fix(context-window-monitor): clamp displayed context status percentages so the directive stays trustworthy (fixes #3655)

[MoerAI]

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:

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 [Bug]: resolveActualContextLimit falls back to 200k when anthropic provider has no explicit model limits in config #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 07064a9. Summary will update on new commits. Review in cubic}

[MoerAI]

I have read the CLA Document and I hereby sign the CLA

[cubic-dev-ai]

No issues found across 2 files

Confidence score: 5/5

• Automated review surfaced no issues in the provided summaries.
• No files require special attention.

_{Auto-approved: The PR implements simple clamping for display percentages to avoid nonsensical values. It includes a regression test and does not change core logic, posing no risk of regression.}