Record A Case

Record a case

Treat the case as an evidence ledger. Preserve what happened before extracting a method. Promotion is a separate process.

Canonical write root

Before writing, resolve the canonical Railly/skills checkout through RAILLY_SKILLS_REPO, ~/Programming/railly/skills, or ~/railly-skills. The canonical resolver is scripts/resolve-source-root.mjs. Never write a case, eval, or log into the current repo's .agents/skills, .claude/skills, another installed skill copy, or the target repository. If the source root is unavailable, return the draft and report the blocked write.

1. Bound one case

Name the maintenance unit: one issue, PR, incident, review, disproven claim, or coherent batch with one transferable lesson. Split unrelated outcomes even when they occurred in the same session.

Complete when: every included outcome shares one mechanism and one transferable lesson.

2. Build the ledger

Inventory evidence before prose. Collect durable retrieval handles:

• repository, commit, branch, issue, and PR
• Issue Contract path, final state, and acceptance IDs when one exists
• exact command and behavior-specific output
• regression test and its fix-absent result
• restored green command
• built, installed, deployed, or serialized artifact check
• review and maintainer decision

Classify every material claim as retrievable evidence, unreviewed report, inference, or unknown. Recheck drift-prone state such as PR status, branch head, merge, and release. Session summaries supply leads.

Complete when: every material claim has a class and every drift-prone state has a dated check or is explicitly unknown.

3. Assign independent statuses

Record technical validation, human review, maintainer acceptance, and delivery independently.

For an agent-generated or synthetic backfill, default to:

Status: observed
Validation: unvalidated
Human review: pending
Maintainer acceptance: pending
Delivery: local

Advance each field only from evidence for that field. Branch existence, test behavior, human review, maintainer acceptance, and delivery remain separate facts.

Complete when: each status cites supporting evidence or retains its conservative default.

4. Extract the lesson

Describe the observed condition, red signal, method, outcome, and limits using the ledger's claim classes. State when the fix covers a failure family or mechanism proxy rather than the reporter's exact environment.

Choose the smallest provisional destination:

• skill method
• reference rule
• exemplar
• deterministic check
• behavior eval
• coverage gap
• no change

Complete when: the lesson states its evidence boundary and exactly one smallest destination is selected.

5. Pass the confidentiality gate

For a public case, retain public repositories and the author's public work. Sanitize secrets, customer data, private review text, internal chat, local paths, neighboring-project identity, and unapproved employer context. Ambiguous provenance routes the case to an approved private destination or a public coverage-gap record.

Complete when: every source is authorized for the chosen visibility and the public draft contains no private retrieval handle or identity.

6. Materialize the case

Read [the case schema and allowed values](references/template.md), then write to cases// under the canonical source root. Use another destination only when the user explicitly names it and it is not an installed skill copy. Follow the canonical corpus naming and validation rules. Without write authority, return the complete draft in the response.

Set human-review status only from explicit human feedback or retrievable review evidence. Present new agent-authored records as pending review.

When the case closes a live mission, update its canonical Issue Contract to Status: closed and record the case path and final delivery state. Do not fill missing proof retrospectively.

Complete when: the artifact passes the destination's validator, all schema fields are resolved, the live contract points to the case when applicable, and a reader can reconstruct the work from retrieval handles without trusting the narrative.

Source & license

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

• Author: Railly
• Source: Railly/skills
• License: MIT

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