What

Adds --source pi to SkillOpt-Sleep, so it can harvest sessions from the pi coding agent (~/.pi/agent/sessions/<slug>/*.jsonl), on par with the existing claude / codex sources. pi follows the open Agent Skills standard and stores sessions as JSONL, so it slots in cleanly alongside Codex.

Why

pi is a terminal-native coding agent with first-class skills + extensions. Sleep currently cannot learn from pi users' sessions; this adds it as a first-class source. pi also exposes toolName on toolResult messages, which corroborates tool-name extraction and catches calls even when the toolCall block's name is absent.

pi schema notes (verified against real transcripts)

• Entry discriminator is type; cwd lives on the single session entry (not on messages).
• Conversational turns are type:"message" with message.role in {user, assistant, toolResult}.
• Content blocks use toolCall (not Claude's tool_use) and carry name; thinking blocks are private reasoning and are skipped (never leak into assistant_finals).
• toolResult messages carry toolName (used to corroborate tool names) and isError (bool). isError is deliberately NOT surfaced as a feedback signal — see the design note below.

Changes

Verification

• ruff check clean on all touched files.
• pytest tests/test_harvest_pi.py tests/test_sleep_engine.py → 54 passed, 1 skipped (no regressions).
• End-to-end against real local pi sessions (Korean + English prompts, tool errors, thinking blocks):

$ python -m skillopt_sleep harvest --source pi --scope all --json
n_sessions: 12 → n_tasks: 11

Per-session digest correctly extracts project (from session.cwd) and tool names (from both toolCall blocks and toolResult.toolName):

session=…019ef47e…
  project=/Users/brandon/Workspaces/hotdeals
  turns: user=7 asst=78
  tools=[bash, edit, read, write]
  feedback=[neg:fix it]   # lexical user feedback only; isError excluded

Design note: why isError is not surfaced as feedback

pi's toolResult carries isError (bool) — whether that one tool invocation failed mechanically. This is deliberately not used as a feedback signal. In agentic coding, intermediate tool errors are normal and are frequently followed by recovery and a successful final result; a successfully completed session can still contain several isError: true entries. Treating recovered errors as neg: feedback would mislabel successful sessions as failures and poison the miner's task-outcome labels. Task outcome is inferred from the user's judgment of the final result (the lexical feedback phrases already in harvest.py), not from transient tool mechanics. See the NOTE in harvest_pi.py and the test asserting isError is not surfaced as feedback.

Other design choices

• Self-contained secret patterns: harvest_pi duplicates _SECRET_PATTERNS (mirroring how harvest_codex keeps its own) rather than importing the underscore-prefixed private tuple from harvest_codex. If a third source lands, it might be worth promoting these into a shared redact module — happy to do that as a follow-up if preferred.
• files_touched is left empty (same as the Codex adapter) — pi tool arguments could be mined heuristically, but that is left out to keep this PR focused.

Follows CONTRIBUTING.md: stdlib-only, type hints, concise docstrings, existing patterns. Looking forward to feedback!