- Scope
- The file
- Values
- Resolution
- Where it is read
- What protects the file
- Cloud launch rule
- Stage-start probe and fallback
- Untrusted-input stages
- Required plugins per stage
- Readers
- Residual risks
- Vendor facts this contract depends on
- Versioning
Owner doc for where a consumer repository's local-lane stages run: on this machine (the default),
in a local background session, or on a Claude cloud host. One setting, execution_target, per stage,
held in one YAML file the repository owner changes by pull request.
execution_target places the stages a local lane runs: pre-ready build work (/work-items:work),
the work loop (/work-items:work-loop), local triage (/work-items:triage) and the babysit loop
(/source-control:babysit-loop, which runs /source-control:babysit-prs). In a repository that runs
the PR pipeline's lanes, post-ready work is placed by the pipeline config
(docs/conventions/pr-pipeline/), not by this file. Merge authority is
not part of this setting and does not change with the host.
docs/conventions/execution-target.yaml in the consumer repository, validated by
execution-target.schema.json beside this README
(ADR 0064 Decisions 1, 5
and 6). It is a policy floor: it has a team layer only, and no plugin declares a userConfig option
for it, so no per-user or personal override can move a stage.
default: local-worktree
class:
untrusted-provenance: cloud-session
skill:
work-items:
work-loop: local-background
source-control:
babysit-loop: local-worktree| Key | Read as | Meaning |
|---|---|---|
default |
default |
Host for any stage no narrower key names. |
class: map, keyed by work-class label |
class.<label>, e.g. class.untrusted-provenance |
Host for items of that class. Labels: read-only, mechanical, scoped, structural, untrusted-provenance. |
skill: map, keyed by plugin, then skill |
skill.<plugin>.<skill>, e.g. skill.work-items.work-loop |
Host for one stage skill. |
Readers use the shared reader lib/parse-concern-value.sh with those dotted keys; each key segment
is [A-Za-z0-9_-], which is why the skill: map nests plugin then skill rather than holding a
<plugin>:<skill> key.
| Value | What runs |
|---|---|
local-worktree (default) |
Today's local launch: for a lane, claude --bg -n <name> --permission-mode auto from the repository root. |
local-background |
claude --bg -n <name> --permission-mode auto from the lane's own linked worktree. Background sessions are a research preview. |
cloud-session |
claude --cloud "<probe preamble + prompt>" under the cloud launch rule. For per-PR fix work in the post-PR stage, the vendor's cloud form is /autofix-pr, which the operator runs on the PR's branch; the launcher does not run it. |
cloud-routine |
A routine the operator creates at claude.ai with the stage's prompt and its own trigger. The launcher prints the setup steps and launches nothing; a stage that may read untrusted input gets no steps and is skipped (see Untrusted-input stages). It holds no routine token: firing a routine through its API trigger needs a token kept in the consumer's CI secret store, outside this setting. |
cloud-project |
A thread in a claude.ai project on the repository. The launcher prints the setup steps and launches nothing; a stage that may read untrusted input gets no steps and is skipped. |
A value outside this table resolves to local-worktree, and the reader reports the key and value.
For a stage, the first key that holds a value wins:
skill.<plugin>.<skill>for the stage skill;class.<label>for the item's work class (a per-item reader), orclass.untrusted-provenancefor a stage whose input is always untrusted (a lane reader; see Untrusted-input stages);default;local-worktree.
An absent file, a file the shared reader rejects, or a failed fetch resolves every stage to
local-worktree, and the reader says why.
The default branch is the one origin's own HEAD names (git ls-remote --symref origin HEAD). A
reader reads no policy when origin's HEAD cannot be read or disagrees with the local origin/HEAD
symref, when a url.<base>.insteadOf or pushInsteadOf rule in any git config scope rewrites
origin's URL, or when remote.origin.url has more than one value. In each case every stage resolves
to local-worktree and a work-items:triage lane is skipped.
These checks catch a stale or moved symref and the plain forms of a redirected origin. They are not a boundary against someone who can write the checkout's git config, who can point origin at any repository; the policy is only as trustworthy as the checkout the launcher runs in.
A reader then runs git fetch origin <default branch>, reads the file with
git show <commit>:docs/conventions/execution-target.yaml from the fetched commit, and prints that
commit's SHA with the value and the key that supplied it. It never reads the working tree, so an
unmerged or uncommitted edit cannot move a stage.
- A change reaches the default branch through a pull request, and AGENTS.md-style rules that ask a person before merging a change to what agents may do unattended apply to an interactive merge.
- A repository that runs the PR pipeline adds this file to
merge.diff-check.denied-pathsin its pipeline config. The pipeline does not force this path today; forcing it is requested in #6202. - Every reader prints the SHA it read, so a run shows which policy placed it.
No CODEOWNERS rule or ruleset is required: in a single-maintainer repository an author cannot approve their own pull request, so a code-owner rule would block every change to the file.
claude --cloud runs only when all of these hold; otherwise the stage runs local-worktree, except a
stage that may read untrusted input, which is skipped (see Untrusted-input stages):
- the stage's input is never untrusted (its row below reads "never"); no current local-lane stage
qualifies, so the lanes launcher sends no lane to
cloud-sessiontoday; - the launch runs from a clean linked worktree (no uncommitted or untracked changes) checked out at the pushed commit the stage needs (the fetched default-branch commit for a lane);
- origin's configured URL is a github.com
owner/repo, andgh api user/installationslists an installation of the Claude GitHub App (claude) that covers that repository (all repositories of its owner, or the repository among the selected ones). The slug comes from origin, the remoteclaude --cloudclones from, never fromgh's default repository; - the lane's telemetry can be read: an explicit numeric
telemetry.issuein that same repository (atelemetry.reponaming any other repository is refused, and no issue is found by title), so a probe fallback can reach the launcher; - the lane does not request the autonomy lane-stop gate, which cannot be armed in a cloud session.
The App check exists because, without the App, claude --cloud uploads a bundle of the local
repository (all its branches, and edits to tracked files not yet committed) instead of cloning from
GitHub.
A stage on a cloud host runs the probe before anything else. The lanes launcher puts it in front of the lane prompt; a per-item dispatcher puts it in front of the item brief. The probe:
- confirms each plugin in the stage's row of Required plugins per stage is loaded in the session, not only declared;
- confirms one read from the work-item tracker succeeds;
- for a stage that merges, confirms the merge wrapper's read-only check passes;
- records the session's permission mode, whether connector (MCP) tools are present, and whether a
request to a host outside the Trusted network allowlist fails, under
execution_target_probein the lane's telemetry state block.
On a gap in steps 1-3 the stage writes execution_target_fallback: {"reason": ..., "at": ...} to the
lane's telemetry state block, claims nothing and exits. A passing probe sets
execution_target_fallback to null.
- Lane fallback. The launcher reads the newest
execution_target_fallbackon the lane's telemetry issue, counting only comments written by thegh-authenticated login or the lane's configuredtelemetry.author; a comment from anyone else,nullincluded, changes nothing. A value it has not yet honored moves exactly one launch tolocal-worktree; the launcher records the value's hash in its plugin data directory, and the next launch probes the cloud again.
Because claude agents --json does not list cloud sessions, the launcher records each cloud launch in
its plugin data directory: start does not send a second cloud session for a lane while that record
stands, and restart sends a new one (the earlier session is archived by hand at claude.ai).
A stage reads untrusted input when it triages raw intake, works a pull request from a fork in the
post-PR stage, or works an untrusted-provenance item. The guard keys on that input's provenance,
not on the item's work class alone (the autonomy isolation ladder's "Axes" section: input
provenance decides whether the L3 bar applies,
plugins/autonomy/reference/guardrails/isolation-ladder.md):
- Enforced by the launcher. The no-connector and denied-egress check has no deterministic form
yet, so the lanes launcher refuses every cloud host (
cloud-session,cloud-routineandcloud-project) for every stage whose "Reads untrusted input" column is anything but "never", for a lane with nostageand for a stage outside the table (lanes.jsonis lane-writable, so a missing or unknown stage counts as untrusted). Such a lane is skipped, not run locally: the launcher prints an error naming/work-items:attend-queueas the escalation route and files nothing. - Repeated in the probe. The probe preamble still tells a cloud session to read untrusted input
only when it finds no connector tools and the step 4 egress request fails, and otherwise to stop
and leave the item for
/work-items:attend-queue, as a second line of defense. /work-items:triagereads raw intake on every run, so the lanes launcher resolvesclass.untrusted-provenancefor it after itsskillkey, and skips a triage lane whenever the policy read is refused (see Where it is read).
Derived from the skills each stage skill invokes on its normal path. A cloud-routine stage may carry
only the connectors its row lists; the probe fails on any other.
| Stage | Required plugins | Reads untrusted input | cloud-routine connectors |
|---|---|---|---|
work-items:work-loop |
work-items, implementation, source-control | when it triages raw intake | none |
work-items:work |
work-items, implementation, source-control | when the item is untrusted-provenance |
none |
work-items:triage |
work-items | always | none |
source-control:babysit-loop |
source-control, work-items | when a PR comes from a fork | none |
source-control:babysit-prs |
source-control | when a PR comes from a fork | none |
The lanes launcher carries a copy of the plugin column (stage_required_plugins in
plugins/harness-ops/skills/lanes/scripts/lane-launcher.sh); change both together.
| Reader | Keys it reads |
|---|---|
/harness-ops:lanes launcher (start, restart, run-once) |
skill.<plugin>.<skill> for the lane's stage, class.untrusted-provenance for work-items:triage, default |
- A direct push to the default branch, or a merge outside the protected paths above, changes the file without a person's review. The printed SHA shows it afterwards; nothing blocks it.
- Anyone who can write the checkout's git config can redirect origin in ways the URL checks above do not see, and so choose the policy a launch reads.
- Until the PR pipeline forces this path (#6202), an unattended
pr-mergein a repository whose pipeline config does not list it can carry a change to it. - The untrusted-input guard is enforced by refusing those stages every cloud host: no
claude --cloudlaunch and no routine or project setup steps. When a deterministic connector and egress check lets the launcher admit them, the guard for them rests on that check and on the probe instruction. - Trusted-input cloud stages run with the account's connectors and the environment's network level; an environment at network level All gives full egress. The probe records this posture and does not block on it.
- A cloud-to-local fallback runs on this machine, below the
L2floor the autonomy isolation ladder sets for unattended runs ("Unattended floor"), the same as every local lane today. The fallback is reported in telemetry and in the launcher's output, never silent. - No security binding reads a per-host isolation floor (ADR 0038); this setting only places stages.
- How
claude --cloudpicks a branch from a linked worktree on a detached HEAD at the pushed commit is not documented; the first live cloud run records it.
- Routine triggers. Pointer: the "Supported events" table on Routines (pull request and release events), and its "Connectors" section (every connected connector is included by default and its tools run without approval). The open conflict is anthropics/claude-code#72290 (an issue-opened routine trigger that did not fire, closed as stale); nothing here relies on an issue trigger.
- Cloud launch and upload. Pointer: Start a cloud session from your terminal and Send local repositories without GitHub.
- Cloud environments. Pointer: What carries over from your setup and Network access.
- Background sessions and their listing. Pointer: the "Limitations" section of
Agent view (research-preview status)
and List sessions as JSON
(which sessions
claude agents --jsonprints).
CHANGELOG.md records each change. Removing or renaming a key or value is a major
bump; a new value, key or reader row is a minor bump; wording alone is a patch.