Add provider-aware metadata digest

[jbarnes850]

Summary

• add an OpenAI adapter digest helper that projects execution metadata and enforces provider budgets
• derive default character budgets from the provider context window (≈10% using 4 chars/token) while keeping overrides simple
• expose digest stats + budget utilisation in the payload/logs and document the new config knobs

Testing

• pytest tests/unit/connectors/test_prompt_digest.py tests/unit/test_openai_adapter.py

[Copilot]

Pull Request Overview

This PR introduces a provider-aware prompt digest system to prevent oversized execution metadata from exceeding LLM context window limits. The digest compresses large metadata sections into summaries while preserving high-signal information.

• Adds MetadataDigestConfig to configure budget limits and trimming behavior per provider
• Implements build_prompt_digest() to intelligently trim metadata sections based on character budgets
• Integrates digest logic into OpenAI adapter's message building pipeline

Reviewed Changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
atlas/connectors/prompt_digest.py	New module implementing metadata digest builder with provider-aware budgets and intelligent section trimming
atlas/config/models.py	Adds MetadataDigestConfig class and integrates it into OpenAIAdapterConfig
atlas/connectors/openai.py	Updates _build_messages to use digest system and handle errors
tests/unit/test_openai_adapter.py	Adds test verifying adapter trims large metadata blobs
tests/unit/connectors/test_prompt_digest.py	New test file with comprehensive digest functionality tests
docs/learning_eval.md	Documents the prompt digest feature, configuration options, and playbook entry schema updates

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Double percent sign '%%' should be a single '%'.

Suggested change

	Defaults reserve roughly 10%% of the provider's published context window, assuming ~4 characters per token.
	Defaults reserve roughly 10% of the provider's published context window, assuming ~4 characters per token.

[Copilot]

This assertion checks if the string 'payload' is not in audit_summary['sample'][0], but based on line 16 where the metadata creates a list of strings ['payload'] * 200, the sample entries would be the string 'payload' itself, not a dict/mapping. The assertion should verify the sample was trimmed, but this logic appears incorrect for the data structure being tested.

Suggested change

	assert "payload" not in audit_summary["sample"][0]
	# The sample should be trimmed; check its length is less than the original (200)
	assert len(audit_summary["sample"]) < 200

[Copilot]

[nitpick] The magic number calculation token_limit * 0.025 * 4 could be clearer. Consider extracting this formula to a named constant or adding an inline comment explaining the relationship (e.g., '2.5% of context window * 4 chars/token').

Suggested change

	if token_limit:
	if token_limit:
	# Character budget: 2.5% of context window (token_limit) * average chars per token (4)

[Copilot]

The step key parsing assumes step IDs follow a 'number-suffix' format but doesn't handle the case where the split produces an empty string before the hyphen (e.g., '-123'). Consider validating that the first part is non-empty before calling int().

Suggested change

	return int(value.split("-", 1)[0]), value
	prefix = value.split("-", 1)[0]
	if prefix == "":
	return math.inf, value
	return int(prefix), value

[Copilot]

The digest JSON is parsed twice: once in build_prompt_digest() during encoding (line 48), and again here (line 53) for logging. Consider having build_prompt_digest() return both the encoded string and stats dict to avoid redundant parsing.

Suggested change

	digest = build_prompt_digest(metadata, self._config.llm, self._config.metadata_digest)
	except PromptDigestTooLargeError as exc:
	raise AdapterError(str(exc)) from exc
	messages.append({"role": "system", "content": digest})
	try:
	stats = json.loads(digest).get("digest_stats", {})
	except json.JSONDecodeError:
	stats = {}
	digest, stats = build_prompt_digest(metadata, self._config.llm, self._config.metadata_digest)
	except PromptDigestTooLargeError as exc:
	raise AdapterError(str(exc)) from exc
	messages.append({"role": "system", "content": digest})