Diff Recap

/diff-recap

Turns a git diff into a recap you can paste into a pull request — a per-file change table whose numbers are transcribed from git, grouped and labelled, rendered as an inline annotated widget with a mandatory ASCII twin, and backed by a diff-recap/1 record. It fills the gap the orient- family leaves open: orient-map shows the shape of the work and orient-status shows where you stand, but neither recaps what a concrete change-set contains for a reviewer. diff-recap does exactly that, and only that.

Its differentiator is true by construction: the structured numbers (per-file added/removed/status and the headline totals) come straight from git diff --numstat and --name-status — the model authors only the prose labels. A record whose totals don't equal the sum of its rows is rejected by a value gate before it can be shown, so the recap cannot quietly lie about the diff. And unlike a hosted recap renderer, it has zero external dependency — the ASCII twin is self-contained and PR-pasteable in any channel.

> Structured execution spec: [protocol.yon](protocol.yon). Read it for the canonical steps, rules, and gates; this file is the human-readable, self-sufficient explanation. The two must stay in sync — if you edit one, update the other and refresh the @STAMP date.

When to use

• "Recap this diff / summarize these changes for a PR" — you want a reviewer-ready change summary, not a fresh read of the whole tree.
• "What changed in this branch" as a change-set, grouped by area with a one-line why per file.
• Right before opening a pull request, when you want a paste-able summary whose numbers you can trust.

How it works (self-sufficient — works without protocol.yon)

Every call recomputes from the real diff; nothing is stored.

1. Take range, repo, and intent as inputs — never prompt. range defaults to the working tree against HEAD (or a A..B range, or --staged); repo defaults to the cwd. intent (the PR's goal) shapes the labels — if it's ungrounded, the recap still emits the mechanical facts but caps the labels to descriptive and never invents a rationale.

1. Gather from git, transcribe verbatim. Run git diff --name-status and git diff --numstat over the range. For each file, copy the status (A/M/D/R) and the added/removed counts exactly as git reports them. A binary file reports -/- in numstat → record 0/0 and mark the row binary in its label. The model never counts by hand — the numbers are git's.

1. Group and label — the only model-written fields. Cluster the files by area (feature / tests / docs / config / …), and write one concise label per file and a one-line summary per group. path, status, added, removed are git-attested; label and group are the model's only freedom.

1. Set the totals as the sum of the rows. total_files, total_added, total_removed are the exact count and sums of the per-file rows — that is the true-by-construction invariant: the headline is, by definition, the sum of the attested rows.

1. Fail closed. No git, or an empty diff → gate_status = barren, attested = false, and a NO DIFF card — never a fabricated recap. A clean recap requires a real git source (attested = true); a partial source caps the verdict at partial.

1. Emit the triple bundle. The YON record (below), the PR-pasteable markdown summary, and the visual face (Claude Code only). diff-recap honors the three render-face branches: a human on Claude Code with the visualize tool present on an explicit invocation gets the widget + ASCII twin; an agent gets the YON record only; any other runtime / no tool / indeterminate handler_type gets the ASCII twin (fail-closed). See the bundled, self-contained [render-face contract](references/render-face-contract.md). diff-recap adopts only this branching, not the orient family's footer roster, "you are here" breadcrumb, or staleness short-circuit. The information-complete ASCII twin is always emitted — it carries every file row, count, group, and total the widget does (worked trio at [examples/diff-recap.{ascii.txt,widget.svg,example.yon}](examples/)).

Record emission (the YON face — reserved tags only)

The record is a diff-recap/1 instance ([recap-schema.yon](recap-schema.yon)). Reserved tags only. The envelope is @CFG id=recap; each changed file is one @CFG id=file.; list fields (groups, provenance) go in a sidecar @MAP, never an in-set bracket-list (an in-set list corrupts the record silently and still validates). Never run yon format on an instance — CANON mode is destructive on set-internal lists. Validate with --profile exec, then run the value gate. Skeleton:

@CFG id=recap | set=[schema_version=diff-recap/1,computed_at:ts=…,ephemeral:bool=true,tool=diff-recap,range="HEAD",repo=".",total_files:int=4,total_added:int=155,total_removed:int=7,attested:bool=true,gate_status=clean,overall_trust=high]
@CFG id=subject | set=[name=…,kind=repo,purpose=…,intent_status=stated]
@CFG id=file.f1 | set=[path="src/rate-limit.ts",status=A,added:int=88,removed:int=0,group=feature,label="new token-bucket limiter"]
@CFG id=file.f2 | set=[path="src/server.ts",status=M,added:int=24,removed:int=6,group=feature,label="wire the limiter in"]
@MAP name=groups | pairs=["feature"->"limiter + wiring (2 files, +112/-6)"]
@MAP name=provenance | pairs=["recap.total_added"->"git-attested:high:git diff --numstat"]

On barren evidence the envelope degrades honestly: gate_status=barren, attested:bool=false, no file rows, and the visual is the NO DIFF card. A full worked instance ships at [examples/diff-recap.example.yon](examples/diff-recap.example.yon).

The value gate — why the recap can't lie

yon validate checks structure, not values: a record can be valid YON yet state total_added=999 while its rows sum to 155. [references/tools/diff-recap-check.mjs](references/tools/diff-recap-check.mjs) travels inside this skill folder and is the missing half — it asserts Σ per-file == totals, enum membership, and the fail-closed/attested gates, and with --numstat it checks every row against the real git numstat. An emitter runs it before emit; the source repository additionally exercises the worked and deliberately broken fixtures as release evidence.

node /references/tools/diff-recap-check.mjs /examples/diff-recap.example.yon --numstat /examples/diff-recap.numstat

Output — worked examples (markdown / ASCII face)

📋 acme-api — add per-route rate limiting.    range: HEAD     [counts: ◆ git --numstat]
## 4 files · +155 / −7                                          ◆ git diff --numstat

feature   +112 / −6
  + src/rate-limit.ts         +88 / −0    new token-bucket limiter (per-route)      ◆ A
  ~ src/server.ts             +24 / −6    wire limiter in; drop old fixed counter   ◆ M
tests     +40 / −0
  + test/rate-limit.test.ts   +40 / −0    burst, refill, per-route isolation        ◆ A
docs      +3 / −1
  ~ README.md                 +3  / −1    document the RATE_LIMIT env var           ◆ M

Σ rows = +155 / −7  ==  headline  ✓  (true by construction)
**→ PR summary:** per-route token-bucket rate limiter + unit tests + env-var doc.
Trust: ◆◆◆ every count git-attested · the model wrote only the labels · gate: clean

Barren (honest degradation is the success case):

📋 ToDo — no git, or an empty diff.
┌───────────────────────────────┐
│  NO DIFF                       │   nothing to recap — no commits, no changes.
└───────────────────────────────┘
**→ Next:** stage or commit a change, and the recap becomes visible.
Trust: ◌◌◌ · [gate: barren · attested: false]

The ASCII twin + the Σ rows == headline line + the trust trailer are mandatory on every output.

Boundaries

• Not /orient-map — that shows the shape + delta of the work (done → here → next); diff-recap recaps a concrete change-set for a reviewer. They never overlap: orient-map is trajectory, diff-recap is a PR table.
• Not /orient-status — that reports position + remaining + ETA; diff-recap reports what a diff contains.
• Not a code review — it summarizes what changed, it does not judge whether the change is good. Pipe it into /cold-review for that.
• Read-only. It runs git diff and reads the tree; it never mutates state.
• No hosted dependency. Unlike a hosted recap renderer, the ASCII twin is the floor and is self-contained — the recap works with no visualize tool, no server, no account.

Rules

• MUST be true by construction — transcribe path, status, added, removed verbatim from git diff --name-status/--numstat; never count by hand. The model authors only label and group.
• MUST set total_files/total_added/total_removed to the exact count and sums of the rows, so the headline equals the sum of the attested rows; the value gate rejects any record where they drift.
• MUST fail closed — no git or an empty diff → gate_status = barren, attested = false, NO DIFF card; a clean verdict requires attested = true.
• MUST always emit the information-complete ASCII twin — it is the only render in a no-SVG channel and is PR-pasteable verbatim.
• MUST render the visual face per the render-face contract (Claude Code only, via mcp__visualize__show_widget): agent → YON record only; human + Claude + tool + explicit invocation → widget + ASCII twin; other runtime / no tool / indeterminate handler_type → ASCII twin (fail-closed). Widget and ASCII twin are projections of one row-structure (same rows + totals), chrome-compliant (no hardcoded color, role="img" with non-empty title/desc, fixed 680 viewBox).
• MUST emit file rows as @CFG id=file. and list fields (groups, provenance) as sidecar @MAP; never a bare in-set bracket-list; never yon format an instance.
• MUST take intent as input and cap labels to descriptive when it's ungrounded — never fabricate a rationale.

> Next skills. On completion, run the Next Skills protocol (next-skills/SKILL.md): surface the next-skills recommendations from front-matter for the caller to pick. Offer only — never auto-invoke.

> Self-improvement. On completion, run the Self-Improvement Protocol (self-improve/SKILL.md): if this run surfaced a concrete, blocking-or-recurring weakness in this skill, propose a specific fix for the handler to approve. Conservative — silent otherwise. Never auto-apply.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: allemaar
• Source: allemaar/open-skills
• License: Apache-2.0
• Homepage: https://allemaar.com

Install and usage instructions live in the source repository linked above.