Harness step 8: Mo agent findings

Orientation

Step 8 of mo-harness-in-mo. Findings M1 to M8 in audit/evidence/2026-09-19/fable-overnight-review/README.md. The code is the Mo coding agent under examples/programs/agent/; learn Mo from the surrounding code and toolchain/PRELUDE.md. Worker: a fresh Claude Opus session, bypass permissions, own worktree based on the lead’s main, which already contains the rebuilt application workspace. The lead owns the wiki, audit and acceptance.

Write scope

examples/programs/agent/*.mo, examples/programs/agent/tests/coding-fixture-v1/ (not its evidence/), and the examples/programs/.mo.ids records and verified: footers those changes regenerate (use the real mo test --write; never hand-edit a hash). No toolchain source, no tests/application-workspace-v1/ except where a changed signature forces a call-site edit, no mo-wiki/, audit/, HANDOFF.md. No machine, no push, no new syntax. Never use tr. No .mo files in any evidence folder: the corpus runs every .mo it finds.

Parts

  1. M1, M2. tests/coding-fixture-v1/boundaries.mo: each mode asserts the outcome it exists for, unguarded, and an error arm fails the test unless that mode expects that exact error. Remove the assertion the code already guarantees. Show each rewritten test failing against a deliberately broken copy of the code it covers (a mutant), and keep that output.
  2. M5, M6. edit-empty and command-failure join CASES in run.py; the four edit refusals assert their reason.
  3. M4. A completed response already in hand is reported as completed even if the deadline has just passed; the run may still stop. Failing test first.
  4. M3. main.mo’s outer ask has real slack over budget plus grace, or the budget is derived from the outer deadline. Failing test first.
  5. M7. exact-edit.mo:43 refuses instead of falling back to "".
  6. M8. The flag-gated budget checks in run.mo (profiled, formerly fixture) fold into asks?/uses?/budget_end so there is one budget rule; resolve the off-by-one at the token bound and say which way and why. The 12 legacy CLI goldens, the coding-fixture matrices (22 cases per runtime) and the application workspace controls must still pass. If one budget rule cannot serve legacy, fixture and application runs, stop and tell the lead with the source evidence; do not invent a fourth mode.

Numbers and done when

Both runtimes (mo run and a mo build binary), every process under toolchain/bench/step36/guard.py. You may run zig build and focused tests with -Dtest-filter; the full zig build test is the lead’s. Report real summary lines and exit codes, Mo source lines before and after per file, the mutant outputs for part 1, decisions the brief did not cover. Evidence under 1 MiB. Small commits as yourself with a Co-Authored-By line naming your model.

Result

Accepted 19 Sep 2026, 9:03 AM ET (decision-log). The worker’s report is committed with its work; the lead’s runs are in audit/evidence/2026-09-19/fable-lead-verification/.