The Dispatcher, for Dummies

A dead-simple, no-jargon guide to running your AI peer-review service.

You don't need to know any code to use this. This guide answers one question: *"I got a card / a message / a notification — what do I DO?"*

This is the everyday guide (what the buttons and cards mean). For *server* stuff
(updating the service, settings, backups) see OPERATOR-GUIDE.md. This guide also
lives on the website at automationation.ai/guide — share that link. To **change
an API key or which AI model a reviewer uses**, sign in at automationation.ai, find
your repo, and click Manage keys & models.

New: you can now run all of this by chatting with Claude instead of the website —
"what needs my approval?", "approve PR 12", "set me away". See **Part 5 — Connect to
Claude**.

30-second version

A robot panel of AI reviewers reads every pull request ("PR") for your customers.
When they need you, you get a card (in Google Chat) and/or an email.
The card has buttons. 99% of the time you'll tap one button and you're done.
If you're ever unsure: tap Open PR, read the comments, and you can always just wait — nothing breaks if you do nothing for a while.

Part 1 — The buttons (this is all of them)

Every card has some of these. That's it. There are no other buttons.

Button	Plain English	When to tap it
🚀 Approve & Merge	"Yes, ship it." Approves and merges the PR.	The card says it's approved / ready and you agree. This is the normal "all good" button.
✅ Approve (or "Approve & mark ready")	"Yes, this is good" — but you still click Merge on GitHub later.	You want to approve but not auto-merge yet.
🔄 Send back (1 more round)	"Not sure — have the robots look again."	The reviewers were split and you want another review pass instead of deciding yourself.
✋ Block	"No. Stop. Don't merge this."	Something's wrong and you don't want it to go in.
Open PR	Opens the actual pull request on GitHub so you can read everything.	Any time you want the full story before deciding.
Check <provider>	A reviewer's AI (Claude/GPT/etc.) didn't answer. This links you to check that service.	Only appears when a reviewer's service had a hiccup.

Golden rule: if a card shows a 🚀 Approve & Merge button, the panel already agreed it's good. Tapping it is the safe, normal thing to do.

Part 2 — "I got a card. What is it telling me?"

Find the line that matches what your card/email says. Left = what it says. Right = what to do.

✅ "All required reviewers approved and CI is green."

Meaning: Everyone's happy, tests pass. It's ready. Do: Tap 🚀 Approve & Merge. Done.

Tired of clicking on the easy ones? On a repo you trust, sign in → that repo →
⚙ Manage keys & models → tick **"Auto-merge when all reviewers approve & CI is
green."** Then a unanimous, green PR just merges itself — no tap, no ping. It still
comes to you for splits, protected-file changes, and high-stakes gut-checks. Off by
default; flip it off any time.

✋ "The reviewers used up their rounds without fully agreeing — your call to break the tie."

Meaning: The robots couldn't reach agreement. It's a judgment call now — yours. Do: Either tap 🔄 Send back (let them try once more), or Open PR, skim the concerns, then Approve (if it's fine) or Block (if it's not).

✋ "Reviewers split and the change has gone quiet."

Meaning: Same as above — split decision, and nobody's pushing new changes. Do: Same as above. Your call: Approve, Send back, or Block.

⚖️ Referee line: on a split card you'll often see a line like *"Referee (high
confidence): the objection is a false positive because … — Recommend APPROVE."* That's a
separate AI double-checking whether the dissent actually holds up against the code, so you
get a recommendation instead of having to read the whole argument. It's **advice, not a
decision** — you still tap. It leans cautious: if a complaint *might* be a real bug, it
says "take a look" rather than waving it through.

🛡️ "This PR changes protected files, so it needs your sign-off before it can merge."

Meaning: Even though the robots are happy, this touched something sensitive, so you must say yes on purpose. Do: Open PR, take a quick look, then Approve & mark ready (or Block).

👀 "Every reviewer approved this fast on a high-stakes change. Worth a quick gut-check."

Meaning: It passed *suspiciously* easily on an important change. Probably fine — but give it a glance. Do: Open PR, eyeball it. If it looks right, Approve & Merge.

🧪 "The automated tests keep failing and the fixes aren't sticking — this one needs you."

Meaning: CI (the automatic test runner) is red and retries didn't fix it. Do: Open PR. This usually means the *code* has a real problem — the developer (or your AI dev assistant) needs to fix it. You can Block until it's fixed.

⏳ "A reviewer's API didn't answer this round, so this PR can't finish on its own."

Meaning: One of the AI services (Claude/GPT/Gemini/Grok) didn't respond. Do: Tap Check <provider> to confirm that service is back, then tap 🔄 Send back to re-run the review. Usually it just works the second time.

⛔ "A reviewer's AI service has been down for a while, so this PR is on hold."

Meaning: Longer outage of an AI provider. Do: Nothing urgent. Wait for the provider to recover, then Send back. (If it drags on for hours, tell your AI assistant — a provider might need a key or billing check.)

💸 "Heads-up — the project hit 80% of its daily AI-review budget." (reviews keep running)

Meaning: Just an FYI about spending. Nothing is paused. Nothing is broken. Do: Nothing. It's informational. (Customers see their own AI spend on their own dashboards.)

⏸️ "Reviews are paused — the project hit its daily AI-review spending limit."

Meaning: Only happens if a spend ceiling was deliberately turned on. Reviews stopped to avoid more cost. Do: If you want them to continue, comment OPERATOR RESUME on the PR (see Part 4). On the latest service, the budget is off by default and never pauses — if you see this a lot, ask your AI assistant to confirm the budget is set to notify-only.

Part 3 — "I tapped Approve & Merge. What does the result page mean?"

After you tap, you land on a little page. Here's what each one means:

🟢 "Merged ✅"

It's done. The PR is merged into the project. Nothing else to do.

🟡 "Approved — merging automatically"

Also good. The PR is approved, but it's waiting on a required check (or a merge queue) to finish. GitHub will merge it by itself the moment that's green. Do nothing — it lands on its own.

🟠 "Approved, but GitHub won't merge it yet — it said: <reason>"

Read the reason — that's GitHub telling you the real blocker:

*"Resource not accessible by integration"* → a one-time permission fix: your

GitHub App can approve but isn't allowed to merge yet. Go to github.com/settings/apps → your app → Permissions & events → Repository permissions, set Contents and Pull requests to Read & write, Save, then approve the new permission on each install. After that it merges fine. (The page itself now tells you these exact steps.)

*"not mergeable"* → there's a merge conflict. The branch needs fixing before it

can merge. Open PR, and the developer / your AI dev assistant resolves it.

*a red required check* → a test is failing. Wait for it to go green (or fix it).
The button already retried for you, so this means it's genuinely blocked, not

just slow. Do: Open the PR, clear the stated problem, then click GitHub's green Merge.

🔵 "Couldn't verify this link"

Meaning: A safety check on the button link failed (rare). Do: Just Open PR and act there, or tell your AI assistant — it's a config thing.

🔴 "This isn't one of your connected projects"

Meaning: The service is running old code, or you're signed into the wrong GitHub account. Do: Make sure you're signed in as your operator account. If it persists, the server needs updating — see OPERATOR-GUIDE.md §2 (or ask your AI assistant to "update the running service").

Part 4 — Typing commands (when there's no button handy)

You can control any PR by posting a comment on it (as your operator GitHub account). Type the command exactly, on its own line:

Type this comment	What it does
`OPERATOR APPROVE`	Approve / mark it ready to merge.
`OPERATOR BLOCK`	Block it — don't merge.
`OPERATOR INVESTIGATE`	Send it back for another review round.
`OPERATOR DISCUSS`	Ask the AI reviewers about something specific.
`OPERATOR PAUSE`	Pause reviews on this one PR.
`OPERATOR RESUME`	Un-pause reviews on this one PR.
`OPERATOR KILL`	Completely stop the dispatcher on this PR.

These are the same actions as the buttons — just typed. Use them when you're on
GitHub already, or a card's button didn't work.

Part 5 — Talk to it from a Claude chat ("Connect to Claude")

You can run this whole thing by chatting with Claude — no website, no GitHub. You connect it once, then just *ask* in plain English.

What you can do (just say it):

Say something like…	What happens
"What needs my approval right now?"	Lists the PRs that are ready-to-merge, split, or paused — across all your repos.
"What projects am I set up on?"	Your connected repos and which AI reviewers are switched on.
"How much time have I worked?"	Your time totals, broken down by repo.
"Approve PR 39 on my-repo — looks good"	Posts your approval on that PR for you.
"Set me away" / "I'm back at my computer"	Turns your chat pings off / on.
"Turn off email alerts for my-repo"	Email alerts on/off, per project.

How to connect it (one time, about a minute):

Sign in at automationation.ai and click Connect to Claude (top menu).
It shows you one URL — copy it.
In Claude: Settings → Connectors → Add custom connector. Paste the URL, click Add, then Connect.
A GitHub sign-in window pops up — approve it. Done. No tokens, no secrets to copy.

Good to know:

Claude only ever acts as you, on your repos.
Your connection stays put across service updates — connect once and forget it. (If a

big update ever disconnects you, just open Connect to Claude and click Connect again.)

Part 6 — Common situations ("help, it's doing X")

"Nothing ever approves / every PR comes to me." That's usually a *split* or *high-stakes* pattern, or a provider that keeps failing. Open a couple of PRs and read the reviewer comments. If a specific AI keeps not answering, tell your AI assistant — a key or billing issue may need fixing.

"A PR is stuck and no card came." Open the PR and look at the latest comments — the reviewers post their verdicts there. If it truly looks frozen, comment OPERATOR INVESTIGATE to kick a fresh round.

"I want to stop everything on one PR right now." Comment OPERATOR BLOCK (don't merge) or OPERATOR KILL (stop the robot entirely).

"I approved by accident." If it hasn't merged yet, comment OPERATOR BLOCK on the PR. If it already merged, you (or the developer) revert it on GitHub like any normal change.

"A reviewer (like Gemini) keeps failing / I need to change or rotate an API key." Sign in at automationation.ai, find the repo, click ⚙ Manage keys & models. There you can paste a new key, remove a reviewer, or change its model (if Gemini keeps failing, set its model to gemini-2.5-flash). Changes apply to new reviews right away.

"It says budget / spending stuff." Ignore it unless it literally says paused. The budget is a heads-up, not a stop sign — work never halts because of cost on the current setup.

"The buttons go to a website and ask me to sign in." That's normal and safe — it's confirming *you're* the one clicking. Sign in with your operator GitHub account and it continues automatically.

Part 7 — When to do NOTHING (this is allowed!)

You do not need to act on:

💸 Budget heads-up cards (the 80% FYI).
🟡 "Merging automatically" result pages (it finishes itself).
A PR that's mid-review with no card (the robots are still working).

Doing nothing never breaks anything. The worst case is a PR waits — and you can always come back to it.

Part 8 — The "I'm stuck, get me help" button

If anything here doesn't match what you see, or a step assumes knowledge you don't have: ask your AI assistant (the same one that built this). Say, in plain English, exactly what's on your screen — for example:

"I got a card that says '<paste the text>' and I don't know what to do."

It will tell you the next step, and if this guide was missing or confusing, it will fix the guide. You should never be expected to "just know."

One-page cheat sheet (screenshot this)

Card says approved / ready → 🚀 Approve & Merge → "Merged ✅". Done.
Result says "merging automatically" → do nothing, it finishes itself.
Result says "won't merge yet: …" → read the reason, fix it, click Merge.
Reviewers split → 🔄 Send back, or read it and Approve / ✋ Block.
Protected files / gut-check → Open PR, look, then Approve.
A reviewer's API failed → Check <provider>, then 🔄 Send back.
Budget heads-up → ignore.
No button? → comment OPERATOR APPROVE / OPERATOR BLOCK / OPERATOR INVESTIGATE.
Prefer chatting? → connect it to Claude once (Part 5) and just say *"what needs my approval?"*
Lost? → Open PR and read, or ask your AI assistant.