Summary

• add persisted execution receipts plus a SQLite receipt ledger
• expose receipt reconcile/prune/maintenance surfaces in tools, CLI, and slash commands
• formalize a narrow direct terminal work-order lane on leased warm workers and surface runtime reuse metadata in receipts

H007 line overview

This PR is part 1 of the H007 line: a narrow, evidence-backed durable execution fabric for direct structured terminal work orders.

The line is intentionally split for reviewability:

• Part 1 (this PR): auditable execution substrate + receipt surfaces + direct lane/runtime metadata
• Part 2 (feat: add durable execution work orders #8403): durable execution work orders with queue/run_due, retry, stale-run reclaim, and runner controls

The key goal is to keep the claim narrow and testable rather than implying that general orchestration is solved.

Motivation

Hermes can already delegate work, but delegated execution is hard to audit and hard to reason about operationally.

This PR makes delegated execution inspectable:

• receipts are persisted under HERMES_HOME
• receipts are indexed in SQLite for query/reconcile/prune
• operator surfaces exist in both CLI and slash commands
• the narrow direct terminal work-order lane now surfaces worker/runtime reuse metadata explicitly

What this PR changes

• add tools/execution_receipts.py
  • durable JSON receipt artifacts
  • SQLite receipt ledger
  • reconcile/prune/query helpers
• add tools/execution_receipts_tool.py
  • list, query, prune, reconcile, maintenance_status, install_maintenance, remove_maintenance
• add operator-facing receipt surfaces
  • hermes receipts ...
  • /receipts ...
• forward execution_envelope and context_package through the delegate path
• persist richer execution metadata
  • execution path
  • envelope digest
  • worker mode / lease key / task ID
  • runtime kind / runtime ID / runtime reuse
• harden the direct terminal work-order substrate and persistent Docker cwd tracking
• add focused test coverage for receipts, delegate integration, receipt maintenance, CLI entrypoints, runtime metadata, and direct work-order execution

Validation

Focused local validation on this branch:

source /home/ubuntu/hermes-agent-dev/extraordinary-prototypes-2026-04-11/.venv/bin/activate
python -m py_compile \
  run_agent.py \
  tools/tool_result_storage.py \
  tools/environments/docker.py \
  tools/delegate_tool.py \
  tools/execution_receipts.py \
  tools/execution_receipts_tool.py \
  hermes_cli/receipts.py \
  model_tools.py \
  toolsets.py \
  hermes_cli/tools_config.py \
  hermes_cli/commands.py \
  hermes_cli/main.py \
  tests/run_agent/test_run_agent.py \
  tests/tools/test_delegate.py \
  tests/tools/test_docker_environment.py \
  tests/tools/test_tool_result_storage.py \
  tests/tools/test_execution_receipts.py \
  tests/tools/test_execution_receipts_tool.py \
  tests/hermes_cli/test_receipts.py

pytest \
  tests/run_agent/test_run_agent.py \
  tests/tools/test_delegate.py \
  tests/tools/test_docker_environment.py \
  tests/tools/test_tool_result_storage.py \
  tests/tools/test_execution_receipts.py \
  tests/tools/test_execution_receipts_tool.py \
  tests/hermes_cli/test_receipts.py \
  -q -o addopts=''

Result:

• 421 passed, 1 warning

External real-run evidence for the supported direct lane that this PR surfaces:

• classic warm mean: 4.35s
• direct warm mean: 0.20s
• about 95.4% faster on the validated supported slice
• 40/40 valid runs
• 40/40 exact-match runs
• one reused concrete Docker runtime ID across valid warm runs

Notes for reviewers

This PR intentionally does not claim:

• general orchestration is solved
• all delegation is dramatically faster
• host-bound git worktree or host-venv pytest workloads already belong in this lane

The claim is intentionally narrow:

• direct structured terminal work orders on leased warm workers are now auditable and operator-visible
• the supported slice has already been benchmarked and validated externally

If the overall H007 direction looks good, the follow-up queue/control-plane PR is already prepared as #8403.