#20861: feat(context): add conversation transcript stats to /context command

## Summary - Parse session JSONL transcript to show post-compaction message count, per-role char breakdown, and combined context window estimate - Adds conversation stats to `/context list`, `/context detail`, and `/context json` views - Combined estimate line shows `system + tools + conversation` vs context window with usage percentage ## Details The `/context` command previously showed system prompt, tools, skills, and workspace files — but had no visibility into the actual conversation messages consuming the context window. This adds a `readConversationStats()` function that: - Reads the raw session JSONL file (no SessionManager overhead) - Resets accumulators on `type: "compaction"` entries so only post-compaction messages are counted - Tracks per-role (user, assistant, toolResult) message count and char totals - Uses chars/4 token estimation (matching the pruner's `CHARS_PER_TOKEN_ESTIMATE`) ### Example output (`/context list`) ``` Conversation (post-compaction): 47 messages, 128,340 chars (~32,085 tok) Compactions: 2 Context estimate: ~45,200 tok (system ~8,100 + tools ~5,015 + conversation ~32,085) / window=200,000 (23%) Session tokens (cached): 51,234 total / ctx=200000 ``` ### Changes Single file: `src/auto-reply/reply/commands-context-report.ts` - `ConversationStats` / `ConversationRoleStats` types - `getEntryContentChars()` — extract char count from message content (handles string, array blocks with text/thinking/tool_use) - `readConversationStats()` — parse JSONL, reset on compaction boundaries, aggregate per-role - Wired into `buildContextReply()` for list, detail, and json sub-commands ## Test plan - [ ] `/context list` — verify conversation line and combined estimate appear - [ ] `/context detail` — verify per-role breakdown appears sorted by chars descending - [ ] `/context json` — verify `conversation` object present in output - [ ] Fresh session (no transcript) — verify "unavailable" fallback - [ ] After compaction — verify only post-compaction messages are counted - [ ] `pnpm tsgo` passes (no new type errors) --- _Re-opened from #19600 (auto-closed during fork maintenance)._ <!-- greptile_comment --> <h3>Greptile Summary</h3> This PR adds conversation transcript statistics to the `/context` command, providing visibility into post-compaction message counts and context window usage. The implementation parses the session JSONL file directly to extract per-role statistics (message count, char totals) while resetting accumulators on compaction boundaries. The stats are integrated into `/context list`, `/context detail`, and `/context json` views with a combined context estimate showing system + tools + conversation vs window size. Key changes: - Added `readConversationStats()` function that parses JSONL transcripts and tracks post-compaction messages only - New `getEntryContentChars()` helper extracts character counts from message content (handles string, array blocks with text/thinking/tool_use) - Uses chars/4 token estimation matching the pruner's `CHARS_PER_TOKEN_ESTIMATE` constant - Graceful fallback when transcript is unavailable - Combined estimate line shows comprehensive context breakdown with usage percentage <h3>Confidence Score: 5/5</h3> - This PR is safe to merge with minimal risk - The implementation is clean, well-contained, and uses appropriate defensive patterns throughout. The function correctly handles missing sessions, malformed JSONL lines, and missing content fields with try-catch blocks and empty fallbacks. The token estimation aligns with existing codebase constants, and the `resolveSessionFilePathOptions` helper is used correctly for path resolution. All error conditions return safe empty states. - No files require special attention _{Last reviewed commit: 2e264bb} <!-- greptile_other_comments_section --> <!-- /greptile_comment -->