Historical forks replay instruction updates that contradict their inherited baseline #61550

Activity

1. opencodejr commented on Nov 4, 2026

   

   opencodejrBot

   on Nov 4, 2026

   Contributor

   More actions

   Thanks for the clear report. I was able to reproduce this on the latest dev (fe0465646), which is the commit you reported, using the session runner tests with a fake model and a real database.

   The steps (the script below runs them):

   1. Prompt three times while the instruction value changes from A to B ("Changed context") to C ("Latest context").
   2. Fork before the second prompt, so B's update is in the copied history and C's is not.
   3. Prompt the fork and look at the next request.

   That request has C in its system prompt, then replays the copied system update for B, and nothing restores C afterwards. The fork's stored instruction state is already C, so no corrective update is sent. For API context entries the replayed text says the value "supersedes the previous value", so the model ends up treating B as current.

   The existing test seeds a fork with the parent's newest instruction values currently expects this stale update to stay. However, the fork is meant to start from the parent's newest values, so I've passed this to the maintainers as a bug, with your proposed fix linked for reference.

   Reproduction script (repro.sh)

   #!/usr/bin/env bash
   # Reproduces #61550 in the real SessionRunner test harness (fake LLM, real SQLite DB).
   # Run from an opencode checkout (dev): bash repro.sh /path/to/opencode
   set -euo pipefail
   REPO=${1:-/workspace/repo}
   cd "$REPO/packages/core"
   F=test/session-runner.test.ts
   cp $F /tmp/opencode/session-runner.test.ts.bak
   trap 'cp /tmp/opencode/session-runner.test.ts.bak '"$F" EXIT
   python3 - "$F" <<'PY'
   import sys
   p=sys.argv[1]; s=open(p).read()
   anchor='  scenario("keeps nested forks self-contained"'
   new='''  scenario("REPRO 61550 historical fork replays superseded instruction update", function* (s) {
       yield* s.runPrompt("First")
       s.systemBaseline = "Changed context"
       // The "Changed context" update is written just before the "Second" user message,
       // so forking before "Second" keeps update B in the copy but not update C.
       const second = yield* s.runPrompt("Second")
       s.systemBaseline = "Latest context"
       yield* s.runPrompt("Third")
       const forked = yield* s.session.fork({ sessionID, before: second.id })
       yield* s.session.prompt({ sessionID: forked.id, text: "Forked", resume: false })
       yield* s.session.resume(forked.id)
       const req = s.requests.at(-1)!
       console.log("REPRO baseline system:", JSON.stringify(req.system.map((p) => p.text)))
       console.log("REPRO system messages:", JSON.stringify(systemTexts(req)))
       console.log("REPRO full messages:", JSON.stringify(req.messages, null, 1).slice(0, 3000))
       // Expected: the baseline is Latest context and no later message overrides it with the stale value.
       expect(req.system.map((p) => p.text)).toContain("Latest context")
       expect(systemTexts(req).join("\\n")).not.toContain("Changed context")
     })
   
   '''
   assert anchor in s
   open(p,'w').write(s.replace(anchor,new+anchor,1))
   PY
   bun test $F -t "REPRO 61550" 2>&1 | grep -E "REPRO (baseline|system|full)|supersedes|\(pass\)|\(fail\)| pass$| fail$" | grep -v console.log

   Run it: bash repro.sh
2. opencodejr

   added

   reproduced

   on Nov 4, 2026