Skip to main content

Only the footer comes back — executor reasoning stops at the boundary

· 3 min read
Carefree Investments LLC
Anchor maintainers

An executor's reasoning, false starts, and quoted templates used to ride along with everything else in its reply — all the way back into the coordinator's own context, and from there into the next task's prompt. Only the structured footer does that now; the rest is archived to disk and discarded from context.

What was crossing the boundary​

Mythos-core rule 8 has always required a fixed footer — ## Result, ## How to verify, ## Deferred / concerns — at the end of every executor reply. But the orchestrator's own footer check only ever asked "is the footer present," then passed the entire raw reply back as the task's result. A verbose executor's scratch work, three abandoned attempts at phrasing an answer, or a quoted copy of the template from its own reasoning all rode along with it — straight into the coordinator's context and the critic's review prompt.

Extract, don't trust​

fleet_metrics.extract_footer pulls just the three required sections out, tolerantly:

extract_footer(text, max_lines=60)
# -> FooterExtraction(ok, footer_text, missing, truncated)

Case and spacing drift are tolerated (##Result, ## HOW TO VERIFY, Deferred/Concerns all match). When a heading appears more than once — an executor quoting the template mid-reasoning before its real answer — the last occurrence wins, so an early false start never gets mistaken for the finished result. All three sections are still required; missing any is a named rejection, not a partial accept. And the reconstructed footer is capped at 60 lines with an explicit [truncated by harness] marker, so smuggling a full transcript inside ## Result doesn't work either.

The transcript goes to disk, not to context​

scripts/orchestrate.py archives every dispatch's full raw reply to var/task-transcripts/<task-hash>.log — a metadata header per attempt, appended across retries and respawns — before anything crosses back to the coordinator. What crosses is relay_text(out), and the retry gate is what guarantees which branch of it fires: a reply missing any of the three required sections never reaches relay_text as an accepted result in the first place, so by the time it's called on a status: ok reply the extraction always succeeds — the coordinator's context and the next task's prompt see the extracted footer, never a raw-capped fallback. relay_text's raw-capped branch still exists for the outcome ledger's audit record on an exhausted-retries escalation, where "log what was said" matters more than the strict contract.

A reply missing any of the three required sections — not just one with no recognizable footer at all — gets exactly the treatment rule 6/8 already specified: one corrective retry with the expected shape quoted, then escalation.

Using it​

Nothing to configure for the footer relay itself. Transcript archival follows the same default/disable pattern as claimed-vs-actual tracking:

python scripts/orchestrate.py --plan-file .plans/features/big-refactor.md \
  --verify "pytest -q"
# transcripts land under <worktree>/var/task-transcripts/ by default;
# --transcripts <dir> to redirect, --transcripts "" to disable