fix(kanban): require tool evidence before worker completion

[LeonSGP43]

Summary

• auto-block dispatcher-scoped workers that call kanban_complete without any prior non-kanban_* tool calls in their recorded session
• emit a first-class protocol_violation event for that path so operators can distinguish fabricated completion from an ordinary block
• cover the new gate with targeted kanban tool tests and keep existing worker session stamping expectations explicit

Closes #32746

Local proof

• uv run --frozen --extra dev pytest tests/tools/test_kanban_tools.py -q
• uv run --frozen --extra dev ruff check hermes_cli/kanban_db.py tools/kanban_tools.py tests/tools/test_kanban_tools.py
• git diff --check

Notes

• direct push to NousResearch/hermes-agent returned 403 for LeonSGP43, so this PR is opened from the fork branch
• final pre-push overlap check found no assignee, no linked closing PR, and no keyword-overlap open PR for Kanban dispatcher accepts kanban_complete with zero tool calls (no tool-evidence gate) #32746

[teknium1]

Thanks for addressing the fabricated-completion path. The underlying gap remains on current main: tools/kanban_tools.py:626-632 completes a dispatcher worker without a non-kanban tool-evidence check.

Problems

• The new gate reads SessionDB.get_messages() in tools/kanban_tools.py:511-513. Those rows are durable history only (hermes_state.py:4016-4026); the agent flushes its in-memory messages through _persist_session at run_agent.py:1678-1695, whose current call sites are after the CLI run returns. Therefore a valid live sequence such as read_file followed by kanban_complete is not yet visible to this query and would be auto-blocked.
• The new tests seed SessionDB before directly calling _handle_complete, so they do not cover that live ordering.

Suggested changes

• Source evidence from active-run tool activity, or explicitly persist it before evaluating the completion gate.
• Add an end-to-end same-run non-kanban-tool → completion test, alongside the no-tool-call rejection case.

Automated hermes-sweeper review.

[teknium1]

This gate reads only persisted session rows, but the current agent flushes its in-memory tool-call messages via _persist_session after the CLI run returns (run_agent.py:1678-1695; call sites are in cli.py). A normal worker's read_file followed by kanban_complete in the same run therefore has no durable evidence yet and will be auto-blocked. Please source active-run activity or persist it before evaluating this gate.

[benperry6]

I prepared a tested follow-up directly on this PR's author branch rather than opening a competing upstream PR:

• follow-up PR: fix: require successful tool evidence before Kanban completion LeonSGP43/hermes-agent#2
• commit: benperry6@e19ff6c

It changes Kanban completion proof from tool-call intent to successful material execution evidence, with a metadata-only current-run receipt in task_events, native Hermes/Codex/MCP coverage, bounded cross-process visibility wait, and fail-closed protocol_violation handling.

Verification: 19 targeted Kanban, 24 Codex projector, and 58 Codex app-server session tests passed; Ruff, py_compile, and git diff --check passed; independent review PASS.

The separate product decision about a reasoning-only/no-tool completion exception remains intentionally with maintainers.

[benperry6]

Current-main refresh, kept separate from every other Kanban fix.

• base: 1bbb6e5bce56e721ab685af4cd87df21bbff4d35
• cherry-pickable commit: benperry6@5de30d1
• branch: https://github.com/benperry6/hermes-agent/tree/fix/kanban-completion-proof-32940

The bug still reproduced on that base: a worker could complete its card after emitting a tool-call intention without any successful material result. The refreshed patch records metadata-only evidence only after successful real Hermes/Codex tool completion, binds it to the current run, and leaves the card in flight on a protocol violation.

Verification on the exact commit: RED reproduction observed; 13 focused tests passed; 123 tests across the touched files passed; the complete Kanban matrix passed with 509 passed, 0 failed, 1 Windows-only skip. Ruff, compileall, and git diff --check passed.

This supersedes my older author-branch follow-up e19ff6c499f0b9832f05af127254c28cc333cf8d. I did not open a competing upstream PR; the commit is available for the existing PR author or maintainers to cherry-pick.