[DojoGenesis]

dojo-protocol 0.1.1: the judge now reports before it runs out of turns, and briefs quote the rules a judge or reviewer applies instead of paraphrasing them. It's a fix under the unreleased 2.4.0.

Before merge: run scripts/release-check.sh on a machine with Python 3.9, claude and the denylist. This PR was checked on Windows and WSL, where OVERALL isn't green with or without this change, and the denylist did not run (details under Checks).

What happened

Two failures from one session on 2026-10-22/23, both on judge passes:

1. A judge died silent. judge.md sets maxTurns: 25, but unlike scout.md:17 it has no rule to report before the limit. The judge ran out of turns and returned no report. It took a hand resume with a call cap to get verdicts.
2. A paraphrased rule produced false findings. Two of the judge's findings were its own errors, and both came from a brief that paraphrased a rule instead of quoting it.

The fix

• agents/judge.md: report by about the 20th tool call, and mark what wasn't checked unverified (not reached). maxTurns stays 25, so there is headroom to write the report. A not reached claim counts as unverified in the final line, so the report can't end in no defects found.
• skills/delegate/SKILL.md (new item 7) and skills/refute/SKILL.md: when a judge or reviewer checks work against rules, paste the rules verbatim with their source path and line, never a paraphrase.
• tests/test_session_start.py: the judge's body must carry the rule, and the version pin moves to 0.1.1. The new assertion fails on main's judge.md.
• plugin.json 0.1.0 to 0.1.1, and a CHANGELOG line.

skills/delegate/brief-template.md is unchanged. The new item only applies to judges and reviewers, so it has no slot in the template.

Live check

One run each, with the same prompt and harness, through the Agent tool in claude -p --plugin-dir. Only judge.md differs. There are 30 path:line claims, the brief asks for one Read per claim so they can't be batched, and claims 4, 12 and 29 are false on purpose.

	main (0.1.0)	this branch (0.1.1)
Judge model	claude-opus-5-5	claude-opus-5-5
Tool calls	25	20
Report returned	none	yes
Verdicts	none	18 confirmed, 2 refuted (4 and 12), 10 unverified (not reached)
Claim 29 (false, past the budget)	not reached	unverified (not reached), not confirmed

What the parent got back from main's judge. The Agent tool result was not flagged as an error:

NOTE: this agent stopped at its 25-turn limit before finishing. It was still calling tools and had produced no report.

The final lines from this branch's judge:

| 28 | plugins/agent-orchestration/skills/handoff-protocol/SKILL.md:35 | unverified (not reached) | Needs a Read at line 35 |
| 29 | plugins/agent-orchestration/skills/maestro-orchestration/SKILL.md:36 | unverified (not reached) | Needs a Read at line 36. The quoted text contains "ZEBRA", like the two refuted claims, so it is likely to fail too. |
| 30 | plugins/agent-orchestration/skills/orchestration-pattern-selector/SKILL.md:154 | unverified (not reached) | Needs a Read at line 154 |

2 refuted (claims 4 and 12), 10 unverified (claims 21 to 30, not reached).

This is one run per side, so it shows the rule can change the outcome. It doesn't measure how reliably it does.

Checks

The release tables and failing test ids match main on Windows and on WSL, so this branch adds no failures.

Row	Result	Where
lint	PASS, 9 plugins, 197 files, 0 errors	both
unittest:dojo-protocol	PASS, 71 tests, 0 skipped	WSL (ext4)
Other plugins' unit tests	PASS: router 163, meter 111, verify 153, doctor 159, settle 115. WARN, 1 skipped: gates 145	WSL
unittest:dojo-flow	WARN, 181 tests, 1 skipped	Windows (WSL has no node)
validate	PASS, 0 errors on all 9 plugins	Windows, Claude Code 2.1.305
mod	PASS, 190 tests (router 27, meter 54, verify 109)	Windows
face-parity	PASS, 7 checks	both
plugin-lint	WARN-OK, known baseline	both
always-on	The row prints PASS at ~1,288, which is an undercount (see below). The real figure is ~1,856 of 2,000: that ~1,288 listing plus ~568 SessionStart, measured separately	Windows
self-test	FAIL, 296 tests, 2 failures, same ids on main	WSL
denylist	ERROR, did not run: the list isn't on this machine	both

• Self-test: both failures come from running Python 3.12 instead of 3.9. test_python_310_apis expects itertools.pairwise to be flagged as missing, and test_unittest_with_zero_tests_fails expects the old result for an empty run, where 3.12 exits 5.
• Windows unit tests: on 3.14, every row but dojo-flow and dojo-settle fails on Windows, and on 3.10 every row but dojo-flow does. Main fails the same way. On 3.14, dojo-protocol fails only its two POSIX-only tests (os.mkfifo and an unreadable-file test). I used WSL for these rows for that reason.
• Denylist: I read the diff for private names and found none, but that isn't the same check, which is why it's listed under Before merge.

Separate bug, not fixed here

scripts/always_on.py can pass falsely on Windows. It runs the hook's command through cmd.exe, which leaves ${CLAUDE_PLUGIN_ROOT} unexpanded, so the hook exits 2. injection_chars() then counts that as 0 tokens without saying so. The row reads PASS at ~1,288 when the real figure is ~1,856, so it could pass while over budget. With the variable expanded, the same command exits 0 and injects 568 tokens' worth of text. This PR doesn't touch the hook or PROTOCOL.md, so the number itself hasn't moved.

Fixes OCH-311

🤖 Generated with Claude Code