printing-press-output-review

name	printing-press-output-review
description	Internal sub-skill: agentic review of a printed CLI's sampled command output for plausibility issues that rule-based checks can't encode (substring-match relevance, format bugs, silent source drops, ranking failures). Invoked via the Skill tool by main printing-press SKILL.md (Phase 4.85) and printing-press-polish SKILL.md during the diagnostic loop. Not for direct user invocation — its actionable wrappers are /printing-press and /printing-press-polish.
context	fork
user-invocable	false
allowed-tools	["Bash","Agent"]
created_by	user

printing-press-output-review (internal)

Review the sampled outputs from a printed CLI for plausibility bugs that dogfood, verify, and the rule-based scorecard --live-check rules can't catch. Wave B policy: all findings surface as warnings, never errors.

This skill is internal-only (user-invocable: false). It's invoked by parents — main printing-press skill at shipcheck Phase 4.85, polish skill during its diagnostic loop. Running it standalone would produce floating findings text with no ship verdict, no fixes applied, no publish offer; the actionable wrappers are /printing-press and /printing-press-polish. The skill carries context: fork so the reviewer agent's diagnostic chatter stays isolated from the calling skill's context.

Input

The caller passes $CLI_DIR as the argument: an absolute path to the printed CLI's working directory.

What this catches

Bugs that rule-based checks miss, typically surfaced by 5 minutes of hands-on testing but slipping past dogfood, verify, and scorecard --live-check rules:

• Substring-match results that coincidentally contain the query but don't match semantically (e.g., a query matches a substring of a larger unrelated term)
• Aggregation commands silently dropping sources when only some of the requested N come back
• Ranking or sort commands returning top-N results that aren't plausibly the best for the query (broken weights, extractor fallbacks)
• URLs in output pointing at category index pages, feed endpoints, or random-selector routes rather than canonical content permalinks
• Format bugs the rule-based layer doesn't catch (mojibake, inconsistent pluralization, truncated/wrapped cell content)

Procedure

Step 1: Gather sample data

# Locate research.json. Adjacent to the binary covers the post-promote
# layout (standalone polish, shipcheck against the library copy). The
# grandparent fallback covers mid-pipeline invocations where $CLI_DIR is
# $PRESS_RUNSTATE/runs/<id>/working/<cli> and research.json lives at
# $PRESS_RUNSTATE/runs/<id>/research.json. Without the fallback, scorecard
# reports `unable: true` mid-pipeline and we SKIP the most informative review.
# Use a bash array so the flag survives paths with spaces.
RESEARCH_ARGS=()
if [ ! -f "$CLI_DIR/research.json" ]; then
  _grandparent="$(dirname "$(dirname "$CLI_DIR")")"
  if [ -f "$_grandparent/research.json" ]; then
    RESEARCH_ARGS=(--research-dir "$_grandparent")
  fi
fi

cli-printing-press scorecard --dir "$CLI_DIR" "${RESEARCH_ARGS[@]}" --live-check --json > /tmp/output-review-livecheck.json 2>&1 || true

If the scorecard call fails or /tmp/output-review-livecheck.json is empty, return the SKIP result (Step 3) without dispatching the reviewer.

Step 2: Dispatch the reviewer agent

Use the Agent tool (general-purpose) with this prompt contract:

Review the sampled outputs from the shipped CLI at $CLI_DIR. You have these ground-truth sources:

• Sampled command output: read /tmp/output-review-livecheck.json and inspect the live_check.features[] array. Each entry has the command, example invocation, redacted stdout evidence (in output_sample, bounded to ~4 KiB), the redacted pass/fail reason, and a warnings array (populated by rule-based checks like the raw-HTML-entity detector). Treat <redacted> markers as privacy scrubbed values, not format bugs.
• Review only status: pass entries. Entries with status: fail either crashed, timed out, or had placeholder args (<id>, <url>) that never produced real output — their sample is empty and there's nothing for you to judge. Phase 5 dogfood handles test-coverage and exit-code concerns.
• $CLI_DIR/research.json novel_features (planned behavior per feature) and novel_features_built (verified built commands).
• The CLI binary at $CLI_DIR/<cli-name>-pp-cli — you may invoke additional commands to gather more output when a finding needs verification.

For each of these checks, report findings under 50 words each. Only report issues a human user would notice in 5 minutes of hands-on testing — not every edge case a thorough QA pass might find:

1. Output semantically matches query intent. For sampled novel features with a query argument, judge relevance beyond what the mechanical query-token check in live-check already enforced. A feature that passed live-check's outputMentionsQuery test still contains some query token somewhere — but "buttermilk" appearing as a substring of "butter" results, or "brownies" returning a chili recipe because the extractor fell back to adjacent content, both slip past the mechanical check. Only flag when a human user would look at the top results and say "this isn't what I asked for." Skip this check when the example has no query argument.
2. No obvious format bugs. Does the output contain raw HTML entities, mojibake (question marks or replacement chars in titles), or malformed URLs (pointing at category index pages, feed endpoints, or random-selector routes rather than canonical content permalinks)? Rule-based live-check catches numeric entities; this layer catches the broader class.
3. Aggregation commands show all requested sources. For commands with a --source/--site/--region CSV flag: if the user requested N sources, does output show N, or does stderr explain the missing ones? Silent drops of failed sources are a top failure mode for fan-out commands.
4. Result ordering/ranking makes sense. For commands that claim to rank or sort, does the top result look plausibly best given the query? Watch for broken score weights, off-by-one sort bugs, and silent fallback to recency when relevance computation fails.

Calibration for learn-loop command samples (recall, learnings, playbook): on a fresh print the local learning store starts empty, so empty candidate lists, zero-count learnings stats, and "no learnings recorded" outputs are plausible-correct. Do not flag them as silent failures or missing data.

Return a list of findings. For each: check name, severity (warning in Wave B; error reserved for Wave C), one-line description, one-sentence fix suggestion. If the CLI passes all four checks, return "PASS — no findings."

Step 3: Emit the structured result block

End the skill response with a ---OUTPUT-REVIEW-RESULT--- block the parent parses:

On clean pass:

---OUTPUT-REVIEW-RESULT---
status: PASS
findings: []
---END-OUTPUT-REVIEW-RESULT---

On warnings:

---OUTPUT-REVIEW-RESULT---
status: WARN
findings:
- check: <check-name>
  severity: warning
  description: <one-line>
  suggestion: <one-sentence>
- ...
---END-OUTPUT-REVIEW-RESULT---

On reviewer failure (timeout, agent-budget exhaustion, missing live-check data):

---OUTPUT-REVIEW-RESULT---
status: SKIP
reason: <one-line description>
findings: []
---END-OUTPUT-REVIEW-RESULT---

Wave B policy (current)

• All findings surface as warning — never error. Shipcheck proceeds regardless.
• The caller logs findings to the run's artifact directory (e.g., manuscripts/<api>/<run>/proofs/phase-4.85-findings.md) and surfaces them to the user. Findings are not persisted to scorecard.json — that path is reserved for Wave C.
• The user decides case by case whether to fix before shipping.

Non-interactive contract (CI, cron, batch regeneration):

• If stdout is not a TTY, callers follow fail-open-with-log: findings recorded, shipcheck proceeds without prompting.
• status: SKIP (reviewer crash, timeout, missing data) is informational — shipcheck does not block on it.
• No --auto-approve-warnings flag yet. The policy is already "warnings don't block" in Wave B, so the flag has no effect to gate.

Wave C (separate future PR) will flip error-severity findings to blocking after calibration data across the library shows false-positive rate below 10%.

Why agentic vs template-only

Output-plausibility questions are not pattern-matchable against source. Rule-based live-check rules cover what regexes can (numeric HTML entities, query-token absence). Everything else — "are these substitution results plausibly correct for the query?", "does the top search result look related?" — is an LLM-shaped question. The token cost is bounded (once per run, not per command) and the catch rate against the bug classes that motivated this phase justifies the dispatch.

Known blind spots

• Can't verify numeric accuracy (prices, ratings, rankings vs ground-truth). If the CLI says a recipe has 4.8 stars and it actually has 4.2, this skill won't catch it.
• Can't detect data-freshness issues (recipe published 2019 vs 2024). These need live comparison against authoritative sources.
• Can't judge subjective preferences ("is this the best recipe for chocolate chip cookies?").
• Sampled outputs only — covers the commands in live_check.features[]. Full command-tree coverage belongs in Phase 5 dogfood.
• Non-English output: the reviewer's query-intent check assumes English-language query/output. For non-English CLIs, calibrate the prompt separately.