Repository navigation
Description
Status: needs-human — this asks for one product decision and no code moves until the owner answers. Do not implement from this issue.
A capability's credential that is typed as a literal string is stored in plain text and handed back to the Settings client — nothing in the product says "use a secret reference" except the placeholder text.
What happens (verified on main e3caccbab4e59ccdf006e1ad5074642d8dddddee)
core/src/host/capabilities.js:270-273—configForbuilds the mount spec withurl: item.urlandheaders: resolveValues(item.headers, botId).resolveValues(:258-267) substitutes only$DSH_SECRET_*references; any other header value is passed through as written.- The item itself is what was typed:
itemInput(:104-120) acceptsurlafter a protocol check andheadersviastringMapwith no shape rule, soAuthorization: Bearer …andhttps://user:pass@host/mcpare legal input. - That item is persisted with the roster/store, so the literal credential sits in the host's own state file, unencrypted, the same way a secret reference does.
core/src/host/capabilities.js:240-245—libraryViewspreads the whole item (url,headers,env) to its caller, andcore/src/host/channel.js:191sendscapabilities: rt.capabilityLibraryView()to the browser, so the Settings form receives the plaintext value back.- Tell a Bot in its system prompt when one of its capabilities is down #3274 (merged
2026-11-21T01:35:40Z,e3caccba) closed the paths that put such a string in front of the model: a new URL carrying userinfo is refused at input, and the stored status reason is masked (withoutUserinfo(redactSecrets(joined)),:321). It did not change what is stored at rest or what the Settings client receives, and it cannot clean an entry saved before it.
Why it is a question and not a patch
Two defensible product rules, and they differ in what a user is allowed to type:
- Refuse a literal credential at input. Only a
$DSH_SECRET_*reference is accepted for a header value that looks like a credential (and userinfo in a URL stays refused). One rule, one place to audit, no plaintext at rest — but a user who pastes a token into Settings gets an error instead of a saved capability, and every capability saved that way today stops working. - Keep literal values, mask them wherever they leave the host, and warn in the UI. The Settings form shows a masked value, the status/prompt paths keep the current masking, and the form states plainly that the credential is stored unencrypted in this host's state file and that a secret reference is the safer choice. Nothing breaks, but the plaintext stays on disk and the product leans on the user reading the warning.
No recommendation is recorded here on purpose: this changes the security surface, so the owner decides. Whichever way it goes, the issue then becomes an ordinary defect/feature issue with a scope.
Out of scope here
The other residual from #3274 — an MCP server echoing a literal header value in its own error body, which lands in the status line — is a separate defect with its own issue (#3275); it does not need this decision.
Dedupe
Searched main and closed issues: literal credential, capability headers secret, userinfo URL name only #3274 (the merged fix that closed the prompt path) and this pair of residuals; #2002/#2007/#2016 are the 300-unit cut, #2699/#2700 keep paths and keys out of room lines. No duplicate.
Bot: FlyBot Lead
Written-by: FlyBot Lead
Activity
HugoXK commentedon Nov 21, 2026
Release
Written-by: FlyBot Lead
Released: 548298c416044b14b44dd2f92fdd5586
Reason: Needs-human decision issue: no implementation claim and no branch. Released so no Dev picks it up; the owner answer is the only next step.
At: 2026-11-21T02:49:35Z