Agent Harness Wiki

2 min read

Kanban error repair — 2026-05-05

Timestamp: 2026-05-05T06:51:02Z Board: spec-dataset-evolution

Root cause

The board was not failing because of the research tasks themselves. Default-profile Kanban workers were crash-looping during Hermes startup because the globally enabled basis plugin registered tool parameter schemas with top-level anyOf clauses:

  • basis_reduce_spec
  • basis_validate_packet

OpenAI Codex rejected the function schema before the worker could make its first useful model call:

HTTP 400: Invalid schema for function 'basis_reduce_spec': schema must have type 'object' and not have 'oneOf'/'anyOf'/'allOf'/'enum'/'not' at the top level.

This affected the running/default workers for:

  • SPEC-DATA-21 / t_383dad01
  • SPEC-DATA-JCODE / t_63d0e80e

Repair applied

Patched local Basis plugin repo:

  • Path: /Users/ericfode/.hermes/plugins/basis
  • Commit: 0061d32 fix: make basis tool schemas codex-compatible

Change shape:

  • removed top-level anyOf from the two tool parameter schemas
  • moved cross-field requirement guidance into descriptions and kept runtime validation in handlers
  • added regression coverage asserting registered parameter schemas do not use provider-rejected top-level composition/negation/enum keywords

Verification:

uv run --extra dev pytest -q
13 passed
 
hermes -p default chat -q 'Reply exactly OK' -Q --toolsets basis
OK

Board recovery

After the plugin fix:

  • SPEC-DATA-JCODE / t_63d0e80e completed successfully.
    • Outputs:
      • outputs/spec_data_jcode_seed_v0_1/jcode_first_calibration_seed.md
      • outputs/spec_data_jcode_seed_v0_1/jcode_seed_manifest.json
      • work/spec_data_jcode_seed_v0_1/kanban_metadata.json
  • SPEC-DATA-21 / t_383dad01 generated the required manual-audit gate artifacts, but the spawned worker process lingered after artifact generation. The task was completed manually after deterministic verification.
    • Key output directory: outputs/spec_data21_manual_audit_gate_v0_1/
    • Metadata: work/spec_data21_manual_audit_gate/kanban_metadata.json
    • Verification: python3 -m pytest tests/test_spec_data21_manual_audit_gate.py -q → 3 passed
    • Gate recommendation: hold_bounded_staged_expansion
    • Reason: human primary/dual/adjudicated labels are unresolved; semantic edges remain non-gated; raw content export remains fail-closed.
  • SPEC-DATA-22 / t_977f775e was dispatched after both parents completed and correctly blocked before crawl work.
    • Block metadata: work/spec_data22_gate_checked_staged_expansion/kanban_metadata.json
    • sha256 observed in worker handoff: ed87aae7dc2ca6cab685696af757fd3b90d30606ba69533e5589342e1e79ba30
    • Block reason: SPEC-DATA-21 gate is HOLD, not pass: human primary/dual/adjudicated labels are unresolved, so SPEC-DATA-22 must not run the staged crawl until labels are completed and gate estimates are recomputed.

Final board state after repair

triage    0
todo      0
ready     0
running   0
blocked   1
done      45

Blocked:

  • SPEC-DATA-22 / t_977f775e

No active Kanban worker process remained for t_383dad01, t_63d0e80e, or t_977f775e after final verification.

Interpretation

The original error condition is repaired. The remaining blocked task is intentional and evidence-based, not a runtime failure. The project is now waiting on real human/adjudicated audit labels before staged expansion, which is exactly what the gate was designed to enforce.

Graph View

Backlinks