Skip to content

revert: undo #37551, merged into release/0.13.0 with failing checks - #37871

Closed
bolichen97 wants to merge 106 commits into
release/0.13.0from
macflaky/revert-37551
Closed

bolichen97 wants to merge 106 commits into
release/0.13.0from
macflaky/revert-37551

Conversation

@bolichen97

Copy link
Copy Markdown
Contributor

PR #37551 was merged into release/0.13.0 by mistake: MacFlackyGuard turned on auto merge, and since this branch has no required checks it merged at once, with 7 checks failing (incl. Security scope adjudication on macOS and Ubuntu, Security Scope Review, Sensitive Change Review). Earlier review found the change refuses ordinary ssh commands. This PR reverts it. Not auto-merged; the owner decides.

bolichen97 and others added 30 commits November 13, 2026 09:28
Start the 0.13.0 insider line on release/0.13.0. The three version
files declare 0.13.0-rc.1, as release.md asks of a release branch that
is cutting RCs, and the tag v0.13.0-rc.1 goes on this commit once it
merges.

CHANGELOG.md gains the 0.13.0 section, written once under its final
heading from the commits main took after the 0.12 branch point, less
the fixes 0.12.1 and 0.12.2 already shipped. The shipped 0.12.2 section
is already on main, so nothing else in the file changes.
…36536)

The local-destructive-rm-rf-root row matches its pattern anywhere in the
text, so `rm -rf /tmp/build && mkdir -p /tmp/build` was refused as a
deletion rooted at the filesystem root. Only the literal -rf spelling is
refused: -fr, -r -f and --recursive --force already pass on main, so the
row protected nothing for these paths and only pushed agents to respell.

The row now reads a view of the command in which the words of each plain
`rm` whose every operand is a direct /tmp/<name> entry are blanked. The
rm must be the unquoted program word (after kiro-cli's "Running:" title
prefix at most), every word literal (no $, glob, brace, paren, quote or
escape), no `--`, nothing glued on either edge (a backtick or paren
there lets bash add operands), and each operand exactly one component
under /tmp, so no folder on the way can be a symlink an earlier command
planted. Redirections stay in the text. Every other row, and the rm
floor for / and ~, still read the full text.

Four existing tests used /tmp/<name> as an always-refused example; they
now use /srv/<name> (or /etc), keeping what each test checks.

Replay of 25 recorded rm-row false positives: 17 now allowed; every
recorded true positive is still refused. No catalog regex changed.

Co-authored-by: Joe Guo <zejiangg@amazon.com>
(cherry picked from commit 3b22e33)
…is (#36538)

Co-authored-by: Joe Guo <zejiangg@amazon.com>
(cherry picked from commit e80a131)
…36603)

The KAS engine sends one extra model request at a session's first
prompt to title the session for kiro-cli's own session list. Crew
titles its chats itself and never reads that title, so apply_spawn_env
sets KIRO_DISABLE_SESSION_TITLE_LLM=true on both routes; an inherited
value wins.

Closes #35799

(cherry picked from commit 5275a5d)
…at (#36530)

Every chat starts Kiro Crew's own MCP servers, and each loads its tool
handlers right after its tool list. kirocrew-guide, kirocrew-mate and
kirocrew-cron imported all of kiro_crew.mcp_core for its HTTP verbs and
identity resolvers, which loads the whole core server: about 970
package modules and 90 MB per process, three times per chat.

The verbs (_send, _get, _post, _patch, _put, _delete) move to the
gateway owner, mcp_core_parts/gateway.py, beside the target resolution
they already used. mcp_core stays their import path and patch surface
through its mirror facade. The three servers import the verbs and the
identity resolvers from mcp_core_parts, so their handler modules load
about 310 to 430 package modules and never the core server. The listing
import-budget test now also bounds each small server's handler import.

The default agent's never-defer row also gains its six most used small
own tools (autonudge_stop, memory_recall, nothing_to_do,
search_chat_history, skill_search, wait), picked by how many default-
agent sessions called each at least once. A first use of one no longer
costs a tool-load model call. The row grows by 1,491 schema tokens; its
budget moves from 3,300 to 4,950.

(cherry picked from commit 5ea941c)
…36362)

While the official catalog is unreachable, or the listing is backing
off after a failed fetch, every store listing raised CatalogUnavailable
into a broad except that logged a full stack trace at WARNING. That is
an expected answer and it recurs on every listing, so the log filled
with identical tracebacks. Both listing paths had it: list_registry's
inventory, and the catalog-first listing's _catalog_installable_rows.

Catch CatalogUnavailable first at both sites and log through one
helper, official_catalog.log_unavailable: one WARNING line naming the
cause, the stack at DEBUG. Any other exception keeps its traceback at
WARNING, since that one is a bug, not an outage. The app-kit-platform
spec records the split.

(cherry picked from commit 62f35ec)
#36558)

The sweep that stops the app backends a gateway spawned is a dashboard
cleanup hook, and aiohttp runs cleanup hooks only after in-flight
requests drain (up to 60s). An open /api/logs stream holds that drain
past the 10s graceful budget, so the budget cuts the shutdown before
the sweep runs and os._exit leaves every backend and its sandbox
wrapper running under PID 1, still holding its port.

The exit path now runs the same sweep after the graceful shutdown,
before its hard exit. When the cleanup hook already ran, the tracking
table is empty and the sweep returns at once. Adopted backends stay
excluded, as before.

(cherry picked from commit ee81308)
* perf(crew): one scheduler for crew-workers reads: no overlap, capped, backoff, If-None-Match (Refs #36154)

A page that is hidden no longer re-reads the list, a re-read never cancels
or doubles one in flight, at most two list requests are out per origin so a
send always finds a free connection, slow answers widen the wait, and the
client sends the last ETag and keeps its cached list on a 304.

* fix(crew): crew-workers reads wait for a free slot; ETag read by lowercase name (Refs #36154)

A read over the cap now waits in a FIFO queue and reads fresh data instead
of resolving with the cached list, cold loads obey the cap too, and a
rejected read frees its slot. Header names are case-insensitive, so the ETag
is read in lowercase, which the untranslated-literal gate does not flag.

---------

Co-authored-by: kirocrew-agent <agent@kirocrew.local>
(cherry picked from commit db39373)
… write (#36534)

Co-authored-by: Joe Guo <zejiangg@amazon.com>
(cherry picked from commit ac45981)
…e prompt (#36593)

* perf(acp): record the process tree beside the start and the turn end

The runtime walked its process tree in front of the person three
times: after the initialize handshake, after session/new or
session/load, and at the end of every turn. On macOS that walk starts
a pgrep per process and a ps per pid, 352 ms with no servers and
954 ms with 64 on the first prompt alone, and 68 to 105 ms at the end
of each warm turn.

Only a kill or a teardown reads that record. So the spawn, every
session start and the shared runtime's turn end now start the same
guarded pass as a task, and every reader that signals or takes a kill
handle waits for it first: _kill_inner before it signals the root,
and refresh_process_tree before its own walk. The identity data and
its checks are unchanged; the record is written before anything acts
on it.

Part of #36582

* perf(kas): read the MCP config scopes once per agent projection

The KAS custom-agent projection asked mcp_server_muted_in_any_store
about each declared server, and each call re-read every MCP config
scope through the sensitive-path fence. Two loops did that per start,
so the projection grew with the user's servers: 214 ms with 30 servers
on Linux against 45 ms with none, and 359 ms with 64 on macOS.

Both loops now read mcp_scope_muted_aliases once and hand the set to
the predicate, through the same reader and patch seam. The answer per
server and the payload are the same.

Part of #36582

(cherry picked from commit aeb98eb)
…35590)

An app agent such as Mochi took a granted server's launch spec only
from its template, the Kiro global mcp.json and the managed servers,
so a server added on the MCP Management page, installed from a
registry or contributed by the edition was skipped with "no launch
spec found" while the app reported the policy applied. A built-in
app's grant now reads the default agent's own spec ahead of the
global file, the rebuild's merge of those sources with the store
winning a tie, and leaves out other apps' own servers. A
user-installed app writes its own grant file and no owner consent
covers what it names, so its grant keeps main's sources. A copied
entry that is muted is withheld: turned off, not yet consented, or a
secret:// reference the rebuild withholds.

Enabling an app, or re-installing an installed one from the registry
(how a registry-sourced app updates), wrote its <app>:<server> entries
into the agent spec and stopped, so on a harness that does not watch
the spec file a chat already open never got the change. These routes
now end with the step an MCP page write ends with, as update, disable
and uninstall already do; the enable decides it from the app's record
as it stands inside the lifecycle lock. A fresh install registers
nothing for the disabled record it writes, so it needs no step.

Closes #33020

(cherry picked from commit 712d24f)
…36194) (#36262)

A slot-detail page stops at 1.5 MB of serialized rows, cut from the old end;
the newest row is always sent, so a row over the budget goes alone and is never
dropped. next_before and has_more point at the first row actually sent. The
finished body goes through conditional_json_response: ETag/304 and gzip, no
change to the response shape.

Co-authored-by: kirocrew-agent <agent@kirocrew.local>
(cherry picked from commit b77f3df)
)

opencode 1.x keeps a saved session on the directory it was created in.
A session/load naming another cwd split the session across two of its
instances: approval replies went to the wrong one and were dropped, and
the MCP servers sent with the load (Kiro Crew's own tools) were mounted
where the model was not.

Add ACP_BACKENDS_CWD_BOUND_SESSIONS (opencode). When the stored cwd of
a member's saved session differs from the folder the chat now starts
in, the allocator does not hand the client the resume sid, so the start
opens session/new in the new folder and the conversation is replayed
as text. kiro-cli takes the cwd from each load and keeps native resume.

Fixes #35459

(cherry picked from commit a67f6ed)
On a KAS main harness with no stored Kiro Crew sign-in, agents run on
kiro-cli's login. When the usage API refused that credential, the Kiro
account popup showed the passive "No active Kiro sign-in" notice, as if
nobody had signed in, with no error and no agent hand-off.

The usage reply now marks a refused credential with
credential_rejected, and the popup shows that state as the error
notice with the agent hand-off. A missing credential keeps the passive
notice, and a stored Kiro Crew sign-in keeps its own lapsed notice.

Closes #35483

(cherry picked from commit 94155a1)
…36818)

The shared reader behind the [AGENT SYSTEM PROMPT] block now takes
kiro-cli's systemPrompt spelling, expands $VAR and ~ in a file:// path,
checks the resolved path with the same rule the KAS reader uses, and
names a prompt file it cannot use in one WARNING. It reuses
acp/kas_prompt.py (spec_prompt, the new expand_prompt_ref) and
acp/kas_prompt_paths.py rather than copying them.

Co-authored-by: Joe Guo <zejiangg@amazon.com>
(cherry picked from commit 4c7a874)
Inside a memory-limited container /proc/meminfo reports the VM or node
the container runs on, so /api/system showed that machine's total and
free memory while spawn admission sized against the container's cgroup
limit. On an 8 GiB Fargate task with a 7.75 GiB container limit the
gauge read 15.4 GB total.

The cgroup walk admission uses now also returns the tightest limit
(_container_cgroup_reading). Session import's (limit, headroom) reader,
_cgroup_memory_bounds, drops its own limit loop and reads that walk, and
/api/system calls it, so the Linux memory figures are bounded by the
limit and the headroom under it. The agents slice stays out of the
gauge: it caps agent processes, not the machine.

(cherry picked from commit 848e002)
…nded (Refs #36194) (#36845)

Turn-end and reconnect refreshes (refreshSlot, warmSlotCache, switchSlot and
the pane hydrate all read the newest bounded page through api.chatSlotDetail)
now ask for a cursor (cursor=1) and echo it on the next read, with
If-None-Match set to the last ETag. A refresh that missed nothing gets a 304;
one that missed rows gets only the rows from since on. A cursor the server
rejects (400) or a delta that does not fit the held page is read whole.

A cursor-mode page is now cut by the same byte budget as a plain page, so the
dashboard's reads keep the bound from #36262.

warmSlotCache no longer reads the whole transcript when the count-matched
bound declines (legacy rows with no timestamp): it reads one clamp-sized page
and walks older pages, as refreshSlot and switchSlot do.

Co-authored-by: kirocrew-agent <agent@kirocrew.local>
(cherry picked from commit 714b0d5)
On KAS, a chat whose project has files in .kiro/hooks got the "hook
files in .kiro/hooks do not run" notice again on every new backend
session, so each reconnect, reload or restart added one more copy.

The chat now keeps a mark of the project folder and the sorted file
names, the way the KAS telemetry notice does, and saves it on the
chat's metadata line. The notice shows again only for another folder
or another set of files.

Closes #36821

(cherry picked from commit 911bde8)
With Kiro CLI saved as the harness on a computer where Kiro CLI runs
on Kiro Agent, the Kiro CLI row in Settings > Agent Harness said "New
sessions switch from Kiro Agent to Kiro CLI." and showed a "Use Kiro
CLI" button. The button saved the value that was already saved, so
nothing changed and the panel showed the same two lines again.

The row now shows no button in that state and says that Kiro CLI runs
on Kiro Agent. Every other row keeps its button, and so does the Kiro
CLI row when Kiro CLI does not run on Kiro Agent.

Closes #36170

(cherry picked from commit ecf7d56)
…#36069)

A shared MCP server that waits for an approval showed nowhere outside
the MCP Servers page, so a pause could go unseen for days. The Customize rail
row and its MCP Servers tab carry the count of such servers, read by
the shell from the same status query the servers list's notice uses;
the tab's label is a plural catalog key. The gateway
records a refusal when it starts, so the socket's connect handler marks
that read stale and the shell's own poll runs only every five minutes.

The paused row's hint compared the approved and the new command lines
as lists. One line is drawn per launch, so one program listed twice
before and three times after read as a changed command although only
its env differed. The lines are compared as a set, and the env hint
says the program is the same and only its environment variables
changed. A line that carries a redaction tag never reads as the same
program.

(cherry picked from commit 21b6680)
)

The harnesses AcpClient starts directly (claude, OpenCode, goose, pi,
Copilot, DeepSeek) do not go through the runtime's retrying spawn factory,
so a CreateProcess refusal with winerror 5 or 1260 reached the chat as a
generic error. The shared launch tail now marks such an error at process
creation, naming the binary; ensure_ready keeps its one retry and then
raises the existing AcpSpawnBlocked for a marked error only, so the chat
shows the policy card. An access denied after the process started (a file
cleanup) is not marked and is unchanged. Detection only: nothing changes
how or whether the binary is started.

Co-authored-by: Joe Guo <zejiangg@amazon.com>
(cherry picked from commit 0b32483)
…ng (#36819)

On a host whose ceiling came from central distribution, KAS-managed MCP
children exited before MCP initialize ("Connection closed"): the sandbox
masked policy_cache, so a cache-only child found no ceiling.

The sandbox's existing read-only cache lift (enterprise_policy_cache_view)
now applies when central_ceiling_installed(), not only in an LTS build, and
keeps its refusals. The gateway derives the cache-only markers
(policy_cache_only_env) for the harness env, managed MCP declarations, the
KAS projection and the discovery probe. A cache-only child accepts a
source-less distribution block.

Closes #36775
Supersedes #36710

Co-authored-by: Joe Guo <zejiangg@amazon.com>
Co-authored-by: Pedro Evangelista <pedro.evangelista@mindera.com>
(cherry picked from commit 918a1c8)
The MCP page's sharing card hid its whole memory row whenever either
figure was missing. With sharing off no stub runs, so "with sharing"
has no stub price and is null, and the row vanished exactly when the
reading of what MCP servers use now matters most.

"Using now" now shows whenever it was read, on its own. "With sharing"
and its arrow show only when that figure was read too. No placeholder
stands in for a missing figure.

Part of #36884

(cherry picked from commit ed9d0b0)
…#36586)

While config.json has a change made outside Kiro Crew pending review,
every trust write the built-in adoption APPLY pass makes is refused
with PendingEditBlocksSave. The pass caught it as an unexpected error
and logged a full traceback at ERROR for each pinned app on every
boot, although nothing is wrong: the built-in keeps running and the
staged ledger is retried at the next start.

Treat that refusal as the expected state it is. The pin's outcome is
waiting_for_review, its ledger is left untouched, and the pass logs one
warning naming the pending file and every waiting app, with each trace
at DEBUG. Recording the legacy nav of an already adopted app waits the
same way, at DEBUG only.

Part of #26967

(cherry picked from commit 4579388)
A Bot's chat is a ChatPane, whose replies go through the shared bubble
(renderAssistantBubble). That bubble passed meta.redactions but never
meta.blocked_links, so a removed link stayed as its raw
"[REDACTED: suspicious URL to <host>]" text: no chip, no card, no
"Allow for this host". The bubble now passes the records on the same
local-slot rule as the lock tags, so a window drawing another gateway's
rows keeps the plain text and what can be allowed does not change.

After Allow or Undo, the chat page reloads only the slot it shows. A
pane now re-reads its own slot on the same event, so the restored link
replaces the chip in a Bot's chat too.

Closes #36862

(cherry picked from commit 8c58e18)
…ts (#31546)

A provider marks its one-run work directory (cron_<job>_<run>, subagent_<id>,
...) once, in start(), and its shutdown reclaim removes it. A runtime started
or resumed again for the same session after that reclaim -- a failed resume
that starts fresh, a killed runtime that respawns -- handed KAS a deleted cwd.
Node's spawn reports a missing cwd as "spawn <command> ENOENT", so every
managed MCP server failed naming a binary that exists, e.g.:

  KAS managed MCP initialization failed: kirocrew-cron: failed
  (spawn .../python3.12/bin/python3.12 ENOENT)

Defense in depth:
- Prevent: on the unbound (non-macOS) path, _session_work_dir re-creates a
  missing directory whose name has a one-run shape and whose parent is the
  workspace root, through mark_run_dir, so the reclaim and the sweep see the
  same provenance. A one-run name anywhere else is never created, only
  logged. Any other cwd skips the check and its thread hop, so a chat or
  title start reaches the admission gate in the same order as before.
  Errors are swallowed; the guard never adds a start failure.
- Diagnose: a managed MCP readiness failure with ENOENT names the
  session cwd when it no longer exists.

Docs: acp-client.md run work directory section.

Co-authored-by: Ray Xu <buluoray@gmail.com>
Co-authored-by: RAVI KUMAR <14358622+kingravi17@users.noreply.github.com>
(cherry picked from commit d1f38d7)
The Agent Harness detail for an absent harness showed its install
command bare. For OpenCode that is `npm i -g opencode-ai@1`, and on a
host without Node.js it fails with "npm: command not found" while the
card never says that npm comes with Node.js. Several other harnesses
send an `npm i -g` command too.

AgentInstallDetail now adds one muted line under an install command
that starts with `npm `: the command needs Node.js and npm on the
gateway host, with a link to the Node.js download page. A curl
installer or an empty command gets no line. The component backs both
Settings > Agent Harness and the first-run agent picker, so both get
the line. The new key is translated in every shipped catalog.

Closes #36377

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 5317def)
buluoray and others added 10 commits November 14, 2026 23:24
…ename

The rename changes the text of the two conductor agent specs and of
_CONDUCTOR_SYSTEM_PROMPT. Main re-pinned the same pins in #36924, but
release/0.13.0 builds those specs from its own prompts, so its digests
differ from main's. Only kirocrew-conductor.json and
kirocrew-ledger-conductor.json change; every other pinned spec keeps its
bytes.
…y destination (#22397, #37183) (#37471)

Backport to release/0.13.0 of three main commits, picked with
git cherry-pick -x in main's merge order and squashed into one commit:

- #22397 (e58c25a): record local usage
  metrics by default, with disk caps and a CLI notice. Existing opt-outs
  are kept: heartbeat off with telemetry.enabled unset, telemetry.enabled
  false (quoted and non-bool values included), KIROCREW_TELEMETRY=0,
  DO_NOT_TRACK=1.
- #33422 (ffd2339): RFC update,
  telemetry.enabled now defaults to true. Docs only.
- #37183 (c1222b9): the public build sends
  the reviewed PRODUCT_METRICS list to Kiro telemetry once privacy notice
  v3 has been shown; notice v3 names it; the OTLP exporter becomes a base
  dependency.

Hand-resolved for 0.13: in docs/guides/telemetry-otlp-export.md the intro,
the telemetry.enabled row and the collection/egress paragraph take the
pick's text, and 0.13's own text is kept elsewhere. The pick's one-line
additions to config_all_settings.py and its test are left out: the All
settings page is not on 0.13.

Also fixes the settings help for telemetry.enabled and beacon_enabled and
the TelemetryConfig docstring, which said default recording sends nothing
off the machine, and regenerates config-baseline.json.


(cherry picked from commit e58c25a)
(cherry picked from commit ffd2339)
(cherry picked from commit c1222b9)

Co-authored-by: Raymond Chen <chenmingwei23@users.noreply.github.com>
…on-head)

The Source Gates "Lint documentation trees" step fails on release/0.13.0
itself, so every 0.13 backport PR goes red on it.

- Link 2026-11-13-the-crew-mode-feature-is-named-conductor-mode.md from
  docs/decisions/README.md with main's index row.
- Bring docs/request-for-change/rfc-kas-session-head-in-agent-prompt.md
  from main (#36964) and its README index row. session_head.py, its test
  and three docs on the branch already cite it.

Docs only; no code changes. scripts/docs-lint.sh passes on the result.
…#37539

Two backend tests on release/0.13.0 were red for reasons main has
already fixed in test-only commits:

- test/tracing/test_tracing_otlp.py expected four spans, but #37005
  (already on the branch) adds kirocrew.step. Cherry-picked #37171
  unchanged.
- TestSelfModuleIndexIsLinear failed three runs in a row on the 0.5 s
  budget (0.522 s, 0.511 s, 0.509 s). Took main's end state of
  test/test_denied_commands_security_cost.py and
  test/denied_commands_security_helpers.py, which carries #37451 (min of
  five readings), #37530 (tracer paused) and #37539 (count work, not CPU
  time), plus the #37539 hunk of
  docs/system-specs/common/testing-conventions.md. The picks conflict
  one by one because the file also carries #28293, a test-only change.

No src/ changes.

(cherry picked from commit 6de8528)
…k runner (#37653) (#37818)

Playwright 1.58.2 has no ubuntu26.04 host entry, and kirocrew-linux-4-core
now boots ubuntu-26.04, so every fork PR's E2E, render benchmark and i18n
render gate fail at 'npx playwright install chromium'. On that image only,
set PLAYWRIGHT_HOST_PLATFORM_OVERRIDE to the ubuntu24.04 build and carry it
through GITHUB_ENV to the steps that launch the browser.

Part of #37647.

(cherry picked from commit b5ac843)

Co-authored-by: Bolin Chen <bolichen@amazon.com>
…itory HEAD (#37686)

* fix(mcp-gateway): fingerprint the backend package tree, not the repository HEAD

code_fingerprint() named the code by the repository's HEAD commit, so a
commit that touched only website/ (no Python) changed the fingerprint and
every Crew MCP stub refused the running gatewayd with "gateway code
fingerprint does not match this stub" until the gateway restarted. On one
host this turned off sharing for core/cron/guide/mate for 13 hours (188
fallbacks), and each fallback made a new session slower than a plain launch.

The git rule now uses the package directory's tree id at HEAD
(git rev-parse HEAD:./ from the package root). Any committed change under
src/kiro_crew still changes it, and the tracked-diff digest still covers
uncommitted edits.

* fix(mcp-gateway): refuse repo-named drivers before the tree lookup, and never lazy-fetch

Resolving HEAD:./ walks tree objects. In a partial clone a missing one is
fetched through the repository's own remote, so a planted
remote.<name>.uploadpack would run before repo_exec_driver_reason() was
consulted. The preflight now runs first, and every fingerprint git call
sets GIT_NO_LAZY_FETCH=1.

(cherry picked from commit 9044daf)
…ervers

A worker session started its own kirocrew-core/cron/guide/mate/work, because
the overlay pass never rewrote the stock worker spec and worker_overlay_spec
skipped every base server. The pass now rewrites kirocrew-mate-worker.json
next to the conductor's, and worker_overlay_spec returns its base-server
stubs (shaped by the governed merged spec: muted, dropped and registry
servers left out, disabledTools from the merged spec, autoApprove dropped)
plus the stubs the Bot's own file adds.

The worker overlay is never served by agent name: without the KAS session
key the lookup answers nothing, so a host that cannot apply a worker's
merged limits keeps declaring the servers itself.

(cherry picked from commit f1b575f)
…37443) (#37551)

The ssh-to-self floor masks quoted separators and parens before it
walks a command line, so it never saw a `$(` inside double quotes:
`echo "$(ssh -p 22 <this host> id)"` and `x="$(...)"` slipped past
the floor, although bash runs that substitution.

The mask stays as it was, so the parent command's walk still reads
the quoted word as one argument. The floor now also takes each
double-quoted `$( )` body from the unmasked text, opened only by an
unpaired `$` and ended at its own matching `)`, and judges it as a
command line of its own. A quoted-tag heredoc that a bare `cat` only
prints is cut from the body first, so a PR body passed through
`"$(cat <<'EOF' ...)"` stays allowed.

Closes #36904

(cherry picked from commit 87b11b5)
@bolichen97
bolichen97 requested review from a team and KiroCrewCaptain as code owners November 16, 2026 07:23
@bolichen97
bolichen97 requested review from iamwhatever and removed request for a team November 16, 2026 07:23
@github-actions

Copy link
Copy Markdown
Contributor

Sensitive Change Review

Sensitive files changed:

  • src/kiro_crew/security/argv_floor.py

  • src/kiro_crew/security/argv_floor_ssh.py

  • No reasoning template for 39d7212 yet.

How to clear this check:

  1. The PR author posts the template below as a PR comment, with
    every field filled (N/A needs a reason). Keep the
    Head commit line: it must name the current head SHA.
  2. The check turns green within a minute, with no push.
  3. A new push resets this check: post it again for the new head.

Only the author's comment counts this way. A reviewer can instead
paste the filled template into an "Approve" review on this head.
Merging still needs a reviewer's approval.

## Sensitive change review
- Head commit: 39d721241af8fa38a2abb7de5cbe65bfb6ceb395
- What rule changed:
- Before -> after (what was allowed/blocked, now):
- Worst case if wrong (breaks commands / leaks secret / locks user out):
- Evidence checked (which sandbox result proves it):
- No regression:
  - Tools still work (which tools/commands were run, result):
  - Backward compatible (old config / old data / old callers still work, how checked):
  - Existing tests (which suites passed on this head):
- How to undo:

@github-actions github-actions Bot added the readiness: checking Automated validation is still running label Nov 16, 2026
@github-actions

Copy link
Copy Markdown
Contributor

UX Review (Opus 5.5) — ⏭️ skipped

Revision 39d721241af8fa38a2abb7de5cbe65bfb6ceb395 touches no user-facing surface (no changes under website/ or committed screenshots), so the UX review was skipped. Does not block merge.

@github-actions

Copy link
Copy Markdown
Contributor

First Principles Review (Opus 5.5) — ✅ PASS

Premise-level review of 39d721241af8fa38a2abb7de5cbe65bfb6ceb395 — 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

This is a clean, complete revert of #37551. Nothing is added, so nothing has to justify itself. The description names the reverted PR, so removing its pins is disclosed.

Notes

  • Item 1 — inherited: the only reason given for the revert is "Earlier review found the change refuses ordinary ssh commands", and no command or log is cited. The description also never says that the fix(security): read a double-quoted $( ) as a command in ssh floor #37443 bypass (echo "$(ssh … localhost id)" slipping past the ssh-to-self floor) is open again once this merges.

What this change ships

Inventory (5 items) — 4 justified

Intent: undo a security-floor change that was merged into release/0.13.0 by accident while checks were failing. This is a FIX (a revert).

  1. The ssh-to-self floor stops checking $( ) bodies inside double quotes, so the fix(security): 0.13 backport, read a double-quoted $( ) as a command in ssh floor (#37443) #37551 refusals return to how they were before it — inherited (refusal claim not shown)
  2. security.md loses the paragraph that described that check — justified
  3. The line-budget entry ssh-self-dquote-subst.txt is removed — justified
  4. The tests fix(security): 0.13 backport, read a double-quoted $( ) as a command in ssh floor (#37443) #37551 added (the deny rows and the stay-allowed rows) are removed — justified
  5. The argv_floor type-checking re-exports for the deleted helpers and read_commands are removed — justified

[FIRST-PRINCIPLES-REVIEWED] 39d7212

Description read

[DESCRIPTION-READ] 7ed2254450f3e95f3d1ef728fa8dfc02bbeab2b7356747123062a56b0a2f81ee

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

Copy link
Copy Markdown
Contributor

Design Review (Opus 5.5) — ✅ PASS

Design-level review of 39d721241af8fa38a2abb7de5cbe65bfb6ceb395 — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS does not.

Design-Verdict: PASS

This is an exact revert of #37551, which was merged by mistake. It puts back the ssh floor as it was before that merge, with its spec text and tests. It can be redone later through the normal PR flow.

Notes

[DESIGN-REVIEWED] 39d7212

Description read

[DESCRIPTION-READ] 32cd27498d8c20014fde70f7e7a31cd957c585475a0a0ca8d832d0f62e061f8d

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

Copy link
Copy Markdown
Contributor

Opus 5.5 Review — 🔴 changes requested (blocking)

Reviewed 39d721241af8fa38a2abb7de5cbe65bfb6ceb395 — this comment is updated in place on each push.

This revert removes the ssh-to-self floor's check on a $(...) inside double quotes and adds nothing to replace it, so an ssh to this machine wrapped in double quotes gets past is_denied and leaves the sandbox.

BLOCKING — src/kiro_crew/security/argv_floor_ssh.py:1045
- if any(map(_is_ssh_to_self, _double_quoted_substitution_bodies(unmasked))): / - return True
Trigger: the agent runs a Bash call echo "$(ssh -p 22 localhost id)" or x="$(ssh -p 22 localhost id)". The -p 22 keeps it off the catalog regex. The call goes is_denied → argv floor → _is_ssh_to_self, where _mask_quoted_separators turns the double-quoted $( into data, and after this revert nothing reads that body as a command line. At this head, is_denied returns None for both of those commands and refuses the unquoted echo $(ssh -p 22 localhost id) with sandbox-escape-ssh-self.
Outcome: bash still runs the substitution, so the agent opens ssh to its own host and gets a shell outside the OS sandbox. The floor exists because the sandbox does not fence this path. The tests that pinned the refusal (test_double_quoted_command_substitution_is_judged) are deleted with no replacement.
Anchor: residual/guard-removal
Fix: keep the _double_quoted_substitution_bodies pass, or another check that reads a double-quoted $(...) body as a command line. If the revert is meant to undo a false denial, narrow that one case and keep the refusal.

[BLOCK-MERGE] 39d7212
[OPUS-REVIEWED] 39d7212

Verdict parsed from the review's SHA-scoped output markers for commit 39d721241af8fa38a2abb7de5cbe65bfb6ceb395.

False positive or not applicable? A repository writer can comment:
/ai-review override fable 39d721241af8fa38a2abb7de5cbe65bfb6ceb395 false-positive: <reason, 20 to 500 characters>

@github-actions

Copy link
Copy Markdown
Contributor

Security Scope Review -- ✅ PASS

Which legitimate operations does 39d721241af8fa38a2abb7de5cbe65bfb6ceb395 newly refuse? Updated in place on each push. A script-confirmed regression blocks; the model proposes candidates and never decides a refusal.

Verdict basis: adjudicated, zero confirmed regressions.

Security scope review -- adjudicated candidates

  • macos-latest (platform posix): 10 classified, 0 newly refused, 6 still allowed.
  • ubuntu-latest (platform posix): 10 classified, 0 newly refused, 6 still allowed.
  • windows-latest (platform windows): 3 classified, 0 newly refused, 1 still allowed.

No candidate flipped from allowed to refused.

Scope-Verdict: PASS

Reason: This diff makes nothing stricter. It is a revert: it deletes _double_quoted_substitution_bodies and _without_quoted_heredoc_text and the check in _is_ssh_to_self that called them, so the ssh-to-self floor refuses less. No operation can be newly refused. Whether the removal is safe is the other lanes' question.

TIGHTENING: none; the ssh-to-self argv floor got looser -- the hunk in src/kiro_crew/security/argv_floor_ssh.py deletes if any(map(_is_ssh_to_self, _double_quoted_substitution_bodies(unmasked))): return True. That check refused an ssh to this host run from inside a double-quoted $(...) (for example echo "$(ssh -p 22 <this host> id)"). After the revert, those spellings are no longer refused by this floor.

ADJUDICATED: 13 candidates, 0 confirmed newly-refused

PLATFORM COVERAGE: posix: 10 rows. They cover the boundary the removed code touched: heredoc-bodied gh pr create and git commit -m "$(cat <<'EOF' ...)", a double-quoted $(ssh ...) to a remote host, quoted substitutions inside ssh -i and scp operands, $$ next to $(, and plain git/date substitutions. windows: 3 rows. They cover the same operations in PowerShell spelling: a $(Get-Content ...) subexpression body, ssh.exe/scp.exe with $env:USERPROFILE, and a drive-letter path with \ separators. The diff adds no path, separator, interpreter or environment assumption, so no platform is left unprobed. No any rows were written, because every candidate names an executable.

[SCOPE-REVIEWED] 39d7212

@github-actions

Copy link
Copy Markdown
Contributor

GPT 6.1 Review — 🔴 changes requested (blocking)

GPT 6.1 found at least one blocking issue that must be resolved before merging 39d721241af8fa38a2abb7de5cbe65bfb6ceb395. 1 of 1 blocking finding(s) are security-class and were withheld from adjudication, so the blocking verdict stands.

This comment is updated in place on each push.

BLOCKING -- src/kiro_crew/security/argv_floor_ssh.py:1044 -- Removing substitution inspection reopens localhost sandbox escape

Removed: if any(map(_is_ssh_to_self, _double_quoted_substitution_bodies(unmasked))):
Trigger: With standard sandboxing and working key-authenticated localhost SSH, the agent submits echo "$(ssh -p 22 localhost id)"; the tool gate calls is_denied → _reaches_self_over_ssh → _is_ssh_to_self.
Outcome: The public deny gate now allows this command, whereas the base refuses it. Bash executes SSH through the host’s unsandboxed sshd, outside the read-only governance mounts. enter_namespaces isolates only user/mount namespaces, the seccomp filter does not block connections, and standard mode leaves SSH keys readable.
Anchor: residual/security
Fix: Retain inspection of executable double-quoted substitution bodies while correcting the false positives, rather than deleting that guard.

[BLOCK-MERGE] 39d7212
[GPT-REVIEWED] 39d7212

Adjudication (Opus 5.5) — is blocking on each finding proportionate?

F1 is fenced, so I can only annotate it. I'm upholding it: the harm stands as written.

  • What the revert removes: It deletes the check that reads a double-quoted $(...) body as a command line of its own. That check used to sit in _is_ssh_to_self at src/kiro_crew/security/argv_floor_ssh.py:1044, and its helpers _double_quoted_substitution_bodies and _without_quoted_heredoc_text sat at about lines 181–236 on the base. The revert also deletes the test that denied echo "$(ssh -p 22 localhost id)" and its variants (test/test_denied_commands_security_ssh_self.py:1211-1270 on the base).
  • How common the trigger is: -p 22 keeps the command off the catalog row. The input is trivial for an agent to produce, and a host running sshd with key auth is ordinary on a developer machine. Nothing in the conditions is extreme or contradictory.
  • Why no FLAG: I could not confirm from code that a sandbox fence covers the escape, and I found no recovery path. Without that record, a FLAG isn't justified.

There were no adjudicable findings, so the first footer is empty.

[ADJUDICATION] 39d721241af8fa38a2abb7de5cbe65bfb6ceb395 total=0 uphold=0 downgrade=0
[GPT-ADJUDICATED] 39d721241af8fa38a2abb7de5cbe65bfb6ceb395

[ADJUDICATION-FENCED] 39d721241af8fa38a2abb7de5cbe65bfb6ceb395 fenced=1 flagged=0
UPHOLD-FENCED F1 src/kiro_crew/security/argv_floor_ssh.py:1044 -- The revert deletes the double-quoted substitution check and its pinning tests, so a trivially producible `echo "$(ssh -p 22 localhost id)"` passes the deny floor again on any host running key-authenticated sshd, a plausible condition with no fence confirmed in code.
[GPT-ADJUDICATED-FENCED] 39d721241af8fa38a2abb7de5cbe65bfb6ceb395

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 39d721241af8fa38a2abb7de5cbe65bfb6ceb395 false-positive: <reason, 20 to 500 characters>

@github-actions github-actions Bot added readiness: action required A blocking check or review needs attention merge conflict Branch has merge conflicts with its base — author must resolve before merge and removed readiness: checking Automated validation is still running labels Nov 16, 2026
@bolichen97 bolichen97 closed this Nov 16, 2026
@github-actions github-actions Bot removed the readiness: action required A blocking check or review needs attention label Nov 16, 2026
@buluoray

Copy link
Copy Markdown
Contributor

@bolichen97 thanks for closing this. One note: release/0.13.0 is now 0.13.0-rc.2, rebuilt from main (head 2fae50aae751029a1c9b2e5879ee762862df1dca), and it still has main's own #37443 commit (87b11b5ef245). This PR targeted the old #37551 merge from before rc.2, so it would not have taken #37443 out. If the ssh refusals you saw should stay out of 0.13, that needs a new revert of #37443 on rc.2. Do you want one?

@bolichen97
bolichen97 deleted the macflaky/revert-37551 branch November 20, 2026 09:18
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

merge conflict Branch has merge conflicts with its base — author must resolve before merge

Projects

None yet

Development

Successfully merging this pull request may close these issues.