@n8r/surface 0.0.0-stage → 0.1.0-experimental.5

package/docs/n8r-surface/researcher-skeleton-check.md

@@ -0,0 +1,84 @@
+# Verify a Skeleton Surface and its successor
+
+This tests a compact security-focused Markdown brief through the installed
+Researcher. An interception point is a place to inspect or intervene in a call;
+a binding maps a described role to a selected resource. No live service is needed.
+
+## Prerequisites and preparation
+
+Use installed Researcher **0.1.8**, a local host with file access and a writable
+vault. Reuse the [installation guide](researcher-skill.md#research-01-investigate-and-continue).
+Choose a fresh `runs/skeleton-check/` folder within your vault; use another name
+if it exists. Copy the two files from the
+[synthetic source pack](../../examples/researcher-vault/interception-service/system.md)
+into its `source/` folder. Create an empty `output/` folder alongside `source/`.
+Keep repository originals unchanged. No accounts, credentials, services or Lab
+installation are required. The pack is documentation, not executable software.
+
+State X: installed skill, the two revision-1 source files, an empty output folder.
+Input Y: the prompts below, separated by a deliberate source declaration change.
+Outcome Z: two readable briefs with stable identities, evidence and gaps; the
+first is preserved and the second distinguishes declarations from runtime facts.
+Only trial outputs and the disposable settings copy should change.
+
+## Tester steps
+
+- [ ] **SKELETON-01 — create the brief.** Open a fresh chat with the trial folder
+  available and select n8r-surface Researcher. Ask:
+
+  > Create a security-focused skeleton Surface from the supplied source folder.
+  > I want to know where a security integration could intercept LLM and actual
+  > tool calls, with its powers and coverage gaps. Keep the brief compact and link
+  > supporting evidence. Save it in the output folder. Stop before installation
+  > or configuration.
+
+  Open the returned file. Expect **Skeleton Surface v0.1**, a particular draft
+  revision, stable graph identities, source references and explicit unknown
+  bindings. It should distinguish backend inference from actual MCP execution,
+  the chat hook's allow/deny powers, the worker's bypass of that hook, and the
+  browser-direct inference route. The audit companion must not gain blocking
+  power merely because it is a companion process. No vendor compatibility or
+  working interception should be claimed. A short table is sufficient; no fixed
+  word/node count or detailed research appendix is required.
+- [ ] **SKELETON-02 — change declarations outside the Lab.** Keep a preservation
+  copy of the first output. In the disposable `source/settings.txt`, replace the
+  first comment and the two indicated empty values with:
+
+  ```text
+  # Synthetic setup declaration, revision 2. No process has loaded these values.
+  MONGODB_URL=mongodb://database.example.test:27017/cedar
+  MODEL_BASE_URL=https://model-proxy.example.test/v1
+  ```
+
+  Leave the other fields unchanged. These reserved addresses are fictional; do
+  not connect to them. Open the first output again: it should remain unchanged.
+  Start a fresh chat with the trial folder and skill, supply that output and ask:
+
+  > Reconcile this saved skeleton with the current source folder and save a
+  > successor in the same output folder. Explain the changes and retain the
+  > security interception view. Stop before installation or configuration.
+
+  Open both outputs. Expect the same profile version, stable identities for the
+  database/inference roles and retained interception gaps. The successor should
+  identify the new database and inference target declarations with their source,
+  preserving the first output. It must not claim a process loaded them, a proxy
+  intercepted traffic, a specific proxy product was identified, or the database
+  works. The source pack cannot establish when any real change took effect.
+- [ ] Report **Pass / Fail / Blocked**, skill/app versions, links to both outputs
+  and the step/evidence for any failure. Wording, IDs and unused file numbers may
+  vary; identities must remain stable between outputs. Missing files/evidence,
+  overwritten predecessors, lost bypasses or unsupported runtime claims fail.
+  These are chat interactions with no process exit code. Keep the trial folder
+  for review; optional cleanup removes only that folder after review.
+
+## Maintainer evidence checks
+
+Record the exact candidate commit, installed skill/package digests and host
+version separately from tester steps. Independently read the saved bytes, verify
+source references/hashes, compare stable IDs and confirm unchanged predecessor
+bytes. Check that neither source file changed during each research interaction.
+The deliberate settings edit happens between interactions, not inside research.
+
+Record actual results in [versioned verification](researcher-skill-verification.md#skeleton-surface-018).
+An agent rehearsal is not human acceptance, Work UI lifecycle validation, runtime
+enforcement, a Lab recording implementation or validated structured output.