devloop continue misrouted as NL task after gate timeout — spawns duplicate architect + Live View

[shaifulshabuj]

Bug Report

Related to: #2 (Approval gate timeout leaves pipeline unrecoverable)

Summary

Two compounding UX failures occur after a diff gate timeout:

1. The user is dropped to a raw bash prompt with no recovery guidance
2. When the user types devloop continue TASK-ID (natural recovery attempt), it is misrouted as a natural language task — spawning a brand-new architect phase on top of the already-completed worker, and opening a duplicate Live View pane

Steps to Reproduce

1. Run devloop run "..." — diff gate appears
2. Let the gate time out (120s)
3. Pipeline stops — user returned to bash-3.2$ with no hint
4. User types devloop continue TASK-20260517-001721
5. DevLoop: Unknown command — interpreting as natural language task
6. DevLoop starts a new pipeline: Architecting... Copilot designing spec: "continue TASK-20260517-001721"
7. A second Live View pane opens alongside the original stuck one

Observed Behavior

After timeout, no guidance — raw bash prompt:

✘ Diff rejected at approval gate — pipeline stopped
Task: TASK-20260517-001721   Inspect: git diff HEAD
bash-3.2$ devloop continue TASK-20260517-001721

devloop continue misrouted as a feature description:

DevLoop v5.1.0
Unknown command — interpreting as natural language task
Tip: devloop do continue TASK-20260517-001721 to be explicit next time

Natural language task: "continue TASK-20260517-001721"
Full pipeline: "continue TASK-20260517-001721"
Stages: arch -> work -> review -> [fix loop x3 max]

[1] Architecting...
Copilot designing spec: "continue TASK-20260517-001721"
Calling copilot for spec generation...

Two Live View panes open simultaneously — original (rejected-at-diff) + new misrouted one.

Why This Is a Bad UX

• continue is the most natural recovery verb — any user unfamiliar with the docs will type it
• The NL router should never treat a TASK-ID pattern as a feature description — TASK-YYYYMMDD-HHMMSS is unambiguous
• Spawning a new architect phase wastes time, creates a duplicate spec, and confuses the session state
• Zero guidance is printed at the point of failure

Expected Behavior

1. Recovery hint printed after gate timeout:

Tip: Run 'devloop resume TASK-20260517-001721' to retry the gate
     Run 'devloop resume TASK-20260517-001721 --approve-diff' to skip gate and go to review

2. devloop continue recognized as alias for devloop resume:

devloop continue TASK-20260517-001721
# same as: devloop resume TASK-20260517-001721

3. NL router detects TASK-ID pattern and routes correctly:

devloop continue TASK-20260517-001721
# NL router sees TASK-ID arg → suggests: did you mean 'devloop resume TASK-20260517-001721'?

4. Resume reuses existing tmux session — no duplicate Live View pane.

Suggested Fix

• Add continue as alias for cmd_resume in the command dispatcher
• Add TASK-ID guard in NL pre-router: if any arg matches TASK-[0-9]+-[0-9]+, route to cmd_resume not cmd_run
• Print recovery hint with exact commands at gate timeout in _approval_gate()
• In cmd_resume, detect existing tmux session and attach rather than creating new Live View

Environment

• DevLoop: v5.1.0
• NL router: unknown-command fallback in main dispatcher
• Gate: diff gate (_approval_gate "diff")

[shaifulshabuj]

Screenshot: NL misroute after plan gate timeout

What this screenshot shows:

Left panel — NL misrouting in action:

• User sees ✘ Plan rejected at approval gate — pipeline stopped
• Types devloop open TASK-20260517-005906 (navigating to inspect spec)
• Then types devloop continue TASK-20260517-001721 — misrouted
• DevLoop prints: Unknown command — interpreting as natural language task
• Spawns full new pipeline: [1] Architecting... → Copilot designing spec: "continue TASK-20260517-001721"

Middle panels — Two simultaneous Live View panes:

• Both show Feature: continue TASK-20260517-001721 / Feature: TASK-20260517-005906
• Both report Status: rejected-at-plan — the original pipeline's state is preserved but unresumable
• User cannot tell which is the "real" session

Right panel — Third Live View pane:

• Feature from a completely different task (https://github.com/earendil-works/pi, here is pi c)
• Also stuck at rejected-at-plan
• Three simultaneous stuck Live View panes — user has no recovery path

This confirms the compounding effect: each devloop continue call opens another stuck view rather than recovering the original.