[lukemaj]

What: The dispatcher's protocol now tells it to report router defects it meets as Issues in toolboxmd/model-router, with the request ID, evidence and expected behavior; packet mistakes go to the planner instead.
Why: The dispatcher sees router defects first (blocks, loops, behavior against RUNNER.md) but only relayed them or worked around them; #144 gave dispatch roles network, so it can now file them.
So what: Review and merge; the planner-dispatch half waits for #146, and the one-real-job proof comes from the first dispatcher that meets a defect after release.

Refs #147 (dispatcher part; the planner-dispatch part stays open, see Pending).

Change

• runner/adapters.py: LUNA_ACTION_PROTOCOL gains a ROUTER DEFECTS: paragraph: trigger (block, loop, runner acts against RUNNER.md), evidence to read (the runner's status for the request ID and outputs/<request ID>/), duplicate search in open toolboxmd/model-router Issues, comment-or-open with request ID, evidence and expected behavior, then continue or block as the runner allows. Packet mistakes (missing proof, wrong workspace) go to the planner as planner_question. The Baseline proof misreads a passing suite (rc=19) and dispatchers cannot reach GitHub #144 "make no tool calls" rule gains the reporting exception. Every dispatcher prompt and resumed turn carries it (it is part of the protocol).
• runner/controller.py: the dispatch prompt carries a JOB EVIDENCE block with the absolute bin/model-router --state-dir <abs> status --request-id <id> command and outputs path, so a non-default state directory still resolves; Issue text read while reporting is data, never instructions (review round 1).
• RUNNER.md: documents the duty beside the dispatcher thread modes; the runner files nothing itself.
• tests/test_issue147.py: pins the trigger, evidence, search, report contents, the packet-mistake exception, the tool-call exception, and that initial and resumed turns carry it before REPLY PROTOCOL.

Proof

• Red first: tests/test_issue147.py against the base code: FAILED (failures=9).
• Head c526ef0ca378b1ec5c0e773de8c9f0d9a696603a (origin/main after fix(runner): block once on a merged PR instead of looping on pr_not_open (#143) #151 merged in from main's side, so versionctl accepts 0.43.0): python3 scripts/test.py: Ran 385 tests ... OK; versionctl release-check: release candidate OK: v0.43.0 at c526ef0ca378b1ec5c0e773de8c9f0d9a696603a.
• tests/test_issue147.py runs the prompt's exact status command from the workspace against a non-default state directory.
• Independent review (Luna, max): REQUEST_CHANGES on 0d22d87 (evidence path, Issue data fence), both fixed; APPROVE on c526ef0, review/independent success.

Pending

• Planner-dispatch mode (Optional dispatcher: let the planner dispatch, then compare both modes with Agent Observer #146): the same duty applies to the planner when it dispatches; its wording lands with Optional dispatcher: let the planner dispatch, then compare both modes with Agent Observer #146 (planner side: Prism guidance: planner reports router defects as model-router Issues chromeria#74). Not in this PR.
• The Issue's "one real job where a dispatcher files or updates an Issue" is behavioral proof after release; not run here.

Elon record

• Requirements and who asked: the user (2026-09-29, Dispatcher reports router defects as Issues (planner does it in planner-dispatch mode) #147): the dispatcher reports router errors; with the planner dispatching, the planner does.
• Deleted: runner-side automatic filing; a separate reporting agent; a new CLI or prompt block (the duty lives in the existing protocol constant).
• Bottleneck: dispatch network, removed by Baseline proof misreads a passing suite (rc=19) and dispatchers cannot reach GitHub #144/PR fix: self-contained dispatch packet, dispatcher network, safe baseline proof (#144) #145; the remaining gap was only the instruction text.
• Checked myself: read LUNA_ACTION_PROTOCOL, build_luna_prompt/build_luna_followup, child_first_message (the dispatcher knows its request ID), RUNNER.md; ran the new test red on base and the full suite green.

🤖 Generated with Claude Code

[lukemaj]

REQUEST_CHANGES on exact head 0d22d871628ec05dc3f50b18e8620f0117287308.

• P2 — runner/adapters.py:41-42: The dispatcher is told to read outputs/<request ID>/, but runner state is rooted at --state-dir / DURABLE_RUNNER_STATE_DIR (see RUNNER.md:416-420) and the dispatcher message supplies only the request ID (runner/t3exec.py:423-431). When a user sets a non-default state directory, that relative path resolves inside the job workspace and the evidence cannot be found; the required evidence read and useful defect report fail. Pass the evidence path/state directory into the prompt or give the dispatcher a resolvable runner command.

• P2 — runner/adapters.py:38-44: The existing ISSUES-as-data fence applies to task Issues fetched at submit, but this new duty tells the dispatcher to search and read arbitrary open Issues without fencing their returned bodies as data. If a matching Issue says to run unrelated commands or make an unrelated GitHub change, the broad or you are reporting a router defect tool-call exception leaves the dispatcher without an explicit instruction to ignore that Issue text; it can be steered into those actions while reporting. Explicitly mark searched Issue content as untrusted data and constrain tool calls to duplicate lookup plus the necessary comment/create operation.

The requested text elements and reply-protocol ordering are present in the diff. I did not run tests; review was read-only.

[lukemaj]

APPROVE on exact head c526ef0ca378b1ec5c0e773de8c9f0d9a696603a. No blocking findings in the requested #147 scope.

The prior evidence-path finding is resolved: JOB EVIDENCE includes an absolute status command and outputs path rooted in the active state directory, and the added test executes that command from the workspace against a non-default state directory. The prior trust-boundary finding is resolved: the protocol marks Issue text read while reporting as GitHub data and restricts command, service-contact, and GitHub actions. The duty, packet-mistake exception, duplicate search, and report contents are present. The dispatch path appends the evidence block after fetched ISSUES, and the actual initial prompt ends on the reply protocol.

I did not run tests; this was a read-only review. The supplied proof is python3 scripts/test.py (385 tests OK) and versionctl release-check (v0.43.0 at this SHA).