## 作成 LocalTicketBackend によって作成されました。 --- ## Intake summary 既存 Ticket は、session JSONL を read-only に解析する再利用可能な `session-analytics` crate と薄い `yoi session analyze --json` CLI を追加する concrete work item として十分に具体化済み。初期対象 metrics、raw content を既定で出さない privacy/secret 境界、non-goals、受け入れ条件、検証項目が明記されており、Orchestrator が実装 routing を判断できる。関連する `prompt-eval-metrics` は評価/改善 offer 寄りであり、本 Ticket は低レベル session/tool analytics 基盤として重複ではない。 --- ## State changed Intake で要件が実装 routing 可能な粒度まで整理済みであることを確認したため、planning から ready にします。 --- ## State changed Ticket を `workspace-panel` が queued にしました。 --- ## State changed Accepted queued implementation after reading the Ticket body/thread and current workspace state. The Ticket has concrete library/CLI/report/privacy acceptance criteria and does not depend on the just-landed Objective work. Implementation can proceed with synthetic fixture tests and read-only analysis semantics. --- ## Decision Routing decision: implementation_ready Evidence checked: - Ticket body and Intake summary. - Current workspace state after Objective records merged, closed, and cleaned up. - Current worktree state: no active implementation worktree remains. - Current queued set: this is the only queued Ticket. Reason: - The Ticket is specific enough for implementation: add a reusable `session-analytics` crate, structured report model, and thin `yoi session analyze --json` CLI. - The privacy boundary is explicit: default output must not dump raw user input, file contents, secrets, or raw session snippets. - The non-goals are clear: no pruning/compaction behavior changes, no TUI dashboard, no prompt/tool guidance auto-rewrite, no LLM semantic summarization. IntentPacket: Intent: - Add read-only session JSONL analytics that extracts structured tool/session metrics and suspicious-pattern diagnostics without exposing raw content by default. Binding decisions / invariants: - Add a reusable library crate, e.g. `crates/session-analytics`, with an API such as `analyze_session(path) -> SessionReport`. - Add a thin product CLI surface, e.g. `yoi session analyze --json`, that calls the library and avoids duplicating analytics logic. - Prefer streaming/bounded parsing for large JSONL files; do not unnecessarily read huge files into memory if line-by-line parsing is practical. - Analysis is read-only and has no runtime side effects. - The analytics crate must stay independent from TUI rendering and Pod runtime side effects; runtime crates must not depend on analytics. - Reuse current session/log schema types only where dependency direction remains healthy. - Structured report should include at least: - tool usage counts by name/kind; - tool calls per turn; - failed tool call counts; - repeated `Read` by path and by path+offset/limit; - whether repeated reads had intervening write/edit observations; - edit/write churn by path; - edit/write argument byte-size estimates and large replacement diagnostics; - `replace_all` usage; - observable tool result sizes/large/truncated indicators where available; - prune/compaction events and bounded correlation with subsequent repeated reads/tool calls; - malformed/unknown entry diagnostics. - Suspicious diagnostics must be framed as observations, not automatic blame. Re-reading after mutation or validation may be correct. - Default CLI/report output must avoid raw user input/file contents/session snippets. It may include paths, counts, byte sizes, turn/timestamp indexes, and hashes if useful. - Tests should use synthetic/minimal session JSONL fixtures rather than relying on private local sessions. - Do not implement a full dashboard, prompt auto-fixer, semantic summarizer, or Worker pruning/compaction behavior changes. Implementation latitude: - Coder may choose exact report struct field names and JSON format if machine-readable and test-covered. - Coder may add a small JSONL adapter tolerant of unknown/malformed records with bounded diagnostics if current session log enum parsing is too strict. - Coder may keep initial path input file-only; `--latest`, `--pod`, directory aggregation, and Markdown output can remain future work. - If current session schema is too hard to reuse without unhealthy dependencies, implement a minimal analytics event parser for the observed JSONL shape and document the limitation. Escalate if: - Implementing this requires changing session writer/history schemas or Worker compaction behavior. - Avoiding raw-content exposure conflicts with required metrics. - Healthy crate dependency direction would require a larger session-schema extraction refactor. - Real local session data would need to be read for tests/validation; use synthetic fixtures instead unless explicitly authorized. Validation: - Focused unit tests for repeated reads with/without intervening edit/write. - Focused tests for large edit argument sizing and replace_all observations. - Focused tests for tool failure counting and tool calls per turn. - Focused tests for large/truncated result observations if representable in fixtures. - Focused tests for compaction/prune event correlation as correlation only. - Focused tests for malformed/unknown JSONL entries producing bounded diagnostics. - CLI JSON output test for `yoi session analyze --json` or equivalent. - `cargo fmt --check`. - `git diff --check`. - `cargo run -q -p yoi -- ticket doctor`. - `cargo check --workspace`. - `nix build .#yoi`. ---