feat(sleep): harvest Claude Skill tool invocations

[bogdanbaciu21]

Summary

SessionDigest records every tool_use block's tool name in tools_used and
discards the block input, so a Claude Skill invocation is indistinguishable
from ordinary tool use after harvesting. This is the first review-sized slice of
the harvest work discussed in #120.

• add an optional, default-empty SessionDigest.skills_used field;
• harvest only well-formed Claude Skill tool-use blocks (type == "tool_use",
  name == "Skill", mapping input, non-blank string input.skill);
• normalize by trimming surrounding whitespace and de-duplicating in first-seen
  order, preserving the skill name's case and punctuation;
• keep tools_used behavior unchanged, including retaining Skill there, so no
  current consumer loses information.

No loader change is needed: skillopt_sleep/__main__.py filters pinned-digest
payload keys through SessionDigest.__dataclass_fields__, and the new field's
default covers older payloads that omit it.

Out of scope for this slice (kept for later slices of #120): TaskRecord
skill hints, per-skill grouping, SKILL.md path resolution, staging/adoption
changes, and Skill extraction in the Codex/Cursor harvesters.

Tests

• python -m pytest -q tests/test_sleep_skill_harvest.py → 11 passed
• python -m pytest -q → 567 passed, 7 skipped (base main at
  8304e6c: 556 passed, 7 skipped)
• python -m ruff check skillopt_sleep/types.py skillopt_sleep/harvest.py tests/test_sleep_skill_harvest.py
  reports only the pre-existing I001 import-block finding already present on
  main for skillopt_sleep/harvest.py.

New coverage: field default and serialization, legacy payload without
skills_used, an older known-field filter ignoring the new key, positive
harvest, duplicate collapse, whitespace trimming without rewriting the name,
ordinary tools never populating skills_used, malformed blocks (lowercase tool
name, non-tool_use type, missing/non-mapping input, non-string or blank
input.skill, non-dict block), non-fatal malformed JSONL, and the guarantee
that other tool arguments and tool outputs stay out of the digest.

Refs #120

[Copilot]

Pull request overview

This PR extends SkillOpt-Sleep’s session-harvesting digest to explicitly track Claude Skill tool invocations (skills_used) so downstream stages can distinguish “which skills were targeted” from general tool usage, supporting the broader multi-skill routing work in #120.

Changes:

• Add SessionDigest.skills_used (default empty) and include it in serialization.
• Harvest only well-formed Claude Skill tool-use blocks, trimming whitespace and de-duplicating in first-seen order.
• Add deterministic stdlib-only tests covering positive/negative/malformed harvesting and backwards compatibility with legacy payloads.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
tests/test_sleep_skill_harvest.py	Adds comprehensive unit tests for skills_used defaults/serialization and robust harvesting behavior.
skillopt_sleep/types.py	Introduces SessionDigest.skills_used as a new default-empty field serialized via to_dict().
skillopt_sleep/harvest.py	Implements skill-name extraction from Claude Skill tool-use blocks and wires it into transcript digesting with de-duplication.

---

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

[bogdanbaciu21]

@microsoft-github-policy-service agree

[bogdanbaciu21]

THANK YOU!!!