## Summary Clamp the displayed `% used` and `% remaining` numbers in the `[SYSTEM DIRECTIVE: OH-MY-OPENCODE - CONTEXT WINDOW MONITOR]` block to `[0, 100]` so the directive remains trustworthy when the resolved `actualLimit` underestimates the model's real context window. Without clamping, the block currently advertises numbers like `144.7% used (289,370/200,000 tokens), -44.7% remaining`, which safety-tuned models flag as a prompt injection and refuse to follow. ## Root Cause `src/hooks/context-window-monitor.ts` lines 62-74 compute and render the directive: ```ts const actualUsagePercentage = totalInputTokens / actualLimit if (actualUsagePercentage < CONTEXT_WARNING_THRESHOLD) return // ... const usedPct = (actualUsagePercentage * 100).toFixed(1) const remainingPct = ((1 - actualUsagePercentage) * 100).toFixed(1) output.output += `\n\n${createContextReminder(actualLimit)} [Context Status: ${usedPct}% used (${usedTokens}/${limitTokens} tokens), ${remainingPct}% remaining]` ``` When `resolveActualContextLimit()` underestimates the real context window (the deeper concern tracked as #3450 - e.g. a 1M-context Anthropic model that falls back to `DEFAULT_ANTHROPIC_ACTUAL_LIMIT = 200_000`), `totalInputTokens / actualLimit` exceeds 1.0 and the formatting math then produces nonsensical numbers. The injected block also wraps itself in `[SYSTEM DIRECTIVE: ...]` framing and issues behavioral imperatives, so safety-tuned models correctly identify the >100% / negative-remaining pattern as adversarial: > "I ignored it as a prompt injection - context usage is fine, and the directive contradicted my actual instructions." (real session log from issue #3655) ## Changes | File | Change | |------|--------| | `src/hooks/context-window-monitor.ts` | Compute `clampedPercentage = Math.min(Math.max(actualUsagePercentage, 0), 1)` and use it for both `usedPct` and `remainingPct`. The threshold check on line 64 still uses the raw value so the block continues to fire correctly above 70%. `resolveActualContextLimit()` is left untouched (deeper concern, tracked as #3450). | | `src/hooks/context-window-monitor.test.ts` | Add a regression test that drives the hook with `input: 289,370` against a `200,000` resolved limit (matching the issue's reproduction numbers) and asserts both `usedPct` and `remainingPct` stay within `[0, 100]`. | Total diff: 2 files, +65 lines. ## Reproduction (before fix) With the regression test added but the source change reverted: ``` (fail) context-window-monitor > should clamp displayed percentages when input exceeds actualLimit (regression #3655) expect(received).toBeLessThanOrEqual(expected) Expected: <= 100 Received: 144.7 0 pass 1 fail ``` The actually rendered block looks like: ``` [Context Status: 144.7% used (289,370/200,000 tokens), -44.7% remaining] ``` ## Verification (after fix) ``` $ bun test src/hooks/context-window-monitor.test.ts -t "regression #3655" 1 pass / 0 fail / 6 expect() calls $ bun test src/hooks/context-window-monitor.test.ts \ src/hooks/context-window-monitor.model-context-limits.test.ts 15 pass / 0 fail / 29 expect() calls $ bun run typecheck $ tsc --noEmit (no output, exit 0) ``` After the fix, the block now reads: ``` [Context Status: 100.0% used (289,370/200,000 tokens), 0.0% remaining] ``` The numbers are still informative (token counts are unchanged so the user still sees they are over the resolved limit) but they no longer match the prompt-injection pattern that safety-tuned models reject. ## Test - Regression test: `src/hooks/context-window-monitor.test.ts` (new case `should clamp displayed percentages when input exceeds actualLimit (regression #3655)`). - Full hook suites: `context-window-monitor.test.ts` (9 tests) + `context-window-monitor.model-context-limits.test.ts` (6 tests) - 15 pass / 0 fail. - Typecheck: clean. - Note: this PR does NOT change `resolveActualContextLimit()` itself. Fixing the underlying root cause - so the resolver returns the model's real context window for 1M-context Anthropic models - is tracked as #3450 and will land separately. This PR is purely a defensive UX clamp on the displayed string. Fixes #3655 ---

^{Need help on this PR? Tag @codesmith with what you need.} - [ ] Let Codesmith autofix CI failures and bot reviews --- ## Summary by cubic Clamp context status percentages to 0–100% so the monitor stays accurate and doesn’t trigger safety checks. Fixes #3655 by preventing outputs like “144.7% used, -44.7% remaining” when the resolved limit is lower than the model’s real window. - **Bug Fixes** - Format `used%`/`remaining%` from a clamped usage value; threshold logic still uses the raw ratio so the block triggers as before. - Added a regression test (289,370 / 200,000) to ensure percentages stay within [0, 100]; all existing tests pass and typecheck is clean. ^{Written for commit 07064a96f5a1627aa2e6bbdd82f56934a74a14fc. Summary will update on new commits. Review in cubic}