Skip to main content

Import past sessions

Until you end a session there is nothing to brief from. daimon import fills that gap from the transcripts already on this machine, through the same pipeline the hooks use, and says what the run will cost before it spends a model call. Run it from the project directory, with a backend configured (daimon configure --test checks one).

daimon import --project . --plan

--plan prints the plan and stops there. Nothing is spent:

import plan for app (claude-code): 4 sessions, about 8 model calls
1 2026-10-07 88 msgs 3 calls 8d2e4b71-c6a9-4f30-b5d2-91e7a3c40f6e note: compact-summary-only-context
2 2026-10-09 30 msgs 2 calls b71f0c3d-2e85-47a6-8c19-d4f6a0b2e715
3 2026-10-12 57 msgs 2 calls 5a9e3d27-18f4-4c6b-a2d0-7e3b9c1f48a2
4 2026-10-14 23 msgs 1 call c4d81f6a-9b3e-4a27-8f05-1e6d2b7c93f4
skipped 0e7b6a52-3c1d-4e8f-9a04-b5c7d2e1f690: modified in the last 10 minutes, may be live
claude-code: 5 of 14 transcripts under ~/.claude/projects recorded this project (7 recorded another project, 2 recorded no directory)
codex: no transcripts under ~/.codex/sessions, ~/.codex/archived_sessions
not discovered: windsurf, kimi (records no project, pass --transcript); gemini (no registered root); hermes (no transcript file)

Discovery covers Claude Code and Codex. daimon lists the transcripts under both hosts' roots and keeps the ones whose recorded working directory is this project. The 5 newest by session end form the window; --sessions N changes the count. Every skip the plan names comes out of that window and nothing older takes its place, so the same command rebuilds the same plan. Above, the newest file was written minutes ago, so it is skipped as possibly live and four sessions are planned. A session that already has a checkpoint, one shorter than DAIMON_MIN_MESSAGES, a missing file, an unreadable one: each is a named skip, before any spend. When the whole window is already captured and older transcripts exist, the plan says all N newest sessions already captured; raise --sessions to reach older history. A run without --plan whose window holds only skips prints nothing to import and exits 0. Windsurf and Kimi record no project in their transcripts, so they are never discovered: pass --transcript PATH with --project <dir> typed on the command line. Gemini CLI and Hermes have no transcript root daimon registers, so import reaches neither, by discovery or by path.

The call count is a plan: one call per chunk of the rendered transcript plus the merges. The run stops before the session whose planned calls would take the spent count past --max-calls. The default is the planned total, so a plain run spends no more than the number it printed, except for the retries and the fallback rescue of one session, which the plan cannot see.

Now run it:

daimon import --project .

The same plan prints, then the sessions go oldest first, two lines each, and a summary closes the run:

wrote checkpoint: /Users/you/.daimon/checkpoints/8d2e4b71-c6a9-4f30-b5d2-91e7a3c40f6e.json (took 231s) [import]
[1/4] 8d2e4b71-c6a9-4f30-b5d2-91e7a3c40f6e: 96,004 tokens
wrote checkpoint: /Users/you/.daimon/checkpoints/b71f0c3d-2e85-47a6-8c19-d4f6a0b2e715.json (took 80s) [import]
[2/4] b71f0c3d-2e85-47a6-8c19-d4f6a0b2e715: 30,118 tokens, 2 quotes downgraded
wrote checkpoint: /Users/you/.daimon/checkpoints/5a9e3d27-18f4-4c6b-a2d0-7e3b9c1f48a2.json (took 101s) [import]
[3/4] 5a9e3d27-18f4-4c6b-a2d0-7e3b9c1f48a2: tokens unreported
wrote checkpoint: /Users/you/.daimon/checkpoints/c4d81f6a-9b3e-4a27-8f05-1e6d2b7c93f4.json (took 44s) [import]
[4/4] c4d81f6a-9b3e-4a27-8f05-1e6d2b7c93f4: 18,902 tokens
imported 4 of 4, 8 calls (8 planned), 145,024 tokens reported, 1 session unreported

The tokens are what the backend reported for that session. tokens unreported means it reported nothing, and the summary counts those sessions instead of guessing. quotes downgraded counts the quotes the verification could not find in the transcript. The note: in the plan is a separate fact: that session opens with a compaction summary, so the model never saw what came before it. Without the [pretty] extra, when stdout is not a terminal, or with NO_COLOR or DAIMON_PLAIN set, an importing <session id>... line prints before each pair.

The first failed serialize of any kind stops the run. The serializer's own error: ... line prints first, then:

[2/4] b71f0c3d-2e85-47a6-8c19-d4f6a0b2e715: error after 3s (the error line is above)
stopped after the first failed session. 1 of 4 imported, 3 left.
run the same command again to resume; done sessions are skipped for free.
imported 1 of 4, 3 calls (3 planned), 96,004 tokens reported

Resuming is the same command. The plan is rebuilt over the same window and the session that landed is now a skip, at no cost. The automatic retry at session start (daimon heal) never touches an import:

import plan for app (claude-code): 3 sessions, about 5 model calls
1 2026-10-09 30 msgs 2 calls b71f0c3d-2e85-47a6-8c19-d4f6a0b2e715
2 2026-10-12 57 msgs 2 calls 5a9e3d27-18f4-4c6b-a2d0-7e3b9c1f48a2
3 2026-10-14 23 msgs 1 call c4d81f6a-9b3e-4a27-8f05-1e6d2b7c93f4
skipped 0e7b6a52-3c1d-4e8f-9a04-b5c7d2e1f690: modified in the last 10 minutes, may be live
skipped 8d2e4b71-c6a9-4f30-b5d2-91e7a3c40f6e: already has a checkpoint
claude-code: 5 of 14 transcripts under ~/.claude/projects recorded this project (7 recorded another project, 2 recorded no directory)
codex: no transcripts under ~/.codex/sessions, ~/.codex/archived_sessions
not discovered: windsurf, kimi (records no project, pass --transcript); gemini (no registered root); hermes (no transcript file)

The first briefing after an import opens with one note under the greeting, so nobody mistakes old work for where they just stopped:

⚠ Imported history: 4 past sessions, the newest ended 12d ago (2026-10-14), before daimon captured anything live here. Check it before acting on it.

The active topic line reads Active topic (imported, 12d ago): ... for the same reason. The note counts the import files still on disk and dates the newest. When a live capture of the project already exists, it says instead that the import is newer than daimon's last live capture here and serialized after the fact; a briefing shown through the global fallback says in that project rather than here. Both the note and the qualifier end on their own the first time a live capture becomes the latest checkpoint. There is nothing to clear.

Every refusal and exit code is in the command notes for daimon import, and each host page says how far import reaches on that host.