Repository navigation
fix(connections): Jira/Confluence start Atlassian sign-in in KAS chats; card says how KAS chats sign in - #30823
Conversation
d697c4b to
19df0b2
Compare
First Principles Review (Opus 5.5) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS Nothing blocks. The fix is a declared stopgap: the description says the card and KAS chats still keep separate grants. The new registry field has a single real value. Notes
What this change shipsInventory (7 items) — 6 justifiedIntent: a KAS chat user asking about Jira or Confluence gets the Atlassian sign-in, and the Connections card tells KAS users their chats sign in separately. This is a FIX (a stopgap).
[FIRST-PRINCIPLES-REVIEWED] ac97acf Description read[DESCRIPTION-READ] 3ad299c6de5bae42b396565327e7967f9458edd138df9cb1cfe99f8f206979ea That sha256 names the PR description this verdict read. Recompute it with |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Nothing blocks: a registry vocabulary entry fed to the one existing matcher, and an optional status flag. It is additive and easy to revert, and screenshots are attached. Notes
[DESIGN-REVIEWED] ac97acf Description read[DESCRIPTION-READ] c42f045fac95d814d854b42c2ed60ce4d34adcad348d30822f916ecae87b6441 That sha256 names everything this verdict read: the PR description, plus the 1 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with |
UX Review (Opus 5.5) — ✅ PASSUX-level review of UX-Verdict: PASS Nothing blocks. The one addition is a passive line of copy, every state it can be in is shown in a screenshot, and the blind reader read it roughly as intended. Notes
[UX-REVIEWED] ac97acf Description read[DESCRIPTION-READ] 56b8221a97808fd1f7854370f6da50f66278202cc7b467205989d4d25aa2bec0 That sha256 names everything this verdict read: the PR description, plus the 9 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with |
Opus 5.5 Review — ✅ no blocking findingsReviewed Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
GPT 6.1 Review — ✅ no blocking findingsGPT 6.1 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: |
19df0b2 to
66b36a1
Compare
66b36a1 to
e158c20
Compare
…s; card says how KAS chats sign in registry: optional sign_in_words per provider (atlassian: jira, confluence); text_names_mcp_server matches them whole-word, like an alias. status: connected rows carry signInInChat while chats run on KAS; the card says to type the slug once in a chat.
e158c20 to
ac97acf
Compare
buluoray
left a comment
There was a problem hiding this comment.
🤖 Kiro Crew Auto-Pipeline [operator: buluoray]:
Verdict: APPROVE at ac97acf.
What this does for users. In a KAS chat (the default harness), "show my Jira issues" or "search Confluence" now offers the Atlassian sign-in. Before this, only the word "atlassian" did. While chats run on KAS, each connected card also tells the user that chats sign in separately and which word to type. That closes the gap where the card said Connected and the chat still had no tools. Kiro CLI chats are unchanged.
Evidence
- KAS identity is checked positively.
chat_backend_is_kas(src/kiro_crew/connections/status.py:191-205) compares the loadedagent.acp_backendtoACP_BACKEND_KAS, and an unreadable config answers no. The loader already resolves a Kiro CLI pick that runs on the v3 relay into KAS (src/kiro_crew/config/loader_builders.py:510-518), so a relayed install gets the line too. - The flag is set only on a connected row that does not carry
backendLacksOauth(status.py:523-524). Pre-registered rows keep their existing caveat. - The matcher extension is scoped.
sign_in_words(name)returns words only when the server name equals a registry slug that declares them (src/kiro_crew/connections/registry.py:743-747). Onlyatlassiandeclares any. Each word is validated as one lowercase word of 3+ characters (registry.py:516-527). Matching reuses the existing whole-word, case-insensitive pattern (src/kiro_crew/acp/session_handle_mcp.py:663-666). - No credential or URL is involved. The new code reads config and a static word list, and it logs nothing.
- The copy is a new string in the catalog,
pages.connectionsPage.chat_use_kas_sign_in, present in en.manual, en-XA and the 12 shipped locales. It gives a fact the card did not show before (chats need their own sign-in, and the word to type) and names no harness. - Tests at this head: test_kas_mcp_sign_in, test_connections_registry and test_connections_backend_oauth gave 167 passed. ConnectionsPage.coverage.test.tsx gave 65 of 65 on two runs. On the first, cold run, the new KAS test failed once and then passed alone and in two full-file runs, so it may be order or warm-up sensitive. CI has 83 success, 11 skipped and 0 failing, and PR Readiness is green. The merge with main is clean.
- I ran four mutations, one at a time, and restored the tree after each:
chat_backend_is_kasreturning True turnedtest_kiro_cli_never_marks_sign_in_in_chatandtest_an_unreadable_config_claims_no_kas_chatred. This pins "Kiro CLI chats behave as on main".- Dropping
sign_in_wordsfrom the matcher turned both newtest_a_server_is_named_by_its_name_or_an_obvious_alias[atlassian-...]cases red. - Dropping the connected-status gate turned
test_kas_leaves_a_not_connected_row_without_the_caveatred. - Removing
signInInChatfromchatUseNotein ConnectionsPageCard.tsx turned "tells a connected row on KAS to type the server name once in chat" red on 2 of 2 runs.
Checked from source only (not run): that the extra words reach the sign-in offer only on KAS. The only caller is _match_mcp_sign_in_ask (session_handle_mcp.py:712), and on Kiro CLI begin_mcp_sign_in refuses every start (acp-client.md). I did not check against a live KAS chat.
Bot lanes: First Principles, Design, UX, Opus and GPT all pass, with no HOLD.
- The Design lane notes that the matcher finds these words only because the MCP entry key equals the slug, and nothing pins that link. I checked this:
mcp_server_alias(slug)is the identity for today's slash-free slugs (connections.md, "Tool-alias key shape"). It does not gate, and it is follow-up 2 below. - The UX lane raises three points: "once" is ambiguous, the word is not set apart in the copy, and on a shared-grant row the line is hidden in a tip. These are copy and placement suggestions, not regressions, and they are follow-up 1 below.
Follow-ups (non-blocking)
- Copy and placement. "Once" holds while
agent.kas_persist_mcp_sign_inis on (the default). With it off, a KAS chat signs in again each time its process restarts, so the line overstates for those users. Either show the line only while persistence is on or drop "once". In the same change, consider the UX lane's points: set the word apart, and keep the line on the card for shared-grant rows. - Add a test that fails if a connection's mounted MCP entry name stops equal to its registry slug, or move
sign_in_wordsintomcp_server_aliasesso the slug and alias link lives in one place. - The new ConnectionsPage coverage test failed once on a cold run and then passed three times. Worth watching for flakiness.
…s; card says how KAS chats sign in (#30823) registry: optional sign_in_words per provider (atlassian: jira, confluence); text_names_mcp_server matches them whole-word, like an alias. status: connected rows carry signInInChat while chats run on KAS; the card says to type the slug once in a chat. Co-authored-by: Joe Guo <zejiangg@amazon.com> (cherry picked from commit 02a7173)
Problem / Motivation
Goal: In a KAS chat, asking about Jira or Confluence offers the Atlassian sign-in, and the Connections card tells KAS users how their chats sign in.
On 0.12.0 a KAS chat offers an MCP sign-in only when the user's typed message names the server (
want_mcp_sign_in_named/text_names_mcp_server, since #26481). Foratlassianthe only alias isatlassian, so "show my Jira issues" gives no Authorize banner and no tools. The Connections card meanwhile shows Connected, because its sign-in wrote a kiro-cli grant that KAS never reads. Users see "Connected" and a chat that cannot use it.Why it matters
Every KAS user of the Atlassian connection (KAS is the default chat harness) hits this: the natural words "Jira" and "Confluence" never start a sign-in, and the card gives no hint that a KAS chat needs one. This is the short-term stopgap; the full fix (card sign-in runs on KAS, so card and chat share one grant) is a separate, larger change.
Not a goal
127.0.0.1: fixing it touches two strings in 13 catalogs, not one, so it is left for its own change.backendLacksOauthcaveat unchanged.What changed (motivation → approach → change)
Symptom → cause → change:
atlassian→ a registry provider may now declare optionalsign_in_words(connections/registry.json:atlassian→jira,confluence; validated as lowercase single words of 3+ characters).text_names_mcp_serveralso tries the words of the provider whose slug is the server name, with the same whole-word, case-insensitive match as an alias. No other provider declares words, so nothing else changes.signInInChaton a connected row whileagent.acp_backendis KAS (chat_backend_is_kas; an unreadable config answers no). It is not set on a row that already carriesbackendLacksOauth, since KAS cannot sign those in by name. The card reuses the existing "In chats" line (or the shared-grant tip): "Chats sign in separately: type<slug>in a chat once." It names no harness, like thebackendLacksOauthcopy: the line shows only while chats run on KAS (UX Review on 19df0b2; "separately" per UX Review on e158c20). The word is the slug: for Atlassian,atlassianalways works andjira/confluencedo too, so the line names one word to keep it short. "Once" is once per gateway whileagent.kas_persist_mcp_sign_inis on (the default). New stringpages.connectionsPage.chat_use_kas_sign_ininen.manual.jsonand all 12 shipped locales;en-XAregenerated.acp-client.md(what names a server) andconnections.md(the new row flag and its copy).Backwards compatibility
Compatible:
sign_in_wordsis optional,signInInChatis an added optional field, and alias matching only gains words foratlassian.Removes nothing: the change only adds an optional registry field, an optional status field, one catalog string and extra matching words.
Tests
test/test_kas_mcp_sign_in.py: "show my Jira issues" and "search CONFLUENCE …" nameatlassian; unrelated text,jiras, and the same words forlinear/notiondo not.test/test_connections_registry.py: malformedsign_in_wordsis rejected; onlyatlassiandeclares words; an unknown slug has none.test/test_connections_backend_oauth.py: KAS + connected DCR row →signInInChat; KAS + not connected → none; Kiro CLI → none; pre-registered row on KAS keepsbackendLacksOauthand nosignInInChat; unreadable config → not KAS.website/src/test/ConnectionsPage.coverage.test.tsx: the line renders withsignInInChat, and no "In chats" line without it.Proof: not provable -- other: prove.py was not run in this worktree; the new matcher and status cases assert behavior that main lacks (no
sign_in_words, nosignInInChat).Local: 220 passed (the three Python files +
test_connections_status*), 65 passed (ConnectionsPage.coverage), i18n parity +i18n-check.mjswithI18N_BASE_REF=origin/maingreen.Manual verification
Rendered the real built SPA against stubbed APIs (Playwright, as
capture-connections-caveat-tip.mjsdoes): connected Atlassian and Notion rows withsignInInChatshow the line, and without it show none, in light and dark. Not checked against a live KAS chat.Screenshots / video
Chat harness KAS: connected Notion and Atlassian rows show the "In chats" line.
KAS with a shared grant: the line joins the tip beside Disconnect (opened).
Chat harness Kiro CLI: no line.
The "Couldn't load your starred Bots" toast is from the stubbed APIs, not this change.
Related Issues
Refs #26481, #26525
Sibling sweep
Swept:
git grep -n "mcp_server_aliases\|text_names_mcp_server" -- src/Siblings: none --
text_names_mcp_serveris the one matcher; its only caller is_match_mcp_sign_in_ask.Pattern harvest
Not generalizable: the gap is a missing product vocabulary entry (brand names of one provider), not a code pattern.
Checklist
Contribution License Agreement
Contributed under the project's Apache-2.0 license.