Repository navigation
docs(mochi): tell the watch checker an auth wall is not an outage - #13020
Conversation
|
Intent: A watch item whose page needs a login must be told it needs a login. The background checker had one tool for a Not a goal: Granting any tool to the unattended |
GPT 5.6 Review — ✅ no blocking findingsGPT 5.6 completed its review of This comment is updated in place on each push. Review detailsFINDING -- src/kiro_crew/apps/builtins/mochi/agents/context/prompt.md:52 -- False positive or not applicable? A repository writer can comment: |
Design Review (Fable 5) — ✅ PASSDesign-level review of Design-Verdict: PASS Confirm kiro-cli's [DESIGN-REVIEWED] e4de7f5 |
First Principles Review (Fable 5) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS Verify end-to-end that a Settings grant scoped to What this change shipsInventory (8 items) — 8 justifiedIntent: stop reporting a login-walled watched page as an outage, and let the checker use a user-granted signed-in reader — a FIX.
[FIRST-PRINCIPLES-REVIEWED] e4de7f5 |
Opus 5 Review — ✅ no blocking findingsReviewed Review detailsA chat-audience MCP grant — the Settings default — never reaches the background checker, so the newly permitted login-walled target can never be checked and the nudge asks for a grant the user already made. FINDING — src/kiro_crew/apps/builtins/mochi/agents/context/prompt.md:52 — [OPUS-REVIEWED] e4de7f5 Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
|
Inherited red, not this PR: The assertion names new Evidence: the test fails on the base commit It is already being fixed in #13022, which vets exactly those sites. No change |
The url kind bound the check to one tool, and the 3-strike nudge had one sentence for every failure. A page behind a login therefore failed three times and was announced to the user as a possible outage. The url row is now a two-step rule: fetch first, and on 401/403 look for a page reader among the tools actually mounted. The nudge takes its wording from the failure reason, so needs-auth asks for a login instead of reporting the site down. The chat prompt no longer refuses to create an item whose target needs a session. No grant changes: mochi-bg's tool list is byte-identical.
f1ec25c to
e4de7f5
Compare
GPT 5.6 Review —
|
GPT 5.6 Review — ✅ no blocking findingsGPT 5.6 completed its review of This comment is updated in place on each push. Review detailsNo findings. False positive or not applicable? A repository writer can comment: |
|
Status on head
This is not a verdict on the change. The four sibling lanes — Reruns tried on this same head, no new push: a The earlier inherited failure is gone: |
|
…rodotdev#13020) The url kind bound the check to one tool, and the 3-strike nudge had one sentence for every failure. A page behind a login therefore failed three times and was announced to the user as a possible outage. The url row is now a two-step rule: fetch first, and on 401/403 look for a page reader among the tools actually mounted. The nudge takes its wording from the failure reason, so needs-auth asks for a login instead of reporting the site down. The chat prompt no longer refuses to create an item whose target needs a session. No grant changes: mochi-bg's tool list is byte-identical. Co-authored-by: Zejiang Guo <zejiangg@amazon.com>
What changed
A watched page that wants a login is no longer reported as a page that is down.
urlcheck,mochi-watch/SKILL.mdweb_fetchmochi-watch/SKILL.mdneeds-authfor an auth refusal,access-deniedfor a bare 403 (bot, rate or region block)mochi-watch/SKILL.mdmochi-watch/SKILL.mdsummarysummarystays inside its 100-char bubble, the guidance moves tochatMessagemochi-watch/SKILL.mdagents/context/prompt.mdtargetMUST be fetchabletargetSHOULD be openable, so a page behind a login is still worth watchingThe checker is an agent reading a skill, so the bug was in the words. The
urlrow named one tool, so a 401 had nowhere to go. The nudge had one sentence for
every kind of failure, so the user was told the site might be down when the site
was fine and only the reader was anonymous. Three strikes later that is the worst
sentence to send: it points at the wrong thing to fix. A 403 gets its own reason
for the same reason — a bot or region block is not cured by signing in either.
Nothing is granted. The background agent's tool list in
mochi-bg.jsonisbyte-identical. A user who wants an authenticated reader there still grants it in
Settings, and
build_policyalready maps that grant onto the background agent —this change only makes the grant usable once it exists. Picking a tool by
description is also where a model could reach for a write tool, so the read-only
rule now covers every server instead of Slack alone.
test/test_mochi_agent_contract.pypins each half: theurlrow must name afallback path rather than a single tool, a bare 403 must not be called a login
wall, every nudge bubble must fit the 100-char cap, the auth wording must not
claim an outage, and the background grant is asserted byte-for-byte so a later
round cannot widen it quietly.
Backwards compatibility
No behaviour change for a page that answers anonymously, and no schema change:
the new reasons ride the free-text
resultfieldhistoryEntryalready carries.Testing
Related backend set green (134 tests). Two mutants, each caught: restoring the
single-tool
urlrow fails the fallback tests, and restoring the over-longbubble fails the cap test. Under both, the grant test still passes — so the tests
measure the skill prose, not the authorisation.
Why no screenshot: agent instruction text only — no rendered surface changes.
Mesh-3761