feat(learn): collect Pi sessions alongside other agents

[tranducquy0]

Adds a pi learn session source that discovers and scans Pi agent session JSONL files, and includes Pi in the default scan sources, source line, empty-state text, and --sources usage.

• New piSessionSource in proxy/internal/store/source_pi.go with usage/cache/billing/tool-call extraction
• piRoot() config (honors CAVEMAN_PI_ROOT, defaults to ~/.pi)
• Registered in learnSessionSources() and default source list
• CLI copy + --sources help updated

[tranducquy0]

Updated in two follow-up pushes:

• 9516db3b fix(learn): use a constant upper-bound check (total > math.MaxInt) for the Pi context conversion in proxy/internal/store/source_pi.go, resolving the CodeQL go/incorrect-integer-conversion finding — CodeQL only recognizes bounds checks against a known constant, not uint64(^uint(0)>>1).
• aee5c9f1 test(learn): update learn-porcelain.runtime.mjs and learn-tui.runtime.mjs expected empty-source message to include opencode, aider or Pi.

generated by OpenCode

[JuliusBrussee]

Reviewed, not adopted — needs a test and two usage-arithmetic fixes. Thanks for this; adding a learn source is real work and the hard part is right.

Verified in your favour: the session path is correct. pi 1.1.0's own bundle writes ~/.pi/agent/sessions/<encoded-cwd>/, which is exactly what piRoot() + filepath.Join(root, "agent", "sessions") walks. The CAVEMAN_PI_ROOT override and repoProvisional handling match the sibling sources.

What needs changing:

1. No test. Every sibling has one — source_{claude,codex,gemini,opencode,aider}_test.go. 225 lines of new parsing with no fixture is the main blocker; this is also the repo's standing rule that a fix needs a test.
2. piCacheUsage and piBillingUsage sum two spellings of one field rather than preferring one: usage["cache_read"] + usage["cacheRead"], usage["input"] + usage["input_tokens"], and the same for cache_write/cacheWrite/cache_creation/cacheCreation. pi's bundle uses both conventions (cacheRead 4279×, cache_creation 40×, cacheCreation 9×), so a payload carrying both double-counts tokens. Prefer first-present, as piContextTotal already does.
3. piTaskSpawns counts every toolCall block as a spawn, so an ordinary tool-using turn reports N spawns. The siblings count subagent/Task spawns only.

Minor: piToolCalls fills the pending map with tool-result text that nothing ever reads, so Pi tool output never reaches the analysis the other sources feed.

Fix 1–3 and this is straightforwardly reviewable — the structure is already right.

---

Generated by Claude Code