auxiliary.title.enabled config is ignored — title generation runs regardless

[meatRoll]

Description

The configuration option auxiliary.title.enabled: false (or auxiliary.title_generation.enabled: false) in config.yaml has no effect. Title generation still triggers even when explicitly disabled.

Root Cause

The code that calls maybe_auto_title() does not check the enabled field from the auxiliary config before invoking title generation.

Affected locations:

• cli.py (around line 11150) — the CLI session handler
• gateway/run.py (around line 15960) — the gateway session handler

In both locations, the code unconditionally calls maybe_auto_title() after a successful response:

if response and result and not result.get("failed") and not result.get("partial"):
    try:
        from agent.title_generator import maybe_auto_title
        ...

The enabled key from the auxiliary config dict is never read.

Expected Behavior

When auxiliary.title.enabled: false is set in config.yaml, title generation should be skipped entirely.

Suggested Fix

Before calling maybe_auto_title(), read the config and check the enabled flag:

_aux_cfg = self.config.get("auxiliary", {}) if hasattr(self, "config") and self.config else {}
_title_enabled = True  # default: enabled
for _key in ("title", "title_generation"):
    _sub = _aux_cfg.get(_key, {})
    if isinstance(_sub, dict) and "enabled" in _sub:
        _title_enabled = bool(_sub["enabled"])
        break
if response and result and not result.get("failed") and not result.get("partial") and _title_enabled:
    try:
        from agent.title_generator import maybe_auto_title
        ...

Environment

• Hermes Agent (latest from git)
• Reproduced on both CLI and Gateway modes
• Affects all users who try to disable title generation via config

[liuhao1024]

This is fixed by #41760 — Respects auxiliary.title_generation.enabled config so title generation is properly skipped when disabled.

[2vn9s1c]

Confirmed: enabled: false still ignored after hermes update (v0.16.0, local Ollama + Qwen3 reasoning model)

---

Adding a reproduction report with concrete log evidence, since this bug has significant impact on local setups running reasoning models.

Environment

• Hermes Agent v0.16.0 (2026.6.5) · upstream 72154ad8
• Ollama 0.30.6
• OS: Pop!_OS 22.04, Kernel 6.17
• GPU: RTX 3090 (24 GB VRAM)
• Backend: Ollama local, model qwen3.6:27b-q4_K_M

Config applied

auxiliary:
  title_generation:
    enabled: false
    provider: custom
    model: qwen2.5-coder:0.5b
    base_url: http://127.0.0.1:11434/v1
    timeout: 30
    extra_body:
      think: false

Observed behavior after hermes update

Title generation fires regardless:

2026-06-09 20:41:56,430 INFO  agent.auxiliary_client: Auxiliary title_generation: using main (qwen2.5-coder:0.5b) at http://127.0.0.1:11434/v1/
2026-06-09 20:41:56,590 DEBUG agent.title_generator: Auto-generated session title: Plan en blanco guardado

The enabled: false key is parsed without error (no config validation warning), but maybe_auto_title() is called unconditionally anyway, confirming the check is still missing post-update.

Why this matters beyond cosmetics: reasoning models + no disable = VRAM accumulation

Before switching to qwen2.5-coder:0.5b, the main model (qwen3.6:27b) was being used as the title generation fallback. Because Qwen3 defaults to extended thinking mode, each title generation attempt would generate 1,000–1,200 thinking tokens before timing out at 30 seconds (HTTP 500). Hermes retried 6 times per session end. Combined effect per session:

• ~6 × 30s = ~180s of unexpected GPU activity after every response
• ~9 failed KV cache entries of ~347 MiB each = ~3.1 GB of accumulated VRAM pressure on top of the model's 17.6 GB footprint

This is directly related to #20305 (reasoning models + max_tokens too low) and was only resolved by routing to a non-reasoning auxiliary model — which is itself a workaround for enabled: false not working.

Diagnosis path that exposed the bug

The sequence that led here:

1. Noticed persistent GPU usage (~180s) after Hermes completed responses
2. Traced via OLLAMA_DEBUG=1 to repeated HTTP 500 completions with temp=0.300 and n_tokens≈282 — clearly an auxiliary, not the main agent
3. Confirmed via hermes logs -f | grep auxiliary that title_generation was the culprit
4. Attempted enabled: false as the natural config-level fix
5. Bug confirmed: the key is silently ignored

Current workaround

Routing title_generation to qwen2.5-coder:0.5b (0.5B, no thinking mode) resolves the VRAM issue — generation completes in ~160ms. But the inability to disable the feature via config remains a gap, especially for users who want zero post-response GPU activity or whose auxiliary model swap causes its own latency (Ollama unloads/loads models on swap).

Waiting for PR #41760 to land. Happy to test once merged.

[GottZ]

This was generated by AI during triage.

Summary

Five supplied PRs address #41744, and the #66049 discussion identifies a sixth relevant proposal, #35416. #37349 and #42098 use the correct central title-generator gate, #41760 and #41951 add incomplete caller-level gates, and merged #66049 salvages #37349 while adding current-main compatibility, canonical configuration, tests, and documentation.

Related pull requests

• Honor disabled title generation config #37349 [closed] fixes — (+65/-0) — fixes, superseded by merged fix(title): honor auxiliary.title_generation.enabled to disable auto titles (salvage #37349) #66049: centrally gates both maybe_auto_title() and generate_title(), covering CLI, gateway, TUI, and ACP while preserving manual titles. Despite the keep_open review on Honor disabled title generation config #37349, it remains closed because its implementation was cherry-picked into fix(title): honor auxiliary.title_generation.enabled to disable auto titles (salvage #37349) #66049 with authorship preserved and follow-up fixes.
• fix(auxiliary): respect auxiliary.title_generation.enabled config #41760 [closed] partial — (+127/-3) — partial fix, superseded by fix(title): honor auxiliary.title_generation.enabled to disable auto titles (salvage #37349) #66049: adds duplicated caller-level checks for CLI, gateway, and TUI plus defaults and TUI tests, but leaves the ACP caller uncovered. Despite the keep_open review on fix(auxiliary): respect auxiliary.title_generation.enabled config #41760, closure is justified because merged fix(title): honor auxiliary.title_generation.enabled to disable auto titles (salvage #37349) #66049 fixes the same cause at the shared funnel and avoids the missing-ACP defect.
• fix(title_gen): honor auxiliary.title_generation.enabled config setting (#41744) #41951 [closed] partial — (+123/-57) — partial fix, duplicate of fix(auxiliary): respect auxiliary.title_generation.enabled config #41760 and superseded by fix(title): honor auxiliary.title_generation.enabled to disable auto titles (salvage #37349) #66049: gates only CLI and gateway through auxiliary task configuration, omitting TUI and ACP, while its tests simulate the gate rather than exercising the actual caller paths. The contributor discussion explicitly identified it as less complete than fix(auxiliary): respect auxiliary.title_generation.enabled config #41760.
• fix(agent): honor auxiliary.title_generation.enabled to disable auto titles #42098 [closed] fixes — (+84/-0) — fixes, superseded by merged fix(title): honor auxiliary.title_generation.enabled to disable auto titles (salvage #37349) #66049: correctly gates maybe_auto_title() centrally and contributed the retained optimization of checking configuration after the cheap first-exchange guard. Despite the keep_open review on fix(agent): honor auxiliary.title_generation.enabled to disable auto titles #42098, its diff also uses the write-capable loader, leaves generate_title() ungated, and implements a default-masked alias; fix(title): honor auxiliary.title_generation.enabled to disable auto titles (salvage #37349) #66049 incorporates the sound central-gating idea without those issues.
• fix(title): honor auxiliary.title_generation.enabled to disable auto titles (salvage #37349) #66049 [merged] best fix — (+72/-0) — best fix, merged reference implementation: centrally gates both automatic entry points with the read-only loader and truthy-value coercion, covering CLI, gateway, TUI, and ACP while defaulting to enabled. It also adds default-config, regression-test, example-config, and web-documentation coverage and salvages Honor disabled title generation config #37349 with attribution.

Duplicates

#35416, #37349, #41760, #41951, #42098, and #66049 target the same auxiliary.title_generation.enabled defect; #37349, #42098, and #66049 substantially share the central-gating approach, while #41760 and #41951 are overlapping caller-level variants. #35416 is a known earlier competing proposal identified in the #66049 contributor discussion, but no diff or Verify verdict for it is included in this triage input.

Suggested consolidation

Merge #66049 (already merged) as the consolidated fix: it is the Verify-selected best fix and the complete central implementation across all four surfaces, with read-only configuration access, canonical-key handling, tests, and documentation. Keep #37349, #41760, #41951, and #42098 closed as superseded duplicates; #37349 was directly salvaged into #66049, and #35416 should be acknowledged as the earliest known proposal but cannot receive a diff-level disposition from the evidence supplied here.

Complex graph

flowchart TD
    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
    I41744(["issue #41744 (open)"])
    subgraph Dup37349 ["PRs duplicating each other"]
        P37349["PR #37349 (closed)"]
        P42098["PR #42098 (closed)"]
        P66049["PR #66049 (merged)"]
    end
    subgraph Dup41760 ["PRs duplicating each other"]
        P41760["PR #41760 (closed)"]
        P41951["PR #41951 (closed)"]
    end
    P37349 -->|fixes| I41744
    P41760 -.->|partial| I41744
    P41951 -.->|partial| I41744
    P42098 -->|fixes| I41744
    P66049 -->|best fix| I41744
    class I41744 open
    class I41744 target
    class P37349 closed
    class P41760 closed
    class P41951 closed
    class P42098 closed
    class P66049 merged
    class P66049 best
    click I41744 "https://github.com/NousResearch/hermes-agent/issues/41744"
    click P37349 "https://github.com/NousResearch/hermes-agent/pull/37349"
    click P41760 "https://github.com/NousResearch/hermes-agent/pull/41760"
    click P41951 "https://github.com/NousResearch/hermes-agent/pull/41951"
    click P42098 "https://github.com/NousResearch/hermes-agent/pull/42098"
    click P66049 "https://github.com/NousResearch/hermes-agent/pull/66049"

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 or no verify verdict yet (state tag in the node label).

Cross-PR triage: Reviewed 5 pull requests and 1 issue in this complex. Each diff was read against this issue; Assessment working set: 39 kB of PR diffs, 11 kB of issue/PR text, 10 kB of discussion (12 comments), 10 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.