fix(memory): align Honcho routing and memory guidance

[nepenth]

Summary

This PR fixes Honcho memory/profile routing and built-in memory guidance.

• Align Honcho peer-card writes with the same observer/target route used by profile reads, so honcho_profile(peer="user", card=[...]) updates the card that honcho_profile(peer="user") returns.
• Align Honcho prefetch/session context routing with manual profile routing when assistant-observes-user mode is enabled, so auto-injected context and explicit tools see the same peer perspective/target.
• Clarify honcho_profile card replacement semantics: passing card overwrites the whole peer card, so callers MUST include facts they want preserved.
• Remove the advertised honcho_context.query parameter for now because the handler did not implement query filtering; this avoids a misleading schema until filtering is added deliberately.
• Improve built-in memory-full guidance so agents use memory.replace / memory.remove, skills, or canonical workspace artifacts instead of claiming to save exact state “mentally” or relying on semantic memory as the only canonical store.
• Keep memory limit defaults aligned across config/runtime/migration surfaces and add broader drift coverage.

Tests

• python -m pytest tests/honcho_plugin/test_session.py tests/tools/test_memory_tool.py tests/tools/test_memory_tool_import_fallback.py -q -o 'addopts=' → 162 passed
• Targeted regressions added/covered:
  • peer-card write/read observer-target alignment
  • prefetch auto-injected context routing for assistant-observes-user mode
  • session-context peer perspective/target routing
  • honcho_profile handler empty-card overwrite behavior
  • honcho_profile schema warning that card overwrites the entire card
  • honcho_context schema no longer advertising an unimplemented query filter
  • memory-full overflow guidance naming memory.replace, memory.remove, skills, and canonical artifacts
  • memory limit defaults checked across DEFAULT_CONFIG, MemoryStore, and the OpenClaw migration helper
• python -m py_compile plugins/memory/honcho/__init__.py plugins/memory/honcho/session.py tools/memory_tool.py hermes_cli/config.py run_agent.py hermes_cli/tips.py optional-skills/migration/openclaw-migration/scripts/openclaw_to_hermes.py
• git diff --check
• Static added-line scan for obvious secret/shell/eval/pickle/SQL-risk patterns → clean
• Independent local review pass performed on the patch set before push.

Security / platform notes

• No credentials, tokens, network endpoints, or provider secrets are added.
• No shell execution, subprocess behavior, SQL generation, filesystem deletion, or migration side effects are added.
• The Honcho behavior change is limited to client-side peer routing/schema/guidance; it does not change Honcho server storage semantics.
• The memory-limit drift test imports the migration helper for constants only; it does not execute a migration.
• This PR intentionally does NOT implement honcho_context(query=...) filtering; the misleading schema field is removed instead. Query filtering can be a focused follow-up.

Review feedback remediation (2026-04-30)

• Added regression coverage for honcho_profile(peer="ai", card=[...]) round-tripping through the same observer/target route used by reads.
• Confirmed existing regression coverage for card=[] clearing peer cards instead of being treated as a read.
• Confirmed honcho_context no longer advertises the unimplemented query schema parameter.
• Updated the PR title/scope language to make Honcho routing and memory guidance/default alignment explicit.

Relationship to context-safety work

Related but independent from the gateway context-safety PRs. This PR fixes Honcho memory/profile routing consistency so long-term context is read and written through the correct observer/target path.

Adjacent PRs:

• fix(matrix): preserve named room identity #17702 fixes Matrix named-room identity.
• feat(gateway): expose normalized channel identity #17711 exposes normalized channel identity.
• feat(gateway): guard repo side effects with workspace binding #17938 uses workspace bindings to guard repo side effects.

[teknium1]

Thanks for the detailed routing coverage. The underlying remaining bugs are still present on current main, but this April patch needs selective salvage rather than a mechanical apply.

Problems

• plugins/memory/honcho/session.py:660 changes prefetch routing but drops search_query=user_message or None. Current main deliberately added that behavior in 0a7cc85eab299667777489a31995eb324d8b7818 and uses it at plugins/memory/honcho/session.py:736; retain it while changing the observer/target route.
• The run_agent.py default hunk no longer reaches the active initialization path. Current main constructs the store in agent/agent_init.py:1341-1344; tools/memory_tool.py:804-818 also has independent no-agent fallback defaults.
• tests/tools/test_memory_tool.py:45-65 snapshots numeric defaults. AGENTS.md:80-83 requires an invariant between surfaces rather than literals expected to change.

Suggested changes

• Preserve search_query=user_message or None and test it together with the observer/target route.
• Update the moved live and no-agent construction paths, then make the drift test compare each surface to DEFAULT_CONFIG.

This is an automated hermes-sweeper review.

[teknium1]

Please retain search_query=user_message or None on this call. Current main intentionally added it in 0a7cc85eab299667777489a31995eb324d8b7818; omitting it would regress topic-relevant prefetch retrieval while fixing the observer route.

[teknium1]

This initialization code has moved on current main to agent/agent_init.py:1341-1344; also update that live path and tools/memory_tool.py:804-818 so the new defaults apply consistently to normal and no-agent memory stores.

[teknium1]

Please avoid freezing these numeric defaults. Derive the expected values from DEFAULT_CONFIG and assert the constructor and migration constants agree, preserving drift coverage without making a routine capacity change fail CI.

[nepenth]

Cleanup triage (2026-07-17)

Status: leave open for selective salvage — do not merge as-is.

• Sweeper (2026-07-12): remaining routing bugs still present on main, but April patch needs selective salvage.
• Blocking notes already on PR: keep search_query=user_message, update live init path (agent/agent_init.py), avoid freezing numeric defaults in tests.

Next if pursued: cherry-pick only routing fixes onto fresh main; drop change-detector default freezes; re-open as a narrow Honcho routing PR if needed.

[nepenth]

Dual-model cleanup decision (2026-07-17)

Reviewers: gpt-5.6-sol xhigh (Codex) + gpt-5.6-sol (Hermes openai-codex).

Decision: SALVAGE_NARROW later — do NOT mechanical-rebase this tip; do NOT roll into Matrix Ship v2 stack.

If/when pursued:

1. Fresh branch from origin/main
2. Port only still-relevant Honcho routing guidance
3. MUST keep search_query=user_message or None (main already uses it)
4. Update live init in agent/agent_init.py (+ memory_tool no-agent path)
5. Invariant tests vs DEFAULT_CONFIG — no frozen numeric defaults

Leave open only as a salvage pointer; no private dogfood of this tip as-is.

[nepenth]

Closing — no longer relevant for our deployment

Closing unmerged.

We are not using Honcho for memory anymore (Hindsight / other paths). This PR was a Honcho routing + built-in memory guidance salvage of an old tip and is not worth re-cutting for us.

If upstream still wants the remaining routing fixes without the April tip’s default freezes, that should be a fresh contribution from current main by someone who still runs Honcho — not this branch.

Not part of Matrix Ship v2. Do not resurrect for our private feature stack.