# Groundtrack > Use for a plan already made or work already done, when a reader needs to see its shape — what calls what, what each part hands back, where it breaks, and what it needs to work. Writes one call graph with recorded walks through it, renders a self-contained page, and prints the same graph as an indented tree on request. Not for a conversation with nothing durable behind it. - Skill: `mephistopheles4/groundtrack` (Agent Skill, multi-file: 20 files) - Install (CLI): `npx skillmds@latest add mephistopheles4/groundtrack` - Raw SKILL.md: https://api.skillmd.com/api/skills/mephistopheles4/groundtrack/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: mephistopheles4 (https://skillmd.com/u/mephistopheles4) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/mephistopheles4/groundtrack --- # groundtrack A reader who did not write a change cannot see its shape. A change arrives as a list of files. A plan arrives as a list of tickets. Neither says what calls what. groundtrack turns durable material into a call graph a reader can step through. You read the material and write one file by hand. A renderer turns that file into one self-contained page. **Stance.** The drawing is a claim, and the material is what holds it. You draw what the material says. You say so when the drawing goes past it. ## When to use - A plan already made, or work already done, and a reader needs its shape. - The reader asks what calls what, what a part returns, or where it can break. - The reader asks by name. **Do not use it on a conversation.** Nothing durable exists to check the graph against, so the one property this skill rests on disappears. ## What it takes in **The rule is a property, not a list.** groundtrack accepts durable material you can read, plus a file set. That admits a written change, an unbuilt plan, a named function or region, and a bare path into a codebase. The source usually states the file set. A change states it. A plan's file list states it. **When the source states none, ask the reader for it.** Never invent a bound the reader did not choose. ## The three channels Every node carries three channels, and they are the point of the drawing. - **A** — what flows out of the node. - **E** — where it breaks: the failure tags it can raise. Each tag is a **retry**, an **escape** or a **die**. - **R** — what it needs to work. The three failure kinds are different facts. A blip that retries, an error that escapes to a caller, and a fault that destroys the process are not one thing. A process killed for running out of memory reaches no handler, so nothing logs it, and the debugging session drops where the data was worth having. The E channel is where a reader sees that in advance. ## Procedure 1. **Read the material.** Follow the calls. Do not draw from the file list. 2. **Cut it into graphs.** One graph per entry point, where an entry point is a changed symbol that no other changed symbol calls. Seventeen files do not fit on one graph. **This rule is argued, not measured.** It was derived from one real change, where it gives the right answer. Nobody has run it against a set of changes and compared the result to a hand-drawn one. So read your own cut before you show it: if a graph comes out with one node, or with everything in it, the rule has told you something you should say out loud rather than draw. 3. **List every graph you found, and let the reader pick.** Say the list before you draw anything, so the reader can disagree with the cut before the work rather than after it. **Rank nothing and suggest nothing.** No rule for ranking graphs has ever been measured, so a score here would put an undecided default in front of every reader. A run that is not told which graph to draw asks. 4. **Write one file for the change.** One file states the change: its title, its blurb, its changed files, its sheet rule, one node map, and a list of graphs. A graph is an entry point and the runs from it, so a second graph is an entry named rather than a second file. A symbol two graphs reach is defined once. Read [`references/flightpath-file.md`](references/flightpath-file.md) for the shape, and [`skills/groundtrack/examples/greet.flightpath.json`](examples/greet.flightpath.json) for a complete legal file. Write the nodes and the graphs first. Validate. Then write the walks, one at a time. 5. **Validate, read the refusal, fix, repeat.** See [`references/writing-walks.md`](references/writing-walks.md) for the loop and for the two mistakes measurement says you will make. 6. **Read every finding, and answer each one.** `--check` exits zero and still prints its findings, because a finding never refuses. A clean checker is not a finished file. Answer a finding in one of two ways. Fix the file, or write one line saying why the finding is what you meant. A node you have written and not yet connected is a finding you write the line for. You usually fix an `E` tag nothing beneath it produces. Draw what raises the tag, or drop a tag the program never raises. Both answers are legitimate. Never skip this step. **Give the reader the findings you kept**, each with the line you wrote for it. A reader who never sees the finding cannot disagree with your answer. 7. **Render the page**, and say where it is. 8. **State the limit.** The validator proves the walk is a legal path. It cannot prove which branch an `if` took or what an effect returned. Those stay your claims. Say this when you hand the page over. ## Where the run writes **Write the file and the page to a scratch directory.** Never beside the input. A page dropped next to the file it was made from is an artifact nobody asked for and nothing cleans up. **Copy either one into the repository only when the reader asks.** The artifact is disposable by default. The renderer will not write a page without an output path, for the same reason. ## The command line ```bash node /scripts/render.mjs .flightpath.json --check node /scripts/render.mjs .flightpath.json --out .html node /scripts/render.mjs .flightpath.json --text [""] [--graph ] ``` `` is the base directory this skill was installed into. Ask the harness for it. Never write a fixed path. | Flag | What it does | | --- | --- | | `--check` | Validates. Refusals on standard error, exit 1. Findings on standard output, exit 0. | | `--out ` | Writes one self-contained HTML file. | | `--text [""]` | Prints the tree to standard output, for one run. | | `--graph ` | Which graph `--text` reads. Refused with `--check` and `--out`, which both read every graph — the check validates them all, and the page carries them all and picks between them. | **A file with several graphs needs the graph named.** `--text` on a one-graph file prints as it always did. On a several-graph file without `--graph` it lists the graphs, one line each, and stops. **Nothing ranks them and nothing suggests one**, which is the same rule as step 3: a run that is not told which graph to read asks. ## The page The page draws one graph and steps a cursor over one recorded walk. **A file with several graphs is one page with several sheets.** The head holds two rows: the title above, and below it the sheet picker, the run picker and the step controls. The sheet picker lists each graph by title. A one-graph file shows none. A sheet draws what its entry reaches through call edges. Write a symbol two graphs reach once; both sheets draw it. Each sheet keeps its own run, cursor, layer, view and open node. Leave a sheet and come back and it is as you left it. Nothing is computed while the reader watches. Every branch an `if` took, every value an effect returned, and every catch is a literal in the file. That is what makes the walk a list of checkable claims. The reader steps forward and back, holds the walk on the next effect or the next error, reads the call stack and the effects ledger, opens one node to see its body, and flips between the drawing and a tree. The cutaway follows the cursor. Each move opens the node the walk is in and brings the step that ran into view. A node opened by hand stays open until the cursor next moves. The files tab reads as a directory tree, so a change of many files reads as a few folders. A directory holding one thing prints on one line. Each file carries a mark for its change — `N` new, `E` edit, `D` delete, `F` forbidden — and the `why` you wrote trails the path the way a comment trails a line of code. The wheel scrolls the drawing. With ctrl held it zooms about the pointer. The contract tab states one number per node: its cyclomatic complexity, as drawn. One, plus one for each `if`, each error handler and each backward jump. It measures the drawing, and the drawing is what you chose to draw. The page makes no network request. Open it by double-clicking, or send it to somebody else. ## Cutting a large change Three things reach the page so a reader can audit the cut. - **The scope rule the run applied**, so the reader can disagree with it. - **Files in the change that no node accounts for**, by name. A documentation-only part of a change is not silently dropped. It reads every node of the change, so a file one graph covers is not reported because another does not. - **Graphs found and not drawn**, each named with why it is worth a draw. Found and in no sheet of this file: a graph the file carries is drawn, so it is never listed here. Those last two stay two separate statements. They are different failures, and merging them loses which is which. ## Layers A layer redraws the same graph under a different set of dependencies. The test layer is the obvious one. **A layer renames a token, never a node.** The geometry is untouched, so the redraw computes nothing. A real double goes in at the call site, through a parameter with a real default, so the requirement is renamed and the node stays. Flip the toggle. A node that still reaches the real network under the test layer is a design defect you can see, rather than a sentence you have to trust. A layer may state its own entry. Everything the call edges cannot reach from it draws as unreached. A file that declares no layer map disables the toggle and says why. ## The text output **Print the tree only when the reader asks for text.** Print it into the reply, in a plain text fence. - **Name the graph when the file states more than one.** `--graph `. Asked without it, the command lists the graphs and stops rather than picking for the reader. - **One row is a call site**, not a node. A node called twice appears twice. - **A repeated node is marked and stopped**, or a cycle never terminates. - **Every E tag on a row carries its failure kind**, derived from the file and never written in it. A tag the file gives no kind for prints bare. - **Suggest the longest walk.** It is the only rule that names exactly one run in every worked example, with no tie. - **List every run you did not print**, by name, with the blurb its author wrote. The reader overrules your suggestion from that line alone. - **Print one line of provenance above everything**, so the reader knows whether they read a claim or a recording. ## The honesty property, and its limit The material is durable, so a sceptical reader can go and check the drawing against it. That is the whole reason this skill prefers real material over a tidy invented example. The validator proves the walk is a legal path through the graph the file declares. **It cannot prove which branch was taken or what an effect returned.** Those stay the author's claims. Say the limit. Do not let a page imply a check it did not make.