Repository navigation
GUI test rounds 2026-10-13: one blueprint grid, working folders, captain modes, Artifacts area (merge after Codex MERGE-FÄHIG + green CI, merge commit) - #6
Merged
Conversation
… use cases Lukas's GUI test on 2026-10-13: the blueprint list was itself a card grid, so every group title and group took a card cell and the page fell apart. The list now stacks titles, groups and notes; each group keeps its own card grid (one column under 600 px). The engine-room side tab moves from Blueprints to My use cases, where the use cases it runs live. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FMTv7ecoTBXem1ndYqUtxj
…tem folder dialog Lukas's GUI test on 2026-10-13, round 2: - Everything is a workflow with one step or several (D-042). Shipped presets, own blueprints and every one-step contract sit as equal tiles in one grid, filtered by the same topic chips as My use cases; the separate action deck and its "Selected · open engine room" highlight are gone. Cards read "WORKFLOW · N STEPS"; the engine-room select is labelled Workflow. - Strict split between full console and public demo (D-043): the full console starts empty or at the operator's first --allow-root, never at the fictional case; the case questions exist only in the demo. - The start page is the captain, the engine room and the folded chart; hero and area cards that repeated the navigation are removed. - Folder dialog: type or paste a path (Enter), or open the system folder dialog on this computer (/api/pick-folder, loopback console only, never in the demo); approving stays an explicit click. - The captain also understands English "contradict"/"disagree"; the captain example is English; the footer shows the served version. Full suite 1175 passed, 2 skipped; public demo 5/5 executed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FMTv7ecoTBXem1ndYqUtxj
…l folders Lukas's GUI test on 2026-10-13, round 3 (D-044): - Working folders: added once on Folders (folder dialog, "Keep as working folder"), they stay approved at every later start of a loopback console (run-reports/web-console/working-folders.json) and are offered as one-click chips at the captain and in the engine room. Removing one forgets it and ends its approval unless the server was started with it; files are never touched. Never in the demo, never on a network-exposed server, which also never widens its roots from the file. - The console starts at the first working folder; the folders chosen last stay chosen across pages (this browser only). - The captain takes several folders (dialog or chips), understands "Fasse ... zusammen" / "summarize" (Synopsis Merge), and a request that names exactly one working folder runs on that folder. - Folders shows "Your folder workflows"; the ship's chart describes the current areas; the four engine-room areas are tinted tiles with large headings; on the start page the embedded engine room no longer swallows the mouse wheel. Full suite 1177 passed, 2 skipped. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FMTv7ecoTBXem1ndYqUtxj
…nsole Lukas's GUI test on 2026-10-13, round 4: - Prepare voyage no longer links to another page with the drawer: step 1 of the prepared voyage loads straight into the engine room below the captain (workflow, folders, questions, parameters), and every further step has its own "Load into the engine room" button. The page-routing helpers that only served the old link are removed. - Governance: the "applies when nobody said otherwise" checkbox sat in the middle of the row, far from its label; a sweep of every view found no other checkbox torn from its label. - README EN/DE and llms.txt describe the console as it now is: everything is a workflow with one step or several, one blueprint grid, a clean full console without the demo case, working folders, path entry and the system folder dialog, the captain with several folders. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FMTv7ecoTBXem1ndYqUtxj
…ngine room
Lukas's GUI test on 2026-10-13, round 5:
- Connections shows what this computer shows per adapter instead of
"Configured" for all: local model server running or not, personal CLI
found or not, API key variable set or not (never its value). The proof
rule names the real 2026-09-02 Nemotron-on-Nebius run in
examples/proven-run; NemoClaw stays unproven; a console run is never the
competition proof.
- The captain is a deterministic keyword planner. A document request with
no task word ("Ich suche eine Information im root des Ordners") now gets
Evidence Analyst with the sentence itself as the question and a note
saying so; anything else is still refused rather than guessed.
- The research notebook (D-045) is the quick investigation mode under the
engine room. It ships as one of My use cases; removing it there hides it,
and Blueprints offers it back.
- Folders: the glance looks at a working folder ("What is on board in X")
and is no second way to choose folders.
- The full console no longer preselects the bare base folder that only
holds the reports.
Full suite 1179 passed, 2 skipped.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FMTv7ecoTBXem1ndYqUtxj
…list
Lukas's GUI test on 2026-10-13, round 6 (his own .SYNC folder, ~8,000 files):
- Synopsis Merge ran out of memory on the real folder (workflow_error:
MemoryError, twice). It now refuses more than 500 documents with a
sentence naming the alternatives, and any MemoryError reads as "this
folder is too large for this workflow on this computer".
- "Fasse ... zusammen" / "summarize" goes to Evidence Analyst with the
sentence as the question, where a reasoning worker such as Nemotron can
be chosen; "Synopse" stays with Synopsis Merge (which has no model step).
- Artifacts lists ledgers up to two folders below the output scope, so a
captain voyage step (voyage/01-<workflow>/ledger) and its failure show up;
after a reload the run dossier names the newest recorded run instead of
"Nothing has run".
- The model field offers the models the local Ollama or LM Studio actually
holds (/api/local-models, loopback, never demo or network-exposed). Models
the local server only relays to a cloud (":cloud") are not offered and are
refused as a local model, so "loopback · local only" stays true.
Full suite 1182 passed, 2 skipped.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FMTv7ecoTBXem1ndYqUtxj
Lukas's GUI test on 2026-10-13, round 7 (D-046): - The captain plans in one of three modes, one per line: keywords (default, local and instant); a model chooses (it reads the sentence and the workflow catalogue, never a document; its choice is checked against the catalogue and the plan is built by the keyword planner's own machinery; an external worker needs --allow-external-models and a one-request approval and gets the sentence pseudonymized; Prepare repeats the choice without asking again); and a brief for Claude Code or Codex CLI, copied to the clipboard with folders, catalogue, job format, commands and rules. Measured: qwen3.8:27b-mlx on the Mac Studio (through an SSH tunnel to 127.0.0.1:11434) chose Coverage Timeline + Cost Timeline for "welche Verträge laufen bald aus und was kosten sie" in 74 s. - The local adapter asks once more without a response format when a model answers "structured output is unavailable" (an MLX build behind Ollama). - Three workflows lacked a planner purpose or title (briefing, document_qa, ocr_pipeline), which would have crashed any plan that chose them. - Artifacts is its own area (/artifacts, after Processes & Workflows); the old tab link redirects; every run shows its date and local time. - Every blueprint tile can be tried once (a shipped multi-step one loads step 1 with its settings); every use-case tile can be removed on the tile, with a second click to confirm. Full suite 1187 passed, 2 skipped. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FMTv7ecoTBXem1ndYqUtxj
Lukas, 2026-10-13: "muss ich einen nach dem anderen durchführen oder unterstützt der engine room auch nacheinander und wo ist die jobs inbox?" Next to Prepare voyage the captain now offers Run all steps: the plan is kept as one of My use cases and run through the existing voyage runner, each step reading the previous output where the plan says so, locally and as a dry run; the result lists every step's status and links to Artifacts. The note under the buttons names where the job inbox is (engine room · Run identity & prepared jobs). Measured in the browser on the fictional case: person_timeline then contradiction_synopsis, both executed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FMTv7ecoTBXem1ndYqUtxj
…wording README EN/DE and llms.txt describe the console of this round: the three captain modes and Run all steps, Artifacts as its own area with failed runs and times, the local model list without relayed cloud models, the JSON fallback, the research notebook under the engine room, and Connections reporting what this computer shows. The brief mode reads "Brief for a CLI agent (e.g. Claude Code or Codex)" instead of naming two products. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FMTv7ecoTBXem1ndYqUtxj
|
Welcome! 👋 Thanks for your first pull request in this repository. A maintainer will review it soon. Please make sure:
Thanks for contributing! |
Member
Author
|
Merged by Codex (OpenAI) as second-model reviewer per owner decision D-047; verdict MERGE-FÄHIG; CI green on head. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Merge condition
Merge only after the second-model review (Codex) says MERGE-FÄHIG and the
cichecks of this PR's head are green. Use a merge commit (no squash). Owner decision D-047: Codex may merge this PR after that review; nobody merges by admin.What this is
Seven rounds of the owner's live GUI test on 2026-10-13 (decisions D-042 to D-046), on top of the merged final acceptance (#5).
Console
/api/working-folders), offered as chips at the captain and in the engine room; never in the demo, never on a network-exposed server, which also never loads them./api/pick-folder, loopback only); approving stays a click.--allow-external-models+ one-request approval, sentence pseudonymized), a brief for a CLI agent copied to the clipboard; Prepare loads step 1 into the engine room below; Run all steps keeps the plan as a use case and runs it step after step./artifacts) with every ledger below the output scope (voyage steps, failed runs) and date/time; the run dossier names the last run after a reload.Robustness found in the owner's real test
/api/local-models); models relayed to a cloud (:cloud) are not offered and are refused as local.Evidence
🤖 Generated with Claude Code
https://claude.ai/code/session_01FMTv7ecoTBXem1ndYqUtxj