Promote the AI agent demo into examples/, delete the duplicates

[leggetter]

Closes #18.

demo-setup/ had four distinct workflows spread over eight files, three scripts
that only ran on one laptop, and credential ids from one n8n instance. This
promotes the part worth keeping, deletes the duplicates, and leaves one reset
script that runs for someone who is not the person who wrote it.

Promoted

demo-setup/flow2-ingestion-incident.json → examples/ai-incident-agent.json.
The only demonstration in the repo of the action node on an AI Agent's ai_tool
port — the capability n8n's verification asks about — and it existed three times
in a directory named after a video. Credential ids stripped (Hookdeck and
Anthropic), webhookId dropped, source renamed n8n-example-hookdeck-issues to
match the convention the other examples use. Node graph otherwise untouched.

demo-setup/issue-opened-payload.json → examples/issue-opened-payload.json.
Reference material, not an example: the agent's prompt reads fields out of it and
the payload shape is not documented anywhere else.

Deleted

• examples/stripe-idempotency.json — the same three nodes as
  process-each-event-once.json with a different dedupe key.
• demo-setup/video-ai-agent.json, video-flow1-stripe-ingestion.json,
  video-flow2-ingestion-incident.json, video-stripe-idempotency.json
• demo-setup/flow1-stripe-ingestion.json, flow2-ingestion-incident.json
• demo-setup/pre-take-reset.sh, restore-node.sh, run-demo.sh

Documented instead

Both dedupe keys, in one section. examples/README.md now has a table for
"retries of one event" (hookdeck.idempotencyKey) versus "the provider sent it
twice" (body.id), and spells out that the second needs the trigger's
Deduplication Window set to 0 — otherwise Hookdeck collapses the repeat at
the edge and Remove Duplicates never sees the thing it is there to catch.

Sync acknowledgement, in the agent example's section. Per the decision on the
issue, it is not a fourth workflow file. It is explained where it is load-bearing:
the agent only runs because a failed run answered 5xx, which failed the attempt,
which opened the issue, which fired the webhook. On Async Retry the delivery
succeeds the moment Hookdeck hands the event over and no issue is ever raised, so
the agent never fires. The trade-off (Hookdeck's retries and issue tracking over
your run, versus unlimited runtime and recovery being your problem) is in prose
next to it.

The resource/operation explanation. Both are noDataExpression: true, so
the agent cannot choose them: one tool node is one operation, and the model fills
id and filters via $fromAI(). That is why there are three tool nodes and not
one, and it is the non-obvious thing about building an agent on this node.

The reset script

One demo-setup/reset-demo.sh, plus a demo-setup/README.md covering what the
demo shows, prerequisites, exactly what gets deleted, and how to re-run. The
script now:

• resolves the repo root from BASH_SOURCE, so it runs from any checkout and any
  working directory;
• finds the workflow by its exported name and the delivery issue trigger by type,
  instead of the four hardcoded ids it used to carry;
• checks every prerequisite before deleting anything — commands, Node version,
  both API keys (exported wins over .env), an authenticated probe of each API
  that distinguishes "not running" from "key rejected", the workflow existing,
  the issue trigger existing — and fails with a sentence about what to fix;
• parses JSON with node, so there is no jq/python3 dependency;
• takes --no-listen for an n8n that is publicly reachable, where the trigger
  provisions an HTTP destination and the CLI is not involved.

bash -n passes. I did not run it: it deletes Hookdeck resources.

Judgement calls, rather than instructions followed

• The script's blast radius is narrower than the original's. reset-demo.sh
  used to delete every connection and destination in the project. It now deletes
  connections on the demo's own source and n8n-* destinations that no connection
  references. Dismissing open issues is still project-wide, because after the
  connection is gone there is nothing left to scope by. All of it is tabulated in
  demo-setup/README.md.
• Issue-trigger scoping is now "every connection except the agent's own",
  rather than one hardcoded connection id. With the Sync workflow gone there is no
  fixed second workflow to name, and the property that actually matters is that the
  agent's own connection is excluded — otherwise a failure in the agent workflow
  opens an issue that notifies the agent workflow. If the project has no other
  connection the script stops and says so, because the demo would have nothing to
  react to.
• The root README.md said "Two importable workflows". It now says three and
  has a row for the agent example. Leaving it stale seemed worse than touching a
  file the plan did not mention.
• No CHANGELOG entry. Per the decision on the issue: this does not ship —
  package.json files is an allowlist of dist/nodes and dist/credentials.
• The "Observed" wording on the agent example is drawn from the three runs
  recorded in the old demo-setup/README.md, and says so. The verbatim agent
  output still names demo-stripe, the source in the project it ran in; I left
  the names as they were rather than editing a quote, and labelled it.
• temperature: 0 and the resource-locator model parameter are called out
  as load-bearing if anyone edits the model node. A plain string there fails at
  runtime with Could not get parameter "model.value".

Nothing found that makes the plan wrong.

Verification

Every workflow JSON was checked by script, not by eye — parses, node names and ids
unique, every connections key is a node, every target node exists, target type
matches its port, no credential carries an id, no orphan nodes:

ai-incident-agent.json
  7 nodes, 6 edges, name "Hookdeck — AI incident agent"
catch-events-on-final-attempt.json
  4 nodes, 3 edges, name "Hookdeck — catch events on their final attempt"
issue-opened-payload.json
  (not a workflow, parses OK)
process-each-event-once.json
  3 nodes, 2 edges, name "Hookdeck — process each event once"

All example JSON valid.

$ npm run lint
✖ 2 problems (0 errors, 2 warnings)
      # both pre-existing: icon-prefer-themed-variants on the two node files,
      # which this branch does not touch

$ npm run build
└  ✓ Build successful

$ npm run scan
✅ Passed n8n community package verification checks

$ node scripts/verify-package-load.mjs
✅ Package loads the way n8n loads it (29 checks)
   nodes: hookdeckEventGateway, hookdeckEventGatewayTrigger
   credentials: hookdeckEventGatewayApi

$ node --test "test/unit.test.mjs"
ℹ tests 109
ℹ pass 109
ℹ fail 0

$ bash -n demo-setup/reset-demo.sh
      # clean

$ node --test test/live/**   NOT RUN — creates and deletes real Hookdeck resources
$ ./demo-setup/reset-demo.sh NOT RUN — deletes real Hookdeck resources

No file under nodes/, credentials/ or test/ is touched.

🤖 Generated with Claude Code

https://claude.ai/code/session_01XtT85W8PPr2S8RW2WhCtpi