FEAT: EVALUATE_PRD command for desktop symphony-loop harness

[aponamarev]

EVALUATE_PRD command for desktop symphony-loop harness

This PR adds the EVALUATE_PRD command to the desktop symphony-loop harness, enabling PRD evaluation through the local gateway. When invoked, the harness writes prd.md from artifacts (or prompt), spawns the Claude CLI with a prompt that activates the judges:run-judges skill, and returns prd-judges.json results to the cloud control plane.

Key Changes

Core Feature:

• Implemented EVALUATE_PRD command in symphony-loop.ts — uses a temp dir (no worktree), writes prd.md via shared writePrdArtifact, and spawns claude -p - --output-format json with an evaluation prompt
• Prompt includes CLOSEDLOOP_WORKDIR, optional REPO_PATH when a repo is linked, and instructs the model to run judges:run-judges --artifact-type prd
• Added readEvaluatePrdOutputs() to read prd-judges.json and return { prdJudges } in the completed event
• Temp dir cleanup after completion (same pattern as DECOMPOSE)

Supporting Changes:

• Extracted writePrdArtifact() for reuse by DECOMPOSE and EVALUATE_PRD
• Extracted spawnClaudeFromFile() to avoid E2BIG when passing large prompts via stdin
• Added CL_TEST_ALLOW_LOOPBACK_API=1 to bypass private-address checks in tests
• Comprehensive test suite for EVALUATE_PRD (prompt contents, repo/no-repo, output parsing, temp cleanup, BINARY_NOT_FOUND)

Impact:

• API: New EVALUATE_PRD command on /api/engineer/symphony/loop — accepts same body shape as DECOMPOSE (artifacts, prompt, optional repo)

[closedloop-ai-stage]

Code Review Summary

Status: Changes Requested

Reviewers: Bug Hunter A, Bug Hunter B, Unified Auditor, Premise Reviewer, gateway-core-architect

Findings

Severity	Count
Blocking	1
High	0
Medium	3

BLOCKING Issues (must fix)

1. [P0] [symphony-loop.ts:126] Test-only SSRF bypass env var baked into production validation code — CL_TEST_ALLOW_LOOPBACK_API=1 disables all SSRF protections in validateApiBaseUrl() with no app.isPackaged guard; if this env var leaks into a production environment the gateway will forward requests to loopback/private addresses including cloud metadata endpoints.

MEDIUM Issues (consider)

1. [P2] [symphony-loop.ts:511] writeArtifactsForEvaluatePrd is a 100%-identical pass-through wrapper of writeArtifactsForDecompose; both bodies are await writePrdArtifact(...). The shared helper writePrdArtifact already exists for this purpose — the extra wrapper adds no value.
2. [P2] [symphony-loop-evaluate-prd.test.ts:479] T-5.4 cleanup assertion is a false positive: all integration test loopIds share the same first 8 chars (00000000), so they all write to the same OS temp dir. T-5.4's cleanup assertion passes because T-5.2 already triggered cleanup — not because T-5.4's own code path works.
3. [P2] [symphony-loop.ts:861] EVALUATE_PRD temp dir is only cleaned up on the success path (exit code 0). Non-zero claude exit and BINARY_NOT_FOUND both leave the symphony-evaluate-prd-<loopId> directory behind.

Validation Stats

• Total findings from agents: 5
• Duplicates merged: 1
• Cross-file grouped: 0

Recommendation: Address the blocking SSRF bypass before merging. The medium items are lower priority but the test false-positive (#2) should be fixed to ensure the cleanup behavior is actually exercised.