# Codex Delegation Prompt — Hostinger Gitea Runner Setup

> Copy-paste the block below to Codex running on the Hostinger VM. It bootstraps Codex into reading and executing the master roadmap.

---

## The prompt

```
You are executing a multi-phase infrastructure setup on the Hostinger VM
you're running on. Read and follow the master roadmap exactly.

REPO: learning_ai_common_plat
  Find it: `find ~ -maxdepth 4 -type d -name "learning_ai_common_plat" 2>/dev/null | head -3`

ROADMAP: docs/devops/gitea-runner/ROADMAP.md

INSTRUCTIONS:

1. Read the roadmap top-to-bottom BEFORE starting any work. It points to
   two companion docs (ACT_RUNNER_SETUP.md and
   PUBLISH_WORKFLOW.md) for implementation detail —
   read those too.

2. Execute phases P0 → P5 in order. Do NOT skip phases.

3. After completing each numbered task (P0.1, P0.2, P1.1, ...), update
   the roadmap file:
   - Change `- [ ]` to `- [x]` for that task
   - Fill in the `Commit:` field with the 7-char short SHA of the
     commit that performed the actual work (NOT the roadmap-update
     commit). For pure system changes (e.g., useradd), write
     "_system change_".
   - Fill in the `Status:` field with a one-line description of what
     you observed.
   - Commit the roadmap update with message:
       chore(roadmap): mark P<phase>.<task> complete — <short note>
   - Push to both remotes:
       git push origin main && git push gitea main

4. STOP and ask the human if ANY of these occur:
   - A pre-flight check (P0) fails or surprises you
   - The cross-Gitea SHA1 comparison (P3.6 or P5.4) does NOT match —
     this is the load-bearing architectural invariant; do not paper
     over it
   - The runner registration token doesn't work or expires before you
     can use it
   - A step in a companion doc doesn't work as written (note in
     "Surprises / deviations" and proceed only if safe)
   - You need to deviate non-trivially from any companion doc
   - The human's pre-execution answers (P0.3, P0.4) turn out to be
     wrong (e.g., they said instance-level scope but you only have
     repo admin)

5. NEVER:
   - Skip P3.6 or P5.4 (cross-Gitea SHA verification)
   - Publish packages outside the workflow you just built
   - Leave the throwaway @bytelyst/_runner-e2e-test package in either
     Gitea registry after E2E
   - Force-push the roadmap file
   - Mark `[x]` if a task didn't fully succeed (use `[ ] FAILED:
     <reason>` and stop)
   - Run the runner daemon as root

6. When P0–P5 are all `[x]`:
   - Fill in the §"Final report" section of the roadmap with NO
     placeholder strings left
   - Mark P6.1 and P6.2 as `[x]`
   - Push the final commit
   - Reply back with: "Roadmap complete. P6.1 + P6.2 ticked. Awaiting
     human review (R1–R9) on commit <sha>."

PRE-EXECUTION QUESTIONS — ASK THE HUMAN BEFORE STARTING IF UNCLEAR:

  a. What user owns `~/.gitea_npm_token` on this VM? (P0 pre-flight
     check #10 in the runner setup doc will discover this; confirm
     with the human.)
  b. Confirm registration scope: instance-level (recommended). If the
     human prefers org-level or repo-level, get that direction.
  c. Confirm consent for the E2E throwaway package
     (@bytelyst/_runner-e2e-test, published + deleted as part of P3).
  d. For P5.1, ask which package to bump for the first real release.
     Suggest a low-risk one (@bytelyst/errors or similar with stable
     downstream consumers).
  e. If `gh` CLI on this VM is not auth'd as saravanakumardb1 (the
     Gitea/GitHub user), ask how to obtain a Gitea admin token for
     the runner registration step (P1.3).

START WITH P0. Read the roadmap. Then go.
```

---

## One-liner handoff (paste this single line to Codex)

```
Read docs/devops/gitea-runner/CODEX_DELEGATION_PROMPT.md in learning_ai_common_plat, then execute the prompt within. Begin with P0.
```

Or, if Codex doesn't have the repo yet:

```
Clone https://github.com/saravanakumardb1/learning_ai_common_plat (or whichever remote you can reach), open docs/devops/gitea-runner/CODEX_DELEGATION_PROMPT.md, and execute the prompt within. Begin with P0.
```

---

## Monitoring from your end (Cascade / corp Mac side)

While Codex works, watch its progress commits flow into `learning_ai_common_plat/main`:

```bash
cd ~/code/mygh/learning_ai_common_plat
git fetch origin
git log --oneline --grep="chore(roadmap)" origin/main | head -30
```

Each `chore(roadmap): mark P<x>.<y> complete` commit = one checkbox ticked. When you see `chore(roadmap): mark P6.2 complete`, run the human review (R1–R9 in the roadmap doc).

---

## If Codex gets stuck

It should stop and surface the question itself. If it doesn't, you can:

1. Check the latest roadmap state on `origin/main`:
   ```bash
   git fetch origin && git show origin/main:docs/devops/gitea-runner/ROADMAP.md | grep -E "^\- \[" | head -30
   ```
2. Find the lowest unchecked `[ ]`.
3. Tell Codex: "Resume from P<x>.<y>. The roadmap state is current as of commit `<sha>`."

The checkbox state IS the resume pointer — no other state file needed.