← graph

fix-shell-pending-eval

Fix: shell tasks stuck in failed-pending-eval (skip rescue path for shell-exec)

Metadata

Statusdone
Assignedagent-1321
Agent identityf51439356729d112a6c404803d88015d5b44832c6c584c62b96732b63c2b0c7e
Created2026-05-01T13:23:16.916942979+00:00
Started2026-05-01T13:38:24.976791111+00:00
Completed2026-05-01T13:50:04.009763563+00:00
Tagsbug, state-machine, eval-scheduled
Eval score0.86
└ blocking impact0.85
└ completeness0.90
└ coordination overhead0.90
└ correctness0.85
└ downstream usability0.80
└ efficiency0.85
└ intent fidelity0.84
└ style adherence0.90

Description

Description

Bug confirmed during verify-document-shell: when the coordinator dispatches a shell task (--exec / --exec-mode shell) and the wrapper script invokes wg fail --class agent-exit-nonzero, the routing code in src/commands/fail.rs:69-71 puts the task into failed-pending-eval even though no .evaluate-X task exists for shell tasks (correctly suppressed by eval_scaffold.rs:174-179). The coordinators rescue resolver in src/commands/service/coordinator.rs:947 then waits forever for the missing eval task, leaving the shell task stuck.

Reproduction (from live smoke)

wg add "Failing shell" --id failtest3 --exec exit 1
wg claim failtest3 --actor agent-1
wg fail failtest3 --class agent-exit-nonzero --reason "Agent exited with code 1"
wg show failtest3 --json
# → status: "failed-pending-eval"   (BUG: should be "failed")

wg exec --shell <task> is unaffected — its dedicated path (src/commands/exec.rs:138) sets Status::Failed directly without going through wg fail. The bug only fires through the agent-spawn wrapper path.

Patch

In src/commands/fail.rs, add a shell-task bypass to the FailedPendingEval routing condition:

// Resolve task once (it is needed for the bypass and downstream).
let is_shell = workgraph::parser::load_graph(&path)
    .ok()
    .and_then(|g| g.get_task(id).map(|t| t.exec.is_some() || t.exec_mode.as_deref() == Some("shell")))
    .unwrap_or(false);

let use_failed_pending_eval = !eval_reject
    && !is_shell                                         // <-- new
    && class == Some(FailureClass::AgentExitNonzero)
    && Config::load_or_default(dir).agency.auto_evaluate;

(Or factor is_shell_task from eval_scaffold.rs:38 into a shared helper and reuse it here.)

Validation

  • Failing test written first: tests/integration_failed_pending_eval.rs adds test_shell_task_skips_failed_pending_eval — claim a shell task, call wg fail --class agent-exit-nonzero, assert final status == Failed, not FailedPendingEval
  • Patch makes the test pass
  • cargo build + cargo test pass with no regressions
  • Live smoke: wg add "f" --exec "exit 1" && wg claim f --actor a && wg fail f --class agent-exit-nonzero → wg show f reports failed
  • cargo install --path . was run

Tags

  • bug, state-machine, shell-mode, agency-pipeline

Depends on

Required by

Log