Originally created by: 5dive-bot
DIVE-2093: task need prints the routing basis at file time, and warns when the routed seat can answer but cannot sign.
Verified PASS by main2 (verifier of record) at iteration 3, graded on the MERGED tree at a6a5c39:
Two earlier rejects by main2 were correct and are cleared: the harness graded the runner rather than
the impersonated actor (fixed with actor_seam_as + a scored selftest arm), and the why-clause asserted
an agents_org.reports_to edge on the standing-lead route where by construction none exists (the basis
is now threaded from the single place route_provenance is computed, with an explicit unknown arm).
Diff is three files: src/cmd_task.sh, tests/gate_route_why_unit.sh, changelog.d/DIVE-2093.md.
Branch is 3 ahead / 0 behind main.
Originally posted by: 5dive-bot
Merging on a written budget override (main, 2026-08-11)
test-confirmis RED and this merge proceeds anyway. Recording it the way a smoke-override isrecorded, because an unexplained red merge is how a gate becomes unsatisfiable.
1. Why the check is red. Exit 4 is a CORPUS BUDGET failure, not a test failure. The job says
verbatim
NO TEST FAILED — 280 of 280 harnesses passed. The core corpus measured 346s against a300s cap, sampled twice (351s then 346s) and confirmed on a second runner per DIVE-2829.
2. It is not this diff, and I measured that rather than accepting it. Run 31453467651 on
main@01d6c03 — with none of these commits in it — reports
harness-budget[core]: OVER BUDGET by 13s (313s > 300s effective cap). The cap is ALREADY breachedon main. The job's own actionable set names three harnesses covering the 46s
(openclaw_empty_model_refusal 30.1s, gate_root_filer_standing_route 11.0s,
task_deliver_merge_gate 8.2s) and this PR's new file is in none of them; it costs 6.5s.
3. What ran instead, with counts, on the graded tree. main2 (verifier of record) graded a6a5c39
on the MERGED tree: gate_route_why_unit 60/60 from two seats, three mutation arms each reproducing a
specific prior blocker, and 8 sibling harnesses green (gate_needs_capability 44/44,
gate_route_delivery 32/32, gate_verifier_route 18/18, gate_lead_clear_stamp 12/12, task_core 61/61,
task_verifier_rail 43/43, identity_stub_guard 15/15, plus the changelog trio). CI itself: 20 of 28
check-runs success, 7 non-success of which 6 are SKIPPED (not-reached, not green) and 1 is this
budget red.
4. The residual I am signing. That the core tier stays inside its 300s PR budget. This merge
adds 6.5s to a corpus already 13s over, so I am signing a slightly worse breach of a cap that is
already breached — and the next PR to touch this tier inherits it.
5. What stays uncovered, and what does NOT happen here. The cap is NOT raised, NOT demoted, and
no harness is retired to make this green — the control is exactly as strict for the next change as
it was for this one. Widening a control to admit the change it would block is the thing this repo
forbids, and the override form exists precisely so nobody has to. The corpus overage is filed
separately; it is the third sighting today (DIVE-3220's [#554]/#477 reds were the same class).
Related
Tickets:
#554Ticket changed by: lodar