feat(core): shared agent-transcript layer core/sessions

[RaghavChamadiya]

What changed

New package core/sessions/: one normalized transcript stream for every consumer. The three distill miners (missed savings, wasteful re-reads, correction mining) each carried a private copy of Claude Code JSONL parsing: munged-cwd discovery, substring gating, tool_use/tool_result pairing, timestamp handling. That parsing now lives once behind a harness-agnostic interface, so the next transcript consumer starts from Events instead of raw JSONL, and supporting another coding agent means one new adapter file.

• events.py: the Event / ToolUse / ToolResult model, the user-interrupt marker, and iter_deduped_usage. One API message spans several transcript lines and repeats its usage object on each; summing raw lines overcounts roughly 2.6x, so usage is deduped by message id.
• adapters/base.py: the HarnessAdapter contract (discover() + normalize()) with shared iter_events(). Consumers pass a raw-string prefilter so multi-hundred-KB lines are still skipped before any JSON parsing, matching the miners' existing perf posture.
• adapters/claude_code.py: first adapter. Field handling verified against real transcripts on this machine: sidechain (subagent) lines stay in the stream and are flagged, isMeta / isCompactSummary lines are flagged so prompt-derived signals can skip them, and the top-level toolUseResult (dict on success, exit-code string on failure) rides on ToolResult.payload.
• cursor.py: per-file byte-offset + mtime cursors with iter_new_events, for incremental scans over append-only transcripts. Truncated or replaced files restart from zero; a trailing line without a newline is a write in progress and waits for the next pass. Ships tested but unconsumed; the miners keep their windowed rescan semantics.
• A README in the package documents the contract, the schema gotchas, and how to add a harness.

The three miners are now consumers, not parsers. Public APIs, report shapes, window semantics, and test seams are all unchanged. One incidental fix: dispatch now keys on the parsed event shape rather than raw substrings, so a tool result whose output happened to contain the literal text tool_use is no longer dropped.

Validation

• Existing distill suite green with zero test edits (569 tests), plus 19 new adapter/cursor tests over a checked-in real-shaped JSONL fixture (split usage blocks, sidechain, compact summary, interrupt marker, malformed lines).
• Full unit suite: 6,943 passed.
• Dogfood equivalence: snapshotted this repo's real transcript directory (342 files, 499 MB), ran all three scans on main and on this branch with a pinned clock; the JSON reports are byte-identical.

[repowise-bot]

✅ Health: 7.6 (unchanged)

📋 At a glance
4 hotspots touched · 4 new findings introduced · 1 file with recent fix history. Scoped to packages.

🚨 Change risk: high _{(riskier than 83% of this repo's commits · raw 9.6/10)}
This change's risk is driven by:

• more lines added than baseline
• more scattered than baseline

🩹 Review priority _{(files here with the most recent bug-fix history — defects cluster, so review these first)}

• .../distill/missed.py — fixed once in the last ~6 months

🔎 More signals (1)

🔥 Hotspots touched (4)

• .../distill/missed.py — 2 commits/90d, 9 dependents · primary owner: Raghav Chamadiya (100%)
• .../distill/session_model.py — 2 commits/90d, 3 dependents · primary owner: Raghav Chamadiya (100%)
• .../distill/corrections.py — 1 commits/90d, 2 dependents · primary owner: Raghav Chamadiya (100%)

1 more

• .../distill/missed_mcp.py — 1 commits/90d, 3 dependents · primary owner: Raghav Chamadiya (100%)

---

📊 Full report · ⭐ Star Repowise · 📥 Install bot · Last updated 2026-07-11 06:31 UTC
_{Silence on a single PR with [skip repowise] in the title · Per-repo toggle on repowise.dev/settings?tab=bot}