[sleegme]

Summary

TEMPORARY measure: disable Excavator until agy native subagents can select
Gemini 4 Argon. This replaces the originally requested runner-dispatched role;
no Excavator runner policy, transport or ledger schema is introduced.

The established AGY 1.3.6 probe showed that native pro and flash both ran
gemini-3.8-flash-tiered, while exact Argon frontmatter failed registration.
Flash does not fit this difficult diagnosis/repair role. Prior model evidence
remains in .omo/evidence/74-argon-routing.md.

Design and temporary workaround

• Keep agents/excavator.md, its full repair contract, tool list and shell
  guard. Set subagent: false; mainAgent: false remains unchanged.
• Remove Excavator from enabled specialist routing, declared delegation
  acceptance and native-delegation telemetry attribution. A packet claiming an
  Excavator delegation fails closed before review.
• Tell Bulldozer not to spawn Excavator or substitute a runner call. Ordinary
  bounded repairs stay with Bobcat and its existing Strix advisor gate.
• For a hard repair with no safe in-scope remediation left, use the existing
  evidenced BLOCKED result and supervisor-owned ledger path. An unsettled
  architectural question uses existing NEEDS_DEEP. Neither path promotes the
  milestone or starts Zen review. This avoids a new status, retry loop or
  dispatch protocol.
• Update routing rules, repository instructions, English/Korean status,
  internals and the canonical contract. Historical live evidence is labeled
  historical, not evidence for current enabled routing.

Re-enable

This is not a permanent removal. When a bounded native-child probe proves
Argon selection in actual executed-model metadata, select that verified native
tier, change subagent back to true, restore the documented Bulldozer/spine
routing and tests, and remove temporary notices together. The preserved agent
file documents these steps. No exact-model runner workaround is shipped.

Verification

• node --test --test-reporter=tap tests/*.mjs: exit 0; 290 tests,
  269 passed, 0 failed, 21 skipped, 107463.033592 ms.
  Skips are 20 existing Windows-specific tests and one existing persistent
  watch-mode test. Bare node --test discovers zero tests in this repository,
  so the explicit filename glob was used.
• python3 -m unittest discover -s tests -p "test_*shell_guard.py":
  15 passed, exit 0.
• Regression coverage verifies disabled frontmatter, adapter rejection,
  declared-delegation rejection for DONE/BLOCKED/NEEDS_DEEP, and durable blocker
  and escalation evidence with no worker dispatch, promotion or Zen review.
• Updated only the two intentionally changed byte-prefix baseline lengths and
  hashes; existing equality and appended-guidance assertions remain intact.
• Node syntax checks and git diff --check passed. Grepping the complete
  committed diff from 8f0e258 for /home/ returned no matches.
• Public runner API rejects Excavator even when an Argon slug is supplied.
  scripts/runner.mjs and scripts/ledger.mjs are unchanged from the base.
• agy plugin validate . passed on installed AGY 1.4.7 with the pre-existing
  warning that root AGENTS.md is not loaded as a plugin rule.
• A fresh private-HOME local plugin installation and agy agents listing
  exited 0 without altering the shared installation.

Not live-verified and remaining risks

• Native delegation rejection is now live-verified by the bounded Flash
  gate follow-up below. Future native Argon selection, real repair quality
  and future AGY versions remain unverified.
• Hard failures that need Excavator's intended capability now stop through the
  existing blocker/escalation path until native Argon is verified and the role
  is deliberately re-enabled.
• LSP diagnostics were unavailable because typescript-language-server is not
  installed. Syntax checks and the full test suite were used instead.
• Independent gate review requested C1 native-boundary evidence. The follow-up
  below supplies that evidence; independent re-approval is not claimed.

The main checkout was not modified. Do not merge this PR.

Gate C1 follow-up: live native rejection and enabled control

Evidence: .omo/evidence/173-native-excavator-probe.md, with the exact prompt,
redacted 33-event native stream and redacted Bobcat transcript alongside it.
Tested production head: e3a7c5ccea3b8d6a5568ceb4b57acda05372619a;
this follow-up adds evidence only.

On AGY 1.4.7, installed that PR head into a fresh private HOME and empty
workspace. Worker files were byte-identical to PR head. Only the private
Bulldozer parent's mainAgent flag was changed to true so it could be selected
natively; its role body and both worker definitions were unchanged.

One bounded gemini-3.8-flash-high native Bulldozer conversation
(7cad473a-ebc9-4bbf-bf1e-de6606a369bd) requested both workers sequentially:

• PASS, Excavator rejected: actual native invoke_subagent step 2 entered
  ERROR. Its stored error_details and payload report
  subagent "excavator" not found or not allowed to be invoked.
  No child ID was created.
• PASS, same-setup Bobcat control: native step 4 created child
  b0b9b93f-cdee-4586-8d72-b071a42ef5d7; its stored completed response was
  NTG173_BOBCAT_CONTROL_READY.
• Stored source URIs identify the isolated installed Bulldozer/Bobcat files.
  The fresh conversation store contained only the parent and Bobcat database,
  with no Excavator child. This is native runtime evidence, not the spine
  callback or a model's unsupported claim.
• CLI exited 0, SUCCESS, reported duration 39.177011262 seconds under a
  150-second AGY timeout and 170-second outer bound. No shared installation
  changed. No disable-mechanism fix was needed.

Exact completed full-suite summary

node --test --test-reporter=tap tests/*.mjs, exit 0, on the unchanged
production tree:

1..290
# tests 290
# suites 0
# pass 269
# fail 0
# cancelled 0
# skipped 21
# todo 0
# duration_ms 107463.033592

Retained shell guards, exit 0:

Ran 15 tests in 5.700s
OK

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota.