@n8r/surface 0.0.0-stage → 0.1.0-experimental.2

package/docs/n8r-surface/command-outcomes.md

@@ -0,0 +1,564 @@
+# Command intent, outcomes and verification
+
+This is the product acceptance specification for `n8r-surface`. It defines what a
+person should be able to accomplish, what changes, what they can observe, and how
+to verify those outcomes. Implementation plans and lower-level tests must serve
+these tasks. A passing codec test does not establish a working research command.
+
+**Status:** product specification approved on 4 November 2026; delivery order
+revised on 5 November 2026 to standalone AI integration, save/reopen/continue verification, then npm distribution. Lab adaptation remains separate.
+[The skill lifecycle and research checks](researcher-skill.md) govern the active
+increment. STATUS-01/CMD-02 are retired before implementation; no status
+subcommand will be added. The existing STARTUP-01 behavior remains the CLI baseline.
+
+The original command specification follows, with the planned inventory updated. The first #16
+increment implements bare startup only, using the #15 experimental artefact
+library. See [startup decisions and scope](startup.md). The setup-only increment now adds accepted local vault writes; other commands
+remain broader acceptance work. The bounded RESEARCH-01 conversation is implemented; see the [Codex evidence](codex-research-verification.md). The first skill
+provides instructions for working notes; it does not complete the structured-output
+or installed-CLI research cases below.
+The [recorded library checks](verification.md) are supporting evidence only.
+
+This specification applies the task-driven development approach requested by the
+user: define a real task, attempt it through the product, observe the gap, develop
+what is missing, then repeat the same task. It does not publish private Lab
+implementation material or require that application to be running.
+
+## Grow the specification with working verification
+
+Work one useful behavior at a time, including behavior that already exists. Add a
+small intent/example, try it, and retain a repeatable check with its observed
+result. If existing behavior meets the intent, adding that evidence is a completed
+increment; no implementation rewrite is required. If it does not, make the bounded
+change and repeat the same check.
+
+The command inventory gives direction. Detailed acceptance examples and evidence
+grow alongside each increment; completing the entire future specification is not
+a prerequisite for verifying the next useful behavior. Keep earlier checks and
+rerun those affected by a change. Extend the spec when a new user need or observed
+gap justifies it, while keeping expectations independent of implementation output.
+
+Keep each human check as short as its proof allows. Reuse established installation,
+runner and fixture prerequisites, with a setup link for anyone who lacks them. A
+small addition can be verified by its exact command, expected result and relevant
+readback; automated checks retain broader regression and edge-case coverage.
+Expand human steps when changed assumptions or observed gaps require it. A short
+proof must still exercise the intended outcome and identify what remains unproved.
+
+Keep progress visible beside the intent: what is specified, what is implemented,
+what was actually verified, and what remains blocked or unattempted. Update the
+example, relevant repeatable checks and evidence together in the same increment.
+A verified claim names its revision and limits; old evidence is not proof of a
+changed path. New scenarios are not verified merely because nearby checks pass.
+
+Current progress, with evidence defining the verified scope and revision:
+
+| User or consumer intent | Current progress | Verification / next increment |
+| --- | --- | --- |
+| Create a security-focused skeleton and reconcile a later declaration change. | Skeleton Surface v0.1 guidance in installed skill 0.1.8; bounded agent brief/successor rehearsal and readback passed. Human acceptance pending; no new CLI command or Lab automation. | [SKELETON-01/02 prompts and readback](researcher-skeleton-check.md); [versioned results and remaining acceptance](researcher-skill-verification.md#skeleton-surface-018). |
+| Read saved research and selected definitions through the library, retaining identities, findings and references. | Implemented and verified for the experimental library profile; no command claim. | [CONTRACT-01 and repeatable fixture/package checks](verification.md#contract-01-retain-useful-attributable-research-and-selected-definitions). |
+| Run `n8r-surface` alone and see settings, saved artefacts and known process status. | Specified, implemented and verified for the read-only overview and guidance question; setup/conversation unavailable. | [STARTUP-01 installed evidence at `b90358a`](startup-verification.md). Historical read-only baseline; setup supersedes its guidance and exit behavior. |
+| Choose a local vault once and reuse it from another source directory. | Setup-only implementation; installed rehearsal and human acceptance tracked separately. | [SETUP-01 evidence and tasks](vault-setup-verification.md). Research and continuation remain pending. |
+| Research supplied sources through the installed skill and save working notes. | Implemented; bounded Lumen and LibreChat agent rehearsals have saved/read back notes. Actual human input survived fresh-agent continuation; Work UI acceptance remains pending. | [RESEARCH-01 evidence and limits](researcher-skill-verification.md). No structured-output or standalone-runner claim. |
+| Offer to retain an initial partial graph when saving is undecided. | Implemented in skill 0.1.7; acceptance and evidence are tracked separately. | [Optional save offer](researcher-skill-verification.md#optional-save-offer-017). |
+| Resume from saved recommendations and retain useful next investigations. | Implemented in skill 0.1.6; installed fresh-agent save/continue and independent readback passed. Human acceptance remains pending. | [Recommendations evidence](researcher-skill-verification.md#saved-recommendations-016) and [prepared interaction](researcher-prepared-check.md). |
+| Try the installed skill using prepared files or another chat-system repository. | [Vault guide](researcher-vault.md) links prepared fictional inputs and descriptive walkthroughs; the prepared comparison has an installed-0.1.5 agent rehearsal. | [Vault evidence](researcher-skill-verification.md#repository-vault-rehearsal-015). Human acceptance and the next user-chosen repository remain pending; fresh chats are not enforced isolation. |
+| Discover an AI Surface graph using available sources and applicable prior knowledge. | Instruction guidance in skill 0.1.5; code/docs-only and mixed evidence are valid starting points. | [DISCOVER-01 interaction](researcher-discovery.md#graph-discovery-reconciliation-and-live-evidence); distinguish guidance trials from live acceptance. |
+| Reconcile a supplied Surface against current graph evidence. | Implemented instruction guidance; a controlled LibreChat declaration trial using installed 0.1.4 detected the target change and preserved prior research. Current 0.1.5 corrects ownership terminology. | [RECONCILE-02 procedure](researcher-reconciliation-trial.md) and [bounded evidence](researcher-skill-verification.md#controlled-declaration-trial-014); runtime and validated structured output remain unproved. |
+| Inspect a graph-relevant live connection with user-assisted sign-in. | Specified through optional host tools; real admin inspection remains unverified. | [LIVE-01 interaction](researcher-discovery.md#graph-discovery-reconciliation-and-live-evidence); require actual UI/API readback without credential exposure. |
+| Research a supplied repository through the standalone command with a human, then save and reopen the result. | Conversation and Markdown saving/reopening implemented; bounded evidence is tracked in SAVE-01. Structured HTML acceptance remains open. | Verify E2E-01 in #17; E2E-02 before claiming standalone LibreChat research. |
+
+## Governing development loop
+
+For each increment, whether verifying existing behavior or adding a capability:
+
+1. Name the useful user task and why it matters.
+2. Define **State X**, the actual **Input Y**, and independently assessable
+   **Outcome Z**. Include the expected changes, observations and evidence source.
+3. Attempt Y through the current installed product in X. Record what it actually
+   does. If the entry point, credentials or environment are unavailable, record
+   the blocker; do not invent a baseline failure or bypass the product.
+4. Compare actual behavior with Z. If it already meets the intent, retain a
+   repeatable check and record the verified result; no code change is necessary.
+   Otherwise bound the next implementation increment to the observed gap.
+5. Add supporting automated checks for that gap and implement the increment when
+   a change is needed.
+6. After a change, repeat the original interaction through the installed product.
+   Reopen/read back its outputs independently, compare them with Z, and record the
+   result and limits.
+
+Update the progress snapshot and linked evidence before choosing the next intent.
+Each completed increment should make another useful behavior demonstrably verified,
+or expose a specific blocker that remains visible.
+
+The developing agent builds the Researcher's capabilities. It must not perform
+the research itself and place prewritten answers in the output as proof that the
+command works. A canned runner can test orchestration, but it is not a real-runner
+research rehearsal. The same distinction applies to mocked storage and source reads.
+
+A useful increment ends in visible, reproducible evidence or a concrete statement
+of what remains unverified. Preserve the user task when implementation changes;
+do not redefine success around whatever the code currently produces.
+
+## Initial public command inventory
+
+This is the approved initial terminal surface for the first standalone milestone.
+CMD-01 implements the overview and explicitly accepted local vault setup;
+all argument forms below remain planned. See the [current parser contract](startup.md#settings-and-parser-decisions).
+It includes startup, discover and explicit help/version behavior; the formerly
+planned CMD-02 status command is retired.
+Corpus operations below are conversation
+intents, not additional undocumented subcommands. Any later command or option must
+be added to this inventory with its own outcome and verification case first.
+
+| ID | Invocation | User intent and delivered outcome | Observable change and verification |
+| --- | --- | --- | --- |
+| CMD-01 | `n8r-surface` | See readable non-private configuration and its file locations, local artefacts one per line, and process status where known. The setup-only increment offers vault selection and explicit write confirmation; configured reuse exits 0 without questions. Research remains unavailable. | Compare the initial overview with prepared settings/files using STARTUP-01. Read back any accepted setup change. Verify the overview itself changes no files and does not inspect/transmit source repositories or invoke inference. |
+| CMD-03 | `n8r-surface discover` | Research the current directory with the human. | Establish cwd as the source scope, follow the identified Researcher instructions, inspect permitted material, ask relevant questions and save an attributable HTML research record. Verify with E2E-01. |
+| CMD-04 | `n8r-surface discover ./product` | Research an explicitly selected directory. | Resolve the path from invocation cwd and report the actual scope; prove the intended source was read, with no fallback to a different repo. Verify source identity/read evidence and the resulting artefact. |
+| CMD-05 | `n8r-surface discover ./documents/architecture.md` | Research one supplied document. | Inspect that file and retain file-only coverage. Reopen the resulting record and confirm it does not claim repository-wide inspection. |
+| CMD-06 | `n8r-surface --help` | Discover the complete supported command surface and its effects. | Display syntax, purpose, source/destination behavior and links to details; exit without setup, source access, inference or writes. Compare with this inventory. |
+| CMD-07 | `n8r-surface discover --help` | Understand the selected command before invoking it. | Display that command's inputs, outputs, side effects and limitations; no operation or setup begins. |
+| CMD-08 | `n8r-surface --version` | Identify the installed tool being used or verified. | Display the package/build version without inference or mutation; compare with the installed package manifest. |
+| CMD-10 | `n8r-surface reopen [FILE] --source SOURCE [--model MODEL]` | Continue a named Markdown note or saved HTML/JSON pair; omission suggests the latest-edited Markdown note. | Renew transmission consent; compare source evidence; see SAVE-01. |
+| CMD-11 | `/save` at `You>` | Explicitly save or create a successor Markdown note. | Read back saved bytes and preserve predecessors; see SAVE-01. |
+| CMD-12 | `/save surface` at `You>` | Save the exact validated pair previewed in the current conversation. | Immutable paired research HTML and Surface JSON, independent readback; see SURFACE-01. |
+| CMD-09 | Unknown command, unknown option or invalid arguments | Receive an actionable explanation without unintended work. | Report the invalid input and applicable help, return failure, and leave source/corpus/settings unchanged. Never interpret a typo as permission to research. |
+
+`discover [source]` also accepts an explicit absolute local source path; the
+resulting portable artefact uses source IDs/portable locators rather than leaking
+that machine's path. A directory need not be a Git repository. Git identity and
+working-copy changes are recorded when present; otherwise file identity, hashes
+and inspected coverage establish the source snapshot.
+
+This inventory does not introduce `init`, `config`, `list`, `show`, `resume`,
+`export`, `deploy` or other standalone subcommands. If real task attempts show a
+need for them, specify and review their intent before adding them. Effective
+settings are configurable as required by the design; the exact override syntax
+must be added here when #16 selects it, before implementing that syntax.
+
+Successful noninteractive informational commands exit successfully. Invalid input,
+missing required setup and unavailable capabilities return a failure with a
+specific reason; they must not wait on a prompt. Bare startup still prints its
+overview without interactive input. If setup or conversation needs interaction,
+explain the required interactive path and finish promptly. Discovery follows the
+same no-wait rule when interaction is unavailable.
+A future unattended-research interface needs an explicit answer/selection policy;
+it is not implied by this first interactive workflow. Exact error codes belong in
+\#16's parser contract and must preserve these observable distinctions.
+
+Until #17 connects the actual Researcher, #16's `discover` reports that research
+is unavailable. Printing a prompt or writing a placeholder HTML file is not CMD-03,
+CMD-04 or CMD-05 success.
+
+## STARTUP-01: run the command and see what is here
+
+**Given:** a readable configuration file at `~/.n8r/n8r-surface.json` selects
+`~/n8r/artifacts` as the local artefact folder and JSON as the Surface output
+format. No AI runner or Lab connection is configured. That folder contains one
+valid synthetic Surface definition named LibreChat, stored as
+`librechat-uniqueid.json`. No process status is known.
+
+**When:** the user runs `n8r-surface` with no subcommands or switches.
+
+**Then:** before any setup question or conversation, the terminal shows a readable
+interpretation of those settings, where they came from, the saved artefact on one
+line, and the fact that process status is unknown. For example:
+
+```text
+$ n8r-surface
+
+Configuration
+  File: ~/.n8r/n8r-surface.json
+  Artefact folder: ~/n8r/artifacts
+  Surface output: JSON
+  AI runner: Not configured
+  Lab connection: Not configured
+
+Processes
+  Status unknown
+
+Local artefacts (1)
+  LibreChat | Surface definition | ~/n8r/artifacts/librechat-uniqueid.json
+
+Setup and AI conversation are not implemented in this build. Research needs a supported AI runner.
+```
+
+This is representative expected output; the process also identifies its own live
+PID and observation time. The interactive guidance question follows this overview.
+The configured
+folder in this example uses `artifacts`; the command displays the actual effective
+path, whatever the folder is named. LibreChat is a synthetic saved-file label here,
+not evidence that the command has researched LibreChat.
+
+**Verify:** prepare those files under a temporary home directory, run the installed
+command with closed input, and capture its output. Check the configuration path,
+the readable values, exactly one artefact row with the right name and location,
+and unknown process status. The command must finish promptly and identify missing
+runner setup. Compare the files before and after: the overview changes nothing.
+Also exercise an interactive launch and confirm the overview appears before the
+first setup question. These are ordinary process, text and file checks; no special
+test language is required. Exact spacing and decorative punctuation are optional.
+
+The overview follows a few straightforward rules:
+
+- Interpret supported non-private settings with useful labels instead of dumping
+  raw JSON. Show every contributing configuration file and explain defaults or
+  overrides when used. Adding a setting includes deciding how it appears here.
+  Omit secret values and private fields, including credentials inside URLs or
+  nested settings. A synthetic private-value marker must never appear in output.
+- List saved research records and Surface definitions in the configured local
+  folder, with one logical line per artefact. Use available name/type metadata and
+  a distinguishing filename or location. Keep embedded newlines/control characters
+  from splitting a row or issuing terminal commands. Reading saved-file metadata
+  does not authorize scanning the source repositories those files describe.
+- Say when the folder is empty. For missing/unreadable configuration or folders,
+  show the relevant location and problem; do not present failed inspection as an
+  empty corpus. An unreadable or unsupported artefact gets a filename and a clear
+  limitation on its own line rather than disappearing from the listing.
+- Report a relevant process's identity and status when reliable status information
+  is available, including when it was observed if it may be stale. Otherwise say
+  unknown. A configured runner or saved PID alone does not establish that a process
+  is running. This overview does not introduce a background service.
+- Produce the overview locally without AI inference, Lab credentials or implicit
+  network calls. It is read-only until the human accepts a setup change.
+
+Repeat the same simple check with two artefacts, an empty folder, inaccessible
+files, a private-value marker, and known versus unknown process status. Compare
+each output to its prepared state. The separately planned status command was
+retired before implementation. Configured vault reuse exits without a prompt;
+first-use setup requires explicit confirmation.
+
+**Current result:** verified through the installed command at `b90358a` on
+Node 22/24/26, including closed input and a real terminal. See the
+[observed output, independent readback and limits](startup-verification.md).
+This is historical evidence. Current setup behavior has its own
+[verification record](vault-setup-verification.md); research remains unavailable. [Startup decisions](startup.md) define the bounded path.
+
+## Conversation intents are part of the product surface
+
+An explicitly requested research conversation must support these user tasks.
+Wording is illustrative; outcomes, not exact phrasing, are the contract. Each
+session establishes a source scope and destination explicitly enough to prevent
+an accidental source or artefact change.
+
+| ID | Human input or intent | Required change/outcome | Independent verification |
+| --- | --- | --- | --- |
+| CHAT-01 | "Research this supplied repository/document." | Establish the actual source, inspect permitted material using the Researcher instructions, retain evidence/gaps, and produce a research record. | Compare source-access evidence and attributed findings with the supplied snapshot; reopen the saved HTML through the reader and in a browser. |
+| CHAT-02 | "Include the archive, but leave callbacks undecided." | Record the human's choice and unresolved option; ask a useful follow-up when the choice matters. | Read the saved decisions/questions and later selected graph. An undecided possibility must not become a selected edge. |
+| CHAT-03 | "Save this research." | Persist the accepted record to the selected destination and report its identity/revision and actual location. | Read the file independently after the command returns; compare identity, revision and digest. A chat statement alone is not save evidence. |
+| CHAT-04 | "Find and open the research I saved for this product." | List only the relevant permitted corpus and let the human select the intended record; no implicit resume or new source upload. | Match the displayed selection to saved IDs/revisions and original bytes; verify no unrelated record changed. |
+| CHAT-05 | "Continue this record with this second document." | Establish access to that document and create a successor retaining earlier identities, evidence, decisions and unresolved work. | Compare before/after records, source hashes and parent digest; reopen both. Earlier revisions remain intact. |
+| CHAT-06 | "I corrected this note/choice." | Preserve the edited bytes, review ambiguity with the human and produce an agreed successor, or report a conflict. | Keep before/edited/after versions; verify the correction is in both readable HTML and data, with no lost extensions or changed historical evidence. |
+| CHAT-07 | "Make a Surface for the options I selected, and save it as JSON/YAML." | Produce a separate partial selected definition with sources, requirements and gaps; preserve the research record. | Reopen the actual output, resolve exact research references, compare selected nodes/edges to the human's decisions, and check supported JSON/YAML equality. |
+| CHAT-08 | "Change the output folder/runner preference." | Show the proposed effective change and apply the human's selection without moving old artefacts or transmitting sources implicitly. | Read effective settings through bare startup; check old files and source-access/inference evidence. Runner and optional Lab destination remain independent. |
+
+A host that can only supply downloads must say so and provide the real download;
+it cannot claim to have saved into a local corpus. This command milestone uses
+local persistence. External skill and Lab hosts have separate verification rows.
+
+## State X, input Y, verified Z
+
+Use plain language and concrete examples. For a small behavior, a short **Given /
+When / Then / Verify** description and representative output, as in STARTUP-01,
+are enough. These are prose labels, not a parser, technical DSL or new test engine.
+Automate the same interaction with ordinary process, output and file assertions.
+Expected values come from the task and controlled sources, not from the
+implementation under test.
+
+The following checklist helps with more involved cases such as repository research;
+it is not a mandatory form to repeat for every small check. Include only relevant
+state and evidence, and record shared run metadata once in the verification record.
+
+| Field | Content when relevant |
+| --- | --- |
+| Intent | The user's useful task and why it matters. |
+| State X | Initial settings/files and interaction mode; for research, also instruction revision, source snapshot/dirty state, permitted scope and runner/model availability. Never include credential values. |
+| Input Y | Exact command from its invocation directory, relevant user questions/answers and any supplied files. |
+| Expected change | Files, records, revisions, decisions or configuration that should be created/changed, and important state that should stay unchanged. |
+| Expected observations | What the terminal/conversation should show, including evidence references, questions, limitations and actual output location. |
+| Verification Z | Independent source checks and persisted-state readback that establish each expected outcome. Name the evidence, not just a test command. |
+| Actual result | What happened, with the exercised entry point and returned output; distinguish it from the expectation. |
+| Assessment | Pass, fail, blocked or not attempted, with criterion-level results and remaining limits. |
+
+The verification record must also identify who performed it, the date, environment,
+installed package revision and retained evidence location. Redact private material
+before publishing evidence; use public/synthetic fixtures for retained repository
+records. A private rehearsal can retain its evidence privately and publish only
+permitted conclusions and limits.
+
+## E2E-01: research a repository and produce a durable record
+
+**Intent:** understand a supplied product repository with Researcher and retain
+findings and human choices for later continuation. This is the minimum complete
+path through source access, instructions, inference, human interaction and storage.
+
+**State X:**
+
+- Install the built package into a temporary consumer environment, independent of
+  its source checkout. Record its package/build, instruction and contract versions.
+- Create a temporary Git repository containing the two tracked Lumen source
+  documents from [the shared fixtures](../../examples/researcher/README.md), with
+  a known commit and matching byte hashes. Copy no research HTML or answer files
+  into this source repository. Run the command from a different directory.
+- Start with valid explicit settings, an empty artefact folder, no Lab connection,
+  a supported configured real runner/model and authorized transmission of only
+  these synthetic documents. The human can answer through the real interface.
+- Preserve a before snapshot of settings, source and corpus. The independently
+  specified facts are: guide limit 10, settings limit 25, no known model version,
+  optional archive delivery, external identity and callback choices. Do not load
+  the expected research record into the runner's context.
+
+**Input Y:**
+
+```sh
+n8r-surface discover ./case-source
+```
+
+The human asks: "Research the components and configuration choices in this repo.
+Tell me where the sources disagree and ask which options I want to include."
+When Researcher presents the consequential choices, the human answers: "Include
+archival delivery and external identity. Leave callbacks undecided and keep the
+batch-limit conflict unresolved. Save the research." Both turns go through the
+command's actual conversation; do not inject the decisions directly into storage.
+
+**Outcome and verification Z:**
+
+| Step | What must happen or change | What we observe | How we verify it |
+| --- | --- | --- | --- |
+| 1. Find the repo | Establish the supplied repo as the source scope. | Selected source identity and inspected coverage; missing/unreadable material is explicit. | Match reported commit/working-copy state and inspected file hashes to the prepared repo. Check source-read events target only the permitted scope. |
+| 2. Follow the Researcher instructions | Load an identified instruction revision and use it to research source evidence. | Instruction version/hash in the session evidence; claims linked to actual source reads and limitations. | Compare the loaded instruction hash with the packaged asset; inspect the actual runner request/tool trace and source-grounded behavior. A version label alone is insufficient. |
+| 3. Interact with the human | Record requested selections, leave callbacks unresolved, and handle a relevant follow-up or correction. | An actual question/answer or correction exchange, with its consequence explained. | Match the transcript to saved decisions/questions. Do not count hard-coded answers or post-hoc edits by the developing agent. |
+| 4. Preserve findings and uncertainty | Retain the 10/25 conflict and unknown model version without inventing a resolution. | Both attributed claims and the unknown are visible in the result. | Check them against the two frozen sources independently, including precise source locators and hashes. Exact generated wording need not match. |
+| 5. Produce the artefact | Create a nonempty HTML research record at the selected destination. | The command reports the actual path, identity, revision and remaining gaps. | Read those bytes from disk after completion; parse with `readResearchHtml`, inspect readable HTML, and verify source references and decisions. A success message/file existence alone is insufficient. |
+| 6. Reopen useful saved work | Exit and open the saved record through a fresh command conversation. | Previous findings, choices and questions remain available; no silent rerun or upload. | Compare reopened identity/revision/digest and retained decisions with the saved output; inspect the real read operation. |
+
+**Pass:** all six steps are evidenced through the installed command with the real
+runner and actual saved bytes. Record unsupported or failed criteria explicitly.
+A fixture-library pass, fake runner, developer-authored artefact, or demonstration
+that skips interaction or reopening cannot pass E2E-01.
+
+**Current result:** not attempted; blocked by the missing CLI, Researcher instruction
+bundle and runner. The supplied synthetic HTML files demonstrate the artefact
+contract only. They are not outputs of this command scenario.
+
+## E2E-02: apply the same path to LibreChat
+
+Before claiming LibreChat research works, repeat E2E-01 on a permitted, explicitly
+selected LibreChat checkout, recording its exact upstream revision and dirty state.
+Use a bounded task: "Describe the authentication path and the components and
+configuration it depends on. Show the evidence and unknowns; ask what I want to
+include, then save the research."
+
+Before running, identify the expected relevant source files and a small set of
+checkable facts directly from that frozen checkout. Record those expectations
+separately from the runner input. Verify that the command really reads those
+sources, distinguishes supported options from the user's chosen setup, and does
+not mistake repository documentation for a running integration. Missing optional
+sources remain coverage gaps. Review unexpected findings against the source;
+do not require a fixed prose answer or silently drop supported new evidence.
+
+Apply all six verification steps above, using actual persisted outputs. Also
+check that the same product-agnostic instructions work without LibreChat-specific
+control flow. The source version, independent expected-fact set and real output
+must be recorded when exercised; none is invented here.
+
+**Current result:** this installed-command path is not attempted. A separate
+[installed-skill LibreChat source rehearsal](researcher-skill-verification.md#predictable-notes-and-real-source-trial-012)
+has saved and reopened working notes. It does not establish E2E-02, runtime
+authentication, human acceptance or validated HTML/Surface output.
+
+## Other required paths and failure outcomes
+
+Each row gets a concrete plain-language example when its increment starts.
+These checks accompany E2E-01; they are not a substitute for the complete path.
+
+| Case | State X and input Y | Expected change and observable outcome | Verification |
+| --- | --- | --- | --- |
+| SETUP-01 | Isolated home without settings, source outside the vault; run bare `n8r-surface`, press Enter then `y`. | Show `Vault directory [~/n8r/]:`, resolved path and `Create vault folders and save this location? [y/N]`. Acceptance creates three folders and saves version-1 `vaultRoot`, exit 0. Refusal/EOF/interrupt exits 1 without writes. | Read settings and `research/`, `artifacts/`, `surfaces/` independently; repeat from another source cwd with no question or settings rewrite. Preserve source and legacy bytes. |
+| INPUT-01 | Unknown argument or nonexistent/unreadable source. | Specific error or explicit coverage limitation; no substituted repo or invented completed record. | Observe terminal/exit result; inspect source-access events and corpus changes. |
+| SCOPE-01 | Explicit file source, or cwd different from installation and source. | Inspect exactly the intended scope, retaining omissions/truncation. | Compare read paths/hashes and saved coverage to prepared inputs; ensure file-only research does not claim whole-repo coverage. |
+| RUNNER-01 | Missing credential, unavailable model, network failure or cancelled interaction. | Report blocked/partial progress honestly, preserve useful accepted work, and avoid fallback. | Inspect actual runner destination/results and saved state; no completed-research claim without evidence. |
+| CONTINUE-01 | Saved incomplete Harbor r1 plus the authentication source; CHAT-05. | Add supported authentication evidence under the existing finding ID and a new revision. | Reopen old/new records; verify hashes, parent, retained decisions and new source evidence. |
+| EDIT-01 | Accepted record changed manually after read; CHAT-06/save. | Preserve the manual edit, report conflict and reconcile only the human-reviewed intent. | Keep before/edited/after bytes and compare both data and visible HTML. Verify the old revision remains recoverable. |
+| SAVE-01 | Unwritable destination or interruption during write. | Report failure accurately and retain recoverable previous work. | Read actual filesystem state after the failed operation; check no partial replacement masquerades as the accepted record. |
+| SELECT-01 | Two researched products and explicit human choices; CHAT-07. | Save a separate selected graph with shared nodes, parallel edges, cycles, references, requirements and gaps. | Compare persisted graph to selections and exact research revisions, reopen both JSON/YAML outputs and verify supported meanings. |
+| NONINTERACTIVE-01 | No interactive input; CMD-01 or discover needs a decision. | For bare startup needing setup/repair: `Vault setup required; run n8r-surface in an interactive terminal.`, exit 1 promptly, no writes. Configured reuse exits 0; invalid settings, filesystem and legacy inspection failures exit 2 before setup. | Exercise the installed process with closed input under a bounded harness and inspect its output/state. |
+| HELP-01 | Any setup state; CMD-06/07/08. | Accurate help/version, no work begun. | Compare to the inventory/manifest and verify unchanged settings/corpus and no source/inference calls. |
+
+Absence-of-side-effect claims need instrumented boundaries and before/after state,
+not just an absence of terminal messages. Verification traces must omit secrets
+and retain only permitted source excerpts. Record input/version metadata, tool
+calls and returned evidence; internal model reasoning is not required. A later independent check can verify
+the recorded readback; the implementer's report does not replace human acceptance.
+
+## Milestone acceptance and evidence status
+
+| Milestone | What its evidence can establish | Required product proof |
+| --- | --- | --- |
+| #15 | Experimental record/graph contracts, synthetic preservation/reference checks and this outcome specification. | Reviewable human-readable examples and concrete inputs for the command work; no successful command claim. |
+| #16 | Installed startup, scoped source operations and safe persistence foundation. | STARTUP-01's readable no-argument overview first, then installed command rehearsals for applicable setup/input/scope/save/help/noninteractive paths. Discovery explicitly remains unavailable until #17. |
+| #17 | The shared Researcher completes standalone discovery with a human. | E2E-01 with a real runner, continued/edited research and selected output; E2E-02 before claiming LibreChat support. External skill behavior is exercised separately. |
+| Later Lab integration | A named consumer imports and continues supported standalone work. | The same facts/choices/gaps survive the named application's actual entry points, with limitations and losses recorded. |
+
+Maintain an evidence record beside the capability being delivered. It contains the
+case ID, initial state, exact interaction, expected/actual changes, evidence/readback,
+revision/environment, verdict and unproved limits. Link it from the issue and PR.
+Tests may automate parts of this process, but the record must explain what those
+checks demonstrated in terms a person trying the feature can assess.
+
+## SETUP-01: setup-only increment
+
+Specified before implementation on 5 November 2026; setup is implemented.
+Candidate verification and its limits are recorded in
+[the setup evidence](vault-setup-verification.md). The [plan](../superpowers/plans/2026-11-05-standalone-vault.md)
+defines exact prompts, path handling, repair, preservation and the user-approved
+cooperating-writer concurrency boundary. No arguments are added. SOURCE-01 remains
+paused. Earlier read-only STARTUP-01 evidence is historical, not evidence for the
+new setup interaction. Missing legacy corpus now permits setup; other legacy
+inspection errors still take precedence (exit 2). No research is started.
+
+## RESEARCH-01: first standalone conversation
+
+Specified on 5 November 2026 before implementation. This is a bounded #17
+increment, not completion of E2E-01. The delivery priority is now standalone AI,
+then save/reopen/continue verification, then distribution (#27). Lab adaptation
+remains separate work in its own repository.
+
+**Given** a configured vault, permitted local source and Codex 0.166.0 signed in
+with ChatGPT, **when** the user runs `n8r-surface discover [source]`, accepts the
+source-transmission confirmation and asks about the source, **then** Codex uses
+bounded host tools to inspect it and returns findings with source locations.
+A correction or follow-up stays in the same conversation. **Verify** by comparing
+the cited source lines and tool digests with the unchanged source bytes.
+
+CMD-03/04/05 accept one optional directory or file (default cwd), plus optional
+`--model MODEL` for an explicit per-session choice. Without it, use the configured
+Codex model; never silently select another. CMD-06/07 display local help and CMD-08
+(`--version`) displays the package version without setup, source reads or Codex.
+Unknown arguments fail with exit 2 without echoing them or reading configuration.
+
+Discovery requires an interactive input/output terminal; otherwise exit 1 without
+loading settings or launching Codex. In a terminal, validate settings, require a
+complete configured vault, resolve the selected scope and display it. Explain
+that inspected source text and the conversation go to OpenAI through Codex using
+ChatGPT sign-in; ask `Start research with OpenAI using Codex ChatGPT sign-in? [y/N]`.
+Only y/yes authorizes inference and source content reads. Refusal/EOF exits 1;
+invalid source/settings, runner failure or failed turns exit 2. Ctrl-C exits 130
+and stops the owned server. No API-key, provider or model fallback is permitted.
+
+After connecting, display the model and bundled Researcher version/digest.
+At `You>`, accept one line per turn. Display tool-read paths/digests and completed
+agent messages as they arrive. `/exit` or EOF at the prompt exits 0. Ctrl-C during
+a turn requests interruption and closes the server. A failed or timed-out answer
+is marked incomplete and exits 2. SAVE-01 adds explicit Markdown saving/reopening;
+structured output, web/live access and extra source selection remain unsupported.
+Without `/save`, the CLI creates no research artefact; Codex owns its own authentication, cache and diagnostic state.
+
+## SAVE-01: save, reopen and continue
+
+Implemented bounded Markdown increment for #17; see [evidence and limits](save-reopen-verification.md).
+
+- Given a configured vault and an authorized research conversation, enter `/save`.
+  Expect a numbered `research/<source-slug>/research-NNN.md` destination, a
+  Researcher-authored working note, and a saved path plus byte digest after readback.
+  The note retains stable component IDs, source references/hashes, human decisions,
+  unknowns and recommended next steps. Ordinary conversation does not write;
+  natural-language save requests receive guidance to use `/save`.
+- Exit, then run `n8r-surface reopen NOTE --source SOURCE [--model MODEL]`.
+  Confirm transmission of the note and source. Expect a fresh conversation which
+  compares recorded evidence against that source, retains decisions separately,
+  and reports changed/missing evidence rather than silently revising history.
+- Alternatively run `n8r-surface reopen --source SOURCE [--model MODEL]`.
+  Expect the most recently filesystem-edited numbered note in that source's
+  research folder, with path and edit time, and a yes/no selection prompt.
+  Refusal exits without launching Codex or writing. Explicit NOTE bypasses ranking.
+- Continue and `/save` again. Expect the next unused number in the predecessor's
+  folder, a predecessor filename/digest, unchanged predecessor bytes and retained
+  identities. If the predecessor changes while drafting, stop and require reopening
+  the edited note. No general write tool or validated structured output is added.
+
+The suggested filename is convenience, not evidence of source identity or a
+selection of a deployment. A fresh process requires explicit `--source`; it never
+restores private source paths from a saved note. Help (`reopen --help`) is read-only.
+
+### Pending-operation feedback
+
+While connecting, answering, reopening or preparing a `/save` draft, show an
+interactive-terminal spinner with elapsed seconds and the operation label. Clear
+it before normal output and stop it on completion, failure or cancellation. Do
+not animate at `You>` or in nonterminal output. It means the request is pending,
+not that server-side progress is known. Verify a delayed answer/save and Ctrl-C;
+expect visible activity, clean answer/prompt output, and no continuing timer.
+
+## SURFACE-01: generate, save and revise a partial Surface
+
+**Implemented in the candidate; verification is tracked in the
+[Surface evidence record](surface-revisions-verification.md).** The installed baseline
+help checked on 5 November 2026 supported Markdown-only saves. Structured output was
+the authorized known missing capability; no failed live generation trial is claimed.
+
+**Given:** a saved numbered note, permitted source, configured vault and supported
+Codex sign-in. **When:** reopen the note, ask “Create a Surface from these notes;
+show me what is selected and still unknown,” inspect the preview and enter
+`/save surface`. **Then:** save and read back linked research HTML and Surface JSON
+without changing the note or source. `/save` continues saving Markdown notes.
+
+Exit and start `n8r-surface reopen FILE --source SOURCE`, selecting either saved
+member of the pair. After renewed transmission consent, change a named decision,
+request an updated Surface and save it. Expect a new pair with retained identities,
+evidence, human decisions and exact parent references. Earlier bytes remain intact.
+Omitting FILE continues suggesting a Markdown note only. Unknown or unsupported
+files fail clearly; no implicit search for missing research records is authorized.
+
+**Verify:** independently parse/read both pairs, check research logical references,
+parent digests and retained IDs/decisions/evidence, and compare note, predecessor,
+source and settings bytes. Repeat a small second-product case. Real installed Codex
+output is required; scripted protocol tests do not establish research correctness.
+See the [design and short human proof](../superpowers/specs/2026-11-05-surface-revisions-design.md).
+
+## DISTRIBUTION-01: install either official package candidate
+
+Given a built checkout and cached dependencies, generate private candidates using
+`npm run pack:candidates -- NEW_DIRECTORY`. Install each tarball independently
+outside the checkout. Expect the same `n8r-surface` command, version, library
+exports and bundled assets; bare startup with a fresh home and closed input exits
+1 with setup guidance and no writes. Configured startup retains existing read-only
+checks. The public command inventory is unchanged.
+
+The [distribution procedure](npm-distribution.md) supplies copyable interactions,
+readback and release blockers. The pre-change registry attempt returned E404 for
+both names and E401 for identity on 5 November 2026. Local candidates cannot prove
+registry installation, permissions or provenance. Candidate checks use scripted
+runner fixtures; no new real-model research claim follows.
+
+## VERSION-01: identify the installed package and its release source
+
+Given an installed candidate in an isolated consumer, run `n8r-surface --version`
+from another directory. Expect the installed manifest's version, exit 0, no setup
+or inference. The pre-change installed command reported `0.0.0-experimental.1`;
+a copied installation with manifest `0.7.3-rc.2` reproduced the stale hardcoded
+output. The regression now reads back `n8r-surface 0.7.3-rc.2`.
+
+The [version policy and human check](versioning.md) describe source/lockfile
+validation and read-only release mapping verification. Synthetic Git repositories
+exercise missing/wrong tags, dirty source, altered archives and release metadata;
+they do not establish an authenticated release or registry provenance. The public
+CLI inventory is unchanged; `check:version` and `check:release` are maintainer npm
+scripts. Real release verification remains blocked until an authorized release
+and publishing setup exist.
+
+## RELEASE-01: prepare and publish the first dual release
+
+Given reviewed version `0.1.0-experimental.2`, explicit
+`npm run pack:candidates -- NEW_DIRECTORY --publishable` creates both equivalent
+publishable archives; the source and default candidates remain private. The
+[human procedure](npm-release.md#human-check-before-publication) checks both
+installed routes and unchanged isolated homes. Public CLI inventory is unchanged.
+
+The manual workflow prepares retained archives before publication and verifies
+existing registry bytes before staging only absent versions for maintainer 2FA approval. Tests cover altered
+bytes, wrong registry identity/integrity, partial-publication recovery and network
+failure. These use synthetic registry boundaries, not authenticated publication.
+Live workflow, registry installation, cryptographic provenance and immutable
+release acceptance remain unverified until the documented release procedure runs.