Fork · replay · sdiff

Because resume is a replay of the JSONL, three operations come almost for free — and together they turn sessions into a comparison instrument no other CLI agent offers.

/fork — branch a conversation#

/fork — checkpoint picker

 fork · pick a checkpoint
  ❯ now add tests                    seq 41 · user-turn      
  ✓ design approved                 seq 28 · design-approved
  ⊟ context checkpoint              seq 19 · compaction
  ❯ fix the auth bug                seq 2  · user-turn

The JSONL prefix up to the chosen event is copied into a fresh session (envelope session ids are rewritten; seq, ids, and payloads are byte-preserved) and a sidecar records the lineage:

.khaelor/sessions/<project>/
  01J8K2….jsonl
  01J8Q7….jsonl            # the fork
  01J8Q7….meta.json        # { "parent": "01J8K2…", "forkPoint": 41 }

KHAELOR opens the fork immediately. The original is untouched — explore the risky approach, keep the safe thread.

/replay — same questions, different run#

/replay <sessionId> extracts the user turns of a finished session and re-runs them sequentially in a fresh session. Two properties matter:

Tool calls re-execute for real — this is a genuine second run, not a tape playback; that is what makes the comparison meaningful.
Sandboxed by construction: the replay runs inside a throwaway git worktree that is deleted afterwards — zero side effects on your working copy.

Switch models first (/model) to answer “what would haiku have done with the exact same requests?”

/sdiff — diff the runs#

/sdiff

 sdiff
                  01J8K2… (sonnet)   01J8K9… (haiku)
 agent turns      12                 17
 tool calls       34                 61
 files modified   2                  3
 tokens (in/out)  145k / 12k         210k / 19k
 cache reads      118k               96k
 verify failures  0                  2
 outcome          completed          completed
 files only in B: src/c.ts
 files in both:   src/a.ts, src/b.ts

Every number is folded from the two durable logs — real usage events, real diff attribution. Combined with /replay, this is the cheapest honest model/prompt comparison you can run: run it twice, diff the runs.

Lineage is queryable#

Forks record parent + forkPoint; replays record replayOf; subtask children record their orchestrator. Your session directory is a family tree, not a pile.