Printing Press Retro

/printing-press-retro

Analyze a Printing Press session to find ways to improve the system that produces CLIs — the Go binary, templates, skills, and workflow docs. Not fixes to the specific CLI that was just printed, but improvements so the next CLI comes out stronger.

It is a non-goal for the Printing Press to produce flawless CLIs without manual tweaks. That's the nature of the system. We expect agents to reason over the generated CLI, customize for the specific API, build novel features, and iterate. Some hand-built work in every run is normal.

The retro's job is to find the subset of manual work where the machine could have realistically raised the floor — given the agent a better starting point, prevented the issue entirely, or eliminated friction that would recur on the next CLI. Two clear cases qualify:

1. The machine could have completely prevented the issue, and the pattern is generalizable across many printed CLIs. File it.
2. The machine could have raised the floor meaningfully — better default, partial scaffold, helper that absorbs the boilerplate — across multiple CLIs you can name with evidence. File it.

Otherwise, the manual work is normal iteration and should not generate a finding. Some items make it back as machine fixes; not all. The retro is the filter that distinguishes the two.

The retro creates a GitHub issue on the printing-press repo with the findings that survive triage and the adversarial check, plus artifacts, so maintainers (or an AI agent) can fix the Printing Press.

Terminology

• The Printing Press: The whole system that produces CLIs. Use this name in all user-facing output (issues, retros, prompts). It has four subsystems:
  • Generator — templates that emit Go code (internal/generator/)
  • Scorer — tools that grade the output: verify, dogfood, scorecard
  • Skills — SKILL.md instructions that guide Claude during generation
  • Binary — the Go CLI itself: commands, flags, parsers (cmd/cli-printing-press/)
• Printed CLI: A CLI produced by the Printing Press for a specific API (e.g., notion-pp-cli). Printed-CLI fixes only help that one CLI.

Use "the Printing Press" when talking about the system. Use the subsystem name when pointing a developer at what to fix — "fix the scorer" and "fix the generator" are different PRs.

Cardinal rules

• Issue bodies and retro docs are public surfaces. Redact every real secret and PII before quoting. Manuscripts contain credentials, account identifiers, real emails, and live API response data — that's why references/secret-scrubbing.md scrubs them before artifact upload. Issue body text goes straight to a public GitHub issue, and the retro doc itself is preserved in manuscript proofs and may be uploaded as a zip. When you quote scanner output, dogfood payloads, Greptile review comments, or API response bodies as "evidence," replace the sensitive substring with <REDACTED:<kind>> BEFORE pasting. This applies hardest to findings about secret/PII leaks: the natural impulse is to quote the actual leaked value to prove the leak exists — that re-leaks it in a public issue. Phase 5 (retro doc write) and Phase 6 (pre-post scrub) enforce this mechanically; this rule is the human-readable charter behind the mechanical enforcement. See references/secret-scrubbing.md "Layer 0" for the redaction patterns and substitution shapes.
• Default is "don't change the machine." The Printing Press is mature — 30+ CLIs printed, most templates exercised across many shapes. The burden of proof is on the finding, not on the Skip path. Most things you encountered while printing one CLI are that CLI's quirks, iteration noise, or upstream API behavior — not generator gaps. Propose a machine change only when cross-CLI evidence is concrete and the finding survives the Phase 3 adversarial check (Step G).
• A retro of three sharp findings is more valuable than ten mixed-quality findings. Each filed finding spends maintainer attention. If you find yourself writing "every finding warrants action" or producing zero drops and zero skips, stop and re-triage — that outcome is the failure mode this skill exists to prevent.
• The retro proposes Printing Press changes that help multiple printed CLIs. Don't propose direct edits to the one CLI that just shipped, and don't propose machine changes whose value is unique to this CLI's quirks — those are printed-CLI fixes wearing a generator costume.
• Never upload un-scrubbed artifacts. All artifacts go through the secrets scrub before upload.
• Never modify source directories. Manuscripts and library directories are read-only. Scrub operations work on temporary copies.
• Never skip the secrets scrub, even if the generation pipeline already ran one. Defense in depth.
• Never work around a scorer bug in the Printing Press. If a scoring tool penalizes something incorrectly, the fix goes in the scoring tool.

Setup

# Path-only setup — no binary detection required.
# The retro skill reads manuscripts and runs gh/curl. It does not invoke the
# cli-printing-press binary. This avoids aborting for users who installed the
# plugin but not the Go binary.

_scope_dir="$(git rev-parse --show-toplevel 2>/dev/null || echo "$PWD")"
_scope_dir="$(cd "$_scope_dir" && pwd -P)"

PRESS_HOME="${PRINTING_PRESS_HOME:-$HOME/printing-press}"
PRESS_MANUSCRIPTS="$PRESS_HOME/manuscripts"
PRESS_LIBRARY="$PRESS_HOME/library"
RETRO_SCRATCH_DIR="/tmp/printing-press/retro"

mkdir -p "$PRESS_MANUSCRIPTS" "$PRESS_LIBRARY" "$RETRO_SCRATCH_DIR"

# Detect whether we're inside the printing-press repo
IN_REPO=false
if [ -f "$_scope_dir/cmd/cli-printing-press/main.go" ]; then
  IN_REPO=true
  REPO_ROOT="$_scope_dir"
  echo "Running from printing-press repo: $REPO_ROOT"
fi

Guard rails

Nothing to retro

if [ ! -d "$PRESS_MANUSCRIPTS" ] || [ -z "$(ls -A "$PRESS_MANUSCRIPTS" 2>/dev/null)" ]; then
  echo "No manuscripts found. Run /printing-press first to generate a CLI."
  exit 1
fi

Resolve which API

If the user passed an API name as an argument, use that. Validate for path traversal:

# Reject names with /, \, or ..
if echo "$USER_API_NAME" | grep -qE '[/\\]|\.\.'; then
  echo "Invalid API name: '$USER_API_NAME'. Names cannot contain path separators or '..'."
  exit 1
fi

# Verify resolved path stays under PRESS_MANUSCRIPTS
RESOLVED="$(cd "$PRESS_MANUSCRIPTS/$USER_API_NAME" 2>/dev/null && pwd -P)"
case "$RESOLVED" in
  "$PRESS_MANUSCRIPTS"/*) ;; # OK
  *) echo "Invalid API name: path resolves outside manuscripts directory."; exit 1 ;;
esac

If no API name was provided and multiple APIs exist, list them with their most recent run dates and ask the user to choose:

echo "Multiple APIs found in manuscripts:"
for api_dir in "$PRESS_MANUSCRIPTS"/*/; do
  api_name=$(basename "$api_dir")
  latest=$(ls -t "$api_dir" 2>/dev/null | head -1)
  echo "  - $api_name (latest run: $latest)"
done

Use AskUserQuestion to let the user pick.

Resolve which run

If the API has multiple runs, default to the most recent. If the user specified a run ID, use that. Otherwise:

API_DIR="$PRESS_MANUSCRIPTS/$API_NAME"
RUN_ID=$(ls -t "$API_DIR" 2>/dev/null | head -1)
RUN_DIR="$API_DIR/$RUN_ID"

echo "Retro for: $API_NAME (run $RUN_ID)"
echo "Manuscripts: $RUN_DIR"

Resolve CLI directory

API_SLUG="$API_NAME"
CLI_NAME="${API_SLUG}-pp-cli"
CLI_DIR="$PRESS_LIBRARY/$CLI_NAME"

if [ ! -d "$CLI_DIR" ]; then
  # Try without -pp-cli suffix (legacy naming)
  CLI_DIR="$PRESS_LIBRARY/$API_NAME"
fi

if [ ! -d "$CLI_DIR" ]; then
  echo "WARNING: CLI directory not found at $PRESS_LIBRARY/$CLI_NAME"
  echo "Proceeding with manuscripts only — CLI source will not be included in artifacts."
  CLI_DIR=""
fi

When to run

Best results come from running in the same conversation where the CLI was generated (post-shipcheck) — the retro can mine the full conversation history for errors, retries, manual edits, and discoveries.

If running in a fresh conversation, the retro proceeds with manuscript evidence only. Phase 2 marks session-dependent findings as "evidence: manuscripts only."

Phase 1: Gather evidence

Read all artifacts from the run:

1. Research brief — $RUN_DIR/research/*brief*
2. Absorb manifest — $RUN_DIR/research/*absorb*
3. Shipcheck proof — $RUN_DIR/proofs/*shipcheck*
4. Build log — $RUN_DIR/proofs/*build-log* (if exists)
5. Live smoke log — $RUN_DIR/proofs/*live-smoke* (if exists)
6. The generated CLI — $CLI_DIR/ (if available)

Also gather the scorecard, verify pass rate, and dogfood report (from the shipcheck proof or by re-running the tools if IN_REPO is true and the binary is available).

Phase 2: Mine the session

Scan the conversation history for six categories of signal and produce a candidate list. The candidate list is not the finding list — Phase 2.5 triage will cull it and Phase 3 will further drop weak survivors. Most candidates will not survive.

While collecting, distinguish:

• Iteration noise — one-off retries, typos, normal trial-and-error during a long generation. Skip these even at the candidate stage; they don't survive triage.
• Per-CLI quirks — behavior tied to this API's shape (auth oddity, undocumented endpoint, vendor-specific envelope) that wouldn't recur on another spec. Add to the candidate list with a "looks per-CLI" tag — most will be dropped at triage.
• Systemic friction — patterns that would plausibly recur on the next CLI (template gap, default that needs to change, skill instruction that misled you). These are what the retro exists to surface.

If running in a fresh conversation without generation history: Note this and proceed with manuscript evidence only. Focus on what the manuscripts reveal — scorecard gaps, verify failures, dogfood issues, and obvious template patterns in the CLI source. Mark session-dependent findings as "evidence: manuscripts only."

2a. Errors and retries

Any time a command failed and was re-run, a build broke, or the Printing Press produced code that didn't compile. What broke and what fixed it?

2b. Manual code edits

Manual edits during iteration are normal — agents reason over the generated CLI and tweak. A single edit to handle this CLI's quirk is the workflow.

For each manual edit, ask: could the machine have raised the floor here?

• Could the machine have completely prevented this edit? Default was wrong for most APIs, template emitted broken code, parser missed a common pattern. If yes AND the same edit would be needed on multiple CLIs you can name with evidence → candidate.
• Could the machine have given a better starting point that made the edit smaller, simpler, or skippable in common cases? Even if you'd still tweak, raising the floor compounds across future CLIs. If yes AND generalizable → candidate.
• Was this just per-API customization the agent was expected to do? Drop.
• Was this iteration noise (typo, retry, transient confusion)? Drop.

The triage question is whether the machine raising the floor would compound across future CLIs — not whether this one CLI would have shipped a few lines lighter.

2c. Features built from scratch

Hand-built features (transcendence commands, novel commands, helper packages for secondary APIs) are part of the workflow — agents build the domain-specific value layer on top of the API surface the machine emits. Building features by hand is not by itself a finding.

For each hand-built feature, ask: could the machine have raised the floor for this kind of feature?

• Could the machine have emitted a working default version, even if you'd still customize it? (E.g., every list+detail API benefits from a summary aggregation that the machine could scaffold from the spec.) Candidate, if generalizable across multiple named APIs.
• Could the machine have emitted scaffolding, types, or helpers that would have cut the build effort meaningfully? (E.g., a typed secondary-client template for combo CLIs, a fanout-aggregation helper.) Candidate, if generalizable.
• Is this genuinely custom domain logic the machine couldn't realistically generate from a spec? (E.g., booking a slot is custom orchestration; the machine can emit the underlying endpoints but not the choreography.) Drop — the SKILL is the right place to share the recipe, not the generator.

The "raises the floor" test separates "machine fix" from "SKILL recipe": if the machine's contribution would still leave significant per-CLI work, the recipe belongs in the SKILL so the next agent knows the pattern; if the machine could absorb the boilerplate cleanly, it's a generator template.

2d. Recurring friction

Work that happens on every generation, not just this one. For each: is this inherent to the approach, or can the Printing Press eliminate it?

Propose at least two possible fixes at different levels (generator templates, binary post-processing, skill instruction) and assess which is most durable.

2e. Discovered optimizations

Improvements noticed during the session — UX ideas, performance improvements, new command patterns, output format improvements. Could this optimization be detected automatically and applied by the Printing Press?

2f. Scorer accuracy audit

Before proposing Printing Press fixes to improve scores, check whether the scoring itself is correct. Changing the Printing Press to satisfy a broken scorer is worse than doing nothing.

For each score penalty from dogfood, verify, and scorecard:

1. Trace the scorer's logic. Read the scoring tool's source code to understand exactly what it checks. Don't guess.
2. Test the scorer's assumption against reality. Does the CLI actually have the problem the scorer claims?
3. Classify the penalty:
   • Scorer is correct — the CLI genuinely has this problem.
   • Scorer is wrong — the CLI is fine; the scoring tool has a bug.
   • Scorer is partially right — both could be better.

Common scorer bugs: name derivation mismatches, grep-based detection missing patterns, file exclusions too broad, section-counting heuristics.

The scorer audit is not optional. Every finding from a score penalty must have a "Scorer correct?" assessment before proposing a fix direction.

2g. Combo CLI priority audit

Only runs when the briefing named 2+ sources. Check $RUN_DIR/source-priority.json (from the Multi-Source Priority Gate in the main skill). If it doesn't exist but the briefing or user command clearly listed multiple services, that's itself a finding: the priority gate didn't fire when it should have.

For runs with a source-priority.json, cross-reference it against the absorb manifest and the shipped CLI:

1. Command count per source. Count commands attributed to each named source in the manifest. The primary should have at least as many as any secondary. If it has fewer, that's a priority inversion and becomes a finding — even if the user approved the manifest, it means the skill's discovery path for the primary failed silently.
2. Auth scoping. If the primary was declared free in the priority gate but the shipped CLI requires a paid key for the primary's headline commands, that's a finding — the economics check either didn't run or didn't route the paid key correctly to secondary-only scope.
3. README leadership. The primary should lead the README and --help. If a secondary is the first thing the user sees, flag it.

Each of these is a skill instruction gap category finding. The durable fix lives in skills/printing-press/SKILL.md (the Multi-Source Priority Gate, the Priority inversion check before Phase Gate 1.5, and the brief's ## Source Priority section) or in the generator if README ordering is template-driven.

Phase 2.5: Triage candidates

Before Phase 3 spends deep analysis on each candidate, run a fast triage to drop candidates that don't justify the deeper look. Most candidates should die here. The retro is a filter, not a funnel — if everything from Phase 2 makes it to Phase 3 unchanged, triage isn't doing its job.

For each candidate, ask in order:

1. Was this iteration noise? Normal trial-and-error during generation — one-off retry, typo recovery, agent forgetting a flag, transient network blip. Drop.
2. Is this a printed-CLI fix? The fix lives in $PRESS_LIBRARY/<api>/ and helps only this one CLI. If the proposed change is "edit this command in this CLI" or "regenerate after fixing the spec," it's not a retro finding — it's a polish pass on that CLI. Drop.
3. Is this an upstream API quirk? The vendor returns null instead of 404, or ignores a query param the docs claim to honor, or has rate limits the spec doesn't declare. The Printing Press doesn't fix vendors. If the only fix is "work around this in the generator for every CLI," that's almost always wrong; if it's "let one CLI work around it," that's a printed-CLI fix. Drop.
4. Is the only evidence "I noticed this once"? A one-time observation that you can't connect to a recurring pattern across other CLIs is a candidate for Drop, not a P3. P3 means "low priority systemic finding," not "I want to record this somewhere."
5. Does the same finding appear in 2+ prior retros without being implemented? Don't re-raise at the same priority. Either drop it (the cost-benefit math has been "no" twice and the retro is becoming a wishlist), or reframe as a smaller incremental fix that addresses part of the friction. Search: grep -l "<finding keywords>" "$PRESS_MANUSCRIPTS"/*/proofs/*-retro-*.md

Survivors of these five questions go to Phase 3. Dropped candidates are recorded as one-line entries in the retro's "Dropped at triage" section — they exist for your own discipline check and for the maintainer to see triage actually ran.

Anti-pattern to avoid. A recent Pagliacci retro produced "Skip: None. Every finding warrants action." That sentence is the failure mode this triage exists to prevent. Two of those findings (snake_case in Use:, root.go Short: rewrite that the SKILL already documents as a manual step) were classic per-CLI / instructional candidates that should have been dropped here. If you find yourself writing "every finding warrants action," stop and re-run triage.

Phase 3: Classify findings

For each candidate that survived Phase 2.5 triage, answer these seven questions. Question 5 has seven sub-steps (A through G); Step G is the adversarial check. Findings that fail Step G drop out — they don't get a priority, they don't go in the Do/Skip tables, they go on the dropped-candidates list with the reason.

1. What happened? One sentence — the symptom, not the fix.

2. Is the scorer correct? (mandatory for score-penalty findings)

• Scorer correct → fix the Printing Press (templates, binary, or skill)
• Scorer wrong → fix the scoring tool, not the Printing Press
• Both → fix both, label which is primary

3. What category?

4. Where in the Printing Press does this originate?

Pick exactly one component. The slug column drives the comp:<slug> label applied to the issue when filed (Phase 6), which is how agents filter related work across retros (gh issue list --label comp:<slug>).

If a finding genuinely spans two components, pick the one where the durable fix lands. Don't multi-label.

5. Blast radius and fallback cost — should the Printing Press handle this?

Step A: Cross-API stress test. Test across API shapes (standard REST, proxy-envelope, RPC-style) and input methods (OpenAPI, crowd-sniffed, HAR-sniffed, no spec).

Step B: Name three concrete APIs from the library with direct evidence. Not "every API with multi-word resources" or "any browser-sniffed CLI." Name three specific APIs already in $PRESS_LIBRARY/ or the public Printing Press Library where you can point to evidence the pattern exists: a path in their spec, a known endpoint shape, a header the vendor documents, an output you can reproduce. "Stripe, Notion, GitHub probably have this" is hand-waving; "Stripe (Stripe-Version header in spec line N), GitHub (X-GitHub-Api-Version on the issues endpoints), Linear (api-version on /v2/*)" is evidence. If you can name only two with evidence — or three with hand-waving — the finding drops to P3 max with a subclass:<name> annotation, or moves to Drop.

Step C: Counter-check question. Ask explicitly: "If I implemented this fix, would it actively hurt any API that doesn't have this pattern?" If yes, the fix needs a guard or condition before being P1/P2 — not a default change. Example: turning on client-side ?limit=N truncation by default would hurt APIs that need server-side pagination for correctness; it stays P2 only because it's gated on profiler-detected absence of a paginator. Without that guard the same finding is unsafe to land.

Step D: Recurrence-cost check. Search prior retros under $PRESS_MANUSCRIPTS/*/proofs/*-retro-*.md for the same finding. If the same finding has been raised in 2+ prior retros without being implemented, the prior cost- benefit math has been "no" twice. Don't re-raise it at the same priority — either move to P3 with a "raised N times, still not justified" annotation, or reframe the finding into a smaller incremental fix that addresses part of the friction. Recurrence at the same priority is a triage failure, not stronger evidence.

Capture matched prior retros. When the search returns hits, record each as a structured tuple — retro CLI name, retro file path (or GitHub issue number if the retro file's frontmatter contains one), and a one-word classification:

• aligned — the prior retro proposed the same fix direction. Strengthens the case; reference it in Step F.
• contradicts — the prior retro proposed an opposing fix or chose a different default. Surface this explicitly: a maintainer reading the new finding must see the disagreement. State in one sentence why this retro reaches a different conclusion (e.g., "prior retro saw single-paginator APIs; this one saw an always-paginated API where the prior default would break").
• extends — the prior retro raised an adjacent finding in the same component area but a different specific fix. Useful context, doesn't change the case.

These tuples flow forward into the per-finding template ("Related prior retros") in the retro doc and merge into the issue body's "Related issues" block alongside the Step 2.5 dedup scan's related-area outputs. GitHub auto-cross-links any #N issue number you write, so contradictions and alignments show up in both retro timelines without further action.

Step E: Assess fallback cost. How reliably will Claude catch and fix this across every future API? A "simple" edit Claude forgets 30% of the time means 30% ship with the defect.

Step F: Make the tradeoff. Default is don't change the machine. The burden of proof is on the finding to justify a machine change. Continue to Step G only when all three of these are true:

(a) Step B named three concrete APIs with evidence (not speculation). (b) Step D's recurrence-cost check didn't disqualify the finding. (c) Step C's counter-check didn't surface a hurts-other-APIs concern that lacks a guard.

If a finding can't clear all three, it doesn't get a priority — it goes to Drop with the specific reason ("only named 2 APIs with evidence" / "raised 3 times, still not justified" / "fix would hurt single-paginator APIs without a guard").

When the finding applies to an API subclass, include: Condition (when to activate), Guard (when to skip), Frequency estimate.

Step G: Construct the case against filing. Before recording the finding, write 1-2 sentences arguing the opposite — what makes this look like a printed-CLI fix, an iteration artifact, or a wishlist item. Why might a maintainer close this as "works as designed" or "too narrow for a machine fix"? What's the strongest version of "this shouldn't be filed"?

If the case-against is stronger than the case-for, drop the finding. If they're roughly even, drop the finding (default direction is don't-file). Only when the case-for is clearly stronger does the finding survive to Phase 4.

This step is not a formality. It is the explicit place where weak findings die. A finding that survives Step G should be able to state, in one sentence, why the case-against fails — and that sentence is worth quoting in the retro entry.

6. Is this inherent or fixable? Push hard on whether smarter templates, a post-processing step, or better spec analysis could eliminate the friction. If inherent, propose the cheapest mitigation.

7. What is the durable fix? Prefer: template fix > binary post-processing > skill instruction.

Mark uncertainty explicitly. If you can't confidently isolate one root cause or one fix, say so — list the candidate causes (or candidate fixes) and how an implementer could disambiguate before committing. The issue body surfaces this uncertainty so the agent picking up the work doesn't lock in a wrong-but-plausible diagnosis. Confidence isn't a virtue when it's manufactured; an honest "either A or B; verify by X" is more useful than a wrong prescription.

Strip API-specific details from the proposed fix. The durable fix must work across APIs, not just the one that surfaced the finding. If the fix includes hardcoded param names (e.g., --sport, --league), date formats (e.g., YYYYMMDD), chunking strategies (e.g., monthly), or domain-specific logic, those are printed-CLI details leaking into the machine recommendation. The machine fix should be parameterized — driven by what the profiler detects in the spec, not by what one API happens to need.

Example of the anti-pattern:

• Finding: "ESPN sync needs --dates for historical data"
• Bad fix: "Add --dates with YYYYMMDD-YYYYMMDD format, --sport/--league flags, and monthly chunking to the sync template"
• Good fix: "When the profiler detects a date-range query param, emit a --dates flag that passes the value through to the API"

The bad fix bakes ESPN's date format, scope params, and chunking strategy into the machine. The good fix lets the profiler drive behavior from the spec.

Phase 4: Prioritize

Sort survivors of Phase 3 into three buckets:

• Do — survived Phase 3 Step G with a clear case-for. Assign a priority (P1, P2, P3) based on frequency, fallback reliability, and complexity. Scorer bugs are just findings like any other — rank them by impact alongside template gaps and parser issues.
• Skip — survived Phase 2.5 triage but didn't clear Phase 3 (Step B couldn't name 3 APIs with evidence, Step D recurrence-cost disqualified, or Step G's case-against was stronger). State the specific step that failed. These are listed in the retro so the maintainer can see what was considered and rejected.
• Drop — rejected at Phase 2.5 triage as iteration noise, printed-CLI fix, upstream API quirk, unproven one-off, or recurring-not-implemented. Listed as one-liners only — they don't need full analysis, they need a record so triage is auditable.

No numerical scoring formulas. State the priority reasoning in words.

Sanity check before moving to Phase 5. Look at the bucket distribution. Almost every retro should have some drops and some skips. A retro with "all Do, no Skip, no Drop" is the failure mode — re-run triage and Step G on the weakest findings. Likewise, if every Do is P1, you're not prioritizing, you're inflating; force yourself to identify the weakest "Do" and ask whether it really beats the Skip bar.

Phase 5: Write the retro

The retro document is the durable audit trail — keep all fields below. The GitHub issue body in Phase 6 will use a slim subset (action-shaped fields only); the full triage rationale lives here, in the doc that gets uploaded as an artifact and linked from the issue. See references/issue-template.md for the issue-body shape.

Write the full retro document using this template:

# Printing Press Retro: <API name>

## Session Stats
- API: <name>
- Spec source: <public-library/browser-sniffed/docs/HAR>
- Scorecard: <score>/100 (<grade>)
- Verify pass rate: <X>%
- Fix loops: <N>
- Manual code edits: <N>
- Features built from scratch: <N>

## Findings

### 1. <Title> (<category>)
- **What happened:** ...
- **Scorer correct?** Yes / No / Partially. [details]
- **Root cause:** Component + what's specifically wrong
- **Cross-API check:** Would this recur?
- **Frequency:** every API / most / subclass:<name> / this API only
- **Fallback if the Printing Press doesn't fix it:** ...
- **Worth a Printing Press fix?** ...
- **Inherent or fixable:** ...
- **Durable fix:** ...
- **Test:** How to verify (positive + negative)
- **Evidence:** Session moment that surfaced this
- **Related prior retros:** *(from Phase 3 Step D; "None" if no matches)*
  - `<api-slug>` retro #<issue-num-if-known> — `aligned` / `contradicts` / `extends`. <one-sentence note on what changed or what's shared>
  - ...

## Prioritized Improvements

### P1 — High priority
| Finding | Title | Component | Frequency | Fallback Reliability | Complexity | Guards |
|---------|-------|-----------|-----------|---------------------|------------|--------|

### P2 — Medium priority
| Finding | Title | Component | Frequency | Fallback Reliability | Complexity | Guards |
|---------|-------|-----------|-----------|---------------------|------------|--------|

### P3 — Low priority
| Finding | Title | Component | Frequency | Fallback Reliability | Complexity | Guards |
|---------|-------|-----------|-----------|---------------------|------------|--------|

*Omit empty priority sections.*

### Skip
| Finding | Title | Why it didn't make it (Step B / Step D / Step G) |
|---------|-------|--------------------------------------------------|

*Findings that survived Phase 2.5 triage but failed Phase 3 — name the specific
step that failed (e.g., "Step B: only 2 APIs with evidence" / "Step G: case-against
stronger; mostly per-CLI"). Empty if every Phase 3 candidate filed.*

### Dropped at triage
| Candidate | One-liner | Drop reason |
|-----------|-----------|-------------|

*Candidates rejected at Phase 2.5. One line each. Reasons: `iteration-noise` /
`printed-CLI` / `API-quirk` / `unproven-one-off` / `raised-N-times`. If this
section is empty, re-check Phase 2.5 — almost every retro has some.*

## Work Units
(see Phase 5.5)

## Anti-patterns
- ...

## What the Printing Press Got Right
- ...

Save the retro to manuscript proofs (always) and to the temp retro scratch directory (always). Do not save retro documents under the source repo's docs/retros/ directory; the skill must work the same way for users who do not have the repo checked out, and retro documents are issue artifacts rather than durable repo docs.

RETRO_STAMP="$(date +%Y%m%d-%H%M%S)"
RETRO_PROOF_PATH="$PRESS_MANUSCRIPTS/$API_NAME/$RUN_ID/proofs/$RETRO_STAMP-retro-$CLI_NAME.md"
RETRO_SCRATCH_DIR="/tmp/printing-press/retro"
RETRO_SCRATCH_PATH="$RETRO_SCRATCH_DIR/$RETRO_STAMP-$API_NAME-retro.md"
mkdir -p "$(dirname "$RETRO_PROOF_PATH")" "$RETRO_SCRATCH_DIR"

Write the full retro document to $RETRO_PROOF_PATH, then copy that file to $RETRO_SCRATCH_PATH. This must complete before Phase 6 Step 1 copies the manuscripts directory to staging.

Scrub the retro doc immediately after writing

The retro doc is preserved in manuscripts/<api>/<run>/proofs/ (durable), copied to /tmp/printing-press/retro/ (scratch), and read by future runs' Phase 3 Step D dedup scan. If a finding's "What we observed" block pasted unredacted scanner output, dogfood payloads, or Greptile review comments, the secret/PII propagates into all three locations. Run the Layer 0 body scrub from references/secret-scrubbing.md immediately after writing the doc, so the scrubbed version becomes canonical:

# Define scrub_body once at the top of the Phase 5/6 bash blocks (full source
# in references/secret-scrubbing.md Layer 0). Then:
RETRO_PROOF_PATH_SCRUBBED="${RETRO_PROOF_PATH}.scrubbed.md"
if ! scrub_body "$RETRO_PROOF_PATH" "$RETRO_PROOF_PATH_SCRUBBED"; then
  echo "" >&2
  echo "ERROR: retro doc contains an unredacted vendor-prefix secret." >&2
  echo "Open $RETRO_PROOF_PATH, redact each match reported above using" >&2
  echo "  <REDACTED:<vendor>-<kind>:<first4>...<last4>:<len>ch>" >&2
  echo "per references/secret-scrubbing.md Layer 0, then re-run /printing-press-retro." >&2
  exit 1
fi
mv "$RETRO_PROOF_PATH_SCRUBBED" "$RETRO_PROOF_PATH"
cp "$RETRO_PROOF_PATH" "$RETRO_SCRATCH_PATH"

Hard-fail behavior is intentional: vendor-prefix secrets are unrecoverable leaks once a retro doc gets archived or uploaded. The agent must hand-redact and re-run rather than silently shipping the leak. PII patterns (real emails, phones, account inboxes) auto-redact in place because the substitution is lossless for the retro's purpose.

Phase 5.5: Plannable work units

Group related findings into coherent work units a planner could pick up directly.

For each "Do" finding or group of related findings:

### WU-1: <Title> (from F1, F3, ...)
- **Priority:** P1 / P2 / P3 *(max priority among absorbed findings — P1 if any
  absorbed finding is P1, else P2 if any is P2, else P3)*
- **Component:** generator / openapi-parser / spec-parser / scorer / skill
  *(must match one of the five fixed component slugs; drives the `comp:*` label
  applied to the issue when filed)*
- **Goal:** One sentence describing the outcome
- **Target:** <component and area, e.g., "Generator templates in internal/generator/">
- **Acceptance criteria:**
  - positive test: ...
  - negative test: ...
- **Scope boundary:** What this does NOT include
- **Dependencies:** Other work units that must complete first
- **Complexity:** small / medium / large

The five fixed component slugs are: generator (internal/generator/), openapi-parser (internal/openapi/), spec-parser (internal/spec/), scorer (verify / dogfood / scorecard), and skill (skills/printing-press/SKILL.md). If a WU genuinely spans two, pick the primary one — the component where the durable fix will land. Pick exactly one; don't multi-label.

If running from inside the printing-press repo (IN_REPO=true): Resolve target file paths using Glob and Grep tool invocations on $REPO_ROOT to make work units more precise. E.g., use Glob to find internal/generator/*.go files, Grep to find where sync code is generated.

If running externally (IN_REPO=false): Describe target components by name (e.g., "Generator templates in internal/generator/") and acceptance criteria without resolved file paths. The fixer will resolve paths when they pick up the work.

Phase 5.6: Issue gate — are there Printing Press improvements?

After prioritization and work units are written, decide whether GitHub issues are warranted. Each WU becomes one flat top-level issue (no parent, no sub-issue hierarchy). The purpose of filing is to give someone (human or agent) something to fix in the Printing Press. If every finding is specific to this one printed CLI with nothing to change in the Printing Press, filing is noise — there's nothing to act on.

Skip filing if:

• Every finding landed in "Skip"
• All findings are printed-CLI-specific (manual edits that only apply to this one API and wouldn't recur across other CLIs)
• The "Do" table is empty

File issues (one per WU) if:

• There is at least one "Do" finding — i.e., something a maintainer or agent could act on in the Printing Press (templates, binary, skills, or scoring tools)

Use judgment. A retro that found three things but all three are "this API has a weird auth scheme no other API uses" is not worth filing. A retro that found one small template gap that would help every future CLI is worth filing.

If filing is skipped, still save the retro locally (manuscript proofs + /tmp/printing-press/retro/), present the findings to the user, then jump directly to Phase 6 Step 6 (present results — adjusted to show local-only paths).

Phase 6: Package, upload, and present

Step 1: Package artifacts into staging folder

Read and apply references/artifact-packaging.md through Step 4 only (create staging dir, copy, scrub, zip). Do not upload or clean up yet — the staging folder stays alive until the end of Phase 6.

The staging folder ($STAGING_DIR) now contains the scrubbed copies and the zips. This is both the review target and the upload source.

Step 2: Compute filing plan + confirm before publishing

This step only runs if the Phase 5.6 issue gate passed (there are Printing Press findings to act on).

Before showing the confirm prompt, run references/issue-template.md Steps 1, 2, and 2.5 to ensure labels exist, sort the work units, and compute the per-WU filing plan via the dedup scan against open retro-tagged issues. Each WU ends up classified as either:

• File new — no matching open issue
• Comment on #N — Step 2.5 found a same match; the new evidence will be added as a comment instead of filing a duplicate
• File new with related issues — Step 2.5 found one or more related-area matches; the new issue's body will reference them via #N in the Related issues block

The dedup scan does not need to be bulletproof. Bias toward "file new" when uncertain — duplicates are recoverable, miscomments on the wrong issue are uglier.

Then show the user a summary including the filing plan and ask for confirmation via AskUserQuestion.

Ready to submit your retro.

Here's what will happen on mvanhorn/cli-printing-press:

Filing plan:

#	Title	Plan	Notes
1		File new (P1, comp:)	No match
2		Comment on #234	Matches ""
3		File new + reference #189	Adjacent open issue

Each new issue carries retro, priority:P<n>, comp:<slug> labels — agents filter related work across retros with gh issue list --label comp:<slug> or gh issue list --label priority:P1.

Scrubbed artifact zips uploaded to catbox.moe and linked from each new issue:

• Retro document — full triage rationale, drops, skips, what went right
• Manuscripts () — research brief, shipcheck proof, build logs
• CLI source () — the generated Go code (no binary, no vendor/) (omit if not available)

Everything is staged at <$STAGING_DIR> if you'd like to insp

…(truncated)