feat: per-model growth-nudge floor (modelMinNudgeLimits) for mixed-context installs

[ranxianglei]

Requested by

@bernhardberger in the PR #343 thread (2026-08-28), as a follow-up to issue #342 / PR #343 (growth-nudge floor via minNudgeContextPercent).

Problem

PR #343 makes the T1 growth-nudge floor the global minNudgeContextPercent (default 15% of the model context window). Mixed-model installations cannot express per-model floors with a single global percent:

Model	Advertised window	Intended floor	Floor at global 37.5%	Floor at default 15%
OpenAI GPT-5.6 (native)	400,000	150,000	150,000 ✓	60,000 (reproduces #342 early nudges)
openrouter/z-ai/glm-5.3	1,048,576 (OpenRouter API)	200,000	393,216 (≈2× intended)	157,286

No single global value satisfies both: 37.5% delays the GLM floor to nearly twice the intended value; 15% puts native OpenAI back at 60K.

Proposed API

An optional per-model growth floor, following the existing modelMaxLimits / modelMinLimits style:

modelMinNudgeLimits?: Record<string, number | `${number}%`>

Keyed by provider/model (same keying as the existing model-limit overrides). Values may be absolute tokens or "X%" of that model's context window.

Precedence order (as proposed by the requester)

1. modelMinNudgeLimits[provider/model]
2. minNudgeContextPercent × model context
3. existing growth-only behavior when model context is unknown

This keeps the simple 15% default and lets mixed-model installations use exact token or percentage floors. The existing modelMinLimits keeps its current turn/iteration-reminder meaning (unchanged).

Implementation notes

• Floor computation lives in injectCompressNudges (minNudgeFloorTokens / overMinNudgeFloor, lib/messages/inject/inject.ts).
• New config field needs validation/merging in lib/config.ts + dcp.schema.json + docs (CONFIGURATION.md, README.md), following the modelMinLimits pattern.
• Tests: extend the issue #342: block in tests/inject.test.ts (per-model floor wins over global percent; percent value resolves against the per-model window; unknown-model-context fallback unchanged).

Scope

Does not block PR #343 — the global floor ships first; this is a follow-up enhancement. Recorded here so the mixed-context limitation is tracked before #343 merges.

[ranxianglei]

🤖 ework agent · vllm-qwen/qwen3.8-27b

Implemented — PR: #345

API (exactly as proposed):

compress: {
  modelMinNudgeLimits?: Record<string, number | `${number}%`>
  // keyed by "provider/model"; absolute tokens or "X%" of that model's window
}

Precedence (as requested):

1. modelMinNudgeLimits[provider/model]
2. minNudgeContextPercent × model context (PR fix: gate T1 growth nudges on minContextLimit (issue #342) #343 behavior — byte-for-byte unchanged when the new field is unset)
3. growth-only behavior when the model context is unknown

Edge-case semantics (mirroring modelMaxLimits/modelMinLimits):

• Absolute per-model floors apply even when the model context window is unknown — the only way to pin a floor for a model that doesn't report limit.context.
• A per-model percent with unknown context falls through to the global percent (also unresolvable → growth-only).
• modelMinLimits is untouched (turn/iteration reminders); overMaxLimit and the emergency override still bypass the floor; T2/T3 independent.

Implementation: floor resolution extracted into resolveMinNudgeFloorTokens() in lib/messages/inject/utils.ts, called from injectCompressNudges (provider/model from the existing getModelInfo). Full plumbing follows the modelMinLimits pattern: CompressConfig field, ?? replace-inherited merge, deepClone, VALID_CONFIG_KEYS + no dynamic-key recursion, validateModelLimits reuse, dcp.schema.json, CONFIGURATION.md + README.md.

Tests: 1048 pass / 0 fail. 9 new issue #344: tests in tests/inject.test.ts (per-model absolute wins over global percent; "20%" resolves against the 1,048,576 window → 209,715; unlisted model keeps global floor; unknown-context fallbacks; production config preserveRecentMessages: 2; full growth cycle baseline→nudge→compress→new baseline→nudge; over-max bypass; modelMinLimits independence) — multi-turn with lastPerMessageNudgeTokens/lastNudgeShownTokens side-effect assertions per AGENTS.md §5.7 — plus 4 config-validation tests. Mutation-verified: disabling the per-model branch fails exactly the 4 per-model-specific tests.

Note: the branch is stacked on PR #343 (the floor it extends only exists there), so the diff against master includes #343's commits until #343 merges — merge #343 first. No version bump; release handled separately.

[ranxianglei]

🤖 ework agent · vllm-qwen/qwen3.8-27b

[bot] Follow-up on PR #345: dual-agent review (AGENTS.md §5.3/§5.6) is complete — both reviewers: no blockers. Code review verified the precedence contract and edge cases ("0%" → floor 0, "150%" clamped, unknown-context fallthroughs, modelMinLimits semantics untouched) and independently reproduced the §5.7 mutation check; test review verified all §5.6/§5.7 requirements (multi-turn, side-effect assertions, production config, growth cycle).

One should-fix was applied: zh-CN doc parity — CONFIGURATION.zh-CN.md and README.zh-CN.md now document modelMinNudgeLimits (commit 4c9d772, pushed). Full review summary is on the PR: #345 (comment)

PR is ready for merge once #343 lands (merge #343 first, then rebase/switch #345's base to master — see the PR body).