decisions

decisions

The host runs this skill once for each repository, on the assembled ref after the root steps. A script finds the decision files and reads what a script can read. You read each decision file and write one context file for it.

1. The step prompt gives the arguments: --source, --source-url, --source-path, and --roots (the root-to-file map, ADR-052 section 6). Run node /ctxpipe/skills/decisions/scripts/decisions.mjs <arguments>.

2. The script writes JSON that matches schema.json to stdout. Each entry in decisions is one decision file of the source (path). It gives the context file to write (target) and how (write):

   • create: a new file.
   • update: a file that this source already wrote, or that describes the same decision. Change it with ctxpipe-stage <target>, and keep what is still true.
   • additive: a file that a human or another source owns. Add only new claims and new lines at the end of its ## Related section.

   Never write a second file for a decision that has existing. The check in step 6 rejects it.

3. For each entry, read the decision file path in /ctxpipe/source/<handle>. Run ctxpipe-stage <target>, then write the file at the path that it prints.

4. For a create or an update, write this shape. Use name from the entry as it is. Write Status as the entry status with a capital letter. The entry date is empty when the file declares none: then write no Date line.

   ---
   kind: Decision
   name: "<name>"
   summary: "<one sentence in your words: what was decided>"
   claims: <the claims of step 5>
   ---
   
   # <name>
   
   **Status:** <Status>
   **Date:** <date>
   **Source:** `<path>`
   
   ## Context
   <why the decision was necessary, in your words>
   
   ## Decision
   <what was decided, in your words>
   
   ## Consequences
   <what follows from it, in your words>
   
   ## Related
   - [<name of a file in related>](<relative path>)
   

   Keep the **Source:** line: the script uses it to find a file whose decision file has left the source. Do not write generated_by, sources, source_sha, skill, bundle, or confidence. The host sets them.

5. Claims. Copy these from the entry, with evidence.kind: script:

   • each path in influences becomes a claim predicate: INFLUENCES with evidence.path set to the entry path;
   • each item of supersedes becomes a claim predicate: SUPERSEDES, to from the item, and evidence.path from the item.

   Do not add an INFLUENCES or a SUPERSEDES claim of your own. Scope and replacement follow fixed rules (below), and a claim from reading prose is not one of them. You may add other claims with evidence.kind: model when the text states them and the target is a file in the context.

6. When the entry has related, add each file as a line in ## Related. For additive, add only the claims and the Related lines, and one Related line that names the decision file: - Decision `<path>` of the source. The script finds the files of earlier pages by the decision file path.

7. For each path in stale, run ctxpipe-delete <path> --reason "The decision file left the source".

8. Run node /ctxpipe/skills/decisions/scripts/decisions.mjs <arguments> --check. Fix each problem that it lists, and run it again until it prints no problem.

9. Run ctxpipe-finish. When the result has more: true, run ctxpipe-finish --more: the host commits the step and runs the skill again on the new tip. The script then gives the next decision files (at most 20 for each page); it finds the ones that this run already wrote.