inceptionstack · royosherove · Apr 7, 2026 · Apr 6, 2026 · Apr 6, 2026 · Apr 6, 2026
diff --git a/deploy/test-templates.sh b/deploy/test-templates.sh
@@ -95,6 +95,73 @@ check_contains "$INSTALL_SH" 'PackName' "install.sh: PackName in PARAM_CFN_NAMES
 check_contains "$INSTALL_SH" 'pack_name' "install.sh: pack_name in PARAM_TF_NAMES"
 check_contains "$INSTALL_SH" 't4g.medium' "install.sh: hermes default size logic present"
 
+# ── Branch detection & SSM doc version ──────────────────────────────────────
+echo -e "${BOLD}Branch & SSM fixes${NC}"
+check_contains "$INSTALL_SH" '[[ "$REPO_BRANCH" == "HEAD" ]]' "install.sh: detached HEAD falls back to main"
+check_contains "$INSTALL_SH" 'REPO_BRANCH=' "install.sh: REPO_BRANCH is set"
+check_contains "$TF_VARS" 'variable "repo_branch"' "TF: repo_branch variable defined"
+check_contains "$TF_MAIN" "repo_branch" "TF main: repo_branch passed to userdata template"
+check_contains "$TF_USERDATA" 'repo_branch' "TF userdata: uses repo_branch for git clone"
+check_contains "$CFN_TEMPLATE" "RepoBranch" "CFN: RepoBranch parameter defined"
+check_contains "$INSTALL_SH" "DocumentDescription.DocumentVersion" "install.sh: SSM update-document captures numeric version"
+check_contains "$INSTALL_SH" 'new_version' "install.sh: SSM update-document-default-version uses captured version"
+
+echo ""
+
+# ── Branch detection unit tests ─────────────────────────────────────────────
+echo -e "${BOLD}Branch detection (unit)${NC}"
+
+# Test: detached HEAD → main
+_test_branch="HEAD"
+[[ "$_test_branch" == "HEAD" ]] && _test_branch="main"
+if [[ "$_test_branch" == "main" ]]; then
+ pass "Detached HEAD resolves to main"
+else
+ fail "Detached HEAD should resolve to main, got: $_test_branch"
+fi
+
+# Test: normal branch → unchanged
+_test_branch="installer-ux-overhaul"
+[[ "$_test_branch" == "HEAD" ]] && _test_branch="main"
+if [[ "$_test_branch" == "installer-ux-overhaul" ]]; then
+ pass "Normal branch name preserved"
+else
+ fail "Normal branch should be preserved, got: $_test_branch"
+fi
+
+# Test: main → unchanged
+_test_branch="main"
+[[ "$_test_branch" == "HEAD" ]] && _test_branch="main"
+if [[ "$_test_branch" == "main" ]]; then
+ pass "main branch preserved"
+else
+ fail "main should be preserved, got: $_test_branch"
+fi
+
+# Test: SSM version regex accepts numeric
+_test_version="5"
+if [[ "$_test_version" =~ ^[0-9]+$ ]]; then
+ pass "SSM version regex accepts numeric version"
+else
+ fail "SSM version regex should accept '5'"
+fi
+
+# Test: SSM version regex rejects $LATEST
+_test_version='$LATEST'
+if [[ "$_test_version" =~ ^[0-9]+$ ]]; then
+ fail "SSM version regex should reject '\$LATEST'"
+else
+ pass "SSM version regex rejects \$LATEST"
+fi
+
+# Test: SSM version regex rejects empty
+_test_version=""
+if [[ -n "$_test_version" && "$_test_version" =~ ^[0-9]+$ ]]; then
+ fail "SSM version regex should reject empty string"
+else
+ pass "SSM version regex rejects empty string"
+fi
+
 echo ""
 echo -e "${BOLD}─────────────────────────────────────────────────${NC}"
 echo -e " Passed: ${GREEN}${PASS}${NC} Failed: ${RED}${FAIL}${NC}"

diff --git a/install.sh b/install.sh
@@ -126,6 +126,7 @@ DEPLOY_TERRAFORM=3
 INSTALLER_COMMIT="${INSTALLER_COMMIT:-$(git -C "$SCRIPT_DIR" rev-parse --short HEAD 2>/dev/null || echo dev)}"
 INSTALLER_DATE="${INSTALLER_DATE:-$(d=$(git -C "$SCRIPT_DIR" log -1 --format='%ci' 2>/dev/null | cut -d' ' -f1,2); echo "${d:-unknown}")}"
 REPO_BRANCH="${REPO_BRANCH:-$(git -C "$SCRIPT_DIR" rev-parse --abbrev-ref HEAD 2>/dev/null || echo main)}"
+[[ "$REPO_BRANCH" == "HEAD" ]] && REPO_BRANCH="main"
 
 # Detect AWS CloudShell (limited ~1GB home dir, use /tmp for large files)
 IS_CLOUDSHELL=false
@@ -1677,16 +1678,20 @@ ensure_ssm_session_document() {
  local doc_content='{"schemaVersion":"1.0","description":"SSM session for Loki - starts as ec2-user and launches TUI","sessionType":"Standard_Stream","inputs":{"runAsEnabled":true,"runAsDefaultUser":"ec2-user","shellProfile":{"linux":"cd ~ && bash --login -c \"loki tui || exec bash --login\""}}}'
 
  if aws ssm describe-document --name "$SSM_DOC_NAME" --region "$DEPLOY_REGION" &>/dev/null; then
- # Update existing document to latest version
- aws ssm update-document \
+ # Update existing document and set new version as default
+ local new_version
+ new_version=$(aws ssm update-document \
  --name "$SSM_DOC_NAME" \
  --content "$doc_content" \
  --document-version '$LATEST' \
- --region "$DEPLOY_REGION" >/dev/null 2>&1 || true
- aws ssm update-document-default-version \
- --name "$SSM_DOC_NAME" \
- --document-version '$LATEST' \
- --region "$DEPLOY_REGION" >/dev/null 2>&1 || true
+ --region "$DEPLOY_REGION" \
+ --query 'DocumentDescription.DocumentVersion' --output text 2>/dev/null) || true
+ if [[ -n "$new_version" && "$new_version" =~ ^[0-9]+$ ]]; then
+ aws ssm update-document-default-version \
+ --name "$SSM_DOC_NAME" \
+ --document-version "$new_version" \
+ --region "$DEPLOY_REGION" >/dev/null 2>&1 || true
+ fi
  ok "SSM session document: ${SSM_DOC_NAME} (updated)"
  return 0
  fi

diff --git a/upstream-fixes.md b/upstream-fixes.md
@@ -0,0 +1,87 @@
+# Upstream Fixes Needed: OpenClaw + pi-coding-agent Bedrock Auth
+
+OpenClaw + pi-coding-agent fail when using AWS Bedrock with EC2 instance-profile auth (IMDS).
+The AWS SDK credential chain works fine, but the auth pre-flight checks reject it before any request is made.
+
+We currently patch these at install time (`packs/openclaw/resources/patch-pi-agent.py`), but
+these patches are overwritten on every OpenClaw update.
+
+## Issue 1: pi-coding-agent `hasConfiguredAuth()` rejects AWS SDK auth
+
+**File:** `src/core/model-registry.ts` (dist: `dist/core/model-registry.js`)
+
+**Problem:** `hasConfiguredAuth()` only checks `authStorage.hasAuth()` and `providerRequestConfigs.apiKey`. Neither is set for AWS SDK auth via instance profile. This causes "No API key found for amazon-bedrock" before any API request is made.
+
+**Fix:** Return `true` when the provider is `amazon-bedrock` and the config uses `auth: "aws-sdk"`.
+
+```typescript
+// In hasConfiguredAuth():
+if (this.providerId === "amazon-bedrock") return true;
+```
+
+**Why:** Bedrock auth uses AWS SDK signing (SigV4), not API keys. The SDK resolves credentials from the instance metadata service (IMDS) at request time. There's no key to pre-check.
+
+---
+
+## Issue 2: pi-coding-agent `_getRequiredRequestAuth()` throws on undefined apiKey
+
+**File:** `src/core/agent-session.ts` (dist: `dist/core/agent-session.js`)
+
+**Problem:** `_getRequiredRequestAuth()` throws when `getApiKeyAndHeaders()` returns `ok: true` but no `apiKey`. Bedrock uses AWS SDK signing (no API key needed), but this code path doesn't account for it.
+
+**Fix:** Early return for `amazon-bedrock` provider with `{ apiKey: undefined, headers }`.
+
+```typescript
+// In _getRequiredRequestAuth():
+if (this.providerId === "amazon-bedrock") {
+ const authResult = await this._getRequestAuth?.() || { ok: true, headers: {} };
+ return { apiKey: undefined, headers: authResult.headers || {} };
+}
+```
+
+**Why:** The Bedrock provider adapter handles auth via the AWS SDK at the HTTP layer (SigV4 signing). It never needs an API key injected by the session.
+
+---
+
+## Issue 3 (Root Cause): OpenClaw auth-controller doesn't inject SDK auth into pi's authStorage
+
+**File:** `src/agents/pi-embedded-runner/run/auth-controller.ts`
+
+**Problem:** OpenClaw's auth controller correctly resolves AWS SDK auth (returns `{ mode: "aws-sdk", source: "..." }` from `resolveAwsSdkAuthInfo()`), but does an early return at ~line 329-337 without calling `setRuntimeApiKey()`. This means pi's `authStorage` never learns about Bedrock credentials.
+
+**Proper fix:** When auth mode is `aws-sdk`, the auth controller should either:
+1. Call `setRuntimeApiKey()` with a sentinel value that pi recognizes as "use SDK signing", or
+2. Set a flag on the provider config that pi checks in `hasConfiguredAuth()` and `_getRequiredRequestAuth()`
+
+This would eliminate the need for patches #1 and #2.
+
+---
+
+## Issue 4: `resolveAwsSdkEnvVarName()` doesn't detect IMDS
+
+**File:** `src/agents/model-auth-runtime-shared.ts`
+
+**Problem:** `resolveAwsSdkEnvVarName()` checks for `AWS_BEARER_TOKEN_BEDROCK`, `AWS_ACCESS_KEY_ID`+`AWS_SECRET_ACCESS_KEY`, and `AWS_PROFILE`. If none are set, it returns `undefined`. EC2 instances with IAM roles don't need any env vars — the SDK discovers credentials via IMDS automatically.
+
+**Current workaround:** We set `AWS_PROFILE=default` in `/etc/profile.d/` and the systemd service unit, which triggers the env var check and lets the SDK fall through to IMDS.
+
+**Proper fix:** The fallback in `resolveAwsSdkAuthInfo()` already returns `{ mode: "aws-sdk", source: "aws-sdk default chain" }` when no env vars are found — this is correct. The issue is upstream in the auth controller (Issue 3) not propagating this to pi.
+
+---
+
+## Environment
+
+- OpenClaw version: latest (as of 2026-04-06)
+- pi-coding-agent version: 0.65.0
+- Platform: Amazon Linux 2023 on EC2 with IAM instance profile
+- Auth method: Bedrock via IMDS (no API keys, no env vars)
+- Config: `models.providers.amazon-bedrock.auth: "aws-sdk"`
+
+## Reproduction
+
+1. Launch EC2 instance with IAM role that has Bedrock access
+2. Install OpenClaw: `npm install -g openclaw`
+3. Configure with `auth: "aws-sdk"` for amazon-bedrock provider
+4. Run `openclaw tui` — fails with "No API key found for amazon-bedrock"
+5. Set `AWS_PROFILE=default` — still fails (auth pre-flight in pi rejects it)
+6. Apply patches to model-registry.js and agent-session.js — works