refactor: remove claude wrapper, run claude directly

cmd/helpers_test.go

@@ -76,7 +76,7 @@ func setupTestAgent(t *testing.T) string {
 	os.MkdirAll(filepath.Join(dir, "agents"), 0o755)
 	os.WriteFile(filepath.Join(dir, "agents", "default.yaml"), []byte(`name: test-agent
 image: quay.io/test:latest
-entrypoint: claude --bare
+entrypoint: claude
 providers:
   - profile: github
   - profile: vertex-local

internal/agent/agent.go

@@ -104,8 +104,6 @@ func (c *AgentConfig) BuildRunSh() string {
 	b.WriteString("PAYLOAD_DIR=\"$(cd \"$(dirname \"$0\")\" && pwd)\"\n\n")
 	b.WriteString("# Prepend payload bin to PATH\n")
 	b.WriteString("export PATH=\"$PAYLOAD_DIR/bin:$PATH\"\n\n")
-	b.WriteString("# Git auth\n")
-	b.WriteString("gh auth setup-git 2>/dev/null || true\n\n")
 	b.WriteString("# Validate entrypoint\n")
 	entrypoint := c.EffectiveEntrypoint()
 	epBin := strings.Fields(entrypoint)[0]

internal/agent/agent_test.go

@@ -17,7 +17,7 @@ providers:
       JIRA_URL: https://issues.redhat.com
   - profile: github
 task: tasks/daily-standup.md
-entrypoint: claude --bare
+entrypoint: claude
 `)
 	cfg, err := Parse(data)
 	if err != nil {
@@ -179,7 +179,7 @@ func TestParseFile_AgentYAML(t *testing.T) {
 	dir := t.TempDir()
 	yamlContent := `name: test-agent
 image: ghcr.io/test:latest
-entrypoint: claude --bare
+entrypoint: claude
 tty: true
 providers:
   - profile: github
@@ -268,7 +268,7 @@ func TestBuildEnvMap_Empty(t *testing.T) {
 
 func TestBuildRunSh(t *testing.T) {
 	cfg := &AgentConfig{
-		Entrypoint: "claude --bare",
+		Entrypoint: "claude",
 		Task:       "tasks/standup.md",
 	}
 	runSh := cfg.BuildRunSh()
@@ -278,13 +278,10 @@ func TestBuildRunSh(t *testing.T) {
 	if strings.Contains(runSh, "env.sh") {
 		t.Error("run.sh should not source env.sh — env vars are injected via --env")
 	}
-	if !strings.Contains(runSh, "gh auth setup-git") {
-		t.Error("missing gh auth setup-git")
-	}
 	if !strings.Contains(runSh, `command -v "claude"`) {
 		t.Error("missing entrypoint validation")
 	}
-	if !strings.Contains(runSh, `exec claude --bare -p "$(cat "$PAYLOAD_DIR/task.md")"`) {
+	if !strings.Contains(runSh, `exec claude -p "$(cat "$PAYLOAD_DIR/task.md")"`) {
 		t.Errorf("missing task exec with -p in:\n%s", runSh)
 	}
 }
@@ -310,7 +307,7 @@ func TestRenderPayload(t *testing.T) {
 			{Profile: "atlassian", Env: map[string]string{"JIRA_URL": "https://jira.example.com"}},
 		},
 		Task:       "my-task.md",
-		Entrypoint: "claude --bare",
+		Entrypoint: "claude",
 	}
 
 	destDir := filepath.Join(t.TempDir(), "payload")

sandbox/CLAUDE.md

@@ -1,41 +1,17 @@
-# Sandbox Agent Instructions
+# Sandbox Environment
 
-You are running inside an OpenShell sandbox. Credentials are injected via the OpenShell provider system — they appear as environment variables automatically.
+You are running inside an OpenShell sandbox. Credentials are injected via the OpenShell provider system and appear as environment variables automatically.
 
-## Tools Available
+## Environment
 
-### GitHub — `gh` CLI
-- Pre-authenticated. Use `gh` for all GitHub operations.
-- Examples: `gh repo clone`, `gh pr create`, `gh issue list`, `gh api`
-
-### Jira & Confluence — mcp-atlassian MCP server
-- Connected via the `atlassian` MCP server (credentials injected by provider).
-- Use MCP tools for Jira searches, issue creation, comments, and Confluence page reads.
-
-### Google Workspace — `gws` CLI
-- Pre-authenticated for Gmail, Calendar, Drive, Docs, Sheets.
-- Path: `/usr/local/bin/gws`
-- Examples:
-  - `gws gmail users messages list --params '{"userId": "me", "maxResults": 5}'`
-  - `gws calendar events list --params '{"calendarId": "primary", "maxResults": 5}'`
-  - `gws drive files list --params '{"pageSize": 10}'`
-- Use `gws schema <service.resource.method>` to discover API parameters.
-
-### AI Coding Agents
-- `claude` — Claude Code (wrapper at `/usr/local/bin/claude`)
-- `opencode` — OpenCode (MCP config at `/sandbox/opencode.json`)
-
-### General Tools
-- `python3`, `pip`, `uv` — Python with a virtualenv at `/sandbox/.venv`
-- `node`, `npm` — Node.js 22
-- `git`, `curl` — pre-installed
-
-## Configuration
-- Inference routes through the gateway proxy at `inference.local`.
-- Model and provider configuration are set by the harness during sandbox creation.
-
-## Conventions
 - Working directory: `/sandbox`
 - Writable paths: `/sandbox`, `/tmp`
-- Network: Outbound allowed to Google APIs, GitHub, Atlassian, npm/pypi.
-- Credentials are managed by the OpenShell provider system and cleaned up on sandbox exit.
+- Inference routes through the gateway proxy at `inference.local`
+- Credentials are managed by OpenShell and cleaned up on sandbox exit
+
+## Tools
+
+- `gh` — GitHub CLI (pre-authenticated). Run `gh auth setup-git` before any git clone/push/pull to configure git credential helper.
+- `gws` — Google Workspace CLI (when available). Use `gws schema <service.resource.method>` to discover API parameters.
+- MCP servers (Jira, Confluence) are configured in `.mcp.json` and connected automatically.
+- `python3`, `uv`, `node`, `npm`, `git`, `curl`

sandbox/Dockerfile

@@ -38,10 +38,6 @@
 #   GWS config could be provider-managed and the CLI could move to a
 #   community sandbox.
 #
-# startup.sh: Baked into the image to wire runtime credentials (Atlassian
-#   MCP config, GWS file placement). See sandbox.sh for the specific
-#   workarounds and their upstream tracking issues.
-
 ARG BASE_IMAGE=ghcr.io/nvidia/openshell-community/sandboxes/base:latest
 FROM ${BASE_IMAGE}
 
@@ -66,27 +62,16 @@ RUN npm install -g opencode-ai@latest
 COPY policy.yaml /etc/openshell/policy.yaml
 
 # Agent instructions + config (Claude Code + OpenCode)
-COPY CLAUDE.md /sandbox/CLAUDE.md
+COPY CLAUDE.md /sandbox/.claude/CLAUDE.md
 COPY settings.json /sandbox/.claude/settings.json
 COPY claude.json /sandbox/.claude.json
 COPY mcp.json /sandbox/.mcp.json
 COPY opencode.json /sandbox/opencode.json
 
-# Runtime script
-COPY startup.sh /sandbox/startup.sh
-
-# Claude wrapper: move the real binary aside, install wrapper in its place.
-# The base image puts claude at /usr/local/bin/claude; we move it to
-# /usr/local/lib/claude-real so the wrapper can call it.
-RUN mv /usr/local/bin/claude /usr/local/lib/claude-real
-COPY bin/claude /usr/local/bin/claude
-
 # Harness config directory (populated at runtime via upload)
 RUN mkdir -p /sandbox/.config/openshell && chown sandbox:sandbox /sandbox/.config/openshell
 
-RUN chmod 755 /usr/local/bin/claude && \
-    chmod +x /sandbox/startup.sh && \
-    chown -R sandbox:sandbox /sandbox/.claude /sandbox/.claude.json /sandbox/.mcp.json /sandbox/opencode.json /sandbox/.config /sandbox/CLAUDE.md /sandbox/startup.sh
+RUN chown -R sandbox:sandbox /sandbox/.claude /sandbox/.claude.json /sandbox/.mcp.json /sandbox/opencode.json /sandbox/.config
 
 USER sandbox
 

test/test-flow.sh

@@ -158,7 +158,7 @@ sandbox_verify() {
   step "sandbox: gws token placeholder" "$CLI" sandbox exec --name "$name" -- bash -c 'echo "$GOOGLE_WORKSPACE_CLI_TOKEN" | grep -q "openshell:resolve:env"'
   step "sandbox: gws api call" "$CLI" sandbox exec --name "$name" -- bash -c 'for i in 1 2 3; do curl -sf https://gmail.googleapis.com/gmail/v1/users/me/profile -H "Authorization: Bearer $GOOGLE_WORKSPACE_CLI_TOKEN" -o /dev/null && exit 0; sleep 3; done; exit 1'
   step "sandbox: mcp config" "$CLI" sandbox exec --name "$name" -- test -f /sandbox/.mcp.json
-  step "sandbox: claude responds" "$CLI" sandbox exec --name "$name" -- bash -c 'echo "respond with ok" | claude --bare --print 2>&1 | head -1'
+  step "sandbox: claude responds" "$CLI" sandbox exec --name "$name" -- bash -c 'echo "respond with ok" | claude --print 2>&1 | head -1'
 }
 
 summary() {