Shared Context Intake — Plugin Project

This repository builds a domain-neutral, pre-specialization Context Intake Plugin. It starts from the owner's goal, known decisions and available evidence; recovers scenarios and consequential gaps; and passes a single authorized docs/intake/<stable-request-slug>/HANDOFF.md to the next specialist factory. It does not implement that factory.

Agent entry and source of truth

Read AGENTS.md, then docs/BUILD_STATE.md. The router selects only the documents relevant to the current authorized stage:

Two separate success gates

1. Behavioral Success: map and assess the relevant UC-01..UC-35 against the augmented Skill. When explicitly authorized, use the already accessible Promptfoo plugin/tool/CLI, with reviewed independent fixtures, provider/host invocation and oracles. Catalog presence or structural tests are not a passing behavior result.
2. Packaging Success: after accepted behavioral evidence, use the existing @plugin-creator and official target-host packaging/installation mechanisms. Reuse current manifests and local tools; do not generate homegrown probes or testing infrastructure without explicit authorization. Promptfoo can verify post-install behavior, but not manifest validity.

Do not silently execute evaluations, package/install, create tools, change authentication, or publish. Latest authorization and evidence are in docs/BUILD_STATE.md; do not mistake this README for execution permission.

Current state

The Skill and augmentation already exist. The recent Codex transcript reports partial packaging and isolated discovery, but no certified Behavioral Success, completed HANDOFF invocation or approved Packaging Success. Exact local Windows checkout state is not visible in this Markdown snapshot: inspect and merge with local updates before replacing files. No commits or pushes are delegated to agents.

Authorized Promptfoo evaluation

All evaluation inputs and outputs are under promptfoo/: the executable config, 35 cases with source-derived rubrics, synthetic consumer fixtures, baseline hashes, local Promptfoo state and result files. The owner explicitly authorized preparing these assets and running them. They are newly derived assets, not a recovered copy of the absent historical evals/cases.yaml.

The suite uses Promptfoo's built-in openai:codex-sdk provider and the existing Codex ChatGPT login. It does not install dependencies, copy credentials, define a custom provider, or run the historical custom probe scripts. The eval config uses invocation-only overrides to disable unrelated MCP servers, plugins and user skills; it does not edit global settings. gpt-5.6-sol is selected from the local Codex model cache because the default gpt-6.1-sol and documented gpt-5.4 example were rejected by this account's CLI endpoint. Actual run status is recorded in docs/BUILD_STATE.md; model-cache presence is not successful inference.

Run from K:\production_plugins\context-intake with the already installed Promptfoo CLI:

$env:PROMPTFOO_CONFIG_DIR = "$PWD/promptfoo/.state"
$env:PROMPTFOO_LOG_DIR = "$PWD/promptfoo/results/logs"
$env:PROMPTFOO_DISABLE_UPDATE = '1'
$env:PROMPTFOO_DISABLE_TELEMETRY = '1'
promptfoo validate config -c promptfoo/promptfooconfig.yaml
$intakeResult = "promptfoo/results/run-$(Get-Date -Format yyyyMMdd-HHmmss).json"
promptfoo eval -c promptfoo/promptfooconfig.yaml -o $intakeResult --max-concurrency 1 --no-cache --no-share --no-write --no-table --no-progress-bar

Use --filter-pattern '^UC-01 ' for the first case, or another UC description to rerun only affected cases. Preserve existing results. UC-32 writes one handoff in its own synthetic consumer directory: move that generated output to a fresh, reviewed run location before rerunning UC-32; the suite never silently deletes or overwrites it. Do not reset the seeded OPS-6/OPS-33 history or baseline hashes to make a failing result pass.

Acceptance is all 35 case rubrics and artifact assertions passing, with zero provider/grader errors. The subject never receives its rubric. A separate Codex grader session assesses semantics; filesystem hashes and the existing canonical intake.py validate check artifacts independently of that grader's claims. The grader uses the same model family, and these fixtures have no independent human review; one run does not prove stochastic reliability, all 519 variants, implicit skill selection, or other-host compatibility. Sandbox/provider errors block the gate and must not be counted as behavioral passes.

Official integration references: Codex SDK provider, testing agent skills, assertions.