Kanban auto-decompose silently reassigns and launches tasks whose assignee is deliberately non-spawnable (containment bypass)

[northernhartdesigns]

Version: hermes-agent 0.18.2, Windows 11, gateway-embedded dispatcher.

Summary

The dispatcher deliberately respects a "non-spawnable assignee" convention: dispatch_once skips ready/review tasks whose assignee doesn't resolve to a real profile (skipped_nonspawnable; the profile_exists filter in has_spawnable_ready). The auto-decomposer does not: it selects triage tasks by status alone, and _normalize_assignee_choice rewrites any non-profile assignee to kanban.default_assignee on every child, while the root task is reassigned to kanban.orchestrator_profile. With auto_promote_children (default true) the children go straight to ready and live workers spawn.

A task that was intentionally parked by assigning it to a non-profile name is therefore silently un-parked, reassigned, and executed.

Steps to reproduce

1. config.yaml: kanban.auto_decompose: true, with orchestrator_profile and default_assignee set to a real profile.
2. hermes kanban create "parked task - do not run" --assignee not-a-profile --triage
3. Wait one dispatcher tick (~60 s).
4. The card fans out into children assigned to real profiles, the root is reassigned to the orchestrator profile, children promote to ready, and workers spawn. The explicit non-profile assignee survives nowhere.

Observed impact

We run shadow/dry-run pipelines contained by non-spawnable assignees. On 2026-07-11 the decomposer fanned a contained triage card into 6 children assigned to a live profile; they auto-promoted, spawned, and ran to completion - live agent tokens spent on work that was parked for human review.

Expected

The decomposer should honor the same convention the dispatcher already does - a task whose assignee is set but doesn't resolve to a real profile stays parked. Any of these would resolve it:

• (a) decompose_task returns ok=False, reason="assignee is not a spawnable profile - parked" when task.assignee is non-empty and profile_exists() is false (mirrors dispatch semantics);
• (b) the same filter in list_triage_ids;
• (c) per-board auto_decompose in board.json (currently global-only);
• (d) a per-card no_decompose/hold flag.

Relevant code (v0.18.2)

• hermes_cli/kanban_decompose.py - triage-only guard ~L288; _normalize_assignee_choice ~L252; list_triage_ids ~L468
• gateway/kanban_watchers.py - _auto_decompose_tick ~L1134
• hermes_cli/kanban_db.py - decompose_triage_task ~L4984; children promoted via recompute_ready ~L5202

[alt-glitch]

This was generated by AI during triage.

Related: #36814 (Kanban triage tasks with an assignee never auto-decompose — the inverse symptom) and #27572 (the merged orchestrator-driven auto-decomposition feature that introduced this path). This is the decomposer failing to honor the non-spawnable-assignee parking convention that dispatch_once already respects (profile_exists / skipped_nonspawnable), not a duplicate of either.

[GottZ]

This was generated by AI during triage.

Summary

One PR addresses #62985. #62994 prevents non-spawnable assignees from being rewritten and executed during auto-decomposition, while also filtering parked cards before they consume the watcher’s per-tick attempt budget.

Related pull requests

• fix(cron): honor non-spawnable assignee containment in kanban auto-decompose #62994 best fix — (+148/-2) — fixes, but needs one remaining validation step: fix(cron): honor non-spawnable assignee containment in kanban auto-decompose #62994 is the best existing fix because its diff adds the containment guard to decompose_task() and switches the watcher to list_spawnable_triage_ids(), directly addressing both the unsafe fan-out and budget-starvation causes. The contributor keep_open review requested pre-budget filtering and a watcher-level regression test; the follow-up diff implements the filtering, but the added regression test exercises the listing helper rather than _auto_decompose_tick itself.

Suggested consolidation

Keep open with a salvage path for #62994, the best existing fix: retain the shared spawnability check, the early decomposition guard, and the watcher’s filtered candidate listing, then add the requested watcher-level regression test proving that parked cards ahead of an eligible card do not consume auto_decompose_per_tick. This follows the keep_open review while reconciling it with the Verify fixes verdict: the production path is corrected in the diff, but the explicitly requested watcher-level coverage is still missing; there are no competing PRs to close as duplicates.

Complex graph

flowchart LR
    classDef open fill:#dbeafe,stroke:#1d4ed8,color:#1e3a8a
    classDef merged fill:#dcfce7,stroke:#15803d,color:#14532d
    classDef closed fill:#e5e7eb,stroke:#6b7280,color:#1f2937
    classDef unverified fill:#f3f4f6,stroke:#9ca3af,color:#374151
    classDef best stroke-width:3px,stroke:#b45309
    classDef target stroke-width:3px,stroke:#4338ca
    I62985(["issue #62985 (open)"])
    P62994["PR #62994 (open)"]
    P62994 -->|best fix| I62985
    class I62985 open
    class I62985 target
    class P62994 open
    class P62994 best
    click I62985 "https://github.com/NousResearch/hermes-agent/issues/62985"
    click P62994 "https://github.com/NousResearch/hermes-agent/pull/62994"

Loading

Graph: solid arrow = fixes / best fix, dashed arrow = partial or unverified (see edge label); boxed group = PRs duplicating each other; amber border = best fix; indigo border = target; gray node = closed (state tag in the node label).

Cross-PR triage: Reviewed 1 pull request and 1 issue in this complex. Each diff was read against this issue; Assessment working set: 8 kB of PR diffs, 6 kB of issue/PR text, 2 kB of discussion (3 comments), 2 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.