Skip to content

session: auto-spawned tasks fire on questions/discussion, not just work orders (4x in one session) #1804

Open

Description

@santoshkumarradha

What happened

In a single session (project -Users-santoshkumar, 2026-10-10), the "this looked like work, so task N started" spawner fired four times on messages that were questions or product-direction discussion — never work orders:

  • task 11 ("ai agent browser") — from a product-direction message; stopped by hand
  • task 19 ("general purpose browser") — same message, second spawn; stopped by hand
  • task 31 ("headless screenshots") — from a question ("can we get screenshots headless…?"); ran 8 minutes / $0.28 before being stopped
  • task 34 ("headless screenshots") — same question, respawned immediately after 31 was stopped; stopped again

Each spawn created a worktree, a branch, and started spending. The user noticed: "seems like it is leaking or not doing good".

Expected

Auto-spawn only on an explicit work order ("do X", "fix Y", "build Z"). A question ("can we…?", "what I am wondering is…") or a direction statement ("we should have…", "the goal is…") must never spawn. A stopped-for-this-reason task must not respawn from the same message.

Impact

Real money and real branch litter on work nobody asked for; user trust in the spawner ("leaking").

Related

#1549 (task workers don't receive standing orders) — same family: the task-spawn path doesn't carry the session's context/rules.

—
Drafted with CodeAF · reviewed and owned by the author

Activity

  1. added
    bugSomething the code does that it should not
    area:sessionThe engine — turns, tasks, the toolbelt, checkpoints
    sev:seriousWrong or missing behaviour a person meets in ordinary use
    on Oct 10, 2026
  2. mikemikimike commented on Oct 17, 2026

    @mikemikimike

    Implemented in PR #1845: #1845

    Both route judges now require a direct, non-question work order before starting a task. A stopped task's result no longer restarts the same request, while a new human instruction remains eligible for normal evaluation.

    The routing regression tests pass. go build ./..., go generate ./internal/manual, and go test -count=1 ./cmd/codeaf-changes passed. go test -count=1 -p 2 -timeout=15m ./... still fails on existing Windows-only test dependencies and platform assumptions; the same failures occur on clean dev.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    area:sessionThe engine — turns, tasks, the toolbelt, checkpointsbugSomething the code does that it should notsev:seriousWrong or missing behaviour a person meets in ordinary use

    Type

    No type

    Projects

    No projects

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions