Skip to content

fix(minimax): drop stale ≤204,800 cache entries for MiniMax-M3 - #36726

Merged
teknium1 merged 1 commit into
mainfrom
hermes/hermes-8df49e6f
Jun 1, 2026
Merged

fix(minimax): drop stale ≤204,800 cache entries for MiniMax-M3#36726
teknium1 merged 1 commit into
mainfrom
hermes/hermes-8df49e6f

Conversation

@teknium1

@teknium1 teknium1 commented Jun 1, 2026

Copy link
Copy Markdown
Contributor

Summary

MiniMax-M3 users on a stale context-length cache no longer see 204K — the cached value is dropped and re-resolved to its real 1M context.

Root cause: pre-catalog builds resolved M3 via the generic minimax catch-all (204,800) and persisted it. Step 1 of get_model_context_length returns a cached MiniMax-M3@<base_url> entry directly, before reaching the minimax-m3 (1M) catalog entry — so anyone who first probed M3 on an older build stuck at 204K forever (reported via Telegram /new showing Context: 204K tokens (detected) on the minimax coding plan, cn).

Changes

  • agent/model_metadata.py: add _model_name_suggests_minimax_m3() + a step-1 cache-invalidation branch mirroring the existing Kimi/Codex guards — drop cached entries <= 204,800 for M3 slugs and re-resolve.
  • tests/agent/test_minimax_provider.py: M3→1M resolution test + stale-cache guard tests (drop stale, preserve correct 1M, leave M2.x 204,800 untouched).

Validation

Scenario Before After
M3, stale 204,800 cache 204,800 1,000,000
M3, correct 1M cache 1,000,000 1,000,000
M2.x, 204,800 cache (correct) 204,800 204,800

tests/agent/test_minimax_provider.py: 47 passed. Related context-length/cache tests: 263 passed.

Infographic

minimax-m3-stale-cache-guard

@teknium1
fix(minimax): drop stale ≤204,800 cache entries for MiniMax-M3
992c451 
M3 is 1M context, but pre-catalog builds resolved it via the generic
'minimax' catch-all (204,800) and persisted that to the context-length
cache. Step 1 of get_model_context_length returned the cached value
directly before reaching the 'minimax-m3' (1M) catalog entry, so users
who first probed M3 on an older build were stuck at 204K forever (e.g.
/new in the Telegram gateway showing 'Context: 204K tokens (detected)').

Mirror the existing Kimi/Codex stale-cache guards: when a cached entry
for a minimax-m3 slug is <= 204,800, drop it and re-resolve. M2.x slugs
(correctly 204,800) are untouched since they don't match the M3 name.
@github-actions

github-actions Bot commented Jun 1, 2026

Copy link
Copy Markdown
Contributor

🔎 Lint report: hermes/hermes-8df49e6f vs origin/main

ruff

Total: 0 on HEAD, 0 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 0 pre-existing issues carried over.

ty (type checker)

Total: 9582 on HEAD, 9582 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 4966 pre-existing issues carried over.

Diagnostics are surfaced as warnings — this check never fails the build.

@alt-glitch alt-glitch added type/bug Something isn't working provider/minimax MiniMax (Anthropic transport) comp/agent Core agent runtime: loop, agent_init, prompt builder, context-compression, responses endpoint P3 Low — cosmetic, nice to have labels Jun 1, 2026
tonydwb
tonydwb approved these changes Jun 1, 2026
View reviewed changes

@tonydwb tonydwb left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Code Review Summary

Verdict: Approved

✅ Looks Good

  • Well-scoped fix: Invalidates stale ≤204,800 cache entries for MiniMax-M3 models that were seeded before the dedicated 1M entry existed
  • Good heuristic: _model_name_suggests_minimax_m3 catches multiple naming variants (MiniMax-M3, minimax/minimax-m3, minimax-m3)
  • Thorough tests:
    • M3 models now correctly resolve to 1M-class context
    • M2.x models still resolve to 204,800 (regression guard)
    • Stale cache guard specifically tested
  • No performance concern: The check only triggers on the cache-hit branch and the regex is cheap

Reviewed by Hermes Agent

mxnstrexgl
mxnstrexgl approved these changes Jun 1, 2026
View reviewed changes

@mxnstrexgl mxnstrexgl left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🤖 Automated PR Review

Security Scan

  • ✓ No hardcoded secrets, injection sinks, unsafe deserialization, or dependency red flags found by this automated scan.

Code Quality

  • ✓ No blocking code-quality issues found by this automated scan.

Summary

Status: APPROVE — security findings: 0, quality suggestions: 0.

Automated review; raw diff content intentionally omitted.

@teknium1
teknium1 merged commit 1ffa22e into main Jun 1, 2026
24 of 26 checks passed
@teknium1
teknium1 deleted the hermes/hermes-8df49e6f branch June 1, 2026 21:59
@ignaciolagosruiz ignaciolagosruiz mentioned this pull request Jun 2, 2026
Closed
@semperaug semperaug mentioned this pull request Jun 11, 2026
Closed
alt-glitch pushed a commit that referenced this pull request Jun 14, 2026
@teknium1
fix(minimax): drop stale ≤204,800 cache entries for MiniMax-M3 (#36726)
3704adb 
M3 is 1M context, but pre-catalog builds resolved it via the generic
'minimax' catch-all (204,800) and persisted that to the context-length
cache. Step 1 of get_model_context_length returned the cached value
directly before reaching the 'minimax-m3' (1M) catalog entry, so users
who first probed M3 on an older build were stuck at 204K forever (e.g.
/new in the Telegram gateway showing 'Context: 204K tokens (detected)').

Mirror the existing Kimi/Codex stale-cache guards: when a cached entry
for a minimax-m3 slug is <= 204,800, drop it and re-resolve. M2.x slugs
(correctly 204,800) are untouched since they don't match the M3 name.
kossteg pushed a commit to kossteg/hermes-agent that referenced this pull request Jun 16, 2026
@teknium1 @kossteg
fix(minimax): drop stale ≤204,800 cache entries for MiniMax-M3 (NousR…
5e1774d 
…esearch#36726)

M3 is 1M context, but pre-catalog builds resolved it via the generic
'minimax' catch-all (204,800) and persisted that to the context-length
cache. Step 1 of get_model_context_length returned the cached value
directly before reaching the 'minimax-m3' (1M) catalog entry, so users
who first probed M3 on an older build were stuck at 204K forever (e.g.
/new in the Telegram gateway showing 'Context: 204K tokens (detected)').

Mirror the existing Kimi/Codex stale-cache guards: when a cached entry
for a minimax-m3 slug is <= 204,800, drop it and re-resolve. M2.x slugs
(correctly 204,800) are untouched since they don't match the M3 name.
T02200059 pushed a commit to T02200059/hermes-agent that referenced this pull request Jun 18, 2026
@teknium1
fix(minimax): drop stale ≤204,800 cache entries for MiniMax-M3 (NousR…
331902b 
…esearch#36726)

M3 is 1M context, but pre-catalog builds resolved it via the generic
'minimax' catch-all (204,800) and persisted that to the context-length
cache. Step 1 of get_model_context_length returned the cached value
directly before reaching the 'minimax-m3' (1M) catalog entry, so users
who first probed M3 on an older build were stuck at 204K forever (e.g.
/new in the Telegram gateway showing 'Context: 204K tokens (detected)').

Mirror the existing Kimi/Codex stale-cache guards: when a cached entry
for a minimax-m3 slug is <= 204,800, drop it and re-resolve. M2.x slugs
(correctly 204,800) are untouched since they don't match the M3 name.
waefrebeorn pushed a commit to waefrebeorn/slermes that referenced this pull request Jul 2, 2026
@teknium1
fix(minimax): drop stale ≤204,800 cache entries for MiniMax-M3 (NousR…
657a108 
…esearch#36726)

M3 is 1M context, but pre-catalog builds resolved it via the generic
'minimax' catch-all (204,800) and persisted that to the context-length
cache. Step 1 of get_model_context_length returned the cached value
directly before reaching the 'minimax-m3' (1M) catalog entry, so users
who first probed M3 on an older build were stuck at 204K forever (e.g.
/new in the Telegram gateway showing 'Context: 204K tokens (detected)').

Mirror the existing Kimi/Codex stale-cache guards: when a cached entry
for a minimax-m3 slug is <= 204,800, drop it and re-resolve. M2.x slugs
(correctly 204,800) are untouched since they don't match the M3 name.
santhreal pushed a commit to santhreal/hermes-agent that referenced this pull request Jul 13, 2026
@teknium1
fix(minimax): drop stale ≤204,800 cache entries for MiniMax-M3 (NousR…
48b1b64 
…esearch#36726)

M3 is 1M context, but pre-catalog builds resolved it via the generic
'minimax' catch-all (204,800) and persisted that to the context-length
cache. Step 1 of get_model_context_length returned the cached value
directly before reaching the 'minimax-m3' (1M) catalog entry, so users
who first probed M3 on an older build were stuck at 204K forever (e.g.
/new in the Telegram gateway showing 'Context: 204K tokens (detected)').

Mirror the existing Kimi/Codex stale-cache guards: when a cached entry
for a minimax-m3 slug is <= 204,800, drop it and re-resolve. M2.x slugs
(correctly 204,800) are untouched since they don't match the M3 name.
donbowman pushed a commit to donbowman/hermes-agent that referenced this pull request Jul 13, 2026
@teknium1 @donbowman
fix(minimax): drop stale ≤204,800 cache entries for MiniMax-M3 (NousR…
5e122ba 
…esearch#36726)

M3 is 1M context, but pre-catalog builds resolved it via the generic
'minimax' catch-all (204,800) and persisted that to the context-length
cache. Step 1 of get_model_context_length returned the cached value
directly before reaching the 'minimax-m3' (1M) catalog entry, so users
who first probed M3 on an older build were stuck at 204K forever (e.g.
/new in the Telegram gateway showing 'Context: 204K tokens (detected)').

Mirror the existing Kimi/Codex stale-cache guards: when a cached entry
for a minimax-m3 slug is <= 204,800, drop it and re-resolve. M2.x slugs
(correctly 204,800) are untouched since they don't match the M3 name.
Gravezzz pushed a commit to Gravezzz/hermes-agent that referenced this pull request Jul 21, 2026
@teknium1
fix(minimax): drop stale ≤204,800 cache entries for MiniMax-M3 (NousR…
2a37fca 
…esearch#36726)

M3 is 1M context, but pre-catalog builds resolved it via the generic
'minimax' catch-all (204,800) and persisted that to the context-length
cache. Step 1 of get_model_context_length returned the cached value
directly before reaching the 'minimax-m3' (1M) catalog entry, so users
who first probed M3 on an older build were stuck at 204K forever (e.g.
/new in the Telegram gateway showing 'Context: 204K tokens (detected)').

Mirror the existing Kimi/Codex stale-cache guards: when a cached entry
for a minimax-m3 slug is <= 204,800, drop it and re-resolve. M2.x slugs
(correctly 204,800) are untouched since they don't match the M3 name.
leewenjie pushed a commit to leewenjie/hermes-agent that referenced this pull request Aug 7, 2026
@teknium1
fix(minimax): drop stale ≤204,800 cache entries for MiniMax-M3 (NousR…
352a602 
…esearch#36726)

M3 is 1M context, but pre-catalog builds resolved it via the generic
'minimax' catch-all (204,800) and persisted that to the context-length
cache. Step 1 of get_model_context_length returned the cached value
directly before reaching the 'minimax-m3' (1M) catalog entry, so users
who first probed M3 on an older build were stuck at 204K forever (e.g.
/new in the Telegram gateway showing 'Context: 204K tokens (detected)').

Mirror the existing Kimi/Codex stale-cache guards: when a cached entry
for a minimax-m3 slug is <= 204,800, drop it and re-resolve. M2.x slugs
(correctly 204,800) are untouched since they don't match the M3 name.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

2 more reviewers
@mxnstrexgl mxnstrexgl mxnstrexgl approved these changes
@tonydwb tonydwb tonydwb approved these changes
Reviewers whose approvals may not affect merge requirements

Assignees

No one assigned

Labels

comp/agent Core agent runtime: loop, agent_init, prompt builder, context-compression, responses endpoint P3 Low — cosmetic, nice to have provider/minimax MiniMax (Anthropic transport) type/bug Something isn't working

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

4 participants

@teknium1 @mxnstrexgl @tonydwb @alt-glitch