Skip to content

Add FCIS shared scientific memory pilot and hosted MCP companion - #15

Merged
TheDarkLightX merged 7 commits into
mainfrom
codex/fcis-shared-memory-pilot
Nov 15, 2026
Merged

TheDarkLightX merged 7 commits into
mainfrom
codex/fcis-shared-memory-pilot

Conversation

@TheDarkLightX

@TheDarkLightX TheDarkLightX commented Oct 1, 2026 •

Copy link
Copy Markdown
Owner

Research Kernel has durable local memory, but shared research needs immutable claim contracts and an explicit boundary between source assertions and checker authority. This adds a private Sites-hosted companion using functional core / imperative shell (FCIS), while preserving the Python kernel and its existing research tools.

The hosted app provides isolated owner/editor/reader workspaces, scoped claim revisions, dependency review flags, negative-result memory, selected CC BY 4.0 package exchange, 17 HTTP MCP tools and sidebar/conversation-panel entrypoints. The bounded integer checker binds the exact claim, recipe, source, environment and aggregate input/output transcript. Receipts cover only the tested finite recipe; external literature, experiments, formal proof claims and imported receipts remain unverified.

D1 atomically commits the version-checked snapshot, exact command/context, results, journal and catalog projection. Actor-bound request IDs provide idempotent retries and recovery after lost responses. Contribution points are non-transferable and non-monetary; usage is recorded without charging.

The Python bridge exports only selected claim content and imports packages as CANDIDATE / triage-only. The package constrains the MCP SDK to 1.x and includes a real SDK discovery regression.

The October 1 continuation also repairs three pilot gaps:

  • Failure search and the interface preserve counterexamples after a corrected claim revision, retaining their original scope and receipts.
  • Negative evidence requests transitive dependency review without granting it checker authority.
  • The notification container no longer consumes a workbench grid column and pushes the main interface below navigation.

The UX redesign applies researched Linear, Elicit, Notion, and Zotero patterns: a compact sidebar, searchable claim list, larger Overview/Evidence/History inspector, contextual review notices, responsive navigation, and a clearly labeled read-only example generated by the existing checker. The UI separates current verification from historical failures and defers fingerprints to expandable details. See research rationale and screenshot.

Validation:

  • 63 Python regression tests passed with MCP SDK 1.30.0.
  • 29 hosted tests passed, including SQLite-backed transactions, recovery, MCP authorization, historical retrieval and WebMCP registry contracts.
  • TypeScript and production Worker build passed.
  • Browser review covered onboarding, claim inspection, revision history, search, empty results, and dependency review; 390px and 320px previews covered reflow and mobile navigation.
  • Owner-only Sites update deployed successfully from c163659f7384b44da7cc373d9b00b42dca6c65bc with MCP enabled and existing D1 storage preserved.

See acceptance evidence and remaining checks.

Live installed-plugin acceptance (October 3):

  • The original 17 hosted tools were exposed in the October 3 live acceptance; the new rk_negative_record tool makes 18 in the merged source, pending deployed-plugin validation; 15 authenticated calls completed successfully.
  • The private synthetic square-of-a-sum sequence passed 121/121, produced the missing-cross-term counterexample x=-5, y=-5: 100 != 50, and passed 121/121 after correction.
  • rk_negative_memory retained the identical failed-revision receipt while the corrected revision remained bounded-checked.
  • An independent Python arithmetic oracle verified all three claim, recipe, checker-source, output-stream and result-transcript fingerprints. The full private call transcript and backup were preserved.
  • The owner-only journal contains only add/revise/check commands: no publications, invitations or package exports. Usage reports pilot-no-charges, financialValue: none, and only non-transferable contribution records.
  • rk_open_memory returned the same workspace; this establishes its authenticated data path.

Limits: FCIS and selected ZenoFCIS patterns are used, without a direct Rust ZenoFCIS mount or formal refinement proof. Real authenticated browser/WebMCP execution and connected ChatGPT panel rendering/actions remain unverified. The live Site's secure browser sign-in attempt returned Session ended / invalid_state; its authenticated application was not reached and the signed-out document exposed no WebMCP tools; this runner exposes no mounted MCP App panel for selecting the claim or checking its scope/receipts. No standalone widget simulation was used to substitute for these host checks. Billing, transferable tokens, external verifier registries and arbitrary experimental/formal replay remain future work.

November 15 — typed negative memory integration

Merged nested PR #16 into this branch (commit 9d541e9). Both Python stdio MCP and the hosted companion now support structured soft negative-memory explanations; hosted checked counterexamples yield scope-bound hard negative insights. The negative-memory query returns original revisions and ranked information gaps without claiming a proof of wider impossibility. No private workspaces or evidence were published.

The integration also resolves three review defects with regression coverage:

  • historical failure search now streams all candidates before top-k ranking, so 1,005 unrelated newer risks cannot hide an older match;
  • exact-integer checks count magnitude bits symmetrically for ±2^255;
  • replacing an existing invitation at full workspace capacity changes its role without consuming a new seat.

The combined PR includes 18 hosted MCP tools, pinned checker/core fingerprints, app/history displays, and the append-only negative-memory contract. Python 3.11–3.13, hosted tests, typecheck, and build are the merge gates.

Release boundary: merging to main updates the source tree, not the separately running Sites-hosted Research Kernel service. The installed plugin must be deployed and revalidated separately before the new rk_negative_record tool can be claimed live.

@TheDarkLightX
TheDarkLightX marked this pull request as ready for review October 1, 2026 22:51
@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Oct 1, 2026 •

Copy link
Copy Markdown

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review Status Commit Review trigger
📝 Code Review ✅ Completed 2026-10-01T22:54:39.586922Z bf5c16a Draft marked ready
ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

@chatgpt-codex-connector chatgpt-codex-connector Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: bf5c16ad74

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Comment thread hosted/lib/core/checker.ts Resolved
Comment thread hosted/lib/core/transition.ts Resolved
Apply researched Linear, Elicit, Notion, and Zotero interaction patterns. Add a responsive evidence inspector and a clearly labeled read-only example generated by the existing checker. Preserve the FCIS authority and persistence contracts.
#16)

Append-only historical failures, scope-bound checked exclusions, informative soft negatives, regression cases, signed magnitude fix, invitation capacity fix.
@TheDarkLightX
TheDarkLightX merged commit 44f3e00 into main Nov 15, 2026
4 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant