[utsab345]

User description

A skills token budget smaller than the truncation marker currently produces output above the configured limit, which can overflow prompts even when the setting is valid. This change returns a safely clipped marker for tiny budgets and adds regression coverage to keep the formatted context within the requested token limit.

---

PR Type

Bug fix, Tests

---

Description

• Clip truncation markers to tiny token budgets.

• Add regression coverage for one-token skill contexts.

---

File Walkthrough

	Relevant files
Bug fix	skills_loader.py Keep truncated skills context within tiny budgets                pr_agent/algo/skills_loader.py Handle positive budgets smaller than the truncation marker. Clip the marker without adding an ellipsis. Remove trailing characters until the token count fits, then return early. +5/-0
Tests	test_skills_context_budget.py Test skills context against one-token budgets                        tests/unittest/test_skills_context_budget.py Add a regression test with a one-token budget. Assert formatted skills context stays within that limit. +6/-0

---

---

⚙️ Agent run details

• Model: openai/gpt-6.1-sol
• Tokens: 4,730 in / 261 out / 4,991 total
• Time cost: 5.1s
• AI calls: 1

[the-pr-agent-bot]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

⏱️ Estimated effort to review: 1 🔵⚪⚪⚪⚪

🧪 PR contains tests

🔒 No security concerns identified

⚡ No major issues detected

---

⚙️ Agent run details

• Model: openai/gpt-6.1-sol
• Tokens: 5,442 in / 94 out / 5,536 total
• Time cost: 4.9s
• AI calls: 1

[IsmaelMartinez]

Thanks @utsab345 for spotting that the truncation marker can overshoot a 1 to 3 token budget. Rather than a clipped marker like "\n\n[t", which is no use in a prompt, the docstring already promises an empty string when nothing fits. One ask: replace the new block with the guard below, and update the test to expect "".

[IsmaelMartinez]

Suggested change

	if max_tokens < marker_tokens:
	marker = clip_tokens(truncate_marker, max_tokens, add_three_dots=False)
	while marker and _count_tokens(marker) > max_tokens:
	marker = marker[:-1]
	return marker
	if max_tokens < marker_tokens:
	return ""

[IsmaelMartinez]

Thanks @utsab345 for switching to the empty-string guard so a tiny budget no longer emits a clipped marker. Output now stays within budget for every value from 0 to 20, and the new test fails without the fix. Approving and merging.