LearnNewsExamplesServices
Frontmatter
id19334
titleUse a 900k compaction threshold for the 1M Codex profile
stateClosed
labels
enhancementai
assigneesneo-gpt-emmy
createdAtSep 30, 2026, 5:38 PM
updatedAtSep 30, 2026, 5:45 PM
githubUrlhttps://github.com/neomjs/neo/issues/19334
authorneo-gpt-emmy
commentsCount0
parentIssuenull
subIssues[]
subIssuesCompleted0
subIssuesTotal0
contentTrust
projected
quarantined0
signals[]
blockedBy[]
blocking[]
closedAtSep 30, 2026, 5:45 PM

Use a 900k compaction threshold for the 1M Codex profile

neo-gpt-emmy
neo-gpt-emmy commented on Sep 30, 2026, 5:38 PM

Disposition — proposal withdrawn before commit

Fresh runtime evidence supersedes the proposed 900k threshold. On 2026-09-30, both active GPT seats report model_context_window: 828400 in their session token metadata. Their current model catalogs advertise max_context_window: 872000 and effective_context_window_percent: 95; the new seat without the project override reports 258400. The earlier 950k calculation incorrectly applied 95% directly to the requested 1M setting without accounting for that ceiling.

The uncommitted 850k → 900k edit was reverted. No peer-local configuration changed and no PR was opened. A future compaction-policy change needs the active threshold/clamping semantics verified against the actual capacity; the operator's 5–10% suggestion is not being converted into an unsupported raw-window assumption. FM's missing repository context defaults remain a separate, confirmed defect.

The original proposal is retained below as history, not an active prescription.


Context

During FM onboarding on 2026-09-30, the operator found that the new Codex seat used the model catalog's 258,400-token effective default while established Neo checkouts configured a 1,000,000-token window. Comparing project files also surfaced the repository's older 850,000-token compaction threshold. The operator requested a 5–10% buffer instead of 15%; this leaf selects 900,000 (10%).

The Problem

.codex/config.template.toml retains model_auto_compact_token_limit = 850000 from #10488 / #10503. The operator now prefers retaining more session context within the existing 1M profile. This is a repository policy choice, not a claim that OpenAI mandates one universal threshold or that a config value increases a model's supported capacity.

The Architectural Reality

The Engine owns the tracked Codex project template. Existing peers' untracked .codex/config.toml files remain local policy. FM's missing-default initialization is a separate Brain provisioner correction; this ticket changes the source policy only.

OpenAI documents model_context_window and model_auto_compact_token_limit separately in the configuration reference. The latter controls automatic history compaction.

The Fix

Change the tracked template threshold to 900,000, retaining the 1,000,000-token window. Preserve the existing smaller-model adjustment guidance and every unrelated setting. Do not silently rewrite other peers' active local configurations.

Contract Ledger

Surface Authority Behavior Fallback Docs Evidence
Neo Codex project template Operator's 5–10% buffer preference 1M configured window, 900k auto-compaction threshold Smaller models still require adjusted values Existing template comments TOML parse and exact diff

Acceptance Criteria

  • AC-1: The template parses with model_context_window = 1000000 and model_auto_compact_token_limit = 900000.
  • AC-2: Unrelated settings and the smaller-model guidance remain unchanged.
  • AC-3: Record that existing local overrides are not automatically migrated; adoption belongs to each seat's coordinated configuration update.

Out of Scope

FM's template inheritance, model availability/pricing, undocumented output-limit behavior, and editing existing peers' private configuration.

Related

#10488 · #10503 · neomjs/neo-agent-institution#12

Sweeps: latest 20 open Engine issues, exact compaction-key history/search, recent all-state A2A and own assignments show no overlapping policy change. Git history identifies the original 850k addition; Memory Core queries returned no relevant policy decision. No new source module or instruction substrate; the turn-memory classifier explicitly excludes ordinary configs. Decision Record impact: none.

Origin Session ID: b0dd802b-6451-48ec-b789-d91e29a2b08e Retrieval Hint: Codex 1M context 900000 850000 compaction buffer.