[LarsArtmann]

Why

The three JSONL adapters (claudecode, codex, pi) each re-implement the same ceremony: resolving ~/.harness directories, opening session files with a deferred close, probing readable directories, and formatting not-a-session errors. Meanwhile the shared adapter.go had grown into a monolith. And result pairing is positional today — when a harness splits a tool_call and its tool_result across messages, index alignment silently drops the outcome.

What changed

• Behavior-preserving helpers, migrated at every call site: HomePath, UserHomeDir, ReadableDir, OpenFile, NormalizePath, NotRecognizedErr.
• adapter.go split into topic files (home.go, input.go, title.go, agent_launch.go).
• Contract additions (consumed by part 3/3): ToolResult.ToolCallID for explicit by-ID pairing, ToolCall.ProviderExecuted, Closer (adapters holding connection pools), DiagnosticsSource + FilesystemDiagnostics (health checks).
• go.mod: go 1.26.5 directive (tests use new(expr)) and go-humanize (PluralWord in agent-launch summaries).

No production behavior changes — the full suite passes unchanged.

Series

Part 2/3, stacks on #25:

• 3/3 crush-adapter-core — the adapter that exercises ToolCallID, Closer, and DiagnosticsSource

Validation

• go build ./...
• go test -count=1 ./... — green at the branch tip
• Diff vs parent is helper extraction + additive types only (verified per-file)

💘 Generated with Crush

---

Reviewing this stacked series

GitHub shows the cumulative diff against master, so "Files changed" also includes part 1 (#25). This PR's own delta is just the second commit, refactor(adapter): shared helpers + tool-call pairing contract — reviewing that single commit gives the actual scope of this change.