee890f and 5cac360.

📒 Files selected for processing (9)

• src/deriver/prompts.py
• src/utils/representation.py
• tests/conftest.py
• tests/deriver/test_deriver_processing.py
• tests/deriver/test_prompts.py
• tests/deriver/test_representation_crud.py
• tests/integration/test_representation.py
• tests/integration/test_token_metrics.py
• tests/utils/test_length_finish_reason.py