Skip to content

Validate immutable effective model capabilities - #26

Merged
Varuas37 merged 2 commits into
mainfrom
codex/effective-capabilities
Sep 12, 2026
Merged

Varuas37 merged 2 commits into
mainfrom
codex/effective-capabilities

Conversation

@Varuas37

@Varuas37 Varuas37 commented Sep 12, 2026 •

Copy link
Copy Markdown
Contributor

A selected model's published capabilities do not describe what a configured binding can accept. Add immutable EffectiveCapabilities snapshots that intersect model and binding support and reject configured token limits above either ceiling. For example, an image-capable model behind a text-only binding rejects image input before admission.

TokenLimits and its typed validation error live in the common domain, shared by metadata, binding declarations, and configured limits. No legacy export remains in model metadata.

The pure domain validator checks explicit feature requirements and caller-supplied input/output budgets. Application DTOs project the same snapshot used for validation. No provider discovery, settings reader, authorization, or conversation lifecycle is introduced.

Validation:

  • 36 SDK tests pass, including 2,401 independent input/output modality combinations, limit boundaries, overflow, diagnostics, and snapshot isolation.
  • All SDK domain code has 100% LLVM line, function, and region coverage: 364 lines, 60 functions, 445 regions. This is not a branch-coverage claim. Application mapping was separately verified at 100%.
  • A checked-in coverage script uses an isolated temporary target; Linux CI enforces all three domain coverage thresholds.
  • SDK formatting, Clippy with warnings denied, and diff checks pass. Two independent review agents found no remaining actionable issues after the review loop.

A whole-workspace formatting check still reports pre-existing differences in untouched desktop files; the SDK formatting check passes. ACP compatibility and SQLite verification reports remain in their separate worktrees.

RetriggerConfidence Score: 5/5

The PR appears safe to merge with no actionable correctness, security, or repository-rule violations identified.

Summary

  • Intersects input/output modalities, tool use, and reasoning support.
  • Rejects configured limits above model or binding ceilings.
  • Validates typed capability requirements and caller-supplied token budgets.
  • Adds an application DTO projection of the validated snapshot.
  • Adds comprehensive domain and application tests plus a Linux coverage gate.
  • Updates SDK and architecture documentation to describe the completed capability layer.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
    M[Selected ModelMetadata] --> E[EffectiveCapabilities::new]
    B[BindingRestrictions] --> E
    C[Configured TokenLimits] --> E
    E -->|intersect features| S[Immutable EffectiveCapabilities]
    E -->|limits exceed either ceiling| CE[ConfiguredLimitExceeded]
    S --> V[validate requirements and budgets]
    R[Capability requirements] --> V
    T[Input tokens and reserved output] --> V
    V -->|supported and within limits| A[Admission succeeds]
    V -->|unsupported feature| U[Unsupported]
    V -->|invalid output budget| O[InvalidOutputBudget]
    V -->|input plus output exceeds window| W[ContextWindowExceeded]
    S --> D[EffectiveCapabilitiesDto projection]
Loading

Reviews (1) · Last reviewed commit: "feat: validate immutable effective model..."

@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Sep 12, 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 2026-09-12T05:12:35.541143Z f1e3ffe PR opened
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: f1e3ffe139

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".

@Varuas37
Varuas37 merged commit 9b27a9e into main Sep 12, 2026
3 checks passed
@Varuas37
Varuas37 deleted the codex/effective-capabilities branch September 12, 2026 05:37
Varuas37 pushed a commit that referenced this pull request Nov 9, 2026
… lined up

- #22 The Agents peek replies with the composer in its compact size
  (`.desktop-composer[data-size="compact"]`): same surface, Send at its
  usual size, instead of a pill of its own.
- #23 The overview's first group starts right under its header, the list
  says Loading (kit LoadingText) until its rows are drawn, and a row is its
  title plus one line: an agent's command, or an app's who-asks-for-what.
- #24 (kit) A trace never cuts the verb, cuts a path from its start, and
  draws a failed call's words in their usual tone.
- #25 (kit) A table's copy button sits above its top-right corner.
- #26 The access menu's right edge meets the composer's (kit `alignTo`).
- #27 Every pane's composer has the same gutter (`--workspace-dock-gutter`,
  the 8px narrow-pane override is gone), and a home beside other panes is
  titled "New session".
- #28 Jump to Latest sits on the composer's top edge, half over it, not on
  the last line of the reply.

nessa-ui-revision pins nessa_ui ef3187f (branch head of nessa_ui#156).
Stress: menu checks the access menu edge, the overview gap, two-line rows
and the compact reply; table checks the copy button and Jump to Latest;
window checks one gutter and the named home beside; trace checks the verb,
the cut side and the failed row's tone.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JcBpmVHPWxktA9QurvUMjM
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