Fix Kanban worker archived-skill detection and diagnostics

[sdgdllc]

Summary

• ignore archived or hidden kanban-worker skill copies when deciding whether to inject --skills kanban-worker
• surface repeated-failure diagnostics when a task-level max_retries cap is lower than the dispatcher default
• add focused regression tests for both behaviors

Validation

• C:\Users\m_t_c\.hermes-sdgd\venv\Scripts\python.exe -m pytest -p no:timeout -o addopts='' tests\\hermes_cli\\test_kanban_cli_dispatch_passthrough.py -q
• C:\Users\m_t_c\.hermes-sdgd\venv\Scripts\python.exe -m pytest -p no:timeout -o addopts='' tests\\hermes_cli\\test_kanban_diagnostics.py -q

Context

This was extracted from SDGD issue SundayGundayLLC/sdgd#807, where Foreman Kanban worker pickups crashed because an archived skills/.archive/kanban-worker/SKILL.md made Hermes preload a missing skill name.

[tonydwb]

Code Review: Approved. Fixes two issues: 1) Kanban worker now ignores archived/hidden skill copies when checking availability. 2) Task-level max_retries is now properly honored in repeated_failures diagnostics. Clean fixes with good tests.

[teknium1]

Thanks for the focused diagnostics coverage. One half of this PR needs updating before it can be salvaged onto current main.

Problems

• The archived kanban-worker detection change is obsolete: commit 84e1d31e5 removed _kanban_worker_skill_available and the automatic worker preload. Current spawn logic forwards only explicit task skills at hermes_cli/kanban_db.py:8036; tests/hermes_cli/test_kanban_core_functionality.py:2768 asserts that no default skill is auto-loaded.
• The retry change records failure_limit_source, but the rendered diagnostic still says "The dispatcher circuit breaker" at hermes_cli/kanban_diagnostics.py:623. A task-level max_retries cap should be described as task-level.

Suggested changes

• Drop the obsolete archived-skill commit/test and reapply only the diagnostics work against current main.
• Use the source field when rendering the detail text and keep the lower per-task-cap regression test.

Automated hermes-sweeper review.

[GottZ]

This was generated by AI during triage.

Summary

Six PRs touch related Kanban dispatch and diagnostics symptoms, but they do not share a single root cause: #34578, #61035, and #65996 expose active-PR respawn guards; #62393 lets an explicit unblock bypass such guards once; #51386 distinguishes interactive claims from reclaimable workers; and #42859 combines obsolete skill-preload detection with task-level retry diagnostics.

Related pull requests

• #34578 [closed] related — (+138/-0) — closed, rejected/superseded: Adds dispatch output and a dedicated warning diagnostic for active-PR guards, but its recorded closure decision says it was superseded by the distinct sticky blocked-task promotion fix on developer/kanban-blocked-promotion-fix, not by #65996; it remains relevant as an earlier observability design reference.
• #42859 related — (+59/-1) — salvage diagnostics only: The archived-skill fix no longer applies because current main removed automatic kanban-worker preloading, while the task-level max_retries change still addresses a real diagnostic threshold mismatch. Consistent with the keep_open review on #42859, retain only the retry work and make the rendered detail distinguish a task-level cap from the dispatcher circuit breaker.
• #51386 related — (+352/-10) — keep open, incomplete: Introduces explicit non-expiring external claims to prevent live interactive sessions from being reclaimed and redispatched, which is distinct from respawn-guard observability. The keep_open review identifies a remaining correctness gap: detect_stale_running can still reclaim these holds, and the public --external syntax also needs documentation and reconciliation with current typed-block cleanup paths before merge.
• #61035 [closed] related — (+99/-9) — closed, wrong upstream: Exposes respawn_guarded in JSON, CLI, daemon, and gateway logs, excludes guarded-only queues from stuck telemetry, and documents review handoffs. It remains relevant as a tested reference implementation, but it was explicitly abandoned because it targeted the wrong upstream.
• #62393 related — (+307/-22) — keep open: Adds a one-shot, evidence-bound override so an explicit unblock permits exactly one guarded claim, then reapplies the guard after that claim, while also fixing guarded-queue health signals and exposing JSON reasons. This is a distinct control-flow fix rather than an observability duplicate; the documentation requested by the keep_open review was subsequently added in commit 110120e6.
• #65996 related — (+252/-32) — merge candidate: Provides the current-main active-PR observability fix by exposing dispatch results, replacing false stranded_in_ready reports with a bounded info diagnostic, and supporting that severity in CLI and dashboard surfaces. This aligns with the keep_open review on #65996, which verified the affected current-main paths and event-order assumptions.

Duplicates

#61035 and #65996 substantially overlap on active-PR dispatch visibility, with #65996 being the current-main implementation and #61035 already closed as wrong-upstream. #34578 overlaps in observability but should not be relabeled as their duplicate because its recorded supersession was the separate sticky blocked-task promotion fix; #62393, #51386, and the retry-diagnostic portion of #42859 address distinct mechanisms.

Suggested consolidation

Merge #65996 for the current-main active-PR observability path; #61035 can remain closed as the wrong-upstream predecessor. Keep #62393 as a separate one-shot-unblock fix, and keep #51386 open until the detect_stale_running, documentation, and cleanup gaps from its contributor review are addressed. Do not merge #42859 as-is: drop the obsolete archived-skill half and salvage the task-level retry diagnostic with source-aware wording. Do not reopen #34578 or treat it as superseded by #65996; preserve its recorded rejection in favor of developer/kanban-blocked-promotion-fix.

Cross-PR triage: Reviewed 6 pull requests and 0 issues in this complex. Each diff was read against this issue; Assessment working set: 96 kB of PR diffs, 8 kB of issue/PR text, 7 kB of discussion (15 comments), 2 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.

[GottZ]

This was generated by AI during triage.

Delta since our previous triage comment

@rosie-skypad reports that #62393 was rebased onto current main at head 13758e0, with the documentation follow-up retained and focused tests passing. This updates the branch-status evidence, but the delta includes no diff for the new head, so the reported retention of the one-shot override, five focused regressions, and exclusion of pruned legacy tests cannot yet be independently confirmed from the diff.

Changed pull requests

• #62393 related — (+314/-22) — updated head, pending diff verification: The author reports that the rebase resolved conflicts, retained the documented one-shot unblock override, and passed focused validation without restoring pruned tests; because no current-head diff is available in the delta, these remain author-reported rather than diff-verified. This does not conflict with the existing keep_open review, which treated the mechanism as salvageable and requested the documentation update that the author says remains included.

Suggested consolidation

The previous consolidation recommendation is unchanged: keep #62393 as a separate one-shot-unblock fix pending current-head diff verification.

Complex graph unchanged since our previous triage comment.

Cross-PR triage: Reviewed 6 pull requests and 0 issues in this complex. Diffs were read for 5 of 6 PRs (rest unavailable); Assessment working set: 75 kB of PR diffs, 8 kB of issue/PR text, 5 kB of discussion (12 comments), 0 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.