incomplete-edit costs about one extra iteration beyond what diff size predicts: find what distinguishes it at declare-done

[prflow-implementer]

Summary

• Retrospective Stage A entries now echo the diff-size fields additions, deletions, and changed_files from the context bundle, making diff size a durable, reproducible control for future analysis instead of requiring a live gh re-fetch.
• Records the incomplete-edit cost investigation as an internal doc page whose answer is that the current durable records are insufficient to decide whether the category is predictable at declare-done.
• Purely additive: the fields ride the existing schema_version 3 (already 3 on main since A retrospective entry records nothing about what evidence the analyst had, so diff-present and diff-absent entries pool indistinguishably #1829); no schema bump, and existing .prflow/learnings/ entries are byte-unchanged.

Changes

Retrospective schema (skills/retrospective/SKILL.md): Documents additions, deletions, and changed_files in the Stage A output schema, the echo-straight-from-the-bundle prose, and the jq example construction.
Entry cleaner (lib/clean-entry.jq): Preserves the three size fields from the bundle, each defaulting to null via // (so a producer lacking them still cleans without error, and the keys are always present under schema_version 3).
Analysis doc (docs/internal/incomplete-edit-cost-analysis.md, new): Records the cohort method, the rework metric (efficiency_runs[].iterations, never post_bot_commits) and its ~41% coverage, and concludes the durable records cannot decide the question, naming what a future pass would need.
Tests (lib/test/run.sh): Adds a #1827 block asserting the size-field cleaning contract.
Coverage map (lib/test/modules/coverage-map.json): Registers issue 1827 as unmodularized.
Changeset (.changeset/issue-1827-incomplete-edit-size-fields.md, new): patch / Added.

Resolves

Resolves #1827

Test Plan

• #1827 block in lib/test/run.sh: additions and deletions survive cleaning when present (120, 34).
• changed_files survives verbatim by contents and order (not merely by length).
• A bundle lacking the fields cleans without error, defaulting each to null, with the keys still present (has() pins the schema_version 3 contract).
• Valid-falsy passthrough: a real 0 / [] survives and is not coerced to null, and changed_files stays an array.

Visual Changes

N/A

Breaking Changes

None

Prompt-surface size

Derived at a64b426bc74f60c31561c5ba6946da06d6948c96 against merge-base 9a2147e221859cbbecff25bf23421487c7cd579b (origin/main). Covered: tracked *.md under skills/, agents/, .prflow/prompt-extensions/.

Path	Before	After	Δ bytes	Δ %
skills/retrospective/SKILL.md	23,760	24,035	+275	+1.2%
Whole covered surface	2,142,597	2,142,872	+275	+0.0%

Generated via /prflow:implement (v2.34.14, claude-opus-4-8, low)

[The01Geek]

/prflow:review

[prflow-implementer]

/prflow:review

[prflow-reviewer]

Verdict: APPROVE (additive schema echo with matching tests; every acceptance criterion met; no findings)

Issue Compliance

Reviewed against issue #1827: incomplete-edit costs about one extra iteration beyond what diff size predicts — criteria from workpad; scope unchanged. Requirement-based checklist items are included in the verification results below.
Criteria came from the /prflow:implement workpad comment — this run's authoritative set, possibly narrowed from the issue.
No divergence between the workpad's criteria and the issue body's.

All four acceptance criteria are satisfied:

• AC1 — fields echoed + preserved; absent entry still cleans. lib/clean-entry.jq adds additions/deletions/changed_files, each (.field // null), under the unchanged schema_version: 3; skills/retrospective/SKILL.md echoes all three in both the output-schema block and the example construction, extending the "echo straight from the bundle — do not recompute" instruction to them. The producer lib/fetch-pr-context.sh already emits all three fields (verified against head), so "echoed from the bundle" holds. A bundle lacking the fields still cleans, defaulting each to null.
• AC2 — analysis doc. docs/internal/incomplete-edit-cost-analysis.md exists and records the rework metric (efficiency_runs[].iterations, max non-null per PR), its coverage (45/109 = 41.3% cohort; 175/489 store-wide), and the §2.3 sweep-selection correlation, concluding with counts that the durable records are insufficient to decide. Table arithmetic is internally consistent.
• AC3 — metric never from post_bot_commits. The doc's Method section states this explicitly and cites #1440 as the reason; no derivation uses that field.
• AC4 — learnings byte-unchanged. No file under .prflow/learnings/ is touched by the commit range (verified).

Verification Checklist Results

0 failed, 0 inconclusive. The verifiable claims (producer emits the three fields in the shapes the consumer/tests assume; jq // treats 0/[] as truthy so valid-falsy values survive; schema_version 3 always emits the keys) were confirmed by the Phase-3 agents against head a64b426. Diff classification: engine_self_modifying (touches skills/** and lib/**) → full checklist profile.

Code Review Findings

No findings from any of the five dispatched Phase-3 agents (code-reviewer, silent-failure-hunter, comment-analyzer, pr-test-analyzer, requesting-code-review). The one executable change (lib/clean-entry.jq) is a fail-safe additive echo; the #1827 test block pins presence passthrough, absent→null defaulting, the has() key-always-present contract, changed_files verbatim-by-contents (not length), and valid-falsy (0/[]) passthrough — the classic //-coercion trap is explicitly guarded.

over-grade annotation: no finding flagged
truthfulness sweep: no finding promoted
intra-diff contradiction scan: no contradiction found

Prompt-surface edit routing evidence gate

The diff touches skills/retrospective/SKILL.md, so the DevFlow-repo routing gate applies. Discharged: issue #1827's workpad carries a Writing-skills evidence: marker with all four required dispositions explicit — skill-loaded=yes guidance-applied=yes pressure-scenario=no micro-tests=no. A recorded no never draws a finding on its own. The Verification evidence: advisory marker is also present (whole-suite CLEAN pass on a64b426), so the advisory clause is silent.

Shipped-prose cost audit

The skills/retrospective/SKILL.md addition (~9 lines) adds only what documents the three new fields — the schema entries, the echo-list extension, and the example-construction lines — each carrying the necessary "echo these from the bundle, do not recompute" instruction. No excess prose; no finding.

Verdict Criteria

No verification FAIL/INCONCLUSIVE, no self-contradicting-diff finding, no review-agent finding at or above the critical threshold, and no findings at all → APPROVE (rule 7).

Note: this standalone cloud review ran with the checkout on main; the reviewed head a64b426b was verified via git show <head>:<path> per the standalone PR-number routing. The Phase 0.6 stale-prose lint is disabled by config (prflow_review.stale_prose.enabled: false).

[prflow-reviewer]

Verdict: APPROVE with notes (small, purely-additive change; two low-value Suggestions, neither gating)

Issue Compliance

Reviewed against issue #1827: incomplete-edit costs about one extra iteration — criteria from workpad; scope unchanged. Requirement-based checklist items are included in the verification results below.
Criteria came from the /prflow:implement workpad comment — this run's authoritative set, possibly narrowed from the issue.
No divergence between the workpad's criteria and the issue body's.

All four acceptance criteria are satisfied:

• AC1 (fields echoed + preserved; absent-field entries still clean): lib/clean-entry.jq adds additions/deletions/changed_files with // null defaults; skills/retrospective/SKILL.md echoes them verbatim from the bundle; the producer lib/fetch-pr-context.sh actually emits all three (additions/deletions as numbers, changed_files = [.files[].path], an array), so the echo carries real data. Satisfied.
• AC2: docs/internal/incomplete-edit-cost-analysis.md exists and records the rework metric (efficiency_runs[].iterations), its 41% cohort coverage (45 of 109), and the §2.3 sweep-selection correlation, concluding with a counts-backed insufficiency statement — the escape hatch AC2 permits. Satisfied.
• AC3: the doc explicitly states the metric "is never derived from post_bot_commits" (citing #1440); no table or finding is sourced from that field. Satisfied.
• AC4: the diff touches no file under .prflow/learnings/ (verified against the merge-base); existing entries are byte-unchanged. Satisfied.

Verification Checklist Results

8 passed, 0 failed, 0 inconclusive — 0 via lite probe, 8 via agent.

✅ Passed items (8 of 8) — click to expand

• VC-1: additions/deletions/changed_files echoed from the bundle, preserved by clean-entry.jq, absent-field entry still cleans [lib/clean-entry.jq:133-137]
• VC-2: incomplete-edit-cost-analysis.md records metric, coverage, and counts-backed conclusion [docs/internal/incomplete-edit-cost-analysis.md]
• VC-3: rework metric never derived from post_bot_commits [docs/internal/incomplete-edit-cost-analysis.md:53]
• VC-4: existing .prflow/learnings/ entries byte-unchanged (diff touches none)
• VC-5: bundle producer emits all three fields [lib/fetch-pr-context.sh:888-890]
• VC-6: changed_files is a JSON array of paths, surviving verbatim [lib/fetch-pr-context.sh:48]
• VC-7: #1827 fixtures match the real bundle field names/types [lib/test/run.sh:166-188]
• VC-8: additive under schema_version 3 (no bump); keys always emitted (has() contract) [lib/clean-entry.jq:17,133-137]

Code Review Findings

🟡 Suggestion / Minor

1. The end-to-end producer→clean-entry.jq join is not pinned anywhere: the #1827 block tests clean-entry.jq in isolation against hand-authored fixtures, and nothing asserts that lib/fetch-pr-context.sh emits changed_files as a path array (vs a scalar count). If the producer later regressed, clean-entry.jq would silently default to null/pass a wrong-typed value with no RED. The producer is pre-existing and unchanged by this PR, so this is a pre-existing-surface gap, not a regression this change introduces. (raised by 1/5 agents)
2. No test distinguishes an explicit "additions": null in the bundle from an absent field; under .additions // null both yield null, so the case is genuinely redundant — noted for completeness only, not recommended. (raised by 1/5 agents)

Notes

• Test evidence (CI, reviewed commit a64b426b): lib + python tests: success, all five shards (monolith, python-pool, modules-pin, modules-large, modules-rest) success, and lint (shellcheck + actionlint + ruff): success. Authoritative per the engine ground-truth block.
• Valid-falsy correctness: the #1827 tests pin that a real 0 and [] survive uncoerced (jq's // treats them as truthy) — the exact #312/#304 off-switch bug class CLAUDE.md flags for //-based extraction, with a regression that fails closed against a naive if .x then rewrite.
• Prompt-surface edit routing gate (repo policy): discharged — the linked issue #1827 workpad carries Writing-skills evidence: for skills/retrospective/SKILL.md with all four slots recorded (skill-loaded=yes, guidance-applied=yes, pressure-scenario=no, micro-tests=no); all dispositions explicit, so the gate passes (a recorded no never draws a finding).
• Verification-evidence advisory (repo policy): Verification evidence: marker present on the workpad (cloud-tier run) — advisory clause silent.
• Shipped-prose cost audit (repo policy): the skills/retrospective/SKILL.md additions are a minimal, factual schema-field/echo edit (3 schema lines + a one-clause echo-list extension + 3 jq-example lines); every added word carries the instruction it names. No wasteful prose.
• over-grade annotation: no finding flagged.
• truthfulness sweep: no finding promoted.
• intra-diff contradiction scan: no contradiction found.
• stale-prose lint disabled by config.
• Deferrals: no Scope-Acknowledged Findings block in the PR body.
• Dirty-tree backstop: disabled this dispatch (the pre-dispatch snapshot redirect was refused by the harness); the advisory review agents are read-only and made no working-tree change.

Verdict Criteria

• No verification checklist FAIL or INCONCLUSIVE.
• No self-contradicting-diff finding (every diff-added claim verified true against HEAD a64b426b).
• No review-agent finding at or above the configured critical threshold — only two Suggestions, both below threshold and one of them pre-existing surface.
• → APPROVE with notes.