diff --git a/.changes/fleet/feat-16541-reach-relay-profile-7c41d2e8.md b/.changes/fleet/feat-16541-reach-relay-profile-7c41d2e8.md
new file mode 100644
index 00000000000..ea7a525823f
--- /dev/null
+++ b/.changes/fleet/feat-16541-reach-relay-profile-7c41d2e8.md
@@ -0,0 +1,8 @@
+---
+bump: minor
+---
+
+### Added
+
+- The reach skill starts a relay turn or session in a WSL lane in the plugin profile the task names: it finds the target's `profiles.mjs` through `claude plugin list --json`, runs `launch --unattended` there with the relay's own settings in `--with`, and starts `claude` with the printed command. With no profile named, no plugin-profiles on the target, or a Windows lane, it starts as before and says the profile was not applied.
+- The reach skill's prompt line says to name the session that holds a PR, issue or branch in the prompt, and never to tell the target to take the work over on its own judgment.
diff --git a/plugins/fleet/evals/reach-start-in-profile/case.yaml b/plugins/fleet/evals/reach-start-in-profile/case.yaml
new file mode 100644
index 00000000000..892c6600b68
--- /dev/null
+++ b/plugins/fleet/evals/reach-start-in-profile/case.yaml
@@ -0,0 +1,28 @@
+schema_version: "1.1"
+name: 'reach-start-in-profile'
+context:
+ scaffold_script: scaffold.sh
+graders:
+ - name: 'finds-profiles-script'
+ type: 'regex'
+ pattern: 'ssh\.exe\s+laptop-wsl\s+''?[^''\n]*\bclaude\s+plugin\s+list\s+--json'
+ flags: 'i'
+ target: 'last_message'
+ - name: 'launch-unattended'
+ type: 'regex'
+ pattern: 'profiles\.mjs\S*\s+launch\s+["'']?build\b[^\n]*--unattended'
+ flags: 'i'
+ target: 'last_message'
+ - name: 'settings-through-with'
+ type: 'regex'
+ pattern: 'profiles\.mjs\S*\s+launch\b[^\n]*--with\s[^\n]*crossSessionInbound'
+ flags: 'i'
+ target: 'last_message'
+ - name: 'starts-with-printed-command'
+ type: 'llm'
+ criteria: 'Starts the laptop session with the command `profiles.mjs launch` prints (its `command` or `argv`) in place of a plain `claude`, and says that when launch exits 1 for a missing required tool nothing is started and the reason is reported'
+ focus: 'trace'
+ - name: 'skill-fired'
+ type: 'tool_used'
+ tool: 'Skill'
+ input_match: '"skill"\s*:\s*"(?:fleet:)?reach"'
diff --git a/plugins/fleet/evals/reach-start-in-profile/prompt.md b/plugins/fleet/evals/reach-start-in-profile/prompt.md
new file mode 100644
index 00000000000..4ef72f74891
--- /dev/null
+++ b/plugins/fleet/evals/reach-start-in-profile/prompt.md
@@ -0,0 +1,9 @@
+---
+tags: ['reach']
+allowed_tools: [Read, Glob, Grep, Skill]
+expected_outcome: 'Starts a named receiver on the laptop''s WSL lane in its build plugin profile: finds profiles.mjs on the laptop through claude plugin list --json (the installPath of the enabled plugin-profiles@ entry), runs profiles.mjs launch build --folder
--unattended --json there with the receiver''s crossSessionInbound settings passed through --with, and starts the receiver with the printed command in place of claude. Says that a launch refused for a missing required tool starts nothing and is reported.'
+---
+
+I'm in the desktop's WSL session (desk-01). Start a Claude session on the laptop's WSL side, in its `build` plugin profile and in `~/repos/app`, that other sessions can keep sending messages to. Give me the exact commands to run; don't run them yourself, and run no shell command to work them out.
+
+In this session the fleet manifest is at `.config/fleet/FLEET.md` in the working directory, not under `~`.
diff --git a/plugins/fleet/evals/reach-start-in-profile/samples/finds-profiles-script.json b/plugins/fleet/evals/reach-start-in-profile/samples/finds-profiles-script.json
new file mode 100644
index 00000000000..ae7dd7f2a1e
--- /dev/null
+++ b/plugins/fleet/evals/reach-start-in-profile/samples/finds-profiles-script.json
@@ -0,0 +1,26 @@
+{
+ "pass": [
+ {
+ "answer": "**1. Find plugin-profiles on the laptop:**\n\n```sh\n/mnt/c/Windows/System32/OpenSSH/ssh.exe laptop-wsl 'claude plugin list --json'\n```\n\nTake the `installPath` of the enabled `plugin-profiles@` entry.",
+ "why": "quoted remote command"
+ },
+ {
+ "answer": "/mnt/c/Windows/System32/OpenSSH/ssh.exe laptop-wsl claude plugin list --json",
+ "why": "unquoted remote command"
+ }
+ ],
+ "fail": [
+ {
+ "answer": "",
+ "why": "empty reply"
+ },
+ {
+ "answer": "claude plugin list --json",
+ "why": "runs locally, not on the laptop"
+ },
+ {
+ "answer": "/mnt/c/Windows/System32/OpenSSH/ssh.exe laptop-wsl 'node ~/.config/plugin-profiles/profiles.mjs launch build'",
+ "why": "guesses the script path"
+ }
+ ]
+}
diff --git a/plugins/fleet/evals/reach-start-in-profile/samples/launch-unattended.json b/plugins/fleet/evals/reach-start-in-profile/samples/launch-unattended.json
new file mode 100644
index 00000000000..c6ebb93f80f
--- /dev/null
+++ b/plugins/fleet/evals/reach-start-in-profile/samples/launch-unattended.json
@@ -0,0 +1,22 @@
+{
+ "pass": [
+ {
+ "answer": "/mnt/c/Windows/System32/OpenSSH/ssh.exe laptop-wsl 'cd ~/repos/app && node /scripts/profiles.mjs launch build --folder ~/repos/app --unattended --json --with \"{\\\"crossSessionInbound\\\":\\\"accept\\\"}\"'",
+ "why": "launch with --unattended"
+ },
+ {
+ "answer": "node ~/.claude/plugins/cache/m/plugin-profiles/1.0.0/scripts/profiles.mjs launch build --unattended --json",
+ "why": "resolved path"
+ }
+ ],
+ "fail": [
+ {
+ "answer": "node /scripts/profiles.mjs launch build --folder ~/repos/app --json",
+ "why": "no --unattended"
+ },
+ {
+ "answer": "node /scripts/profiles.mjs apply build --unattended",
+ "why": "apply, not launch"
+ }
+ ]
+}
diff --git a/plugins/fleet/evals/reach-start-in-profile/samples/settings-through-with.json b/plugins/fleet/evals/reach-start-in-profile/samples/settings-through-with.json
new file mode 100644
index 00000000000..e7825a8737b
--- /dev/null
+++ b/plugins/fleet/evals/reach-start-in-profile/samples/settings-through-with.json
@@ -0,0 +1,18 @@
+{
+ "pass": [
+ {
+ "answer": "/mnt/c/Windows/System32/OpenSSH/ssh.exe laptop-wsl 'cd ~/repos/app && node /scripts/profiles.mjs launch build --folder ~/repos/app --unattended --json --with \"{\\\"crossSessionInbound\\\":\\\"accept\\\"}\"'",
+ "why": "settings merged through --with"
+ }
+ ],
+ "fail": [
+ {
+ "answer": "node /scripts/profiles.mjs launch build --unattended --json\nthen: -p -n app-inbox --settings '{\"crossSessionInbound\":\"accept\"}'",
+ "why": "a second --settings on the start command, which replaces the profile's"
+ },
+ {
+ "answer": "node /scripts/profiles.mjs launch build --unattended --json --with /tmp/relay.json",
+ "why": "--with without the receiver's setting"
+ }
+ ]
+}
diff --git a/plugins/fleet/evals/reach-start-in-profile/samples/skill-fired.json b/plugins/fleet/evals/reach-start-in-profile/samples/skill-fired.json
new file mode 100644
index 00000000000..9c7398fac3d
--- /dev/null
+++ b/plugins/fleet/evals/reach-start-in-profile/samples/skill-fired.json
@@ -0,0 +1,54 @@
+{
+ "pass": [
+ {
+ "answer": [
+ {
+ "tool": "Skill",
+ "input": {
+ "skill": "fleet:reach"
+ }
+ }
+ ],
+ "why": "qualified name"
+ },
+ {
+ "answer": [
+ {
+ "tool": "Skill",
+ "input": {
+ "skill": "reach"
+ }
+ }
+ ],
+ "why": "bare name"
+ }
+ ],
+ "fail": [
+ {
+ "answer": [],
+ "why": "no Skill call"
+ },
+ {
+ "answer": [
+ {
+ "tool": "Skill",
+ "input": {
+ "skill": "other-plugin:other-skill"
+ }
+ }
+ ],
+ "why": "another skill"
+ },
+ {
+ "answer": [
+ {
+ "tool": "Skill",
+ "input": {
+ "skill": "fleet:reach-other"
+ }
+ }
+ ],
+ "why": "a skill whose name only starts the same"
+ }
+ ]
+}
diff --git a/plugins/fleet/evals/reach-start-in-profile/scaffold.sh b/plugins/fleet/evals/reach-start-in-profile/scaffold.sh
new file mode 100755
index 00000000000..da2e8cc1001
--- /dev/null
+++ b/plugins/fleet/evals/reach-start-in-profile/scaffold.sh
@@ -0,0 +1,3 @@
+#!/usr/bin/env bash
+set -euo pipefail
+bash "$(dirname "${BASH_SOURCE[0]}")/../fixtures/fleet-manifest.sh"
diff --git a/plugins/fleet/evals/reach-start-without-profile/case.yaml b/plugins/fleet/evals/reach-start-without-profile/case.yaml
new file mode 100644
index 00000000000..defea5ab27f
--- /dev/null
+++ b/plugins/fleet/evals/reach-start-without-profile/case.yaml
@@ -0,0 +1,25 @@
+schema_version: "1.1"
+name: 'reach-start-without-profile'
+context:
+ scaffold_script: scaffold.sh
+graders:
+ - name: 'one-shot-turn-over-wsl-hop'
+ type: 'regex'
+ pattern: '(?/scripts/profiles.mjs launch --unattended --json'",
+ "why": "runs launch with no profile named"
+ }
+ ]
+}
diff --git a/plugins/fleet/evals/reach-start-without-profile/samples/one-shot-turn-over-wsl-hop.json b/plugins/fleet/evals/reach-start-without-profile/samples/one-shot-turn-over-wsl-hop.json
new file mode 100644
index 00000000000..1a6bc3a1450
--- /dev/null
+++ b/plugins/fleet/evals/reach-start-without-profile/samples/one-shot-turn-over-wsl-hop.json
@@ -0,0 +1,18 @@
+{
+ "pass": [
+ {
+ "answer": "```sh\nprintf '%s\\n' 'List the local branches in this repository and reply with the list.' | /mnt/c/Windows/System32/OpenSSH/ssh.exe laptop-wsl 'cd ~/repos/app && claude -p'\n```\n\nNo plugin profile is applied: you named none, so the turn starts as the route gives it.",
+ "why": "stdin prompt over the WSL hop"
+ }
+ ],
+ "fail": [
+ {
+ "answer": "",
+ "why": "empty reply"
+ },
+ {
+ "answer": "/mnt/c/Windows/System32/OpenSSH/ssh.exe laptop-wsl 'cd ~/repos/app && git branch'",
+ "why": "a script, not the laptop's Claude"
+ }
+ ]
+}
diff --git a/plugins/fleet/evals/reach-start-without-profile/samples/says-profile-not-applied.json b/plugins/fleet/evals/reach-start-without-profile/samples/says-profile-not-applied.json
new file mode 100644
index 00000000000..7ac9d24a2e7
--- /dev/null
+++ b/plugins/fleet/evals/reach-start-without-profile/samples/says-profile-not-applied.json
@@ -0,0 +1,22 @@
+{
+ "pass": [
+ {
+ "answer": "No plugin profile is applied: you named none.",
+ "why": "no ... profile ... applied"
+ },
+ {
+ "answer": "The turn starts without a plugin profile; the profile was not applied because none was named.",
+ "why": "profile ... not applied"
+ }
+ ],
+ "fail": [
+ {
+ "answer": "printf '%s\\n' 'List the local branches.' | /mnt/c/Windows/System32/OpenSSH/ssh.exe laptop-wsl 'cd ~/repos/app && claude -p'",
+ "why": "command only, says nothing about the profile"
+ },
+ {
+ "answer": "The build profile is applied.",
+ "why": "claims a profile"
+ }
+ ]
+}
diff --git a/plugins/fleet/evals/reach-start-without-profile/samples/skill-fired.json b/plugins/fleet/evals/reach-start-without-profile/samples/skill-fired.json
new file mode 100644
index 00000000000..9c7398fac3d
--- /dev/null
+++ b/plugins/fleet/evals/reach-start-without-profile/samples/skill-fired.json
@@ -0,0 +1,54 @@
+{
+ "pass": [
+ {
+ "answer": [
+ {
+ "tool": "Skill",
+ "input": {
+ "skill": "fleet:reach"
+ }
+ }
+ ],
+ "why": "qualified name"
+ },
+ {
+ "answer": [
+ {
+ "tool": "Skill",
+ "input": {
+ "skill": "reach"
+ }
+ }
+ ],
+ "why": "bare name"
+ }
+ ],
+ "fail": [
+ {
+ "answer": [],
+ "why": "no Skill call"
+ },
+ {
+ "answer": [
+ {
+ "tool": "Skill",
+ "input": {
+ "skill": "other-plugin:other-skill"
+ }
+ }
+ ],
+ "why": "another skill"
+ },
+ {
+ "answer": [
+ {
+ "tool": "Skill",
+ "input": {
+ "skill": "fleet:reach-other"
+ }
+ }
+ ],
+ "why": "a skill whose name only starts the same"
+ }
+ ]
+}
diff --git a/plugins/fleet/evals/reach-start-without-profile/scaffold.sh b/plugins/fleet/evals/reach-start-without-profile/scaffold.sh
new file mode 100755
index 00000000000..da2e8cc1001
--- /dev/null
+++ b/plugins/fleet/evals/reach-start-without-profile/scaffold.sh
@@ -0,0 +1,3 @@
+#!/usr/bin/env bash
+set -euo pipefail
+bash "$(dirname "${BASH_SOURCE[0]}")/../fixtures/fleet-manifest.sh"
diff --git a/plugins/fleet/skills/reach/SKILL.md b/plugins/fleet/skills/reach/SKILL.md
index 63998b94280..4088dbd5e20 100644
--- a/plugins/fleet/skills/reach/SKILL.md
+++ b/plugins/fleet/skills/reach/SKILL.md
@@ -109,7 +109,13 @@ dated measurement and how to re-measure it). When a script can do the job, run t
| Background session | `--bg --name ""`, not `-p`, with `< /dev/null` on an SSH row; manage with `claude agents --json --all`, `stop `, `rm ` | none: it is in `` | tested on R2, trusted directory only |
- Give every prompt the whole task: what done looks like, and what should make it stop and report.
- Nobody answers a question a headless turn asks.
+ Nobody answers a question a headless turn asks. When the task touches a PR, issue or branch a
+ session holds, get that session's name and put it in the prompt, and never tell the target to
+ take the work over on its own judgment.
+- **Start in a profile.** When the task names a plugin profile, start the target's turn or session
+ in it as [reference/relay.md § Start in the target's plugin profile](reference/relay.md#start-in-the-targets-plugin-profile)
+ says. With no profile named, or none applied, say with the result that the profile was not
+ applied.
- **Query and wait has no one-turn form.** Never wait on a `notify_when_idle` notice from a `-p`
sender: in our probe it never reached the sender
([reference/relay.md § Verified behavior](reference/relay.md#verified-behavior)). Send the
diff --git a/plugins/fleet/skills/reach/reference/relay.md b/plugins/fleet/skills/reach/reference/relay.md
index eb6d121dbe1..9e3247feac4 100644
--- a/plugins/fleet/skills/reach/reference/relay.md
+++ b/plugins/fleet/skills/reach/reference/relay.md
@@ -11,6 +11,7 @@ plain prompt.
- [A named receiver](#a-named-receiver)
- [Multi-turn and the open pipe](#multi-turn-and-the-open-pipe)
- [Background sessions](#background-sessions)
+- [Start in the target's plugin profile](#start-in-the-targets-plugin-profile)
- [Mechanisms not used here](#mechanisms-not-used-here)
- [Getting a reply back](#getting-a-reply-back)
- [Signed-out Windows lane](#signed-out-windows-lane)
@@ -165,6 +166,36 @@ For a session that outlives the hop, the target lane's own supervisor hosts it:
- **Pointer**: [Agent view: Limitations](https://code.claude.com/docs/en/agent-view#limitations),
fetched live.
+## Start in the target's plugin profile
+
+When the task names a plugin profile, a turn or session started in a WSL lane (R2, R4, R5) starts
+in that lane's profile, through plugin-profiles' `profiles.mjs launch` on the target. Each step
+runs over the row's hop; `` is the directory the session starts in.
+
+1. Find the script: run `claude plugin list --json` on the target and take the `installPath` of
+ the entry whose `id` starts with `plugin-profiles@` and whose `enabled` is true. The script is
+ `/scripts/profiles.mjs`.
+2. Run `launch` there with `--unattended`, since nobody answers a relay start, and put the verb's
+ own `--settings` JSON in `--with`, so the started session carries one settings file
+ (`profiles.mjs` merges it into the profile's):
+
+ ```console
+ 'cd && node launch --folder --unattended --json --with "{\"crossSessionInbound\":\"accept\"}"'
+ ```
+
+ Leave out `--with` when the verb carries no `--settings`. On exit 1 (the profile switches on a
+ plugin missing a required tool) start nothing and report the reason it printed.
+3. Start the session with the printed `command` in place of `claude`, and the verb's other
+ arguments after it, without its `--settings`:
+
+ ```console
+ printf '%s\n' '' | 'cd && -p -n '
+ ```
+
+With no profile named, no enabled `plugin-profiles@` entry in step 1, or a Windows lane, start the
+session as the row does without this section, and say with the result that the profile was not
+applied and why. A Windows lane is left out because the printed `command` is a POSIX `env` line.
+
## Mechanisms not used here
Each fails the fleet's per-lane account split or needs a human; see the linked page before
@@ -243,6 +274,7 @@ also rests on documented flag behavior.
| Port 22 script hop | Not probed in this pass |
| `notify_when_idle` from a `-p` sender | Message delivered; the notice never reached the sender, whose turn ended first |
| `-p --remote-control` | Did not connect; Remote Control is interactive only |
+| [Start in the target's plugin profile](#start-in-the-targets-plugin-profile) | Untested on a live hop |
| Cost | Most of each one-line `-p` turn's cost was SessionStart hooks and context loading; measure a turn per |
## What the relay does not change