Round 10 results (2026-10-14) #
Round-10 code on a preview with production's flags plus the round-10 flags named per colony. All runs are labeled MEASURED; logs pass through node packages/bench/src/evidence.ts redact.
| Check | Result |
|---|---|
Agent2Agent (checks/a2a.log) |
The colony's card is public and names the JSON-RPC URL, one skill and the bearer scheme; SendMessage without a token (and with a wrong one) answers 401 with a sentence saying writes need the admin token; with it an intent was declared and read back submitted without a token; a scripted agent took it over MCP and landed it inline; GetTask then reported completed with the landing and its commit and named ft why; ListTasks listed it; a second intent was declared and canceled (withdrawn with the reason); a completed task cannot be canceled (-32002); unknown method -32601, bad params -32602, bad JSON -32700; a matching message answered the existing intent's task (twin intents). |
Tests before landing (checks/preland.log) |
With FT_PRELAND naming the colony and prelandPaths set: the breaking change came back before landing (63 ms) naming the test and its error, with the run in its evidence (failed, on its base, 1 test file, 0 of 1 passed); nothing landed; the fixed change landed with the run (passed, 1 of 1) in its lineage note, and certification stayed green; a change outside the named paths ran nothing; the colony counted one failed and one passed run. On the colony the flag does not name, the breaking change landed with no run and certification went red after landing. |
Quoted acceptance (checks/acceptance.log) |
With FT_QUOTED_ACCEPTANCE naming the colony and review on: the change meeting both acceptance lines landed on the second opinion, its note keeping both lines met with a quoted changed line each (const lower = text.toLowerCase();, return lower.replace(/ +/g, "-");); the change meeting only line 1 came back, the second opinion naming line 2 as not met. The keep-line case was sent to a person by the reflex before the second opinion ran (needs_human 0.91), so keep lines rest on the unit tests; the script reports that as skipped. |
First-red demo on production (checks/firstred_demo_production.log) |
The same demo on the live service, colony first-red: the search found the breaking landing in 3 test runs (202 ms) over 6 landings, the scripted repair landed, and trunk certified green again on 294d7d5. Its replay opens at t=3&speed=2. |
First-red demo (checks/firstred_demo.log) |
scripts/bisect.ts --demo: a node:test project certified green, six changes landed while certification was held (the fourth broke sub()), the red result's search found the breaking landing in 3 test runs (216 ms) over the 6 landings since green, naming its commit, proposal and intent; the repair intent's why names the commit; a scripted agent took the repair and landed the fix; certification was green again on the new head. The landing page's firstred card reads it from the public certifications list. |
Import-name revalidation, the cascade on Hono (rv-grow.json, rv-edit.json, checks/rv-grow.log, checks/rv-edit.log) |
With FT_REVALIDATE naming both colonies and review on: when the writer added an export to src/hono.ts, 15 of 17 readers landed on their first proposal (2 went back on the review model's verdict, not as stale, and were re-proposed by the scenario); when the writer edited the body of an existing declaration, 17 of 17 went back, 17 naming the file and the changed statement's line (16). |