EPS Dashboard

Overview / Motivation

Auto-filed by the /daily 2026-07-26 problem sweep (route 2). Surfaced by 4 independent miner group(s) over the 2026-07-26 session transcripts.

Goal

Close four ergonomic defects in scripts/select_step9c_tests.py — --map-files silently ignoring --json, --json output corrupted by stderr NOTEs, an unhelpful --map-files read error, and a silent degrade to the invariant-only fallback on uncommitted edits — each of which cost a live session a wasted turn today.

Workflow gap

Bug observed: (a) the --map-files branch is TSV-only and returns before any --json handling, so a consumer that passes both dies on json.load; (b) the script's informational NOTEs go to stderr by design, so a caller redirecting 2>&1 into json.load gets JSONDecodeError; (c) an unreadable --map-files argument reports only the raw Errno 2 without saying the flag wants a PATH to a newline-separated file, not a comma-separated list; (d) run on UNCOMMITTED edits the selector silently degrades to the workflow-invariant set only, and its NOTE names the wrong likely cause.
Why it is a workflow gap: the selector is the mandated pre-report step for every implementer (implementer.md L174) and the Step 10d merge-gate mapping mode, so each of these fires on the copy-paste path rather than on an improvisation, and (d) silently voids the #1288 scope narrowing the step exists to buy.
Confidence (emitter): high
verified-at-filing: presence claims, per-target, in the named target scripts/select_step9c_tests.py — (a) Read L1644-1729: the if args.map_files is not None: branch (L1644) prints TSV at L1727-1728 and return 0s at L1729; grep -n 'args.json' scripts/select_step9c_tests.py → the ONLY consumer is L1786 (if args.json:), i.e. AFTER the map-files early return — --json is unreachable in mapping mode. 1 hit, confirming the silent-ignore. (b) every NOTE/WARN/sizing line in the file is file=sys.stderr (L1662-1664, L1670-1673, L1707-1711, L1722-1726, L1747-1751, L1777-1783); the JSON is the only stdout write at L1786+. Confirmed by construction. (c) sed -n '1658,1665p' → print(f"select_step9c_tests: cannot read --map-files input: {exc}", file=sys.stderr) — no PATH/format hint. The --map-files help= at L1601-1618 does say "newline-delimited repo-relative paths", but the ERROR does not. 1 hit. (d) grep -n 'empty diff\|falling back to the workflow-invariant' scripts/select_step9c_tests.py → 2 hits (L190 docstring, L1747-1751 the live NOTE); the NOTE text is "…falling back to the workflow-invariant set only. If this task's changes live in an issue worktree, re-run from that worktree (Step 9c contract)." — names ONLY the wrong-worktree cause, 0 hits for any uncommitted-work mention. Secondary target: grep -nic 'commit your edits first\|commit first' .claude/agents/implementer.md → 0 hits (absence claim); the recipe at L174 prescribes uv run python scripts/select_step9c_tests.py --json with no commit-first warning and no 2>/dev/null. Landed-fix check: git log --oneline --since='7 days ago' -- scripts/select_step9c_tests.py → most recent touch d79fa07b0e (#1699, the pin-sweep mode); none of the 4 defects addressed. (2026-07-26)

Evidence

Defect (a). Session 7df6ce4c, 2026-07-26T09:30:42Z: ran --map-files <file> --json and piped stdout into json.load; the consumer died with a bare traceback — "Exit code 1 ⏎ Traceback (most recent call last): ⏎ File \"<string>\", line 3, in <module> ⏎ File \".../json/__init__.py\", line 293, in load" — and the session had to re-derive the output shape by hand. Cost: one wasted turn plus a manual re-derivation.
Defect (b). Sessions 0e2c3b21 (implementer subagent, 09:27:33Z) and 35d7c0fa (implementer subagent, 09:36:16Z): both ran select_step9c_tests.py --json 2>&1 | uv run python -c "json.load(sys.stdin)"; the stderr NOTE lines landed ahead of the JSON and both crashed — "json.decoder.JSONDecodeError: Expecting value: line 1 column 1 (char 0)". Both agents then independently hand-wrote the same workaround: "# grab the json line\nlast_json = None\nfor line in raw.splitlines(): …". Counting: 2 tool_result FIRING events (is_error=true results carrying json.decoder.JSONDecodeError, deduped per tool call, one per session's implementer). Cost: 2 wasted turns + 2 bespoke ad-hoc parsers in two independent sessions on the same day.
Defect (c). Session c0a2df1b, 2026-07-26T06:35:26Z: passed the touched-file list to --map-files as a comma-joined string; the error surfaced only the raw Errno 2 on the whole comma-blob — "select_step9c_tests: cannot read --map-files input: [Errno 2] No such file or directory: '.claude/agents/experiment-implementer.md,.claude/agents/implementer.md, .claude/rules/crash-fix-rounds.md,scripts/workflow_lint.py,tests/test_issue_skill_marker_contract.py'". The session then wrote the list to /tmp/m1682.txt and re-ran. Cost: 1 wasted probe (~10 s).
Defect (c), second shape. Session a2c4bae3, 2026-07-26T17:18:23Z: passed two bare paths instead of a list file → "select_step9c_tests.py: error: unrecognized arguments: tests/test_autonomous_session_watch.py | === rc=0 ===". The wrapper printed === rc=$? === after a … 2>&1 | tail -20 pipeline, so $? captured tail's status and the turn reported rc=0 on a failed command. Cost: one wasted turn; the masked rc=0 could have been read as success.
Defect (d). Sessions 0e2c3b21 (implementer, 09:27:38Z) and 35d7c0fa (implementer, 09:36:21Z): both ran the selector on uncommitted edits and got "select_step9c_tests: NOTE — empty diff vs 'origin/main' in …/worktrees/issue-1694; falling back to the workflow-invariant set only. If this task's changes live in an issue worktree, re-run from that worktree" — reasons_summary: {'invariant': 65}, i.e. the pin-sweep universe was NOT the gate's universe. Counting: 2 tool_result FIRING events of the empty-diff NOTE, one per session's implementer. The NOTE's own hint blames the wrong cause (wrong worktree) rather than the actual one (uncommitted edits). Cost: the pin-sweep ran against a possibly-wrong file set in both sessions; the Step 9c gate caught nothing this time, so no downstream damage — but the #1288 narrowing was not obtained.

Proposed change

(a) In the --map-files branch (scripts/select_step9c_tests.py L1644), fail loud when --json is also passed: parser.error("--json is not supported with --map-files (mapping mode emits TSV: '<test>\\t<matched_path>' per line)"). Emitting JSON from mapping mode is the alternative, but the consumers are TSV-shaped today (implementer.md L174 reads col-1; the Step 10d TG legs sort -u the stdout) — fail-loud is the smaller, safer change and matches the "gate must fail CLOSED when it cannot classify" comment already at L1652-1657.
(b) Add an explicit stderr warning line printed alongside the JSON path (or in --help) stating: with --json, never redirect stderr into stdout — the NOTE / WARN / sizing lines are stderr BY DESIGN and will corrupt the JSON. Optionally add a --json-only / --quiet mode that suppresses the informational stderr lines for machine consumers. Then correct the copy-paste recipe in .claude/agents/implementer.md L174 to the safe invocation: uv run python scripts/select_step9c_tests.py --json 2>/dev/null | ….
(c) In the --map-files OSError handler (L1660-1665), when the argument is unreadable AND contains a comma, append (--map-files takes a PATH to a newline-separated file list, not a comma-separated list) to the error. Independently, the argparse unrecognized arguments shape (bare paths instead of a list file) is worth a usage hint: either accept nargs="+" of paths, or keep the single-FILE contract and add the list-file hint to the flag's error path.
(d) Extend the empty-diff NOTE at L1747-1751 to name uncommitted work as the FIRST likely cause: "empty diff vs '<base>' in <root> — the selector diffs COMMITTED state against fetched origin/main, so uncommitted edits produce an empty diff. Commit first; if this task's changes live in an issue worktree, re-run from that worktree." And in .claude/agents/implementer.md § gate-matched scope (L174), insert: Commit your edits first — the selector diffs committed state against fetched origin/main; run it on uncommitted edits and it silently degrades to the invariant-only set.
(e) Related, from the same session's evidence: .claude/rules/code-style.md (or the CLAUDE.md piped-exit-code bullet, which today covers only git push) should carry the general rule that $? after a pipe is the LAST stage's status — use set -o pipefail or check ${PIPESTATUS[0]} before echoing an rc. Optional here; name it so the planner can decide whether to fold it in or leave it for its own filing.

Scope / surfaces

Primary target: scripts/select_step9c_tests.py
.claude/agents/implementer.md (commit-first line + the safe 2>/dev/null invocation in the L174 recipe)
tests/test_select_step9c_tests.py (pins for the new parser.error on --map-files --json, the widened empty-diff NOTE text, and the comma-hint error)
.claude/rules/code-style.md (optional, per (e) — the post-pipe $? rule)

Constraints / invariants

Workflow-surface only — never experiment code, configs/, or tasks/.
uv run python scripts/workflow_lint.py passes (no-flags); ruff clean on touched files.
scripts/select_step9c_tests.py is in tests/test_ruff_policy.py's live-workflow-helper scope class — run uv run pytest tests/test_ruff_policy.py::test_live_workflow_helpers_clean_under_full_ruleset -x in addition to bare ruff check (verify the roster membership at plan time).
Do NOT change the mapping mode's stdout SHAPE or its exit-code contract: empty stdout + exit 0 is the Step 10d gate's skip signal; exit 1 (unreadable input) and exit 2 (#1613 zero-resolution) are load-bearing fail-closed codes.
Do NOT move the informational lines from stderr to stdout — the TSV/JSON stdout purity is what the consumers depend on; the fix is a warning plus a corrected recipe, not a stream swap.
This session runs under a workflow_fix_target: Provenance line — it MUST NOT auto-route its own subagents' workflow-fix candidates (recursion guard).

Provenance

fingerprint: 7e501bf395c5
workflow_fix_target: scripts/select_step9c_tests.py
fingerprint: PENDING

/daily 2026-07-26 route-2 filing. Miner refs: G-P8, H-P4, C-P12, B-P12, H-P5.