Workflow improvements

kind: infra

Right now are the agents able to use SSH MCP when they create a new pod? Have concrete hypothesis, falsifiable claims before running experiment --> integrate into workflow, search web for best way Does the manager give full plan to the sub agent? USE LESS VERY CONFUSING ACRONYMS (e.g. P1, H1 -- be clearer about everything in the github issues/clean results) The Clean Result should be clear enough that even someone with very little context on the project can understand it Link to plan in chat when asking for approval from user For the daily skill --> we first need to review clean results and then decide next steps, and this should be summarized in the daily gist (what was done today, blockers, next steps) -- daily gist needs to be approved by user Add useful vs non useful results column The TLDRs refer to H1 and stuff without ever defining them. Each Clean Result claim should stand on its own without referring to past results. In the background section we should refer to past results to say why we ran this Can we add a step (or augment a step) to look at the raw results/ text generations and make sure nothing fishy is going on Also add to always include some examples from each dataset used or generated in an experiment to the TLDR, with a link to the wandb with the full data The issue should be moved to "Planning" as soon as the /issue command gets run

This bug should not happen: ● Bash(uv run python scripts/gh_project.py set-status 260 "In Progress") ⎿ Error: Exit code 1 unknown column 'In Progress'. valid: Archived, Awaiting promotion, Blocked, Clean results, Done, Followups running, In flight, Plan awaiting review, Planning, To do

Called happy (ctrl+o to expand)

● Project board uses more granular column names than the skill doc. Mapping to "Planning" for the planning phase.

Timeline · 13 events

epm:clarify2026-05-06T01:22:04.000Z· system

## Clarifier — blocking ambiguities (v1) Issue body lists ~15 distinct workflow items. After co…

<!-- epm:clarify v1 -->
## Clarifier — blocking ambiguities (v1)

Issue body lists ~15 distinct workflow items. After context-gathering, several are clear and several need scoping decisions before drafting a plan.

**Context resolved from project knowledge:**
- Items 14 + 15 are the same bug: `.claude/skills/issue/SKILL.md` lines 46/53/58/283/285 reference column name `"In Progress"`, but the actual board (per `gh_project.py list-options Status`) has `Planning / Plan awaiting review / In flight / Awaiting promotion / Followups running / Clean results / Done / Blocked / To do / Archived`. `LABEL_TO_COLUMN` in `gh_project.py` already maps `status:planning → Planning` and `status:approved/implementing/...→ In flight`. Fix is mechanical.
- Items 4 + 9 are the same request: `H1`, `H2`, `P1`, `P2` acronyms appear in published clean-results (e.g. #181, #102) and TL;DRs without definitions. Concrete fix in `.claude/skills/clean-results/template.md` + `principles.md`.
- Item 5 (clean-result accessibility) and item 11 (background should cite past results to motivate) are already partially in the template (Background subsection minimum 30 words, citing prior issue numbers) — the question is whether to strengthen these constraints.
- Item 12 (sanity-check raw text generations) — analyzer agent already pulls every number from raw JSON (analyzer.md L38) and cites Sample outputs section in template. Whether to add an explicit "fishiness scan" step is a real new ask.
- Item 13 (dataset examples in TL;DR with WandB link) — template currently has `## Sample outputs` section with ≥3 triplets per condition, but the request is for examples in the TL;DR (not just the detailed report).

**Blocking ambiguities (need answers before adversarial planning):**

1. **Scope.** This bag covers (a) one bug fix, (b) clean-result template/principles edits, (c) /issue skill flow tweaks, (d) /daily skill changes, (e) open investigative questions. Do you want this all bundled into one issue/PR, or split into several focused issues so each adversarial-planning round stays tight? My recommendation: split into 3 issues — (i) Bug+Template Polish, (ii) Workflow flow tweaks (link to plan in chat, raw-output sanity step, daily skill review-then-decide), (iii) Open investigations (items 1, 2, 3, 8). But your call.

2. **Item 1 — "Are agents able to use SSH MCP when they create a new pod?"** Pod creation goes through `pod.py provision` (local script) which calls bootstrap_pod.sh over SSH; downstream the experimenter agent uses `ssh_execute(server="epm-issue-N", ...)` (per agents/experimenter.md L56-116) — so agents DO use SSH MCP after the pod exists. Is the request: (a) verify this is actually happening / audit a recent pod creation; OR (b) extend SSH MCP to the provisioning step itself; OR (c) something else?

3. **Item 2 — "Have concrete hypothesis, falsifiable claims … search web for best way."** The clarifier already enforces hypothesis + quantitative prediction + kill criteria for `type:experiment` (clarifier.md L130-134) and the plan template requires a falsifiable success criterion. What additional check do you want?  (a) An adversarial-planner step that scores the hypothesis quality before user approval; (b) a research-best-practices doc summarizing pre-registration / falsifiability literature integrated into the clarifier; (c) both.

4. **Item 7 — "Daily skill: review clean results, decide next steps, gist needs approval."** Concretely, do you want: (a) the daily skill to prompt the user via `AskUserQuestion` BEFORE publishing each gist (review of the gist body); OR (b) a separate "review clean results" sub-task that lists open `clean-results:draft` issues + asks the user which to promote/reject, then summarizes that into the gist; OR (c) both?

5. **Item 8 — "Add useful vs non-useful results column."** Where? (a) New project-board column ("Useful" vs "Non-useful" in addition to "Clean results"); (b) label on clean-result issues (e.g. `usefulness:high|medium|low`

epm:clarify-answers2026-05-06T01:33:17.000Z· system

## Clarifier — user answers (v1) 1. **Scope:** **Bundle everything into one big issue.*…

<!-- epm:clarify-answers v1 -->
## Clarifier — user answers (v1)

1. **Scope:** **Bundle everything into one big issue.** All 15 items go in this issue's plan + PR.
2. **Item 7 (daily skill):** **Both** — add a "review clean-results draft" sub-task AND prompt user to approve the gist before publishing.
3. **Item 8 (useful column):** **New project-board column** ("Useful" / "Not useful").
4. **Items 1, 2, 3:** **Investigate only** — audit current behavior, post findings as a comment; file follow-up issues for any concrete fixes that emerge.

Proceeding to adversarial planning. Plan will cover: (a) bug fix for `In Progress` → `Planning`/`In flight` column-name divergence, (b) clean-result template + principles edits (acronym definitions, accessibility, dataset examples in TL;DR, claim self-containment, background motivation), (c) /issue skill flow (link plan in chat at approval gate; raw-output sanity-check step), (d) /daily skill (clean-results review sub-task + gist approval), (e) new "Useful"/"Not useful" project-board columns, (f) audit report for items 1/2/3.
<!-- /epm:clarify-answers -->

epm:plan2026-05-06T02:03:39.000Z· system

## Plan v2 — bundled workflow improvements (post 3-critic review) **Cost gate:** ~5.8 hours implem…

<!-- epm:plan v1 -->
## Plan v2 — bundled workflow improvements (post 3-critic review)

**Cost gate:** ~5.8 hours implementer time. **No GPU, no pod.** All changes are local edits + 3 GitHub API migration commands. Reply `approve` to dispatch.

**Plan source of truth:** the canonical, line-and-diff-level plan is cached at `.claude/plans/issue-275.md` (1438 lines, exceeds GitHub comment limit). This comment is the decision-grade summary.

**Adversarial review:** 3 critics (design / verification / alternatives) reviewed v1 → all returned REVISE → v2 incorporates every consolidated must-fix → 3 critics re-reviewed v2 → **all returned PASS**.

---

### 1. Goal

Bundle 15 workflow-improvement items from #275 into one PR. Items split into 6 design groups (A bug fix, B clean-result template, C /issue skill, D /daily skill, E project-board column, F audit comment for items 1/2/3 which are investigate-only per user direction). No experiment, no scientific claim.

### 2. Acceptance criteria (every row must be observably true)

| # | Item from #275 | Observable check |
|---|---|---|
| 1 | SSH MCP for pod creation (audit) | `<!-- epm:audit v1 -->` posted on #275 with fresh-read evidence; ≥1 follow-up issue if gap, else "no gap" stated. |
| 2 | Falsifiability enforcement (audit) | Same audit comment covers it; concrete gap list. |
| 3 | Manager→subagent plan handoff (audit) | Same comment; states whether full plan body or path is passed. |
| 4 | Less acronyms | `verify_clean_result.py` rejects undefined `H1/H2/H3/P1/P2/P3` in TL;DR; passes if defined inline. Code blocks + inline backticks exempt. |
| 5 | Outsider-friendly clean results | New `check_background_motivation` requires ≥1 `#<issue>` ref distinct from current issue. |
| 6 | Plan link in approval prompt | Step 2c captures `gh issue comment` URL via inline shell variable in the same Bash block (no cache file). |
| 7 | /daily review + gist approval | New `Review clean-result drafts` sequential subagent runs FIRST; single end-of-run AskUserQuestion gate stacks all gist previews (`keep_all` / `retract` / `edit_and_repost`). `INTERACTIVE=false` skips both. |
| 8 | Useful vs Not useful column | `NEW_COLUMN_SPEC` gets 2 new entries (`Useful` BLUE, `Not useful` GRAY) inserted BETWEEN `Clean results` and `Done`. **NO labels, NO routing edits.** |
| 9 | Defined H1 acronyms | Same validator as item 4. |
| 10 | Each claim stands alone | New `check_results_self_contained` flags `### Results` paragraphs whose only quantitative anchor is a `#<issue>` cross-reference. |
| 11 | Background cites prior work | `check_background_motivation` (#5). `no-dataset` label is the soft-skip; `--skip-checks` is the per-invocation escape hatch. |
| 12 | Raw-output sanity check | `analyzer.md` Step 1.5 augmented: pastes `### Raw-output spot check (5 random rows)` H3 at TOP of `epm:interpretation` body. **No new marker, no `block` verdict** (interpretation-critic catches absence naturally). |
| 13 | Dataset examples in TL;DR | New `check_tldr_dataset_example` requires ≥1 fenced/backticked example AND ≥1 link of {`wandb.ai/...`, `wandb://...`, `huggingface.co/<owner>/<repo>/...`}. Soft-skipped via `no-dataset` label. Literal `**Dataset example:** N/A` rejected. |
| 14 | Issue moves to Planning at `/issue` start | Already correct in `LABEL_TO_COLUMN`. SKILL.md doc fix replaces every `In Progress` literal with the live column name. |
| 15 | `set-status … "In Progress"` bug | Same fix as #14 + regression test (`test_skill_set_status_calls.py` globs every `set-status … "<col>"` literal in skill docs and asserts `<col>` ∈ `NEW_COLUMN_SPEC`). |

PR ships when: (a) every row above observable, (b) `pytest -q` green (≥9 new tests + regression test), (c) `ruff check + format --check` pass, (d) audit comment posted on #275, (e) migration commands run on live board (snapshot first), (f) /daily dry-run smoke passes (`INTERACTIVE=false`), (g) placeholder-text guard returns no matches before audit posts.

### 3. Design — file-and

epm:code-review2026-05-06T08:13:23.000Z· system

# Audit (items 1, 2, 3) This comment closes the audit-only sub-asks of #275. The user asked for investigate-only on the…

# Audit (items 1, 2, 3)

This comment closes the audit-only sub-asks of #275. The user asked for
investigate-only on these three boundaries; behaviour does NOT change in
this PR. Concrete findings + follow-up issues filed below.

Reading order: each item ends with a "Verdict" line so a skim of this
comment leaves the reader knowing whether a follow-up exists.

---

## Item 1 — Are agents able to use SSH MCP when they create a new pod?

**Files read:** `scripts/pod.py`, `scripts/pod_lifecycle.py`,
`scripts/pod_config.py`, `scripts/runpod_api.py`, `scripts/bootstrap_pod.sh`.

**Lifecycle of pod creation, with SSH-MCP availability annotated:**

1. `pod.py provision --issue N` (`scripts/pod.py:121`) dispatches to
`pod_lifecycle.cmd_provision` (`scripts/pod_lifecycle.py:189`).
2. `cmd_provision` calls `create_pod(...)` from
`scripts/runpod_api.py:193`, which submits the GraphQL mutation.
Pod begins booting on RunPod side. **SSH MCP unusable** — the new
pod isn't in `~/.claude/mcp.json` yet.
3. `wait_for_ssh(pod_id, timeout=600)` polls until raw SSH responds
(`scripts/pod_lifecycle.py:228`). **SSH MCP still unusable** — same
reason.
4. `_upsert_pods_conf(pod)` (`scripts/pod_lifecycle.py:246`) appends
the row to `scripts/pods.conf` and calls `cmd_sync(pods)` from
`scripts/pod_config.py:450`, which regenerates `~/.ssh/config` AND
writes the new pod's host/port into the SSH MCP server's `env`
block in `~/.claude/mcp.json` (`scripts/pod_config.py:275-340`).
**SSH MCP STILL unusable** for the current Claude Code session —
the file is written, but Claude Code's MCP runtime caches the
config at session start. The user must `/mcp` to reload before the
new pod's SSH tools resolve.
5. `_bootstrap(name)` runs `scripts/bootstrap_pod.sh` over **raw
SSH** (`scripts/bootstrap_pod.sh:45-47` defines `ssh_cmd()` using
`ssh -p $PORT root@$HOST`). The bootstrap script does NOT use any
`mcp__ssh__ssh_execute` calls — it can't, because it's a Bash
script invoked from `subprocess.run`.

**The gap.** Between provision-success and `/mcp` reload, agents
working in the same Claude Code session cannot use SSH MCP for the
freshly-provisioned pod. This is a Claude Code runtime constraint, not
a project-side bug — the project-side code DOES write to the right
file. The user-visible message at `pod_config.py:468` (`If MCP config
changed, restart the MCP server (/mcp).`) acknowledges this.

**Practical impact for `/issue` flow:** Step 6b provisions the pod and
EXITS. The /issue skill exits before the experimenter spawns, so by
the time the experimenter starts in a NEW Claude Code session, MCP is
loaded fresh and the new pod is reachable via `mcp__ssh__ssh_execute`.
This is a non-issue in production. The gap only bites when an agent
provisions a pod and then tries to immediately SSH into it without an
intervening session boundary.

**Verdict.** No code change in this PR (clarifier answer #4 said
investigate-only). Filed follow-up: **#287** — `[Investigate] Add a
post-provision MCP-reload nudge to pod.py provision`. The follow-up
proposes either (a) a `/mcp` reminder line emitted at the end of
`provision` output, or (b) a one-shot `pod.py mcp-reload` command that
invokes Claude Code's reload via the `claude` CLI if available.

---

## Item 2 — Does the workflow REQUIRE a falsifiable hypothesis with a kill criterion for `type:experiment` issues?

**Files read:** `.claude/skills/issue/clarifier.md` (the
`type:experiment` section, lines 126-160), `.claude/skills/adversarial-
planner/SKILL.md` (planner phase, lines 38-60).

**Clarifier (`clarifier.md:130-134`).**

```
### Hypothesis + prediction
- What specific hypothesis does this test? State it as an `if X then Y`.
- What is the quantitative prediction? (e.g., "EM coupling drops by ≥30% under
intervention A vs. baseline")
- What result would FALSIFY the hypothesis? (kill criterion)
```

The clarifier ASKS these as questions for the user to answer in the
issue body

epm:results2026-05-06T08:19:10.000Z· system

## Completion Report **Task:** #275 workflow improvements — bundled type:infra changes coverin…

<!-- epm:results v1 -->

## Completion Report

**Task:** #275 workflow improvements — bundled type:infra changes covering 15 user-listed asks across 6 design groups (A-F).
**Status:** SUCCESS

### Summary by group

| Group | Items | Status |
|---|---|---|
| **A** — column-name divergence fix + regression test | 14, 15 | DONE |
| **B** — clean-result template/principles/checklist + verifier upgrades + 17 new tests | 4, 5, 9, 10, 11, 13 | DONE |
| **C** — /issue plan-link UX + analyzer raw-output spot check H3 | 6, 12 | DONE |
| **D** — /daily review-clean-result-drafts subagent + single end-of-run gate | 7 | DONE |
| **E** — Useful / Not useful columns (column-only, no labels) + live migration | 8 | DONE |
| **F** — audit comment for items 1, 2, 3 + 3 follow-up issues filed | 1, 2, 3 | DONE |

### Files modified / created

| Path | Change | LOC delta |
|---|---|---|
| `.claude/skills/issue/SKILL.md` | Group A column-name table rewrite + Step 1/Step 10 set-status fixes; Group C Step 2a/2c PLAN_URL capture | +87 / -38 |
| `.claude/skills/clean-results/template.md` | Author-note admonition, Background placeholder rewrite, `Dataset example`/`Full data` bullets, claim-self-containment sentence | +30 / -7 |
| `.claude/skills/clean-results/principles.md` | New `## #275 additions` section with 4 operational principles | +34 / 0 |
| `.claude/skills/clean-results/checklist.md` | 4 new items (Background ref, no bare H1-P3, self-contained takeaways, dataset example + full-data link) | +4 / 0 |
| `.claude/skills/daily/SKILL.md` | New `Review clean-result drafts` sequential subagent (D.3), Step 4.5 single end-of-run gate (D.2), Daily-summary sourcing of `/tmp/daily-clean-result-review.md`, doc heading disambiguation | +152 / -3 |
| `.claude/skills/weekly/SKILL.md` | Doc heading disambiguation (`Running / In Progress` → `Running experiments`) | +1 / -1 |
| `.claude/agents/analyzer.md` | Step 1.5 raw-output spot-check H3 requirement | +49 / 0 |
| `scripts/verify_clean_result.py` | Three new check functions (`check_undefined_acronyms`, `check_background_motivation`, `check_tldr_dataset_example`), `--skip-checks <names>` CLI flag, `--current-issue <N>` flag, `_parse_current_issue_from_body` helper, `run_all_checks` rewrite to dispatch via `_maybe(name, fn)` | +311 / -36 |
| `scripts/gh_project.py` | `NEW_COLUMN_SPEC`: insert `Useful` (BLUE) + `Not useful` (GRAY); hotfix `cmd_add_status_option` + `cmd_remove_status_option` to use `_replace_options` (typed JSON) instead of `-f options=<string>` (silently broken); add `--description` flag | +25 / -47 |
| `tests/test_verify_clean_result.py` | 17 new tests + GOOD_BODY fixture updated with TL;DR Methodology bullets | +261 / 0 |
| `tests/test_skill_set_status_calls.py` | NEW FILE — regression test (`set-status` literal sweep) | +39 / 0 |
| `.claude/cache/issue-275-audit.md` | NEW (gitignored) — audit comment body | +189 / 0 |

Diff stat: **989 insertions, 106 deletions across 11 files** (audit cache file excluded since it is gitignored).

### Commit list

| SHA | Message |
|---|---|
| `1814b6e` | [#275 group A] Fix In Progress→Planning column-name divergence |
| `b39c97b` | [#275 groups B+E] Clean-result verifier upgrades + Useful/Not useful columns |
| `b9d27c9` | [#275 group C] /issue plan-link UX + analyzer raw-output spot check |
| `73d9824` | [#275 group D] /daily: review clean-result drafts + end-of-run gate |
| `a05d445` | [#275 group E hotfix] Fix add/remove-status-option silent breakage |
| `1e428e7` | [#275 group A follow-up] Disambiguate doc subheadings from board-column literal |

**PR:** https://github.com/superkaiba/explore-persona-space/pull/286 (draft)
**Branch:** `issue-275` (pushed)

### All 12 plan §6 smoke commands

#### 1. Lint + format

```
$ uv run ruff check scripts/verify_clean_result.py tests/test_verify_clean_result.py tests/test_skill_set_status_calls.py
All checks passed!
$ uv run ruff format --check scripts/verify_clean_result.py scripts/gh_project.py tests/test_verify_cle

epm:code-review2026-05-06T08:26:40.000Z· system

## Code-Reviewer Verdict — FAIL **Diff size:** +989 / −106 across 11 files **Plan adherence…

<!-- epm:code-review v1 -->
## Code-Reviewer Verdict — FAIL

**Diff size:** +989 / −106 across 11 files
**Plan adherence:** COMPLETE (every group A-F item shipped)
**Tests:** FAIL — 2 pre-existing tests in `tests/test_gh_project.py` regress on the worktree
**Lint:** PASS for the new files; 5 pre-existing ruff errors in `scripts/gh_project.py` carry forward (verified pre-existing on `origin/main`)
**Security sweep:** CLEAN

### Plan adherence
- (A) bug fix: ✓ — `In Progress` literals removed from skill docs (`grep -rnE '"In Progress"|In Progress' .claude/skills/` returns clean), regression test added.
- (B) verifier upgrades: ✓ — 3 new validators with grandfather guards, `--skip-checks` flag with stderr log, code-block strip, self-ref filter, `no-dataset` label skip, literal `N/A` rejection. 17 new tests (well above the ≥9 floor).
- (C) /issue plan-link + analyzer spot check: ✓ — Step 2a captures `PLAN_URL=$(gh issue comment ... | tail -1)` inline (`SKILL.md:362`); Step 2c interpolates `${PLAN_URL}` (`SKILL.md:420`); analyzer Step 1.5 augmented with the `### Raw-output spot check (5 random rows)` H3 requirement (`analyzer.md:58-105`). No new marker, no `block` verdict — as required.
- (D) /daily: ✓ — sequential `Review clean-result drafts` runs FIRST (`daily/SKILL.md:74`), parallel fan-out preserved, single end-of-run gate at Step 4.5 stacks all gist previews with `keep_all` / `retract` / `edit_and_repost`. `INTERACTIVE=false` skips both.
- (E) Useful / Not useful columns: ✓ — inserted BETWEEN `Clean results` and `Done` in `NEW_COLUMN_SPEC` (positions 8-9). NO `LABEL_TO_COLUMN` / `PRIORITY_LABELS` edits.
- (F) audit: ✓ — fresh-eye evidence with concrete `file:line` cites, follow-ups #287/#288/#289 filed, no `<!-- epm:audit -->` marker (NIT 1 honoured).

### Issues

#### BLOCKER — `tests/test_gh_project.py` regresses

```
tests/test_gh_project.py::test_add_status_option_preserves_existing_colors  FAILED
tests/test_gh_project.py::test_remove_status_option_preserves_surviving_colors  FAILED
```

Verified by checking out `origin/main:scripts/gh_project.py` against the worktree's tests — the same tests **PASS** on main. The regressing commit is `a05d445` (group E hotfix), which made two changes that the existing test fixtures don't cover:

1. **`scripts/gh_project.py:390`** — `cmd_add_status_option` now reads `args.description or ""` unconditionally:
   ```python
   "description": args.description or "",
   ```
   The existing regression test at `tests/test_gh_project.py:162-167` constructs a `Namespace(owner=..., project=..., option=..., color=...)` with NO `description` attribute → raises `AttributeError`.

   Fix: `getattr(args, "description", "") or ""`, OR set `description=""` on the test fixture. The latter is cleaner since the CLI default is `""` (`gh_project.py:1029`), so the test was simulating an out-of-band caller.

2. **`scripts/gh_project.py:400` and `:427`** — both `cmd_add_status_option` and `cmd_remove_status_option` now route through `_replace_options` → `_graphql` → `subprocess.run(["gh", "api", "graphql", "--input", path], ...)` (line 530). The existing tests `monkeypatch.setattr(gh_project, "_gh", rec)` only intercept the `_gh` helper; `_graphql` calls `subprocess.run` **directly**, bypassing the monkeypatch. The test then actually shells out to live `gh api`, hits an unknown field id (`PVTSSF_test_status_field_id`), and aborts with `SystemExit: 1`.

   Fix options (any one):
   - Refactor `_graphql` to call `_gh` instead of `subprocess.run` (single-source the gh shell-out path).
   - Update the failing tests to also `monkeypatch` `_graphql` (or `subprocess.run`) — the cleaner choice if `_graphql` deliberately stays separate.
   - Add `getattr` to skip `--input` for the test path.

   The first option is preferable: single source of truth for "shell out to gh" is exactly the kind of invariant a reviewer should enforce, and it preserves the test contract.

   Note: the implementer's commit message claims `cmd_add_

epm:results2026-05-06T08:42:01.000Z· system

## Round 2 Completion Report **Task:** Fix the gh_project test regression introduced by round-…

<!-- epm:results v2 -->

## Round 2 Completion Report

**Task:** Fix the gh_project test regression introduced by round-1's group-E hotfix, and address the 3 code-review NITs.
**Status:** SUCCESS

### Round-2 changes (≤5-bullet summary)

- **BLOCKER** — Refactored `_graphql` in `scripts/gh_project.py` to delegate the actual subprocess call to `_gh`, so `monkeypatch.setattr(gh_project, "_gh", ...)` still intercepts the mutation path while keeping the `--input <tempfile>` form (required for typed JSON arrays).
- **BLOCKER** — Hardened `cmd_add_status_option` against synthetic Namespace fixtures missing the `description` attribute (`getattr(args, "description", "")`).
- **BLOCKER** — Updated `_GhRecorder` in `tests/test_gh_project.py` to read the `--input <tempfile>` body during the call (the file is deleted in `_graphql`'s `finally`); helpers `_mutation_payload` / `_mutation_call` now operate on parsed bodies. Test bodies and assertions unchanged.
- **NIT-1** — `--skip-checks` now validates each token against the registered check names. Unknown tokens trigger `parser.error(...)` instead of silently passing.
- **NIT-2 / NIT-3** — Documented the extended F2 placeholder regex in `.claude/skills/issue/SKILL.md` (Step 0b audit-comment guard) and promoted the `INTERACTIVE` short-circuit string to a documented top-level variable in `.claude/skills/daily/SKILL.md`.

### Files modified vs round 1

| File | LOC delta (round 2) |
|---|---|
| `scripts/gh_project.py` | +14 / -12 (refactor `_graphql`; `getattr` defaulting on `description`) |
| `tests/test_gh_project.py` | +56 / -15 (new `_read_input_body` helper; recorder now stashes parsed body; `_mutation_payload` / `_mutation_call` operate on body) |
| `scripts/verify_clean_result.py` | +34 / -0 (`KNOWN_CHECKS` frozenset + `parser.error` validation) |
| `.claude/skills/issue/SKILL.md` | +8 / -0 (audit-comment placeholder guard doc) |
| `.claude/skills/daily/SKILL.md` | +10 / -0 (Top-level variables section documenting `INTERACTIVE`) |

`git diff --stat`:
```
 .claude/skills/daily/SKILL.md  | 10 ++++++
 .claude/skills/issue/SKILL.md  |  8 +++++
 scripts/gh_project.py          | 26 +++++++++-------
 scripts/verify_clean_result.py | 34 ++++++++++++++++++++
 tests/test_gh_project.py       | 71 ++++++++++++++++++++++++++++++++----------
 5 files changed, 121 insertions(+), 28 deletions(-)
```

### New commit

```
5229b24 [#275 round 2] Fix gh_project.py test regression + code-review NITs
```

### `pytest tests/ -q` output

Full suite (excluding two pre-existing unrelated failures, both verified on `origin/main`):

```
162 passed, 4 skipped, 20 deselected, 2 warnings in 19.37s
```

Pre-existing failures excluded:
- `tests/test_data_validation.py` — module collection error (`ModuleNotFoundError: No module named 'explore_persona_space.data'`).
- `tests/test_hub.py::TestUploadDataset::test_upload_file` — flaky / depends on real HF API behaviour.

Both reproduce on `origin/main` (verified via `git stash` round-trip).

The targeted regression-test suite for this PR:

```
$ uv run pytest tests/test_gh_project.py tests/test_verify_clean_result.py tests/test_skill_set_status_calls.py -q
56 passed in 0.19s
```

Specifically the two previously-failing tests (now PASS):
- `tests/test_gh_project.py::test_add_status_option_preserves_existing_colors` — PASS
- `tests/test_gh_project.py::test_remove_status_option_preserves_surviving_colors` — PASS

### Lint output

```
$ uv run ruff check scripts/verify_clean_result.py tests/test_gh_project.py
All checks passed!
```

`scripts/gh_project.py` carries forward 5 pre-existing ruff errors from `origin/main` (E501×2, E741, C901, SIM108) — verified pre-existing in round 1; my round-2 hunks introduce no new errors. `ruff format --check` PASS for all 5 touched files (the format pass is verified for the touched files specifically; the rest of the repo has 11 pre-existing format-only divergences unrelated to this PR).

### Confirmation: BLOCKER + 3 NITs addressed

| Item |

epm:code-review2026-05-06T08:46:39.000Z· system

## Code-Reviewer Verdict (round 2) — PASS Narrowly scoped to round-2 commit `5229b24` (NOT …

<!-- epm:code-review v2 -->
## Code-Reviewer Verdict (round 2) — PASS

Narrowly scoped to round-2 commit `5229b24` (NOT re-reviewing previously-approved plan-aligned commits).

### BLOCKER fix verified

**Cause #1 — `args.description` AttributeError on synthetic Namespace.**
`scripts/gh_project.py:391` now `getattr(args, "description", "") or ""`. Defensive accessor + comment cite the test fixture. ✓

**Cause #2 — `_graphql` bypassed `_gh` interception.**
`scripts/gh_project.py:540` now `raw = _gh(["api", "graphql", "--input", path])` (was `subprocess.run(...)`). The `--input <tempfile>` form is preserved (load-bearing for typed JSON variables — explained in docstring). ✓

**Test recorder updated correctly.** `tests/test_gh_project.py` adds `_read_input_body` that reads the tempfile during the `_gh` call (before `finally` deletes it). `_mutation_payload` / `_mutation_call` now operate on the parsed body — assertion semantics unchanged (verified `by_name["X"]["color"] == "..."` etc. all unchanged at lines 213-226 / 274-282).

### Test results

- `tests/test_gh_project.py` — **5 passed** (was 3 passed / 2 failed in round 1).
- Full suite (worktree, ignoring pre-existing `test_data_validation.py` collection error): **162 passed, 4 skipped, 1 failed**. The single failure is `test_hub.py::test_upload_file`, **also failing on `origin/main`** (verified). No new regressions.

### NIT 1: `--skip-checks` validates check names — ✓

`scripts/verify_clean_result.py:898-916` defines `KNOWN_CHECKS` frozenset (15 names). Confirmed all 15 match the `_maybe(...)` calls at lines 948-988. `check_tldr_structure` is intentionally non-skippable (documented at line 943-946 — downstream checks depend on its return value). Validator at `:1037` raises `parser.error(...)` listing both unknown and known names. Useful failure message.

### NIT 2: Placeholder regex documented — ✓

`.claude/skills/issue/SKILL.md:257-263` adds an "Audit-comment placeholder guard" block citing the extended regex `(^|\\s|>)(TBD|TODO|placeholder|\\[X\\]|implementer fills)(\\s|$|<)` and explaining why `^`-anchored form was insufficient.

### NIT 3: `INTERACTIVE` documented — ✓

`.claude/skills/daily/SKILL.md:23-31` adds a "Top-level variables" section explicitly defining `INTERACTIVE` as a boolean string set from `--autonomous` parsing, referenced literally in Steps 0/0.5/2.5.

### Scope check — ✓

`git diff 1e428e7..5229b24 --stat` touched only the 5 expected files. No scope creep.

### Lint

5 pre-existing ruff errors in `scripts/gh_project.py` (E501/E741/C901/SIM108) verified to also exist on `origin/main`. No new errors introduced. Round-2-changed files all `ruff format` clean.

### Findings

None.

---

**Verdict: PASS** — BLOCKER cleanly fixed (both causes addressed without regressing test assertions), all 3 NITs addressed, no scope creep, no new test failures, no new lint errors. Ready to merge.
<!-- /epm:code-review v2 -->

epm:test-verdict2026-05-06T08:50:12.000Z· system

## Test-verdict gate (Step 9c) — PASS `type:infra` issue, code-review (round 2) PASSed, no…

<!-- epm:test-verdict v1 -->
## Test-verdict gate (Step 9c) — PASS

`type:infra` issue, code-review (round 2) PASSed, no pod phase. Test-verdict run inline from the worktree at `5229b24`.

### Pytest

| Pool | Result |
|---|---|
| 17 new validator tests (`tests/test_verify_clean_result.py`) | **17/17 pass** |
| 1 new regression test (`tests/test_skill_set_status_calls.py`) | **1/1 pass** |
| Pre-existing `tests/test_gh_project.py` (round-1 BLOCKER target) | **5/5 pass** |
| Full worktree suite (excluding 2 pre-existing failures verified on `origin/main`) | **144/144 pass** |

### Pre-existing failures (NOT introduced by this PR)

Verified on a fresh `git clone -b main origin` (commit `189a247`, identical to the worktree's base):

1. **`tests/test_data_validation.py` collection error** — imports `explore_persona_space.data.wrong_answers_deterministic`, but `src/explore_persona_space/data/` does not exist on `origin/main`. The local main checkout has the directory as uncommitted/untracked work.
   - Source verified: `git show origin/main:src/explore_persona_space/data/__init__.py` → `fatal: path ... exists on disk, but not in 'origin/main'`.
   - This PR touches no `data/` files and no `test_data_validation.py`. Not a regression.

2. **`tests/test_hub.py::TestUploadDataset::test_upload_file`** — `AssertionError: assert 'test/data.jsonl' in ''` from a hub-upload assertion. Reproduced on `origin/main`.
   - This PR touches no `hub.py` or `test_hub.py`. Not a regression.

### Ruff

| Pool | Result |
|---|---|
| `ruff check` PR-touched files (`scripts/gh_project.py`, `scripts/verify_clean_result.py`, `tests/test_*.py`) | **5 pre-existing errors in `gh_project.py`** (`SIM108`, `B008` — same on `origin/main`); 0 new from PR |
| `ruff format --check` PR-touched files | **0 format regressions** in any PR-touched file |
| Pre-existing baseline (origin/main) | 595 errors + 2 unformatted scripts; same baseline carries forward |

Cross-check: `comm -12 <pr-files> <files-needing-format>` returns **empty set**.

### Conclusion

This PR introduces **zero new test or lint regressions**. All test-verdict criteria from `/issue` Step 9c §1-2 are satisfied for the diff under review:

- Unit tests for new code: pass.
- Lint for changed files: pass.
- Integration tests: not applicable (diff does not touch `train/`, `eval/`, `orchestrate/`).

Pre-existing failures and lint debt on `origin/main` are documented above and out-of-scope for #275.

**Verdict: PASS.**
<!-- /epm:test-verdict -->

epm:done2026-05-06T08:51:14.000Z· system

## Auto-complete (Step 10) — `status:done-impl` `type:infra` issue #275 has cleared every gate of …

<!-- epm:done v1 -->
## Auto-complete (Step 10) — `status:done-impl`

`type:infra` issue #275 has cleared every gate of the `/issue` pipeline.

### Pipeline summary

| Stage | Result |
|---|---|
| Clarifier | All blocking ambiguities resolved (1 round, in-chat answers persisted as `epm:clarify-answers v1`) |
| Adversarial planner | v1 → 3 critics → REVISE → v2 → 3 critics → **all PASS** |
| User approval | **Approved** |
| Implementer (round 1) | All 6 design groups (A-F) shipped per plan; audit comment + 3 follow-up issues filed |
| Code-review (round 1) | **FAIL** — 1 BLOCKER: `tests/test_gh_project.py` regression from group-E hotfix |
| Implementer (round 2) | BLOCKER fixed in commit `5229b24` (refactor `_graphql` to delegate to `_gh`); 3 NITs addressed |
| Code-review (round 2) | **PASS** — no findings |
| Test-verdict gate | **PASS** — 144/144 (excluding 2 pre-existing failures verified on `origin/main`); 0 lint regressions on PR-touched files |

### What shipped (8 production files + 2 new test files + 1 audit cache)

**(A) Bug fix (items 14, 15):** `In Progress` → live column names in `.claude/skills/issue/SKILL.md` (and `weekly/SKILL.md` follow-up); regression test `tests/test_skill_set_status_calls.py` globs every `set-status … "X"` call site and asserts X ∈ `NEW_COLUMN_SPEC`.

**(B) Clean-result template + 3 new validators (items 4, 5, 9, 10, 11, 13):** `check_undefined_acronyms`, `check_background_motivation`, `check_tldr_dataset_example`. All date-gated. Acronym regex strips fenced + inline code before matching. New `--skip-checks` flag with `KNOWN_CHECKS` validation. 17 unit tests in `tests/test_verify_clean_result.py`.

**(C) /issue + analyzer (items 6, 12):** Step 2c captures `PLAN_URL=$(gh issue comment ... | tail -1)` inline (no cache file). `analyzer.md` Step 1.5 augments interpretations with a mandatory `### Raw-output spot check (5 random rows)` H3.

**(D) /daily skill (item 7):** New sequential `Review clean-result drafts` subagent runs FIRST; single end-of-run AskUserQuestion gate stacks all gist previews. `INTERACTIVE` documented as a top-level variable.

**(E) Project-board column (item 8):** `Useful` (BLUE) and `Not useful` (GRAY) columns inserted between `Clean results` and `Done` in `NEW_COLUMN_SPEC`. No labels, no routing edits. Migration done with snapshot-first; smoke-tested via `set-status 275 "Useful"` then revert.

**(F) Audit (items 1, 2, 3 — investigate-only):** Comment posted on this issue (4386196053) with fresh-read evidence; 3 follow-up issues filed:
- **#287** — SSH MCP usability boundary during pod creation
- **#288** — falsifiable hypothesis enforcement gaps in clarifier + adversarial-planner
- **#289** — manager→subagent plan handoff: full body vs path-only

### Artifacts

- **PR:** https://github.com/superkaiba/explore-persona-space/pull/286 (ready for review)
- **Branch:** `issue-275` (7 commits)
- **Worktree:** `.claude/worktrees/issue-275` (preserved; user decides cleanup)
- **Audit comment:** https://github.com/superkaiba/explore-persona-space/issues/275#issuecomment-4386196053
- **Test-verdict:** https://github.com/superkaiba/explore-persona-space/issues/275#issuecomment-4386433363

### Project board

Moved to **Done** on the Experiment Queue project board.

### What's next

Step 10d (worktree merge prompt) fires next — user decides whether to rebase-merge `issue-275` into `main` now or defer.

No `RESULTS.md` / `eval_results/INDEX.md` updates — `type:infra` is workflow plumbing, not headline science.
<!-- /epm:done -->

epm:merged2026-05-06T09:16:18.000Z· system

## PR #286 rebase-merged into `main` 7 commits landed individually on `main` (each independently…

<!-- epm:merged v1 -->
## PR #286 rebase-merged into `main`

7 commits landed individually on `main` (each independently revertible via `git revert <sha>`):

| New SHA on `main` | Commit |
|---|---|
| `1fef02c` | `[#275 round 2] Fix gh_project.py test regression + code-review NITs` |
| `7bc0686` | `[#275 group A follow-up] Disambiguate doc subheadings from board-column literal` |
| `afa6ef3` | `[#275 group E hotfix] Fix add/remove-status-option silent breakage` |
| `563b192` | `[#275 group D] /daily: review clean-result drafts + end-of-run gate` |
| `8e3e33d` | `[#275 group C] /issue plan-link UX + analyzer raw-output spot check` |
| `8444171` | `[#275 groups B+E] Clean-result verifier upgrades + Useful/Not useful columns` |
| `e07f47f` | `[#275 group A] Fix In Progress→Planning column-name divergence` |

`origin/main` advanced from `189a247` → `1fef02c`. Worktree `.claude/worktrees/issue-275` removed; branch `issue-275` preserved (per `--delete-branch=false`).

Issue stays OPEN (project board column = **Done**).
<!-- /epm:merged -->

epm:worktree-cleanup-deferred2026-05-06T09:16:51.000Z· system

## Worktree cleanup deferred `.claude/worktrees/issue-275` was NOT auto-remov…

<!-- epm:worktree-cleanup-deferred v1 -->
## Worktree cleanup deferred

`.claude/worktrees/issue-275` was NOT auto-removed because it contains uncommitted modifications to files OUTSIDE the PR scope:

- `.claude/agent-memory/{experimenter,retrospective,reviewer}/MEMORY.md` (sub-agent memory writes)
- `.claude/agents/{analyzer,planner}.md` (planner.md is unrelated to this PR)
- `.claude/skills/{adversarial-planner,paper-plots/*}` (skill docs unrelated to this PR)
- `CLAUDE.md`
- `figures/causal_proximity/*` (PDF/PNG/JSON regenerations)

These are likely auto-writes from sub-agents or metadata regeneration during the multi-round implementation, not code intended for #275. To clean up safely:

```bash
cd .claude/worktrees/issue-275
git status                # inspect what's there
git stash                 # if any of it is worth preserving
cd /home/thomasjiralerspong/explore-persona-space
git worktree remove --force .claude/worktrees/issue-275
```

The merge itself is complete — main is at `1fef02c` with all 7 commits from PR #286.
<!-- /epm:worktree-cleanup-deferred -->

state_changed2026-05-13T13:20:57.683Z· user· completed → archived
Moved on Pipeline board to archived.
```
Moved on Pipeline board to archived.
```

Comments · 0

No comments yet. (Auth + comment composer land in step 5.)