DevRecap

View saved version →

---
name: devrecap
description: Reconstruct developer work from explicitly authorized local Codex/Claude history and read-only Git evidence, then use the host AI to write polished, evidence-backed recaps.
---

# DevRecap

DevRecap is an AI-native work-reconstruction skill. The bundled CLI is the factual layer; the current Codex/Claude model is the writing and synthesis layer.

The preferred experience is not a deterministic CLI report. Use the host AI to transform structured Activities and Workstreams into a coherent report, while preserving evidence-backed status exactly.

## What to do when invoked

Interpret requests such as:

- "what did I work on this week?"
- "prepare my daily"
- "help me remember the last 14 days"
- "prepare my sprint review"
- "give me a detailed recap of last month"

Resolve the requested period and reporting intent, then follow the AI-first pipeline below.

## Use the bundled DevRecap runner

When this Skill is installed from the DevRecap plugin/marketplace, always use the runner bundled with the same installed plugin version. Do not prefer a `devrecap` executable found on PATH: it may be an older global install, npm link, or source checkout and can render with stale code.

Determine the plugin root from this skill file: the plugin root is two directories above `skills/devrecap/SKILL.md`.

For every factual CLI step in marketplace/plugin usage, run:

`node <plugin-root>/scripts/devrecap-plugin.mjs <args>`

The bundled runner prepares the local workspace links automatically. Node.js 24+ is required.

Before `prepare`, always run this preflight:

`node <plugin-root>/scripts/devrecap-plugin.mjs --plugin-info`

The returned JSON must report `renderer` as `editorial-v2`. If it reports `legacy-or-unknown`, stop and tell the user the installed plugin is stale or inconsistent. Do not silently continue with a legacy renderer. Use the reported `pluginRoot` when diagnosing installation problems.

Only when developing DevRecap itself from its source repository may the repo-local commands be used explicitly:

- Windows PowerShell: `npm.cmd run recap -- <args>`
- other shells: `npm run recap -- <args>`

Never mix an installed marketplace Skill with a different `devrecap` binary on PATH during one report run.

## Consent comes first

Before any local collector is used, DevRecap setup must already exist.

Run the factual preparation step with the bundled runner. If it reports that setup is required, stop collection and ask the user to authorize sources by running:

`node <plugin-root>/scripts/devrecap-plugin.mjs setup`

Never bypass setup. Never inspect Codex history, Claude history, or Git repositories yourself as a workaround. Never enable a source the user did not authorize.

## AI-first report pipeline

1. Run the `--plugin-info` preflight and require `renderer: editorial-v2` before collecting or rendering anything.
2. Resolve the requested period faithfully. If the natural-language resolver produces a shorter range than the user asked for, rerun `prepare` with an explicit equivalent period such as `last 14 days` or explicit `--from/--to` dates.
3. Run the bundled runner's `prepare` command and write `.devrecap/run.json` in the current project/workspace.
4. Read `.devrecap/run.json`.
5. Analyze only `contract.facts` and obey `contract.rules`.
6. Use the host model to synthesize a human-quality report and write one JSON object matching `contract.outputShape` to `.devrecap/analysis.json`.
7. Every analysis item must reference one or more IDs from `contract.allowedActivityIds`.
8. Render the validated AI analysis to `reports/derecap-YYYY-MM-DD.html`, where the date is the report range end date. Always pass this exact `--out` path explicitly; do not rely on a CLI default filename. Add a PDF with the same basename only when requested.
9. Return the generated path plus a concise natural-language recap.

Typical marketplace/plugin commands:

`node 

Preview of the first 4,000 characters. Read full instructions →