Debug session skill
Skill lubochka/xiigen-mvp-engine/.agents/skills/debug-session-skill
Self-building AI code generation engine that generates application flows instead of implementing them. AGPL-3.0.
npx -y skills add lubochka/xiigen-mvp-engine --skill debug-session-skillAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 16 days oldThe repository was created 16 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
3.3 KB, 784 tokens by cl100k_base, as published. Nobody here has run it
Debug Session Skill
Maintains a persistent debug session file that survives /clear. Every hypothesis, eliminated path, and confirmed API shape is appended — so new sessions know exactly where debugging stopped and what NOT to re-investigate.
The Problem
When a write-time fix fails at execution (e.g., getQueueDepth is async but plan assumed sync), Codex has no protocol. It guesses, stalls, or writes a comment. In the next session, the same investigation restarts from scratch. The debug file prevents this.
Debug Session File
Location: .Codex/debug/active-session.md
Lifecycle: Created when a fix fails. Appended to throughout the investigation. Archived (renamed to YYYY-MM-DD-description.md) when the fix is confirmed working.
File Format
# Debug Session: [issue title]
**Date:** YYYY-MM-DD
**Session:** S[N]
**Step:** Step [N] of SESSION-[N]
**Issue:** [one-line description]
## Hypotheses
### H1: [hypothesis title]
**Status:** ELIMINATED / CONFIRMED / INVESTIGATING
**Evidence for:** [what supports this]
**Evidence against:** [what disproves this]
**Elimination command:**
\`\`\`bash
[command that proved/disproved this]
\`\`\`
**Result:** [exact output]
### H2: [next hypothesis]
...
## Confirmed Facts (do NOT re-investigate)
- [symbol]: export shape is [shape], params: [params], returns: [return type] (line [N] of [file])
- [API]: is [sync/async], signature: [exact signature]
## Current Best Theory
[What we currently believe is happening]
## Next Action
[Exact next step to take when this session resumes]
When to Create the File
- A WF fix fails during execution (test fails after applying the fix)
- An API shape differs materially from the plan (WF-3 stop condition triggered)
- A TypeScript error cannot be resolved in ≤ 2 attempts
When to Append to the File
After every hypothesis test: record the command, the result, and the new status.
When to Archive the File
When the fix is confirmed working: rename to .Codex/debug/YYYY-MM-DD-[issue].md.
When to Load the File
At the START of any session where debugging was in progress:
ls .Codex/debug/active-session.md # check if debug session is active
# If exists: read it BEFORE reading STATE.json
Integration
- Invoked by self-verification (SK-416) when a fix attempt fails
- Invoked by code-examination-skill (SK-415) before modifying a file with a known issue
- Pairs with context-overflow-skill (SK-428): if context emergency stop occurs mid-debug, the debug file preserves the investigation state
XIIGen Example — S3 ElasticsearchProvider Issue
# Debug Session: ElasticsearchProvider constructor API
**Issue:** Plan uses `new ElasticsearchProvider(mockCls, {url})` but this fails
## Confirmed Facts
- ElasticsearchProvider constructor: `(cls: ClsService, client: IAsyncElasticsearchClient, config?)`
- `{url}` is NOT a valid second arg — must be a client implementing IAsyncElasticsearchClient
- `@elastic/elasticsearch` is NOT installed — fetch-based wrapper is correct approach
- `IAsyncElasticsearchClient` is defined in `server/src/fabrics/database/base.ts`
- `delete()` returns `Promise<Record<string,unknown>>` — NOT a typed result struct
What ships with it
Read from the repository
Just SKILL.md. No reference files, no scripts.