fix(memory): preserve system prompt in honcho_reasoning truncation (#59469)

[SquabbyZ]

Fixes #59469

Problem

honcho_reasoning tool results longer than dialecticMaxChars (default 600) were silently truncated to the auto-injection budget. The model deliberately asked Honcho for a full synthesized answer but the response came back clipped mid-word with a trailing " …" and no warning — the system-prompt injection guardrail was wrongly applied to an explicit tool result.

Fix

dialectic_query() in plugins/memory/honcho/session.py now accepts apply_injection_cap: bool = True:

• Auto-injection path (_run_dialectic_depth → dialectic_query) keeps the default True so the per-turn system-prompt supplement is still bounded — this is the path the cap was originally designed for.
• honcho_reasoning tool handler in plugins/memory/honcho/__init__.py passes apply_injection_cap=False so an explicit call returns Honcho's full synthesized answer untouched.

When the cap is applied and truncation occurs, the result now ends with … [truncated, full result in logs] instead of a bare ellipsis, and a WARN log names the original length and cap so operators can tune the budget or fix the over-long prompt. The full pre-truncation text is also emitted at DEBUG level. No more silent failure.

Backward compatible: the default kwarg preserves the existing auto-injection behavior, so the only caller-visible changes are the more explicit truncation marker and the warning log.

Tests

tests/plugins/test_honcho_truncation.py — 7 tests, all passing:

1. Short result → no truncation regardless of cap setting.
2. Long result with cap on (auto-injection path) → truncated, marker visible.
3. Long result with cap off (honcho_reasoning tool path) → returned untouched.
4. Truncation emits WARN with original length and cap value.
5. Short result does not log a false-positive WARN.
6. Long result with cap off does not log a WARN.
7. Default apply_injection_cap=True preserves prior auto-injection behavior.

Existing honcho tests (22 passed, 1 skipped, no regressions) confirm the default-path semantics are unchanged.

Files

• plugins/memory/honcho/session.py — dialectic_query() signature + truncation logic
• plugins/memory/honcho/__init__.py — honcho_reasoning handler passes apply_injection_cap=False
• tests/plugins/test_honcho_truncation.py — new test file

---

AI-assisted fix by https://github.com/SquabbyZ/peaks-loop

[alt-glitch]

This was generated by AI during triage.

Duplicate of #59471 — same mechanism (apply_injection_cap flag on dialectic_query() scoping the dialecticMaxChars truncation to the auto-injection path only; the honcho_reasoning tool handler passes apply_injection_cap=False) and same code sites fixing #59469; #59471 was opened earlier and is canonical. The clearer truncation marker + WARN log here don't change the underlying fix. Related: #59470 / #59665 (the distinct 250-token tier-selection facet).

[teknium1]

Thanks for the focused fix. The underlying premise is live on current main: plugins/memory/honcho/session.py:658-660 applies dialecticMaxChars to all dialectic queries, while plugins/memory/honcho/__init__.py:1358-1362 sends explicit honcho_reasoning through that method.

Problems

• In plugins/memory/honcho/session.py:683, the code slices up to cap and then appends … [truncated, full result in logs]. That can exceed the configured injection budget, whose documented contract is a maximum at website/docs/user-guide/features/honcho.md:124.
• tests/plugins/test_honcho_truncation.py tests manager calls directly, but does not prove the tool handler supplies apply_injection_cap=False. The existing handler-level assertion is tests/honcho_plugin/test_session.py:488-506.

Suggested changes

• Reserve suffix space before slicing, or retain the compact existing suffix, and assert the capped result remains within the configured budget.
• Extend the existing honcho_reasoning dispatch test to assert apply_injection_cap=False.

Automated hermes-sweeper review.

[teknium1]

dialecticMaxChars is documented as the maximum injected result length, but this slices to cap and then appends a much longer marker. Please reserve suffix capacity before slicing (and add a cap-boundary assertion) so the auto-injection guardrail remains bounded.

[teknium1]

This is now fixed on main via #62290 (commit 8d1c96f), which adopted the earliest submission for #59469 (#59471 by @vizi0uz). dialecticMaxChars is now scoped to automatic context injection, so explicit honcho_reasoning results return Honcho's full answer rather than being truncated at all. Thanks for the fix, sorry it collided with an earlier duplicate.