The agent’s report cap and the clamped command’s margin

Orientation

report_cap() in examples/programs/agent/application.mo is 262,144 bytes because a toolchain defect corrupted larger reports. That defect is fixed in both runtimes (toolchain-raw-memory-report, accepted 19 Sep 2026), so the number now describes nothing. The lead’s decision: a report stays bounded, but by what the profile can legitimately produce, so the cap is a defence that a correct run never meets. Worker: a fresh Claude Opus session, bypass permissions, own worktree and Herdr tab, based on the main that holds the end-to-end slice. Small: one sitting.

Write scope

examples/programs/agent/application.mo, its tests and sidecars, examples/programs/agent/tests/application-workspace-v1/ (matrix, README items 10 and the 256 KiB sentence, new evidence under 2 MiB). No toolchain, no wiki, no audit/, no HANDOFF.md, no machine, no push. Never use tr; ls is aliased, use /bin/ls.

Parts

  1. RED first. A matrix row with a transcript between 256 KiB and the new bound that must render whole and byte-identical in both runtimes (it fails today with transcript_too_large); sizes 300,000, 1,000,000 and the largest the profile allows. Commit the failing output.

  2. The bound. Derive report_cap() from the profile’s own limits (steps times the request and response bounds, plus the model text bound), as an expression of the existing functions, not a new literal. The bounds line of the profile reports the derived value; update matrix.py’s expectation. Rewrite the comment: no mention of the defect as a reason.

  3. Still refused. One row past the bound still gives the proved transcript_too_large error with persistence: proved.

  4. Whole-report check. For the largest size, diff the entire report between mo run and the mo build binary (the raw-memory worker compared only the last line).

  5. The clamped command’s collection margin (end to end v1’s defect D1). examples/programs/agent/workspace-adapter.mo timed sends a command the whole of what the run has left, and posted then waits for the reply for that same time; the service needs 0.5 to 1.4 s more to collect and prove cleanup (measured), so the last clamped command is always reported transport_timeout, execution unknown, though it completed. RED first against the loopback service with a delayed reply; then send remaining - margin with a margin of a few seconds stated as a named function, refuse below the floor as now, and show the reply received and the execution known. workspace-adapter.mo and its tests join the write scope.

Numbers

Wall time and peak resident memory of the largest report, both runtimes, best of five, load average beside them.

Done when

Every process under toolchain/bench/step36/guard.py. The focused matrix and the agent’s mo test green in both runtimes with real summary lines and exit codes; the full suite and the machine are the lead’s. Write your final report to examples/programs/agent/tests/application-workspace-v1/REPORT-CAP.md and commit it. While anything runs, wait in the foreground.

Result

Accepted 3:06 PM ET, 19 Sep 2026, merged to main through lead/verify-reportcap (78b093f3). Worker: Claude Opus 5, report examples/programs/agent/tests/application-workspace-v1/REPORT-CAP.md. report_cap() is 16 × (851,968 + 524,288) + 1,048,576 = 23,068,672 bytes, derived from the profile; transcripts of 300,598, 1,000,598 and 3,669,468 bytes render whole and identical across runtimes outside took_ms; the bound renders and one byte past it is a proved transcript_too_large. Defect D1 is fixed with candidate_margin_ms() of 5 s, and the lead saw it hold against the real service on the machine (the last clamped command received, execution known). Darwin full suite 263 of 263. Found and not caused by this work: real_bridge.py fails 2 of 3 cases on the base tree too, because the executor’s test double changed in harness steps 1 and 2; queued. Evidence: audit/evidence/2026-09-19/fable-lead-verification/reportcap/.