[GitJarno]

Problem

With Auto-compact after set below the model's context, the context meter says Claude "compacts automatically at about N tokens", where N is the setting itself. Claude compacts 33,000 tokens before that. A 150,000 setting compacts at 117,000, 22% sooner than the meter says, and a 300,000 setting compacts at 267,000.

Change

claudeAutoCompactThreshold in the Claude adapter now reports the window minus 33,000. Settings at or above the model's context still report nothing, as they do today. The setting only accepts 100,000 to 1,000,000, so the threshold is always positive.

Two things read this value: the meter's tooltip, and the handoff budget, which caps imported history at the lower of the context window and this threshold. Both now get the point where Claude actually compacts.

Scope and approval

This follows up #18307 / #18332. They report this setting as the threshold and label it approximate because Claude compacts somewhat below the window. That gap is a fixed 33,000 tokens, and Claude reports it itself, so the meter can show the real number. The fix is one constant in the adapter plus the matching test values. Contracts, UI, defaults and when compaction happens all stay the same. I think that makes it a very small, focused fix for an obvious bug: the meter states a threshold Claude never compacts at.

Verification

• Measured where Claude compacts with Claude Code's own getContextUsage({ detail: "summary" }) through the Agent SDK (0.3.276), passing autoCompactWindow in the SDK's settings the way the adapter does. This needs no model call. The CLI was 2.1.305, with 2.1.302 spot-checked and giving the same results.

  Model	autoCompactWindow	Claude's autoCompactThreshold
  claude-sonnet-4-6	150,000	117,000
  claude-sonnet-4-6	180,000	147,000
  claude-sonnet-4-6	200,000 or 650,000 (at/above context)	167,000 (not reported, unchanged)
  claude-opus-5-5	300,000	267,000
  claude-opus-5-5	500,000	467,000
  claude-opus-5-5	1,000,000	967,000
  claude-haiku-5-5, claude-haiku-4-5, claude-opus-4-5	150,000	117,000
  claude-fable-5-1	300,000	267,000

• vp test run apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.test.ts: 203 passed. With the adapter change reverted, the three threshold assertions fail (150,000 instead of 117,000, and 650,000 instead of 617,000).

• vp lint and vp fmt --check on both files, and tsc-rs --noEmit in apps/server, came back clean. The only lint output is an existing no-unused-vars warning near the end of ClaudeAdapterV2.ts, which main has too.

• Not checked: a live auto-compaction in the app, and CLI versions older than 2.1.302. The 33,000 offset belongs to the CLI, so a later CLI could change it. No screenshots: the meter's UI doesn't change, only the number the server reports.

Written with Claude Opus 5.5 in Claude Code.

[macroscopeapp]

Approvability

Verdict: Approved at 2d55610

Macroscope's review found this PR approvable — This focused fix aligns the reported threshold and handoff history cap with Claude's observed auto-compaction point by subtracting 33,000 tokens for explicitly configured windows. Existing defaults and settings at or above the model context remain unchanged, with matching tests updated.

^{You can add or adjust custom eligibility rules. Learn more.}

[coderabbitai]

No actionable comments were generated in the recent review.

Recent review info

Run configuration

• Source mode: Auto
• Repository: pingdotgg/t3code
• Branch: main
• File: .coderabbit.config.ts
• Commit: 9b1d351
• Layers: Repository TS config
• Review profile: CHILL
• Plan: Advanced
• Run ID: 56a0cd98-b8a2-4b93-8a42-7a49aa56efd3

Commits

Reviewing files that changed from the base of the PR and between 2d55610 and 4525d2e.

Files selected for processing: 2

• apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.test.ts
• apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts

Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 8 remain after this review.

Walkthrough

The Claude adapter now subtracts 33,000 tokens from eligible auto-compaction windows. Updated tests check the adjusted thresholds for catalogued and uncatalogued models.

Changes: Claude auto-compaction threshold

Layer / File(s)	Summary
Apply and verify the compaction reserve apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts, apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.test.ts	The adapter subtracts 33,000 tokens when the configured window is positive and below the known model context limit, or when that limit is unknown. Tests expect thresholds of 117,000 and 617,000 for the corresponding windows; other listed cases remain unchanged.

Estimated code review effort: 2 (Simple) | ~10 minutes

Change: Bug fix

Possibly related PRs

• pingdotgg/t3code#22318: Both changes use Claude context-limit information in ClaudeAdapterV2.ts and its tests.

Suggested labels: bugfix

Suggested reviewers: akbarakma

Merge Risk: Minimal · up to 4525d

The threshold reserve is consistent with the adapter’s consumers and does not introduce an actionable merge-blocking risk.

Security architecture risk: Low · up to 2d556

The change does not expand permissions or alter when automatic compaction occurs. It can reduce the allowance for imported conversation history, although the end-to-end effect on delivery was not fully confirmed. No introduced security finding was established.

Retained concerns
No architecture-level concerns identified.

Security review details

Security Blast Radius

• inferred — Where the adjusted value reaches handoff budgeting, its direct effect is to restrict imported history, not increase accessible data or execution authority. The effective delivery impact remains conditional on live propagation.

Resilience and Maintainability Implications

• observed — Existing delivery controls reject insufficient budgets before injection, persist pending status before injection, and refuse an ambiguous repeat injection. Successful injection and inline acceptance have separate completion markers; refusal recovery restores unsent handoffs. These controls are unchanged by the PR, though concurrent caller serialization was not established.

Pre-merge checks: 4

Check name	Status	Explanation
Title check		The title clearly and concisely describes the Claude context-meter fix and uses an appropriate conventional-commit format.
Description check		The description includes all required sections, explains the problem and change, documents scope and approval rationale, and provides detailed verification results and limitations.
Linked Issues check		Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check		Check skipped because no linked issues were found for this pull request.

Finishing Touches

🧪 Generate unit tests (beta)

• Create a new PR

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}

[maria-rcks]

Note

Written by gpt-6.1-sol on behalf of Maria

Verified in a real dev environment

verified as-is at 2d55610a1c989c0d1a86f8086d4d237b4ad3bc5a, using linux chrome 151, playwright from the repo, isolated dev state, and real providers through the local proxy. no source changes or mocks.

• reproduced on github/main (32ecd71726) in a separate worktree: set claude's "auto-compact after" to 150000, created a project pointing at a real git repository through the app, and sent a real sonnet 4.6 turn. the meter showed 150,000.
• stopped main, installed and ran the exact pr head. the same project path, model, setting and prompt showed 117,000. the persisted completed provider turn also contains autoCompactThreshold: 117000. claude code 2.1.314 independently reports 117000 through the real sdk's getContextUsage({detail:"summary"}).
• new claude sessions with an empty setting and 200000 both completed real turns and showed the generic automatic-compaction message, with no numeric threshold.
• clicked "compact context": the real claude provider compacted 33.5k → 4.67k tokens. the meter updated and retained 117,000; a subsequent claude turn completed.
• created a codex gpt-6.1-sol thread through the app. its real turn completed and displayed context usage (23,499 / 258,400) with no claude-specific threshold.
• blacksmith: vp test run apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.test.ts passed all 203 tests. no commits or pushes were needed.

short runtime transcript

Not verified

automatic compaction triggered by reaching 117,000 tokens, and imported-history handoff truncation at the reduced budget boundary. the runtime checks above prove the reported threshold, manual compaction, and continuation; they do not prove those two boundaries. native ios/android and electron clients, macos/windows, relay/tunnel connections, and other claude cli versions were not exercised on this linux vm.

[maria-rcks]

Note

Written by gpt-6.1-sol on behalf of Maria

merged current github/main (9b1d351ba0f6) into the existing contributor head without conflicts, rebase, or source fixups. pushed 4525d2eb02d6b7fe5488aeaac3bbf8cc1ca94763 to the same branch without force. the diff remains two files, 10 additions and 6 deletions.

blacksmith vp run check:changed passed: format, lint (zero warnings/errors), both knip checks, graphql validation, typechecks for server/scripts/evals, and 258 tests across ClaudeAdapterV2 and AgentContextBudget. two independent scoped code reviews found no issues on this exact head.

prior real-provider evidence remains above for the original source change. live provider/client behavior was not rerun after this main merge; automatic compaction at the threshold and imported-history truncation remain unverified. new-head CI is pending.