Repository navigation
Closed as duplicate
Labels
bugSomething is not workingSomething is not workingchannel: stableReported from a stable releaseReported from a stable release
Description
Existing issues
- I searched open issues and this is not already reported.
KiroCrew version
0.13.0
Release channel
Stable
How is it installed?
Desktop app
Platform
Not platform-specific
What happened
Title: KAS secret store rejects the engine's "kiro-path-client" key kind, so every remote MCP OAuth sign-in fails ("key is outside the MCP sign-in namespace")
Environment
- Kiro Crew 0.13.0 (desktop app, darwin-arm64, macOS 26.7.3)
- Harness: Kiro Agent (KAS) via kiro-cli relay, agent.acp_backend = kas, agent.kas_bundle = false
- kiro-cli 3.5.0, KAS engine 0.66.119 (kas.server.starting in kiro-chat.log)
- agent.kas_persist_mcp_sign_in = true (default)
- Custom agent in ~/.kiro/agents with a remote MCP server:
"awsevents": { "type": "http", "url": "https://api.awsevents.com/mcp",
"oauth": { "clientId": "...", "redirectUri": "http://127.0.0.1:8976",
"oauthScopes": ["openid", "email", "events/access"] } }
Steps to reproduce
- Start a chat on the custom agent. The MCP report shows awsevents as "waiting for sign-in"
and the notice "awsevents (an MCP server you added) needs you to sign in ... Send a message
that names awsevents" appears. - Send the message "Sign in to awsevents" (as its own turn, agent idle).
Expected
- Crew sends _kiro/mcp/resetServer with startOAuth, the lock banner with the Authorize link
appears, tokens land in <data_home>/kas/mcp-grants/.vault/.
Actual
- No banner. The server flips to: awsevents: failed: key is outside the MCP sign-in namespace
- The grant vault only ever holds the ".client" entry for the server's connection hash,
never ".tokens".
Root cause (from reading both sides)
- kiro_crew/acp/kas_secret_storage.py line 89:
_KEY_RE = r"\Akiro.mcp.[0-9a-f]{64}.(client|tokens|verifier|discovery)\Z"
Any other key is refused at line 227 with SecretStorageRefused("key is outside the MCP
sign-in namespace"). The failure is not logged with the key, so it is invisible in
gateway.log. - KAS 0.66.119 (@kiro/agent dist/server/acp-server.js, class CredentialStorageManager) uses
five kinds: ["client", "kiro-path-client", "tokens", "verifier", "discovery"].
clientKind = "kiro-path-client" when the oauth config has no redirectUri, no clientId and
no clientMetadataUrl; otherwise "client". - On startOAuth the engine calls oauthProvider.invalidateCredentials("all"), which runs
clearCredentials("all") = Promise.all(delete over all five kinds). The delete of
kiro.mcp..kiro-path-client is refused by Crew, the Promise.all rejects, and the
sign-in fails with Crew's refusal text as the server error. - For a remote OAuth server WITHOUT a pre-registered client the primary kind is
kiro-path-client, so the very first readStored() is refused and the server fails before
any sign-in. - The engine also keys enterprise IdP sessions as kiro.mcp.idp..session (IdpSessionStore
storageKey()), which the regex rejects as well. - mcp_grant_kas.py notes the key scheme was checked against KAS 2.36.0 and 2.37.1; the
0.66.x engine added the new kind.
Suggested fix
- Accept the new kind in _KEY_RE (and in mcp_grant_kas's presence check):
(client|kiro-path-client|tokens|verifier|discovery)
and decide how kiro.mcp.idp..session should be handled (persist like client/tokens,
or memory like verifier/discovery). - Log the refused key's kind (never the value) so a namespace mismatch is diagnosable.
- Optionally: when a delete is refused, the engine's clear-all should not abort the whole
sign-in; that part is on the KAS side.
Workaround confirmed in code (not yet applied here)
- agent.kas_persist_mcp_sign_in = false: Crew stops declaring the secret store, the engine keeps
grants in memory and the sign-in flow works, at the cost of re-signing in per session.
Related log lines
- Every KAS process start logs:
[WARN] [mcp-sdk] SEP-2352: stored OAuth credential has no 'issuer' stamp (pre-upgrade storage
or provider not round-tripping the value). - gateway.log: "kas backend: installed kiro-cli 3.5.0 is newer than the releases the KAS
tool-name tables were verified on (2.24.0, 2.24.1, 2.27.1)"
Steps to reproduce
kirocrew-diagnostics-20261119-141432-e30e0e26.zip
Relevant log output
Anything else
Diagnostics bundle: kirocrew-diagnostics-20261119-141432-e30e0e26.zip
0 secret(s) auto-redacted and 137 path(s) collapsed to ~ or host name(s) replaced before packaging.
kiro-cli: kiro-cli 3.5.0
Host: macOS-26.7.3-arm64-arm-64bit
Metadata
Metadata
Assignees
Labels
bugSomething is not workingSomething is not workingchannel: stableReported from a stable releaseReported from a stable release
Activity
bolichen97 commentedon Nov 19, 2026
[operator: bolichen97] #39425 reports the same refusal of the
kiro.mcp.<hash>.kiro-path-clientkey with kiro-cli 3.5.0, also names thekiro.mcp.idp.<hash>.sessionkey and the missing log detail, and its comments cover the delete refused during a sign-in reset; the open fix there checks get, store and delete the same way. Your engine-side analysis and theagent.kas_persist_mcp_sign_in = falseworkaround are useful input for that issue. This is the same report as #39425, so we track it there and close this one.kiro-path-clientkey, so remote OAuth MCP servers fail with "could not be read from the secret store" and never get a sign-in banner #39425