core: interrupted tools cannot retain partial results #39565

Activity

1. Dante-dan commented on Sep 10, 2026

   

   Dante-dan

   on Sep 10, 2026

   Contributor

   More actions

   I traced the current V2 boundary and this still looks broader than Shell: Tool.Context has no checkpoint/progress capability, while Effect interruption prevents a normal tool result from reaching settlement. The session layer already has replayable tool.progress events, and failed tool states can retain prior structured content, but the runner does not connect those pieces.

   I propose a bounded design before implementation:

   • let a running tool publish a checkpoint through its context;
   • persist that checkpoint through the existing progress event path;
   • keep interruption as a failed/interrupted terminal state rather than converting it to success;
   • project the latest checkpoint into that terminal state, including managed-output references where applicable.

   I would validate this with an interruptible Shell fixture that emits output before cancellation, plus a generic tool fixture: both should remain interrupted, retain the latest checkpoint after replay, and preserve explicit removal semantics. I also checked the closed narrow Shell PR #39562; its author closed it specifically in favor of this broader design, and I found no active replacement PR.

   Would the core team approve this boundary and assign the issue before I implement it?
2. opencodejr commented on Oct 28, 2026

   

   opencodejrBot

   on Oct 28, 2026

   Contributor

   More actions

   Thanks for the report. I reproduced this on the latest dev (c63a939) with a core test that runs the real Shell tool on a loop printing tick-N every 200 ms, then interrupts the call after about a second:

   • Before the interrupt: shell.output(id) returned tick-1 through tick-7.
   • The tool call's result: a pure interrupt cause, with no output or metadata attached.
   • After the interrupt: shell.get(id) returned NotFound and the shell's .out file had been deleted, so the partial output was lost.

   The cause matches your description. On interrupt, tool/plugin/shell.ts runs shell.remove(info.id) (around shell.result), and the jobs.block interrupt handler only calls jobs.cancel. Neither path records anything on the cancelled outcome. This code is the same in v2.2.0 and on dev.

   I couldn't make a TUI recording: in my opencode-drive attempt, the simulated model's shell call never showed on screen. The script below shows the bug directly. It temporarily appends the test to packages/core/test/tool-shell.test.ts and runs it.

   Reproduction script (repro.sh)

   #!/usr/bin/env bash
   # Repro for #39565: interrupting a foreground V2 Shell tool call loses its partial output.
   # Run from an opencode checkout (dev or v2.2.0) after `bun install`:
   #   REPO=/path/to/opencode bash repro.sh
   set -euo pipefail
   REPO=${REPO:-/workspace/repo}
   cd "$REPO/packages/core"
   cp test/tool-shell.test.ts /tmp/tool-shell.test.ts.bak
   trap 'cp /tmp/tool-shell.test.ts.bak test/tool-shell.test.ts' EXIT
   cat >> test/tool-shell.test.ts <<'TEST'
   describe("repro #39565", () => {
     it.live("interrupting a foreground shell discards its partial output", () =>
       Effect.acquireUseRelease(
         Effect.promise(() => tmpdir()),
         (tmp) => {
           reset()
           return withSession(tmp.path, (registry) =>
             Effect.gen(function* () {
               const shell = yield* Shell.Service
               const scope = yield* Scope.Scope
               const fiber = yield* executeTool(
                 registry,
                 call({ command: "for i in $(seq 1 60); do echo tick-$i; sleep 0.2; done" }, "call-interrupt"),
               ).pipe(Effect.exit, Effect.forkIn(scope, { startImmediately: true }))
               let info: Shell.Info | undefined
               for (let i = 0; i < 200 && !info; i++) {
                 info = (yield* shell.list())[0]
                 if (!info) yield* Effect.sleep(Duration.millis(10))
               }
               if (!info) throw new Error("shell never started")
               yield* Effect.sleep(Duration.millis(1200))
               const before = yield* shell.output(info.id)
               console.log("REPRO partial output before interrupt:", JSON.stringify(before).slice(0, 200))
               yield* Fiber.interrupt(fiber)
               const exit = yield* Fiber.await(fiber)
               console.log("REPRO tool exit:", JSON.stringify(exit).slice(0, 300))
               const after = yield* shell.get(info.id).pipe(Effect.exit)
               console.log("REPRO shell.get after interrupt:", Exit.isFailure(after) ? "NotFound (record removed)" : "present")
               const fileExists = yield* Effect.promise(() => fs.stat(info!.file).then(() => true, () => false))
               console.log("REPRO output file exists after interrupt:", fileExists, info.file)
             }),
           )
         },
         (tmp) => Effect.promise(() => tmp[Symbol.asyncDispose]().then(() => undefined)),
       ),
     )
   })
   TEST
   bun test test/tool-shell.test.ts -t "repro #39565" 2>&1 | grep -E "REPRO|pass|fail"

   Run it: bash repro.sh
3. opencodejr

   added

   reproduced

   on Oct 28, 2026
4. jlongster

   closed this as completedin #62977on Nov 10, 2026