Auto-compaction triggers at ~200k tokens instead of model's actual context window (1M)

Summary

When using context1m: true with Claude Opus 4.6 or Sonnet 4.6, the Memory Flush hook computes its threshold from a hardcoded contextWindow: 200000 in models.generated.js instead of the actual 1M context window. This results in compaction at ~144k tokens (200k − 50k reserve − 6k soft threshold).

Environment

OpenClaw 2026.3.2 (9288d93)
Model: anthropic/claude-opus-4-6 with context1m: true
Compaction mode: safeguard
reserveTokensFloor: 50000, softThresholdTokens: 6000

Reproduction

Configure anthropic/claude-opus-4-6 with params.context1m: true
Use the session until context reaches ~144k tokens
Memory Flush triggers compaction

Expected: No compaction below ~944k (1M − 50k reserve − 6k soft)
Actual: Compaction at ~144k (200k − 50k − 6k)

Evidence

6 consecutive compactions, all fromHook: True (Memory Flush):

Timestamp (UTC)	tokensBefore
2026-02-28 03:41	157,265
2026-03-01 22:54	271,529
2026-03-03 05:28	264,061
2026-03-03 15:33	249,187
2026-03-03 16:09	163,063
2026-03-03 20:32	150,573

Root Cause

1. `models.generated.js` hardcodes `contextWindow: 200000`

File: /app/node_modules/.pnpm/@mariozechner+pi-ai@0.55.3_ws@8.19.0_zod@4.3.6/node_modules/@mariozechner/pi-ai/dist/models.generated.js

Both claude-opus-4-6 and claude-sonnet-4-6 have contextWindow: 200000. This value is cached in MODEL_CACHE at startup via applyDiscoveredContextWindows().

2. `resolveMemoryFlushContextWindowTokens` ignores `context1m`

// pi-embedded-DgYXShcG.js, line 47871
function resolveMemoryFlushContextWindowTokens(params) {
    return lookupContextTokens(params.modelId) ?? params.agentCfgContextTokens ?? DEFAULT_CONTEXT_TOKENS;
}

This calls lookupContextTokens() which returns the cached 200k. It does not call resolveContextTokensForModel() (line 22146), which correctly checks context1m:

function resolveContextTokensForModel(params) {
    // ...
    if (resolveConfiguredModelParams(params.cfg, ref.provider, ref.model)?.context1m === true
        && isAnthropic1MModel$1(ref.provider, ref.model))
        return ANTHROPIC_CONTEXT_1M_TOKENS; // 1,048,576
    // ...
}

3. `applyDiscoveredContextWindows` takes minimum for duplicates

// pi-embedded-DgYXShcG.js, line 22021
if (existing === void 0 || contextWindow < existing) params.cache.set(model.id, contextWindow);

If a model appears with both 200k and 1M entries, 200k wins.

4. `lookupContextTokens` has a race condition

// pi-embedded-DgYXShcG.js, line 22107
function lookupContextTokens(modelId) {
    ensureContextWindowCacheLoaded();  // async — NOT awaited
    return MODEL_CACHE.get(modelId);   // may return undefined if cache not loaded
}

On startup, the cache may be empty → undefined → falls back to DEFAULT_CONTEXT_TOKENS (200k).

5. `DEFAULT_CONTEXT_TOKENS = 2e5`

// model-selection-Zb7eBzSY.js, line 30 (6 copies across dist/)
const DEFAULT_CONTEXT_TOKENS = 2e5;

The last-resort fallback is also 200k.

6. Heartbeat model bypasses `context-1m` header

// pi-embedded-DgYXShcG.js, line 88969
if (extraParams?.context1m === true)
    if (isAnthropic1MModel(modelId)) betas.add(ANTHROPIC_CONTEXT_1M_BETA);

isAnthropic1MModel() matches only claude-opus-4 and claude-sonnet-4 prefixes. If the heartbeat uses a different model (e.g. Haiku), the context-1m header is not sent — but the full session history (built at 1M) is. Anthropic enforces 200k → overflow compaction.

7. Compaction uses triggering model, not session model

// pi-embedded-DgYXShcG.js, line 94460
model: modelId  // uses the model that triggered the overflow

The compaction API call uses the triggering model (e.g. Haiku). Since Haiku is not a 1M model, the compaction request also lacks the context-1m header — so Anthropic enforces 200k on the compaction call itself. Additionally, a weaker model writes the summary instead of the session's primary model.

Suggested Fixes

resolveMemoryFlushContextWindowTokens should use resolveContextTokensForModel — this already correctly handles context1m and returns 1M
Update models.generated.js — claude-opus-4-6 and claude-sonnet-4-6 should have contextWindow: 1048576
applyDiscoveredContextWindows should take maximum, not minimum, for duplicate entries
lookupContextTokens should await ensureContextWindowCacheLoaded() or ensure synchronous population
DEFAULT_CONTEXT_TOKENS should be 1M or derived from the configured model
Heartbeat should use the session's context-1m header if the session was built with a 1M model
Compaction should use the session's primary model, not the triggering model

Workaround

1. Patch `models.generated.js`

find /path/to/openclaw -name "models.generated.js" \
  -exec sed -i 's/contextWindow: 200000/contextWindow: 1048576/g' {} +

2. Patch `DEFAULT_CONTEXT_TOKENS` (defense-in-depth for race condition)

find /path/to/openclaw/dist -name "*.js" \
  -exec sed -i 's/const DEFAULT_CONTEXT_TOKENS = 2e5/const DEFAULT_CONTEXT_TOKENS = 1048576/g' {} +

Both patches must be reapplied after every update. Restart OpenClaw after patching.

Files Affected

File	Line	Issue
`models.generated.js`	various	Hardcoded `contextWindow: 200000`
`pi-embedded-DgYXShcG.js`	47871	Memory flush ignores `context1m`
`pi-embedded-DgYXShcG.js`	22021	Min-value for duplicate entries
`pi-embedded-DgYXShcG.js`	22107	Race condition (async not awaited)
`model-selection-Zb7eBzSY.js`	30	`DEFAULT_CONTEXT_TOKENS = 2e5`
`pi-embedded-DgYXShcG.js`	88969	Heartbeat bypasses `context-1m` header
`pi-embedded-DgYXShcG.js`	94460	Compaction uses triggering model

This issue might be a duplicate of existing issues. Please check:

Beta header context-1m-2025-08-07 not sent for Claude Opus 4.6, causing 200k hard limit #12507: Same root cause — context-1m beta header not being sent for Claude Opus 4.6, resulting in the 200k token hard limit being enforced instead of the model's actual 1M context window. Issue 1 in your root cause analysis appears to already be tracked there.
Auto Compact Failed when recovery compaction request is also over context limit (WSL + Claude Opus 4.6 via CLIProxyAPI) #15849: Related compaction failure when the recovery compaction request also exceeds the context limit (same Opus 4.6 / 200k interaction).

If your report covers additional distinct bugs not addressed in those issues (Issues 2–6 in your analysis), it may be worth linking to them and focusing this issue on the remaining gaps.

The premature compaction at ~200k is caused by two things: (1) missing context-1m-2025-08-07 beta header, and (2) default model limits in OpenCode being set to 200k.

For Anthropic API users, I made a plugin that fixes both. Add to your opencode.json:

{
  "plugin": ["opencode-anthropic-context-1m"],
  "provider": {
    "anthropic": {
      "models": {
        "claude-opus-4-6":          { "limit": { "context": 1000000, "output": 128000 } },
        "claude-sonnet-4-6":        { "limit": { "context": 1000000, "output": 64000 } },
        "claude-sonnet-4-5":        { "limit": { "context": 1000000, "output": 64000 } },
        "claude-sonnet-4-20250514": { "limit": { "context": 1000000, "output": 64000 } }
      }
    }
  }
}

The plugin injects the beta header via chat.headers hook (appends, doesn't overwrite existing SDK flags). The model overrides push compaction threshold from ~168K to ~968K.

https://github.com/DusKing1/opencode-anthropic-context-1m

fix: resolve 8 open bug reports from upstream tracker

I can add one more high-confidence data point from a real opencode v1.2.26 session.

Evidence (DB)

Session: ses_3...
max_total = 382,889
First >=200k: 2026-03-15 13:07:07
First >=300k: 2026-03-15 15:40:32
Compaction markers in the same session:
- part.type='compaction' = 0
- message.mode='compaction' = 0

Runtime sanity

OPENCODE_DISABLE_AUTOCOMPACT was unset
OPENCODE_DISABLE_PRUNE was unset
Overflow compaction is only disabled when compaction.auto === false (explicit false check)

Overflow-threshold math used by core

In session/compaction.ts, overflow uses:

usable = input ? input - reserved : context - maxOutputTokens(model)
maxOutputTokens = min(model.limit.output, 32000) (from provider/transform.ts)

Local ~/.cache/opencode/models.json in this environment reports:

anthropic/claude-opus-4-6: context=1,000,000, output=128,000
anthropic/claude-sonnet-4-6: context=1,000,000, output=64,000

So derived overflow threshold for both is:

usable = 1,000,000 - 32,000 = 968,000

Observed implication

Given those values, a session at 382,889 is still below the overflow threshold, which is consistent with observing zero compaction markers.

Next-session checklist (I plan to run and report back)

Reproduce on a fresh session with the same model and collect the same SQL evidence.
Capture the exact runtime model-limit inputs used by overflow calculation on that path.
Compare behavior in a minimal setup (without extra plugin layers) to isolate core behavior.

Would also be nice if we can just set this manually per model in the opencode.json file, ie set opus 4.6 to 200K instead of 1M because token usage shoots up beyond that.

The issue is still existing in v2026.3.13.

this issue seems to cite pi mono not opencode or some other tool

Navigation Menu

Repository navigation

Auto-compaction triggers at ~200k tokens instead of model's actual context window (1M) #15871

Summary

Environment

Reproduction

Evidence

Root Cause

1. `models.generated.js` hardcodes `contextWindow: 200000`

2. `resolveMemoryFlushContextWindowTokens` ignores `context1m`

3. `applyDiscoveredContextWindows` takes minimum for duplicates

4. `lookupContextTokens` has a race condition

5. `DEFAULT_CONTEXT_TOKENS = 2e5`

6. Heartbeat model bypasses `context-1m` header

7. Compaction uses triggering model, not session model

Suggested Fixes

Workaround

1. Patch `models.generated.js`

2. Patch `DEFAULT_CONTEXT_TOKENS` (defense-in-depth for race condition)

Files Affected

Evidence (DB)

Runtime sanity

Overflow-threshold math used by core

Observed implication

Next-session checklist (I plan to run and report back)

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Auto-compaction triggers at ~200k tokens instead of model's actual context window (1M) #15871

Description

Summary

Environment

Reproduction

Evidence

Root Cause

1. models.generated.js hardcodes contextWindow: 200000

2. resolveMemoryFlushContextWindowTokens ignores context1m

3. applyDiscoveredContextWindows takes minimum for duplicates

4. lookupContextTokens has a race condition

5. DEFAULT_CONTEXT_TOKENS = 2e5

6. Heartbeat model bypasses context-1m header

7. Compaction uses triggering model, not session model

Suggested Fixes

Workaround

1. Patch models.generated.js

2. Patch DEFAULT_CONTEXT_TOKENS (defense-in-depth for race condition)

Files Affected

Activity

github-actions commented on Mar 3, 2026

DusKing1 commented on Mar 7, 2026

jason2077 commented on Mar 15, 2026

Evidence (DB)

Runtime sanity

Overflow-threshold math used by core

Observed implication

Next-session checklist (I plan to run and report back)

imqqmi commented on Mar 15, 2026

logos007 commented on Mar 21, 2026

rekram1-node commented on May 13, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

1. `models.generated.js` hardcodes `contextWindow: 200000`

2. `resolveMemoryFlushContextWindowTokens` ignores `context1m`

3. `applyDiscoveredContextWindows` takes minimum for duplicates

4. `lookupContextTokens` has a race condition

5. `DEFAULT_CONTEXT_TOKENS = 2e5`

6. Heartbeat model bypasses `context-1m` header

1. Patch `models.generated.js`

2. Patch `DEFAULT_CONTEXT_TOKENS` (defense-in-depth for race condition)