fix: harden context compaction pipeline — Markdown summaries, read-only action history, and non-executable output rejection

[JasonW404]

Problem

The context compaction pipeline had multiple issues that caused models to produce degraded or incorrect outputs after context compression:

1. JSON round-trip in compression pipeline

The compression pipeline was doing a pointless JSON round-trip:

Prompt asks JSON → LLM outputs JSON → format_summary_output() converts to Markdown
→ HistoryCompressor json.loads() back to dict → renderer converts back to text

This caused:

• HistoryCompressor regression: json.loads(summary_text) always failed on Markdown output, causing history compression to silently degrade to mechanical truncation
• _render_summary() flat format: Dict-typed summaries rendered as key: value lines without Markdown headings
• _render_current_action() JSON injection: Compacted action items fell back to json.dumps(item.content), injecting raw JSON that models mimicked in final answers

2. Action history in assistant role caused mimicry

Compacted action history was rendered as assistant-role messages with Step N / Called tool format. Models treated these as response templates and reproduced the same structured format instead of generating real answers.

3. Non-executable outputs returned as final answers

When the model produced non-executable structured output (e.g. Step N / Called tool, Markdown action records, action/tool_calls JSON, fenced action JSON, incomplete <code> or <RUN> protocol), the agent loop treated them as valid final answers and terminated — instead of recognizing them as format errors and continuing the loop.

4. python_interpreter code loss

String-form tool_calls.arguments were not preserved during compaction. python_interpreter tool calls degraded to empty arguments, losing the actual code.

5. Missing compression schema sections

The compression schema lacked unresolved_issues and next_steps sections, losing important forward-looking context.

Fix

1. Eliminated JSON intermediate format

The LLM now outputs Markdown directly, and all consumers work with plain text.

Before (4-step conversion):

Prompt: "Output strict JSON" → LLM: JSON → format_summary_output(): Markdown
→ HistoryCompressor: json.loads() back to dict → renderer: back to text

After (1-step):

Prompt: "Output Markdown with ## headings" → LLM: Markdown → used directly

2. Markdown heading quality gate

Replaced the old json.loads() gate with a lightweight "##" in summary_text check that rejects plain text while accepting structured Markdown.

3. Read-only action history in user role

Compacted action history is now injected as a user-role message wrapped in <completed_action_history> tags. This makes it clearly read-only context that the model should not mimic or continue.

4. InvalidActionFormatError and non-executable output rejection

Added InvalidActionFormatError to the parser. The following non-executable outputs are now caught, recorded as step errors, and the agent loop continues instead of terminating:

Pattern	Example
Step N / Called tool	Step 3:\nCalled tool search...
Markdown action record	## Action\n- tool: search
action/tool_calls JSON	{"action": "search", ...}
Fenced action JSON	```json\n{"action": ...}\n```
Incomplete <code> / <RUN>	<code>print("hello") (no closing tag)

Natural language final answers remain fully compatible.

5. String tool_calls.arguments preservation

String-form tool_calls.arguments (e.g. python_interpreter code) are now preserved verbatim during compaction instead of being dropped to empty.

6. Expanded compression schema

Added two new sections:

New Field	Description	Complements
unresolved_issues	Problems encountered but not yet resolved	key_decisions (resolved findings)
next_steps	Concrete planned actions and expected outcomes	pending_items (blockers/waiting)

Full schema (7 sections):

# Compact Result of History

## Task Overview
## Completed Work
## Key Decisions
## Unresolved Issues
## Pending Items
## Next Steps
## Context to Preserve

Changes

File	Change
sdk/.../context/config.py	Prompts request Markdown; added unresolved_issues and next_steps to schema
sdk/.../context/llm_summary.py	Prompt builds Markdown section descriptions; uses _strip_code_fences()
sdk/.../context/history_compression.py	summary: str; "##" in summary_text quality gate; incremental passes Markdown directly
sdk/.../context/rendering.py	_render_summary() handles legacy dict; _render_current_action() renders natural language; <completed_action_history> user-role format
sdk/.../context/budget.py	format_summary_output() simplified to code-fence stripping
sdk/.../core_agent.py	InvalidActionFormatError; non-executable output detection; step error recording + loop continuation
test/.../test_context_helper_contracts.py	Updated assertions for new format
test/.../test_context_item.py	Summary fixtures: dict → string
test/.../test_history_context_runtime.py	Mock response updated with all 7 sections
test/.../test_core_agent.py	Tests for non-executable output rejection and loop continuation
test/.../test_guardrail_checkpoints.py	Full degradation chain tests

Testing

• Context-related tests: 69 passed
• Renderer + parser/core tests: 145 passed
• Guardrail + agent loop tests: 20 passed
• py_compile: passed
• git diff --check: passed

New functional tests verify the full degradation chain:

1. Model outputs Step N / Called tool on first attempt
2. Parser raises InvalidActionFormatError
3. Agent loop does not terminate
4. Enters next step
5. Returns real semantic answer

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[codecov]

Codecov Report

❌ Patch coverage is 96.90722% with 3 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
sdk/nexent/core/agents/core_agent.py	90.00%	2 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!