Skip to content

fix(mcp-gateway): never pool a credential vendor - #24672

Merged
buluoray merged 2 commits into
mainfrom
fix/mcp-unpoolable-credential-vendors
Oct 27, 2026
Merged

buluoray merged 2 commits into
mainfrom
fix/mcp-unpoolable-credential-vendors

Conversation

@buluoray

@buluoray buluoray commented Oct 27, 2026 •

Copy link
Copy Markdown
Contributor

Problem / Motivation

Goal: A credential-vending MCP server is never shared across chats, by any path.

On main, nothing stops it. UNPOOLABLE_SERVERS in src/kiro_crew/mcp_gateway/rewriter_spec.py is empty, so every stdio server reports can_stub: true (mcp_gateway_routes.py: can_stub = is_stdio and not denylisted). No other check knows what a credential vendor is: the shareability assessment only withholds a share for declared env names with the AWS_SECRET, AWS_SESSION or OAUTH prefixes, and the page's bulk "share all" picks every can_stub && !stub row. So a single toggle or a bulk press can put creds-agent behind one shared backend.

Why it matters

It bypasses the chat sandbox:

  • Each chat's own copy of an MCP server is a child of that chat's kiro-cli, so it runs inside the chat's sandbox. The sandbox tiers hide the credential directories and the ~/.aws profile files (sandbox_tiers.py).
  • A shared backend is spawned by gatewayd outside every sandbox, as the user (docs/architecture/mcp.md: "The sandbox is not applied to a pooled backend at all").
  • So sharing a credential vendor hands every chat, including a prompt-injected agent, a vendor that can read what its own sandbox withholds. Same user, same rights: the gap is the sandbox, not another account.

This is a gap on main today, not something a pending PR introduces. #19661 makes sharing one click, which is why it should close first.

Not a goal

  • A general rule for detecting credential vendors by behaviour. This lists the two known server names.
  • Changing how any other server is pooled.

What changed (motivation → approach → change)

  • UNPOOLABLE_SERVERS = {"creds-agent", "local-creds-agent-mcp"}, with the reason next to it. The rewriter leaves them unwrapped, the servers endpoint reports denylisted: true, can_stub: false, and the batch stub endpoint skips them as cannot_stub. The existing denylist plumbing does the rest; no new path.
  • Comments and one design note that described the set as empty now say it lists only the credential vendors.
  • _FINGERPRINT_SCHEMA 10 → 11, so kept overlays regenerate with the denylist applied.
  • Every reader asks the same rewriter_spec.is_unpoolable: pooled_session_servers and injection_server_names skip a vendor stub, filter_target_env drops its target, and LaunchApprovals.admits_launch (the check gatewayd runs before a spawn) refuses it whatever its recorded approval. So an overlay, sidecar or approval written before this change, or kept by a failed regeneration, cannot pool a vendor.

Security

Narrows only. A listed server keeps running once per chat, inside that chat's sandbox, exactly as it does when it is not in mcp_gateway.stub_servers.

Backwards compatibility

An install that listed either name in mcp_gateway.stub_servers keeps the entry, but the rewriter no longer wraps it: from the next overlay regeneration each chat runs its own copy again, and the dashboard row shows it as not shareable. No config migration is needed.

Tests

  • test/test_handlers_mcp_coverage_gateway_servers.py: for both names, against the live set (not patched), the servers endpoint reports denylisted and can_stub: false even when the operator lists the server, and _stub_eligibility skips it as cannot_stub even when the assessment says share.
  • test/test_mcp_gateway_rewriter_fingerprint.py: a schema 10 overlay that pooled creds-agent is rebuilt on the next pass with unchanged inputs and leaves it unwrapped (_FINGERPRINT_SCHEMA is now 11, because the denylist is not a fingerprint input).
  • test/test_mcp_gateway_session_inject.py, test/test_mcp_launch_approval_store_records.py: a stale overlay with a vendor stub injects nothing for it; a recorded approval admits an ordinary server but not a vendor, in the target filter or admits_launch.
  • Revert-reddens: with rewriter_spec.py restored, all 4 denylist cases fail; with the schema left at 10, the overlay test fails.
  • Local: 308 rewriter, fingerprint, control-plane and gateway-servers tests, and 347 tests across the gateway servers, pool identity, rewriter and stub-eligibility files pass.

Manual verification

Not run on a live gateway. What was not measured: whether a per-chat creds-agent can vend inside a sandboxed chat on this host (that needs a signed-in chat in a pod with a working OS sandbox).

Related Issues

Found while reviewing #19661 (one-click MCP sharing).

Pattern harvest

Rule candidate: a server whose job is to hand out what the chat sandbox withholds (a credential vendor) is never pooled, because a shared backend runs outside every sandbox; exclude it by name until there is a capability-based refusal.

Checklist

  • Two commits, with Conventional Commits titles
  • Regression tests and revert-reddens proof
  • Comments and design note updated
  • No secrets added

Contribution License Agreement

Contributions are accepted under the project's Apache-2.0 license.

creds-agent and local-creds-agent-mcp join UNPOOLABLE_SERVERS, so the
rewriter leaves them unwrapped and the dashboard reports them as
can_stub: false. A single toggle, a bulk "Enable all" and the batch stub
endpoint therefore all refuse them.

Each chat's own copy of a credential vendor runs inside that chat's
sandbox, which hides the credential directories and the ~/.aws profile
files (sandbox_tiers). A shared backend is spawned by gatewayd outside
every sandbox. Sharing one would hand every chat a vendor that can read
what its own sandbox withholds.

Comments that described the set as empty now say it lists only the
credential vendors.
@buluoray
buluoray requested review from a team and KiroCrewCaptain as code owners October 27, 2026 01:28
@buluoray
buluoray requested a review from iamwhatever October 27, 2026 01:28
@github-actions github-actions Bot added the readiness: checking Automated validation is still running label Oct 27, 2026
@github-actions

github-actions Bot commented Oct 27, 2026 •

Copy link
Copy Markdown
Contributor

UX Review (Opus 5.5) — ⏭️ skipped

Revision ec24bf647b49c4f6fb051044889bdb4f57af6844 touches no user-facing surface (no changes under website/ or committed screenshots), so the UX review was skipped. Does not block merge.

@github-actions

github-actions Bot commented Oct 27, 2026 •

Copy link
Copy Markdown
Contributor

Design Review (Opus 5.5) — ✅ PASS

Design-level review of ec24bf647b49c4f6fb051044889bdb4f57af6844 — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS does not.

Design-Verdict: PASS

Nothing blocks: this only narrows what gets pooled. Every reader checks the same is_unpoolable predicate, and the schema bump regenerates overlays written before the change.

Notes

  • Untested: whether a per-chat creds-agent can still vend from inside a chat's sandbox, which hides the credential directories it reads. If it can't, the vendor stops working for operators who used to share it. A revert undoes that, but say so to those operators.

[DESIGN-REVIEWED] ec24bf6

Description read

[DESCRIPTION-READ] 4e8a2110282ee6c86c90e648f488c04eaef0b4a9b745996cb192dcd93686efa2

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 .github/scripts/pr-description-capture.sh over the same inputs; if it differs, the description or that evidence changed after this verdict was formed and any finding drawn from either is unproven.

@github-actions

github-actions Bot commented Oct 27, 2026 •

Copy link
Copy Markdown
Contributor

First Principles Review (Opus 5.5) — ✅ PASS

Premise-level review of ec24bf647b49c4f6fb051044889bdb4f57af6844 — why this exists and whether the shipped surface is the smallest honest version. Updated in place on each push. A BLOCK verdict blocks PR readiness; PASS does not.

First-Principles-Verdict: PASS

Nothing blocks. The PR fixes a real sandbox-bypass defect, and tests that fail on base back it up. The one gap is the claim that every reader goes through one predicate: the diff adds a second spelling, and five sites still test the set directly.

Notes

  • Item 5 — duplicate of rewriter_spec.is_unpoolable: the description says "Every reader asks the same rewriter_spec.is_unpoolable", but launch_approval._unpoolable_stems rebuilds the name-or-alias test by hand. Separately, I grepped for in UNPOOLABLE_SERVERS and found 5 sites that still test the set directly and skip the helper: mcp_gateway_routes.py:566, mcp_gateway_stubs.py:190, rewriter_spec.py:499, rewriter_spec.py:713 and rewriter_spec.py:891. Subtraction: replace _unpoolable_stems with {stem for stem in … if is_unpoolable(...)}-style use of the one predicate, or drop the "every reader" sentence.

What this change ships

Inventory (6 items) — 5 justified

Intent: stop anyone from sharing a credential-vending MCP server across chats, so it always runs inside each chat's own sandbox. This is a FIX.

  1. creds-agent and local-creds-agent-mcp show as not shareable; the single toggle and bulk share skip them — justified
  2. Overlays kept from before this change regenerate on next boot (schema 10→11) — justified
  3. A stale overlay's vendor stub is no longer injected into sessions — justified
  4. gatewayd refuses to spawn a vendor backend whatever its recorded approval — justified
  5. New is_unpoolable predicate, plus a separate stem-mapped copy in launch_approval — duplicate of rewriter_spec.is_unpoolable
  6. Comments that called the set "empty" are reworded — justified

[FIRST-PRINCIPLES-REVIEWED] ec24bf6

Description read

[DESCRIPTION-READ] e9da9cce5b400ccb5a2dcd09efd927d6ddde03a85d30207d8ef532acf82df495

That sha256 names the PR description this verdict read. Recompute it with .github/scripts/pr-description-capture.sh; if it differs, the description changed after this verdict was formed and any finding drawn from the description is unproven.

@github-actions

github-actions Bot commented Oct 27, 2026 •

Copy link
Copy Markdown
Contributor

Opus 5.5 Review — ✅ no blocking findings

Reviewed ec24bf647b49c4f6fb051044889bdb4f57af6844 — this comment is updated in place on each push.

Review details

No findings.

[OPUS-REVIEWED] ec24bf6

Verdict parsed from the review's SHA-scoped output markers for commit ec24bf647b49c4f6fb051044889bdb4f57af6844.

False positive or not applicable? A repository writer can comment:
/ai-review override fable ec24bf647b49c4f6fb051044889bdb4f57af6844 false-positive: <reason, 20 to 500 characters>

@github-actions

github-actions Bot commented Oct 27, 2026 •

Copy link
Copy Markdown
Contributor

GPT 6.1 Review — ✅ no blocking findings

GPT 6.1 completed its review of ec24bf647b49c4f6fb051044889bdb4f57af6844 and found no blocking issues.

This comment is updated in place on each push.

Review details

No findings.
[GPT-REVIEWED] ec24bf6

False positive or not applicable? A repository writer can comment:
/ai-review override gpt ec24bf647b49c4f6fb051044889bdb4f57af6844 false-positive: <reason, 20 to 500 characters>

@buluoray

Copy link
Copy Markdown
Contributor Author

Re GPT 6.1 BLOCK on 1e7273d (kept overlay keeps a pooled vendor): real, fixed in bbf08e1.

  • _FINGERPRINT_SCHEMA 10 → 11 in rewriter.py, with the reason next to it, so the next boot rebuilds every kept overlay with the denylist applied, whatever the package version.
  • New test test_a_kept_overlay_from_before_the_vendor_denylist_drops_the_vendor_stub: a schema 10 pass with an empty denylist pools creds-agent; the next pass with unchanged inputs leaves it unwrapped. It fails with the schema left at 10.
  • 308 rewriter, fingerprint, control-plane and gateway-servers tests pass locally.

@buluoray

Copy link
Copy Markdown
Contributor Author

Re GPT 6.1 BLOCK on bbf08e1 (a failed regeneration keeps an overlay that still holds a vendor stub; its verdict was withheld from this thread, read from the run artifact): real, fixed in e2a7802.

The rewriter was the only place the denylist was applied. Every reader now asks the same rewriter_spec.is_unpoolable (name or alias):

  • pooled_session_servers and injection_server_names skip a vendor stub, so a kept or stale overlay injects nothing for it and the session resolves the vendor from its agent spec, inside its sandbox.
  • filter_target_env drops a vendor's KIROCREW_MCP_TARGET_* entry before it becomes gatewayd's env (not recorded as a refusal).
  • LaunchApprovals.admits_launch refuses a vendor whatever its recorded approval; gatewayd runs it before every spawn, directly or from its snapshot.

Other spawn paths checked: daemon/launch.py reaches launch_approved/admits_launch; the slack broker and the rewriter cache path go through filter_target_env. Tests: a stale overlay with a vendor stub injects nothing for it; a recorded approval admits an ordinary server but neither the target filter nor admits_launch admits a vendor. All 4 fail without the reader checks.

@buluoray
buluoray force-pushed the fix/mcp-unpoolable-credential-vendors branch from e2a7802 to 482bc11 Compare October 27, 2026 02:51
The rewriter applies UNPOOLABLE_SERVERS while it writes an overlay, but an
overlay, env sidecar or approval written before the vendors were listed (or
kept by a failed regeneration) could still pool one.

- _FINGERPRINT_SCHEMA 10 -> 11. The denylist is not a fingerprint input, so
  a kept overlay from before it (same version, unchanged inputs) would keep
  the vendor's stub; the bump rebuilds every kept overlay.
- Every reader asks the same rewriter_spec.is_unpoolable (name or alias):
  pooled_session_servers and injection_server_names skip a vendor's stub, so
  a session resolves it from its agent spec, inside its own sandbox;
  filter_target_env drops a vendor's KIROCREW_MCP_TARGET_* entry before it
  becomes gatewayd's env, without recording a refusal; and
  LaunchApprovals.admits_launch, the check gatewayd runs before it spawns a
  backend, refuses a vendor whatever its recorded approval.

Tests: a schema 10 overlay that pooled creds-agent is rebuilt unwrapped; a
stale overlay that still holds a vendor stub injects nothing for it; a
recorded approval admits an ordinary server but not a vendor, in the target
filter or admits_launch. Each fails without its fix.
@buluoray
buluoray force-pushed the fix/mcp-unpoolable-credential-vendors branch from 482bc11 to ec24bf6 Compare October 27, 2026 03:31
@github-actions github-actions Bot added readiness: passed Eligible automated validation passed for the current revision and removed readiness: checking Automated validation is still running labels Oct 27, 2026
@iamwhatever

Copy link
Copy Markdown
Contributor

Holding ec24bf6 for human review because it changes how credential vendors run and what gatewayd may launch outside the chat sandbox.

I found no defect. The change only narrows pooling: creds-agent and local-creds-agent-mcp join UNPOOLABLE_SERVERS, and the readers (pooled_session_servers, injection_server_names, filter_target_env, admits_launch) refuse them even from a stale overlay or approval (launch_approval.py L468-L478). The two earlier GPT 6.1 BLOCKs have author replies and the current head has no GPT or Opus findings.

For the human reviewer:

  • Design Review notes it is untested whether a per-chat creds-agent can still vend inside a sandbox that hides the credential directories. If it cannot, operators who shared it lose a working vendor. This has no reply.
  • First Principles notes that the description's "every reader asks the same is_unpoolable" is not quite true: _unpoolable_stems re-implements it and five sites test the set directly. Both names are slash-free, so there is no practical gap, but the claim or the code should match. This also has no reply.
  • PR Readiness is still pending on this head.

— crew-pr-reviewer (r128)

@buluoray
buluoray merged commit 2464225 into main Oct 27, 2026
108 checks passed
@buluoray
buluoray deleted the fix/mcp-unpoolable-credential-vendors branch October 27, 2026 14:03
@github-actions github-actions Bot removed the readiness: passed Eligible automated validation passed for the current revision label Oct 27, 2026
bolichen97 added a commit that referenced this pull request Oct 29, 2026
Cherry-picks these main commits onto release/0.11.0, in merge order:

- fix(dashboard): tell the crewmate about unresolved fields, draw no stale band (#24069)
  (cherry picked from commit f9d8e19)
- fix(mcp-gateway): never pool a credential vendor (#24672)
  (cherry picked from commit 2464225)
- fix(acp): keep MCP tools direct on KAS when the spec grants no tool_search (#24297)
  (cherry picked from commit ddc1fea)
- fix(crew): move a working icon avatar as one still image so it cannot tear (#24678)
  (cherry picked from commit e062153)
- fix(crew-mode): keep a crewmate's own chat on the conductor (#24894)
  (cherry picked from commit 7582829)
- fix(memory): name a damaged memory.db instead of restore advice (#24567)
  (cherry picked from commit efe7624)
- fix(env): search Builder Toolbox's Windows bin dir for agent binaries (#24563)
  (cherry picked from commit 686d39a)
- fix(discord): re-admit bot-opened threads after a gateway restart (#21776)
  (cherry picked from commit 072a37b)
- fix(web): keep a hidden model hidden after a harness switch respells its id (#24564)
  (cherry picked from commit 501f455)
- fix(gateway): say the gateway is running once boot finishes (#25029)
  (cherry picked from commit 02627ec)
- fix(mcp): say why an mcp.json settings file could not be read (#22169) (#24572)
  (cherry picked from commit ed75593)
bolichen97 added a commit that referenced this pull request Oct 29, 2026
Cherry-picks these main commits onto release/0.11.0, in merge order:

- fix(dashboard): tell the crewmate about unresolved fields, draw no stale band (#24069)
  (cherry picked from commit f9d8e19)
- fix(mcp-gateway): never pool a credential vendor (#24672)
  (cherry picked from commit 2464225)
- fix(acp): keep MCP tools direct on KAS when the spec grants no tool_search (#24297)
  (cherry picked from commit ddc1fea)
- fix(crew): move a working icon avatar as one still image so it cannot tear (#24678)
  (cherry picked from commit e062153)
- fix(crew-mode): keep a crewmate's own chat on the conductor (#24894)
  (cherry picked from commit 7582829)
- fix(memory): name a damaged memory.db instead of restore advice (#24567)
  (cherry picked from commit efe7624)
- fix(env): search Builder Toolbox's Windows bin dir for agent binaries (#24563)
  (cherry picked from commit 686d39a)
- fix(discord): re-admit bot-opened threads after a gateway restart (#21776)
  (cherry picked from commit 072a37b)
- fix(web): keep a hidden model hidden after a harness switch respells its id (#24564)
  (cherry picked from commit 501f455)
- fix(gateway): say the gateway is running once boot finishes (#25029)
  (cherry picked from commit 02627ec)
- fix(mcp): say why an mcp.json settings file could not be read (#22169) (#24572)
  (cherry picked from commit ed75593)
iamwhatever pushed a commit that referenced this pull request Oct 29, 2026
Cherry-picks these main commits onto release/0.11.0, in merge order:

- fix(dashboard): tell the crewmate about unresolved fields, draw no stale band (#24069)
  (cherry picked from commit f9d8e19)
- fix(mcp-gateway): never pool a credential vendor (#24672)
  (cherry picked from commit 2464225)
- fix(acp): keep MCP tools direct on KAS when the spec grants no tool_search (#24297)
  (cherry picked from commit ddc1fea)
- fix(crew): move a working icon avatar as one still image so it cannot tear (#24678)
  (cherry picked from commit e062153)
- fix(crew-mode): keep a crewmate's own chat on the conductor (#24894)
  (cherry picked from commit 7582829)
- fix(memory): name a damaged memory.db instead of restore advice (#24567)
  (cherry picked from commit efe7624)
- fix(env): search Builder Toolbox's Windows bin dir for agent binaries (#24563)
  (cherry picked from commit 686d39a)
- fix(discord): re-admit bot-opened threads after a gateway restart (#21776)
  (cherry picked from commit 072a37b)
- fix(web): keep a hidden model hidden after a harness switch respells its id (#24564)
  (cherry picked from commit 501f455)
- fix(gateway): say the gateway is running once boot finishes (#25029)
  (cherry picked from commit 02627ec)
- fix(mcp): say why an mcp.json settings file could not be read (#22169) (#24572)
  (cherry picked from commit ed75593)
bolichen97 added a commit that referenced this pull request Oct 29, 2026
Parts 2 and 3 of the 0.11.0 list, combined after part 3 (#25666) was merged into part 2's branch.

Cherry-picks these main commits onto release/0.11.0, in merge order:

- fix(desktop): say why a pop-out cannot open instead of blaming the browser (#23922)
  (cherry picked from commit 99c2474)
- fix(crew): list a crewmate's own script and command schedules (#24023)
  (cherry picked from commit d2d8d14)
- fix(mcp-gateway): show a refused share outside Developer Mode (#23917)
  (cherry picked from commit c612613)
- fix(apps): negative-cache failed app.json fetches in the store listing (#23925)
  (cherry picked from commit 455e8ec)
- fix(power): say that Windows Modern Standby can still sleep on lock (#18259)
  (cherry picked from commit 0480058)
- fix(chat): show nothing in the transcript when the context cap trims (#24492)
  (cherry picked from commit 1411fcc)
- fix(frontend): seed settings rows from the served config during render (#24625)
  (cherry picked from commit e198b63)
- fix(schedule): keep the empty-folder chip rename open and return focus after it (#23904) (#24632)
  (cherry picked from commit 9f51b91)
- feat(website): swipe to dismiss the notification banner on mobile (#24674)
  (cherry picked from commit e4fad39)
- fix(agent): drop a gateway-inherited scratch temp before an agent spawn (#24645)
  (cherry picked from commit 1f28708)

- fix(dashboard): tell the crewmate about unresolved fields, draw no stale band (#24069)
  (cherry picked from commit f9d8e19)
- fix(mcp-gateway): never pool a credential vendor (#24672)
  (cherry picked from commit 2464225)
- fix(acp): keep MCP tools direct on KAS when the spec grants no tool_search (#24297)
  (cherry picked from commit ddc1fea)
- fix(crew): move a working icon avatar as one still image so it cannot tear (#24678)
  (cherry picked from commit e062153)
- fix(crew-mode): keep a crewmate's own chat on the conductor (#24894)
  (cherry picked from commit 7582829)
- fix(memory): name a damaged memory.db instead of restore advice (#24567)
  (cherry picked from commit efe7624)
- fix(env): search Builder Toolbox's Windows bin dir for agent binaries (#24563)
  (cherry picked from commit 686d39a)
- fix(discord): re-admit bot-opened threads after a gateway restart (#21776)
  (cherry picked from commit 072a37b)
- fix(web): keep a hidden model hidden after a harness switch respells its id (#24564)
  (cherry picked from commit 501f455)
- fix(gateway): say the gateway is running once boot finishes (#25029)
  (cherry picked from commit 02627ec)
- fix(mcp): say why an mcp.json settings file could not be read (#22169) (#24572)
  (cherry picked from commit ed75593)
bolichen97 added a commit that referenced this pull request Oct 29, 2026
Backports parts 2-4 of the 0.11.0 insider list (#25665, #25666, #25667) onto
release/0.11.0. Parts 0-1 landed in e379f74 (#25663, #25664).

- fix(desktop): say why a pop-out cannot open instead of blaming the browser (#23922)
  (cherry picked from commit 99c2474)
- fix(crew): list a crewmate's own script and command schedules (#24023)
  (cherry picked from commit d2d8d14)
- fix(mcp-gateway): show a refused share outside Developer Mode (#23917)
  (cherry picked from commit c612613)
- fix(apps): negative-cache failed app.json fetches in the store listing (#23925)
  (cherry picked from commit 455e8ec)
- fix(power): say that Windows Modern Standby can still sleep on lock (#18259)
  (cherry picked from commit 0480058)
- fix(chat): show nothing in the transcript when the context cap trims (#24492)
  (cherry picked from commit 1411fcc)
- fix(frontend): seed settings rows from the served config during render (#24625)
  (cherry picked from commit e198b63)
- fix(schedule): keep the empty-folder chip rename open and return focus after it (#23904) (#24632)
  (cherry picked from commit 9f51b91)
- feat(website): swipe to dismiss the notification banner on mobile (#24674)
  (cherry picked from commit e4fad39)
- fix(agent): drop a gateway-inherited scratch temp before an agent spawn (#24645)
  (cherry picked from commit 1f28708)

- fix(dashboard): tell the crewmate about unresolved fields, draw no stale band (#24069)
  (cherry picked from commit f9d8e19)
- fix(mcp-gateway): never pool a credential vendor (#24672)
  (cherry picked from commit 2464225)
- fix(acp): keep MCP tools direct on KAS when the spec grants no tool_search (#24297)
  (cherry picked from commit ddc1fea)
- fix(crew): move a working icon avatar as one still image so it cannot tear (#24678)
  (cherry picked from commit e062153)
- fix(crew-mode): keep a crewmate's own chat on the conductor (#24894)
  (cherry picked from commit 7582829)
- fix(memory): name a damaged memory.db instead of restore advice (#24567)
  (cherry picked from commit efe7624)
- fix(env): search Builder Toolbox's Windows bin dir for agent binaries (#24563)
  (cherry picked from commit 686d39a)
- fix(discord): re-admit bot-opened threads after a gateway restart (#21776)
  (cherry picked from commit 072a37b)
- fix(web): keep a hidden model hidden after a harness switch respells its id (#24564)
  (cherry picked from commit 501f455)
- fix(gateway): say the gateway is running once boot finishes (#25029)
  (cherry picked from commit 02627ec)
- fix(mcp): say why an mcp.json settings file could not be read (#22169) (#24572)
  (cherry picked from commit ed75593)

- fix(mcp-gateway): render a refused share through ErrorNotice (#25660)
  (cherry picked from commit d895a90)
- fix(crew): write a crewmate's hidden welcome and check-ins in the user's language (#24966)
  (cherry picked from commit d5a76e6)
- feat(crew-mode): move every used crewmate to its conductor on upgrade (#24954)
  (cherry picked from commit 76e89ae)
- fix(hooks): Reads mode auto-approves Kiro Crew's own read-only MCP tools (#24910)
  (cherry picked from commit d8f2038)
- fix(members): list crewmates the person created, even with an empty chat (#24987)
  (cherry picked from commit 163652b)
- feat(cron): let a script cron feed its crewmate's Dashboard page (#24692)
  (cherry picked from commit 959b35a)
- fix(kas): honour a server's autoApprove through the derived permissions (#25106)
  (cherry picked from commit 073c83a)
- fix(chat): keep a scrolled-up reader still when a reprice clamps them (#24591)
  (cherry picked from commit 9a73292)
- fix(security): restore every allowed link on a host, not only identical ones (#20487)
  (cherry picked from commit b161144)
- fix(website): swipe the mobile notification banner up to dismiss it (#25151)
  (cherry picked from commit e68a716)
- test: carry the _card_outcomes helper #24910's tests read
  (cherry picked from commit 9df4d01)

The #24987 decision entry is new on this branch, so its Evidence lines
now link the owner restatement and the maintainer approval on #24987
instead of describing them. Main keeps its shipped (immutable) copy.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants