Harness session state lives on disk in each harness’s own store, so a conversation is a sequence of discrete invocations naming the same session:

  1. Turn 1: run (claude-code and copilot: preset --session-id to know the ref up front).
  2. Get the session ref in-band from stdout with JSON output: the claude-code and antigravity envelopes and the codex and copilot event streams all carry the session/conversation ID. Antigravity’s envelope landed in agy 1.1.8 (the upstream request for an in-band ref); on older releases, or in text mode, the ref is only discoverable post-hoc with agentminutes sessions.
  3. Turn N: run --resume <ref>.

Resume appends rather than forks on all four harnesses: same transcript file (claude-code, codex, copilot) or conversation directory (antigravity), same session identity. The ref from turn 1 stays valid for every later turn.

One antigravity trap when discovering refs post-hoc: a single -p invocation writes two conversations (a warm-up plus the real one), so discovery must match candidates on content (e.g. the recorded prompt), never on recency alone.

agentsummons requires an explicit ref; the harnesses’ continue-most-recent forms are deliberately unmodeled, because they’re racy against any concurrent agent session. However, if you prefer to use the harnesses’ continue-most-recent forms, instead, they remain reachable with --extra-arg.

Example: preset the ref (Claude Code)

Claude Code accepts a session ID up front, so the simplest multi-turn shape generates the ref first and never has to parse anything:

  SESSION=$(uuidgen | tr '[:upper:]' '[:lower:]')

# Turn 1: preset the session identity.
agentsummons run --harness claude-code --session-id "$SESSION" \
  -p "Read the failing test and propose a fix" --auto-approve

# Turn 2: resume the same session.
agentsummons run --harness claude-code --resume "$SESSION" \
  -p "Apply the fix and run the tests" --auto-approve

Because the ref is preset, the transcript is addressable before turn 1 even starts, which is useful when something else (archiving, a watchdog) needs to find the session while it’s still running.

Example: capture the ref in-band (Antigravity)

Antigravity CLI doesn’t expose the ability to preset an ID, so turn 1 must capture the ref from the harness’s own JSON envelope instead. --json-output asks the harness for machine-readable stdout, which passes through untouched, so jq can read the conversation_id field directly:

  # Turn 1: run with harness JSON output and capture the conversation ref.
REF=$(agentsummons run --harness antigravity \
  -p "Outline a test plan for the parser" \
  --auto-approve --json-output | jq -r '.conversation_id')

# Turn 2: resume that conversation.
agentsummons run --harness antigravity --resume "$REF" \
  -p "Expand the edge cases into concrete test names" --auto-approve

The behaviors these examples encode:

  • The caller stores one string and passes it back as --resume. Which flag that becomes (--resume, --conversation, copilot’s joined --resume=<id>, or codex’s exec resume subcommand) is agentsummons’s problem.
  • The ref extraction differs per harness, and only where it has to: the codex and copilot variants of turn 1 read the session ID from their JSONL event streams rather than a single envelope.
  • Result.SessionID (and the session_id field in run --json’s envelope) echoes a preset identity. agentsummons never parses refs out of harness output for you; that stays in-band, in the harness’s own format.