refactor(gateway): use get_custom_provider_context_length() helper + pass custom_providers to @context path by quinnmacro · Pull Request #18844 · NousResearch/hermes-agent

quinnmacro · 2026-05-02T14:00:22Z

Summary

Three targeted fixes for the custom_providers context-length resolution chain:

1. Bug fix: `@context` expansion path ignores `custom_providers`

The @context reference expansion path calls get_model_context_length() without passing custom_providers. This skips step 0b (per-model context_length from custom_providers), so users who configure custom_providers[].models.<m>.context_length without a top-level model.context_length get wrong injection limits.

2. Refactor: replace inline parsing with `get_custom_provider_context_length()` helper

The hygiene-run path has a 27-line inline custom_providers parsing loop. PR #15844 introduced get_custom_provider_context_length() as the single source of truth. This replaces the inline loop with a helper call.

3. Bug fix: `run_conversation()` result dict missing `context_length` field (NEW)

The gateway runtime_footer reads context_length from agent_result, but run_conversation() never included this field in its return dict. This caused the footer's context_pct display to always be skipped (None).

The CLI status bar works fine because _get_status_bar_snapshot() reads context_length directly from context_compressor. But the gateway footer only sees the result dict — so agent_result.get("context_length") was always None.

Fix: Add "context_length" to the result dict returned by run_conversation(), derived from context_compressor.context_length, matching the same source the CLI uses.

New commit: fix/context-length-result-dict branch (1 line change in run_agent.py).

Supersedes fix(context): pass config_context_length to all get_model_context_length() call sites #12630 (old branch had 1035 commits of divergence)
Builds on fix(context): honor custom_providers context_length on /model switch + bump probe tier to 256K #15844 (which added get_custom_provider_context_length() and fixed the main /model switch paths)
Closes Bug: /model switch to named custom provider ignores custom_providers model context_length #15779

Test Plan

@context with large file: Before → hard limit 64K; After → hard limit 100K
Hygiene run: Before → 128K; After → 200K
/modelinfo: Already correct (uses custom_providers= param from fix(context): honor custom_providers context_length on /model switch + bump probe tier to 256K #15844)
/model switch: Already correct (uses resolve_display_context_length from fix(context): honor custom_providers context_length on /model switch + bump probe tier to 256K #15844)
Gateway runtime footer context_pct: Before → always blank; After → displays percentage

alt-glitch · 2026-05-02T14:03:40Z

Related to #15844 (merged), which introduced the helper this PR now uses in the remaining code paths.

alt-glitch · 2026-05-02T14:04:01Z

Related to #15844 (merged), which introduced the helper this PR now uses in the remaining code paths.

quinnmacro · 2026-05-09T15:25:47Z

New commit: `fix/context-length-result-dict`

While investigating why the gateway runtime footer still shows blank context_pct after #15844, I found a deeper bug:

Root cause: run_conversation() in run_agent.py never includes "context_length" in its return dict. The gateway footer reads context_length from agent_result.get("context_length") (run.py L6297), which is always None.

The CLI status bar works because _get_status_bar_snapshot() reads context_length directly from context_compressor — it never goes through the result dict. So #15844 fixed the CLI path but left the gateway path broken.

Fix: 1-line addition in run_agent.py L14637:

"context_length": getattr(self.context_compressor, "context_length", 0) or 0,

Branch: fix/context-length-result-dict on quinnmacro/hermes-upstream-fork

Upstream 678a87c (May 3) replaced the entire __init__.py, losing: 1. Ebbinghaus decay + domain-aware half-life (volatile/normal/stable) 2. Contradiction detection (SUPERSEDE/EXTEND/ADD with keyword-first logic) 3. Freshness tags on search results (🔴/⚠️/⏳ lifecycle states) 4. Self-hosted host detection (_is_self_hosted, added earlier this session) Restored from commit 128b3a820 (Ebbinghaus) + 55ace09b0 (contradiction). Added _enrich_results() for search freshness + contradiction warnings on conclude. 400 lines → 710 lines.

… bypass

…tructor)

…er_id

quinnmacro · 2026-05-30T07:52:50Z

Rebased onto latest `upstream/main` (v2026.5.29)

Clean single commit: +1 line in agent/conversation_loop.py (moved from run_agent.py per the 3-module refactor).

What this fixes (real-world impact)

The gateway runtime footer shows context usage percentage (context_pct) for every message. Since v0.15.x, it's been blank for all users — not because the data is missing, but because run_conversation() returns a dict that doesn't include context_length.

The CLI status bar works fine because it reads context_compressor.context_length directly. But the gateway path goes through the result dict, so agent_result.get("context_length") returns None, and the footer skips the display entirely.

The fix

One line, zero side effects:

"context_length": getattr(agent.context_compressor, "context_length", 0) or 0,

This matches the exact same source the CLI uses (context_compressor.context_length), so the gateway and CLI now show consistent numbers.

What I verified

Path	Before	After
Gateway footer `context_pct`	Always blank (None)	Shows correct %
CLI status bar	Already correct	Unchanged
`/modelinfo` command	Already correct	Unchanged
`/model` switch display	Already correct	Unchanged

This is the last remaining gap from the custom_providers context-length work (#15844 merged). Ready for merge whenever.

quinnmacro · 2026-06-11T08:12:46Z

Rebased onto v2026.6.5 (commit `c94e93a`)

Adapted to the turn_finalizer.py extraction refactor:

context_length field added to result dict in agent/turn_finalizer.py (was conversation_loop.py)
1-line fix, zero side effects

Gateway footer context_pct now shows correct value. Diff: +1/-0.

…l memory Pattern 1: BM25 + Vector dual-path search (bm25_index.py) - Local BM25 index with CJK bigram tokenization - RRF fusion when both paths return results - Fallback to BM25-only when vector API fails - Persistent cache at ~/.hermes/state/mem0_bm25_cache.json Pattern 3: Supersession chains (supersession.py) - Version memories instead of overwrite - Auto-create chain on SUPERSEDE contradiction - Annotate search results with chain info - Persistent storage at ~/.hermes/state/mem0_s...ions/ Inspired by agentmemory BEAM analysis (rohitg00/agentmemory, 22.6K stars)

teknium1

Thanks for identifying the missing footer metadata field. The underlying footer issue is still present on current main, but the PR needs to be narrowed and moved to the current architecture.

Problems

The normal result dictionary moved to agent/turn_finalizer.py:399-428; it still omits context_length after last_prompt_tokens at line 423. The PR edits the former location in agent/conversation_loop.py, so its one-line change does not apply to current main.
The claimed @context and hygiene changes are absent from this PR's changed-file list. Current @context expansion still omits custom_providers at cli.py:12177-12181, and gateway hygiene still has its inline lookup at gateway/run.py:11062-11088.
The bundled Mem0 work is unrelated and predates the current Mem0Backend architecture (plugins/memory/mem0/_backend.py:9-36), making it unsuitable for direct salvage.

Suggested changes

Split the Mem0 commits and move the footer field to agent/turn_finalizer.py, with a regression test through gateway/run.py:11706-11713.
Submit the @context and hygiene changes as explicit, tested current-main edits if they remain intended.

Automated hermes-sweeper review.

teknium1 · 2026-07-12T21:10:58Z

@@ -4528,6 +4528,7 @@ def _stop_spinner():
        "completion_tokens": agent.session_completion_tokens,


This result dictionary was extracted after the PR branch point. On current main the normal return is built in agent/turn_finalizer.py:399-428, so please place this field there and add a regression that verifies the gateway consumer at gateway/run.py:11706-11713 receives it.

alt-glitch added type/bug Something isn't working P2 Medium — degraded but workaround exists comp/gateway Gateway runner, session dispatch, delivery labels May 2, 2026

dinedal mentioned this pull request May 3, 2026

fix(cli): honor custom context lengths in model switch display #19034

Closed

19 tasks

quinnmacro force-pushed the refactor/custom-providers-context-length-helper branch from f0918e8 to b764842 Compare May 15, 2026 00:58

quinnmacro added 5 commits May 29, 2026 23:20

feat(mem0): SDK 2.0.2 upgrade + selfhost_patch for Project validation…

df8c576

… bypass

fix(mem0): move selfhost_patch to plugin dir (survives pip upgrades)

4c3fef2

fix(mem0): drop org_id/project_id kwargs (SDK 2.0.2 removed from cons…

58c9211

…tructor)

fix(mem0): self-hosted forces config user_id over gateway platform us…

e009efc

…er_id

quinnmacro force-pushed the refactor/custom-providers-context-length-helper branch 2 times, most recently from 1549dba to 1c7ebc2 Compare May 30, 2026 07:50

banditburai mentioned this pull request Jun 1, 2026

fix(custom_providers): resolve per-model context_length for publisher/slug model ids #36852

Open

quinnmacro force-pushed the refactor/custom-providers-context-length-helper branch from 1c7ebc2 to d4f02d1 Compare June 11, 2026 08:10

quinnmacro added 2 commits June 13, 2026 22:03

fix(gateway): add context_length to run_conversation result dict

97a77d4

quinnmacro force-pushed the refactor/custom-providers-context-length-helper branch from d4f02d1 to 97a77d4 Compare June 20, 2026 10:00

teknium1 reviewed Jul 12, 2026

View reviewed changes

GottZ mentioned this pull request Aug 3, 2026

Context fill rate calculation uses hardcoded 256k instead of config limits #43462

Open

Navigation Menu

Uh oh!

refactor(gateway): use get_custom_provider_context_length() helper + pass custom_providers to @context path - #18844

refactor(gateway): use get_custom_provider_context_length() helper + pass custom_providers to @context path#18844
quinnmacro wants to merge 7 commits into
NousResearch:mainfrom
quinnmacro:refactor/custom-providers-context-length-helper

quinnmacro commented May 2, 2026 •

edited

Loading

Uh oh!

alt-glitch commented May 2, 2026

Uh oh!

alt-glitch commented May 2, 2026

Uh oh!

quinnmacro commented May 9, 2026

Uh oh!

quinnmacro commented May 30, 2026

Uh oh!

quinnmacro commented Jun 11, 2026

Uh oh!

teknium1 left a comment

Uh oh!

teknium1 Jul 12, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

		@@ -4528,6 +4528,7 @@ def _stop_spinner():
		"completion_tokens": agent.session_completion_tokens,

Uh oh!

Conversation

quinnmacro commented May 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

1. Bug fix: @context expansion path ignores custom_providers

2. Refactor: replace inline parsing with get_custom_provider_context_length() helper

3. Bug fix: run_conversation() result dict missing context_length field (NEW)

Related

Test Plan

Uh oh!

alt-glitch commented May 2, 2026

Uh oh!

alt-glitch commented May 2, 2026

Uh oh!

quinnmacro commented May 9, 2026

New commit: fix/context-length-result-dict

Uh oh!

quinnmacro commented May 30, 2026

Rebased onto latest upstream/main (v2026.5.29)

What this fixes (real-world impact)

The fix

What I verified

Uh oh!

quinnmacro commented Jun 11, 2026

Rebased onto v2026.6.5 (commit c94e93a)

Uh oh!

teknium1 left a comment

Choose a reason for hiding this comment

Problems

Suggested changes

Uh oh!

teknium1 Jul 12, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

quinnmacro commented May 2, 2026 •

edited

Loading

1. Bug fix: `@context` expansion path ignores `custom_providers`

2. Refactor: replace inline parsing with `get_custom_provider_context_length()` helper

3. Bug fix: `run_conversation()` result dict missing `context_length` field (NEW)

New commit: `fix/context-length-result-dict`

Rebased onto latest `upstream/main` (v2026.5.29)

Rebased onto v2026.6.5 (commit `c94e93a`)