fix(honcho): disable AI self-observation by default

[lebedyncrs]

Problem

With ai_observe_me: true, Honcho treats the AI peer's own messages as input for its self-representation. When the model answers questions about the user, it repeats user facts in its replies and those facts land under AI Self-Representation instead of the user peer.

Example: user says "I play tennis on Tuesdays" → AI replies "Got it, you play tennis on Tuesdays" → Honcho stores tennis under the AI persona, not the user.

Solution

Set ai_observe_me default to false in:

• HonchoClientConfig dataclass default
• "directional" observation preset
• HonchoSessionManager fallback when no config is passed

AI still observes the user via ai_observe_others: true. Opt back in explicitly:

"observation": { "ai": { "observeMe": true, "observeOthers": true } }

Also maps lebedyncrs in AUTHOR_MAP for contributor attribution CI.

Test plan

• scripts/run_tests.sh tests/honcho_plugin/test_client.py tests/honcho_plugin/test_session.py tests/honcho_plugin/test_empty_profile_hint.py - 207 passed
• test_directional_preset_disables_ai_self_observation
• test_ai_self_observation_off_without_config

[alt-glitch]

Related: #48092 (earlier closed attempt at the same fix, same author). This open PR is the active version — disables ai_observe_me by default so user facts echoed in AI replies no longer land under AI Self-Representation.

[lebedyncrs]

Concrete example: what goes wrong with Honcho roles

Honcho models a conversation as two peers — a user peer and an ai peer — each with its own representation. Two flags control who observes whom:

• observeOthers — model the other peer from their messages
• observeMe — model myself from my own messages

The bug is the AI peer's observeMe: true.

The conversation

User:  I play tennis on Tuesdays.
AI:    Got it — you play tennis on Tuesdays.

What should happen

The user stated a fact about themselves → it belongs to the user peer:

User Representation
  Explicit Observations
    [..] user plays tennis on Tuesdays
AI Self-Representation
  (empty — the AI has no tennis schedule)

What actually happens with ai_observe_me: true

Because the AI observes its own reply ("you play tennis on Tuesdays"), Honcho extracts the fact from the AI's message and files it under the AI's self-representation — as if the assistant plays tennis. This is copied verbatim from a real pre-fix Langfuse trace (session 20260604_230747):

## AI Self-Representation
## Explicit Observations
[2026-05-23 08:06:28] hermes has a tennis event scheduled from 08:00 to 09:00 on May 23, 2026.
[2026-05-23 08:06:28] hermes has a haircut appointment scheduled from 11:00 to 11:45 on May 23, 2026.
[2026-05-23 08:06:28] hermes is providing calendar information for Saturday, May 23, 2026.

Every user fact got attributed to hermes (the AI peer). On later turns this polluted self-model is injected back into context, so the agent reasons about itself having the user's tennis match and haircut.

[teknium1]

Thanks for tracing the AI-peer attribution path and covering both config resolution and the no-config manager fallback. The production premise is confirmed on current main: plugins/memory/honcho/client.py:249-251 sets directional ai_observe_me=True, and plugins/memory/honcho/session.py:205-210 forwards it to the assistant SessionPeerConfig.

Problems

• Several shipped surfaces would remain contradictory after this change: the setup prompt still says directional has all observations on (plugins/memory/honcho/cli.py:830), the optional Honcho skill says all four defaults are on (optional-skills/autonomous-ai-agents/honcho/SKILL.md:93-110), and the checked-in Chinese page does too (website/i18n/zh-Hans/docusaurus-plugin-content-docs/current/user-guide/features/honcho.md:124,166).
• website/docs/user-guide/features/honcho.md:214 still recommends ai.observeMe=true for avoiding reply-derived user facts, which retains the self-observation path this PR is fixing.

Suggested changes

• Update those remaining user-facing descriptions and replace the common-pattern entry with ai.observeMe=false, ai.observeOthers=true; document all-four-flags-on as an explicit opt-in.

Automated hermes-sweeper review.

[GottZ]

This was generated by AI during triage.

Summary

Two PRs address the same Honcho attribution bug by changing the directional/default AI peer configuration from self-observation enabled to ai_observe_me=false while retaining ai_observe_others=true. #48092 contains the core config, fallback, and regression-test changes; #48096 carries the same fix plus partial documentation updates and is the active version.

Related pull requests

• #48092 [closed] duplicate — (+35/-6) — superseded duplicate of #48096: The diff changes the directional preset, HonchoClientConfig, and the no-config session fallback to disable AI self-observation, with tests covering those defaults. Although closed, it remains relevant as the earlier implementation of the same fix and was explicitly superseded by the active #48096.
• #48096 related — (+42/-13) — keep open with a salvage path: The diff preserves the complete core fix and regression coverage from #48092 and additionally updates several English documentation surfaces. This agrees with the maintainer-bot keep_open review, but the salvage path must address its concrete blockers: update plugins/memory/honcho/cli.py:830, optional-skills/autonomous-ai-agents/honcho/SKILL.md:93-110, and website/i18n/zh-Hans/docusaurus-plugin-content-docs/current/user-guide/features/honcho.md:124,166, and correct the contradictory common-pattern guidance at website/docs/user-guide/features/honcho.md:214 to use ai.observeMe=false, ai.observeOthers=true.

Duplicates

#48092 and #48096 implement essentially the same production and test changes; #48096 is the active, broader version, while #48092 is the closed superseded duplicate.

Suggested consolidation

Keep #48096 open with a salvage path: retain its config, session-fallback, tests, and partial documentation changes, then reconcile the remaining user-facing surfaces identified by the maintainer-bot review before reassessment. Keep #48092 closed as a duplicate superseded by #48096.

Complex graph

flowchart LR
    classDef open fill:#dbeafe,stroke:#1d4ed8,color:#1e3a8a
    classDef merged fill:#dcfce7,stroke:#15803d,color:#14532d
    classDef closed fill:#e5e7eb,stroke:#6b7280,color:#1f2937
    classDef unverified fill:#f3f4f6,stroke:#9ca3af,color:#374151
    classDef best stroke-width:3px,stroke:#b45309
    classDef target stroke-width:3px,stroke:#4338ca
    subgraph Dup48092 ["PRs duplicating each other"]
        P48092["PR #48092 (closed)"]
        P48096["PR #48096 (open)"]
    end
    class P48092 closed
    class P48096 open
    class P48096 target
    click P48092 "https://github.com/NousResearch/hermes-agent/pull/48092"
    click P48096 "https://github.com/NousResearch/hermes-agent/pull/48096"

Loading

Graph: solid arrow = fixes / best fix, dashed arrow = partial or unverified (see edge label); boxed group = PRs duplicating each other; amber border = best fix; indigo border = target; gray node = closed (state tag in the node label).

Cross-PR triage: Reviewed 2 pull requests and 0 issues in this complex. Each diff was read against this issue; Assessment working set: 19 kB of PR diffs, 3 kB of issue/PR text, 2 kB of discussion (3 comments), 0 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.