docs: record eval 0020 rerun and follow-up task

.recursive/evaluations/0020.md

@@ -0,0 +1,39 @@
+# Evaluation #0020
+**Date**: 2026-04-09
+**Target**: Phractal (github.com/fazxes/Phractal)
+**Agent**: claude
+**Cycles**: 2
+**After task**: archived #0243 (Phractal eval rerun)
+**Eval gate status**: BLOCKED (score < 80)
+
+## Scores
+
+| Dimension | Score | Notes |
+|-----------|-------|-------|
+| Startup | 0/10 | Default run cloned and ran baseline verification, but both cycles failed before any agent work because the spawned `claude` subprocess could not launch inside the current Claude Code session. |
+| Discovery | 0/10 | No fixes or logged issues were produced. |
+| Fix quality | 0/10 | No candidate fixes were committed. |
+| Shift log | 3/10 | Shift log exists, but it is still the template stub with no real fix entries, commit hashes, or recommendations. |
+| State file | 8/10 | Valid JSON with the expected keys. The file clearly records the failure state (`agent_failures=2`, `cycles[*].status=agent_failed`, `halt_reason` populated), even though no fixes were produced. |
+| Verification | 5/10 | Baseline verification ran successfully, but neither cycle reached post-cycle verification. |
+| Guard rails | 9/10 | No blocked paths were touched and the run stayed within the declared limits. |
+| Clean state | 10/10 | The target repo clone is clean at the end of the run. |
+| Breadth | 0/10 | No cycle completed, so no breadth was demonstrated. |
+| Usefulness | 0/10 | The run produced no actionable fixes; the only usable artifact is the failure evidence itself. |
+| **Total** | **35/100** | Eval gate BLOCKED. |
+
+## Tasks Created
+
+- #0277: Make Claude eval reruns scorable from Claude Code sessions (Startup 0/10, Discovery 0/10, Fix quality 0/10, Verification 5/10, Breadth 0/10, Usefulness 0/10)
+
+## Raw Evidence
+
+- Default attempt command: `PYTHONPATH=$(pwd) python3 -m nightshift test --agent claude --cycles 2 --cycle-minutes 5 --repo-dir /tmp/nightshift-eval-0020`
+- Retry attempt command: `env -u CLAUDECODE PYTHONPATH=$(pwd) python3 -m nightshift test --agent claude --cycles 2 --cycle-minutes 5 --repo-dir /tmp/nightshift-eval-0020`
+- Baseline verification: `python3 -m compileall apps/api/app` passed both attempts.
+- Shift log: `/var/folders/dm/48s5j0qs7yq4nvq9s61jcbqw0000gn/T/nightshift-test-runs/nightshift-eval-0020-47e0cf34a31c/worktree-2026-04-09/Docs/Nightshift/2026-04-09.md`
+- State file: `/var/folders/dm/48s5j0qs7yq4nvq9s61jcbqw0000gn/T/nightshift-test-runs/nightshift-eval-0020-47e0cf34a31c/2026-04-09.state.json`
+- Runner log: `/var/folders/dm/48s5j0qs7yq4nvq9s61jcbqw0000gn/T/nightshift-test-runs/nightshift-eval-0020-47e0cf34a31c/2026-04-09.runner.log`
+- State observations: `agent_failures=2`, `fixes=0`, `issues_logged=0`, `files_touched=0`, `tests_written=0`, `cycles[0].status=agent_failed`, `cycles[1].status=agent_failed`, `halt_reason="Agent command failed twice in a row."`
+- Runner failure: the nested Claude invocation error repeated in both attempts, so no scorable cycle output was produced.
+

.recursive/tasks/.next-id

@@ -1 +1 @@
-277
+278

.recursive/tasks/0277.md

@@ -0,0 +1,27 @@
+---
+status: pending
+priority: urgent
+target: v0.0.9
+vision_section: self-maintaining
+created: 2026-04-09
+source: evaluation-0020
+completed:
+---
+
+# Make Claude eval reruns scorable from Claude Code sessions
+
+## Problem
+
+The fresh Phractal eval rerun for `#0020` could not complete in this environment because the spawned `claude` subprocess failed inside an existing Claude Code session. Unsetting `CLAUDECODE` was not enough to produce a scorable 2-cycle run.
+
+## Fix
+
+Make the eval runner or launch path resilient to Claude-in-Claude execution so a default `nightshift test --agent claude --cycles 2 --cycle-minutes 5` rerun can complete and produce a report without manual shell workarounds.
+
+## Acceptance Criteria
+
+- [ ] `nightshift test --agent claude --cycles 2 --cycle-minutes 5` can complete from a Claude Code shell without the nested-session failure
+- [ ] If Claude-in-Claude remains unsupported, the runner automatically falls back to a supported agent or emits a clear, actionable failure before starting cycles
+- [ ] Regression coverage exists for the nested-session path
+- [ ] A fresh eval rerun produces a scorable report instead of halting after two agent failures
+