V0.9.2/one prompt eval

.github/workflows/validate-skill-templates.yml

@@ -11,7 +11,7 @@ jobs:
   validate:
     name: Validate (${{ matrix.name }})
     runs-on: ubuntu-latest
-    timeout-minutes: 5
+    timeout-minutes: 10
     strategy:
       fail-fast: false
       max-parallel: 13

CHANGELOG.md

@@ -4,6 +4,33 @@ All notable changes to this project will be documented in this file.
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.9.2] - 2026-09-08
+
+This patch release adds optional one-shot external handoff for explicit eval requests, complete with deterministic workflow helpers and test coverage. Agents can now accept `yolo` or `auto` modifiers on explicit evaluation requests to authorize a single fresh external Eval Orchestrator handoff; preparation remains deterministic and repository-local, while only an explicit, authorized harness receives the sealed package. Skill eval preparation gains a PassThru parameter for returning prompt paths, runner names are normalized for user-facing consistency, and validation tooling gains comprehensive eval-request-workflow checks.
+
+### Added
+
+- `scripts/eval-request.ps1` with deterministic helpers for optional one-shot external handoff: `Get-EvalHandoff` reserves and transitions between manual and external handoff states, while `Invoke-EvalRequest` normalizes harness names, discovers available models, returns preparation decisions, and never launches a model itself,
+- Full test coverage for eval-request workflow in `scripts/eval-runners/tests/test-eval-request.ps1` exercising all runner/model normalization paths, handoff state transitions, already-started detection, unavailable-host fallback, and case-insensitive normalization,
+- Runner normalization in eval-request helpers converting user-facing harness names (`GitHub Copilot`, `Copilot CLI`) to internal canonical runner ids (`github-copilot`) before model discovery and profile generation,
+- PassThru parameter in `scripts/prepare-skill-evals.ps1` for returning the prepared prompt paths without file output, enabling downstream helpers to collect decisions and route handoffs,
+- `-Yolo` flag in `scripts/eval-request.ps1` Invoke-EvalRequest authorizing one-shot external handoff after explicit user eval request, while keeping preparation and validation deterministic,
+- Optional-one-shot-external-handoff section in `AGENTS.md` documenting the explicit user authorization, harness selection, model discovery flow (including OpenCode's explicit-model requirement), and the deterministic one-handoff boundary,
+- External evaluation authorization guidance in `AGENTS.md` clarifying that `yolo`/`auto` modifiers on explicit eval requests enable handoff without widening repository automation, CI, completion gates, or model-backed execution,
+- Eval-request-workflow validation checks in `scripts/validate-skill-templates.ps1` ensuring eval-request.ps1 presence, test coverage, runner normalization completeness, and handoff-state-reservation correctness.
+
+### Changed
+
+- `AGENTS.md` eval preparation guidance now distinguishes between manual handoff (normal path that prepares and returns `RUN-THIS.prompt.md`) and external handoff (when user explicitly requests eval with yolo/auto), clarifying that only the latter authorizes orchestrator delegation,
+- Runner-model discovery and selection in `scripts/Get-HarnessModels.ps1` now fully documents the -Runner requirement and exposes current available models without auto-selecting a previous or default model,
+- `README.md` eval section expanded with documentation of optional one-shot handoff feature, explicit authorization scoping, and reference to `AGENTS.md` for the complete workflow,
+- `CONTRIBUTING.md` now references the optional one-shot eval handoff capability and directs users to full preparation and handoff guidance in `AGENTS.md`.
+
+### Fixed
+
+- Eval request workflow now clearly separates deterministic preparation (always model-free, always stops for manual handoff) from explicit external handoff (only with user-supplied yolo/auto, never retried if already launched),
+- Model discovery no longer accepts harness names as ambiguous shortcuts; runner normalization explicitly converts `Copilot` and `GitHub Copilot CLI` to the canonical `github-copilot` runner before passing to model-discovery logic.
+
 ## [0.9.1] - 2026-09-07
 
 This patch release adds harness-agnostic Eval Runner execution boundary infrastructure without changing the paired evaluation methodology or existing report schemas, while optimizing skill descriptions and refactoring repository-level authoring guidance. Prepared packages now carry `execution-profile.json`, package-local runner protocol tools, and normalized `execution-result.json` evidence. The deterministic fake runner is the conformance reference, with Codex, GitHub Copilot CLI, and OpenCode as supported real adapters. Repository automation remains model-free; only a human-directed external Eval Orchestrator may invoke the selected runner, and unsupported isolation fails closed.
@@ -638,6 +665,7 @@ This is a minor release that introduces two complementary git workflow skills, e
 
 - Improved scaffold fidelity with hidden `.bot` asset preservation, explicit UTF-8 and BOM handling, and checks aimed at preventing mojibake or incomplete generated output.
 
+[0.9.2]: https://github.com/codebeltnet/agentic/compare/v0.9.1...v0.9.2
 [0.9.1]: https://github.com/codebeltnet/agentic/compare/v0.9.0...v0.9.1
 [0.9.0]: https://github.com/codebeltnet/agentic/compare/v0.8.2...v0.9.0
 [0.8.2]: https://github.com/codebeltnet/agentic/compare/v0.8.1...v0.8.2

CONTRIBUTING.md

@@ -98,7 +98,9 @@ Run it after the last skill edit and before `scripts/sync-skill-install.ps1`, wh
 pwsh -NoProfile -NonInteractive -File ./scripts/prepare-skill-evals.ps1 -Skill <skill-name> -Runner <runner-id> -Model <runner-native-model>
 ```
 
-Before running the script, choose a Harness + Model when the user did not already do so. Normalize explicit harness wording immediately: `Codex` maps to `codex`; `GitHub Copilot`, `GitHub Copilot CLI`, and `Copilot` map to `github-copilot`; `OpenCode` maps to `opencode`; matching is case-insensitive. Use `scripts/Get-HarnessModels.ps1 -Runner <runner-id>` to list current selectors; it fails immediately with the supported runner IDs when `-Runner` is omitted. OpenCode mirrors every model exposed by all configured providers, with availability retained as metadata only, while GitHub Copilot and Codex list all currently available models. When OpenCode is selected without an explicit model, present the discovered exact `provider/model` selectors, ask the user to choose one, and wait; never select the first, free, recommended, previous-iteration, previous-successful, or previous-failed model automatically. Preserve an explicitly supplied OpenCode selector verbatim in `execution-profile.json`; discovery failure or incomplete metadata must not substitute another model. The Codebelt Reference shortcut is GitHub Copilot CLI + `claude-haiku-4.5`, and Codex defaults to `gpt-5.6-luna` with low reasoning; package preparation validates the resolved model against current discovery before writing `execution-profile.json`. The script writes `.bot/<skill-name>-workspace/iteration-<n>/` with one directory per eval. Each holds the grading key `eval-metadata.json` and result stubs under `results/` at the eval-case level, plus two paired run directories, `with_skill/` and `without_skill/`. A run directory is the worker's run root: `prompt.md`, a `run.json` contract, a `repo/` working tree materialized from the fixtures, an isolated `home/`, and - for `with_skill` only - a `skill/<name>/` copy of the candidate. The grading key and results sit outside both run directories. At the root it writes `manifest.json`, `execution-profile.json`, the package-local Eval Runner protocol, the package report adapter, the exact Anthropic skill-creator grader/aggregator/viewer assets, and `RUN-THIS.prompt.md`, the one prompt you hand to the external Eval Orchestrator. That orchestrator resolves and preflights the selected runner, reads `delegation.dispatch_owner`, and either dispatches the declared orchestrator-owned native worker or starts the declared runner-owned one-arm native surface directly. It stores genuine transport-produced `execution-result.json` evidence, bridges the results, grades only after execution, and runs the adapter, which invokes `aggregate_benchmark.py` and `eval-viewer/generate_review.py --static`. It never runs an eval prompt in the coordinator context, never chooses runner/model policy, and never reuses a worker. Both worker prompts carry the same task, materialized repository, and response contract; only the operating instructions and the presence of `skill/` differ, and neither prompt identifies itself as an eval. `.gitignore` covers `.bot/*`, so nothing there reaches git. The script refuses an `-OutputRoot` inside the repository but outside `.bot/`; pass an explicit temp path when the harness does not need repository-local storage.
+Before running the script, choose a Harness + Model when the user did not already do so. Normalize explicit harness wording immediately: `Codex` maps to `codex`; `GitHub Copilot`, `GitHub Copilot CLI`, and `Copilot` map to `github-copilot`; `OpenCode` maps to `opencode`; matching is case-insensitive. Use `scripts/Get-HarnessModels.ps1 -Runner <runner-id>` to list current selectors; it fails immediately with the supported runner IDs when `-Runner` is omitted. OpenCode mirrors every model exposed by all configured providers, with availability retained as metadata only, while GitHub Copilot and Codex list all currently available models. When OpenCode is selected without an explicit model, present the discovered exact `provider/model` selectors, ask the user to choose one, and wait; never select the first, free, recommended, previous-iteration, previous-successful, or previous-failed model automatically. Preserve an explicitly supplied OpenCode selector verbatim in `execution-profile.json`; discovery failure or incomplete metadata must not substitute another model. The Codebelt Reference shortcut is GitHub Copilot CLI + `claude-haiku-4.5`, and Codex defaults to `gpt-5.6-luna` with low reasoning; package preparation validates the resolved model against current discovery before writing `execution-profile.json`. For `github-copilot` without an explicit model, discovery is only an availability check for the repository-defined default; never rank or replace it with a subjective stronger/newer choice. The script writes `.bot/<skill-name>-workspace/iteration-<n>/` with one directory per eval. Each holds the grading key `eval-metadata.json` and result stubs under `results/` at the eval-case level, plus two paired run directories, `with_skill/` and `without_skill/`. A run directory is the worker's run root: `prompt.md`, a `run.json` contract, a `repo/` working tree materialized from the fixtures, an isolated `home/`, and - for `with_skill` only - a `skill/<name>/` copy of the candidate. The grading key and results sit outside both run directories. At the root it writes `manifest.json`, `execution-profile.json`, the package-local Eval Runner protocol, the package report adapter, the exact Anthropic skill-creator grader/aggregator/viewer assets, and `RUN-THIS.prompt.md`, the one prompt you hand to the external Eval Orchestrator. That orchestrator resolves and preflights the selected runner, reads `delegation.dispatch_owner`, and either dispatches the declared orchestrator-owned native worker or starts the declared runner-owned one-arm native surface directly. It stores genuine transport-produced `execution-result.json` evidence, bridges the results, grades only after execution, and runs the adapter, which invokes `aggregate_benchmark.py` and `eval-viewer/generate_review.py --static`. It never runs an eval prompt in the coordinator context, never chooses runner/model policy, and never reuses a worker. Both worker prompts carry the same task, materialized repository, and response contract; only the operating instructions and the presence of `skill/` differ, and neither prompt identifies itself as an eval. `.gitignore` covers `.bot/*`, so nothing there reaches git. The script refuses an `-OutputRoot` inside the repository but outside `.bot/`; pass an explicit temp path when the harness does not need repository-local storage.
+
+For an explicit eval request only, `yolo`/`auto` permits the interactive agent to hand the prepared `RUN-THIS.prompt.md` to one fresh external Eval Orchestrator and wait for its completed result. Without the modifier, preparation still stops at the manual handoff. `external_handoff` means authorization is complete: do not ask again. The helper encodes that as `confirmation_required = false`, `dispatch_immediately = true`, and `max_new_external_orchestrators = 1`. In GitHub Copilot CLI, `task` + fresh `general-purpose` agent availability is a valid external handoff capability for this one-shot flow: delegate exactly one fresh orchestrator context, pass only the generated absolute `RUN-THIS.prompt.md` path, and follow that same returned agent/task handle to terminal completion. After launch, keep that same handle across bounded native waits; a wait that returns no completed agent or otherwise no terminal result means still running, so wait again on the same handle until a true terminal `completed` or `failed` result exists. Follow [Optional one-shot external handoff](AGENTS.md#optional-one-shot-external-handoff); `scripts/eval-request.ps1` computes and reserves the handoff but never launches a model. Host unavailability returns the manual path, preparation/model failures launch nothing, and an attempted handoff must never be dispatched again. This exception does not apply to repository automation or completion gates.
 
 Repository preparation, validation, CI, hooks, deterministic tests, and automatic completion gates never run those prompts or invoke a model. That boundary is the Priority 1 rule in `AGENTS.md`, and preparing a prompt is not permission to execute one. A human-selected external Eval Orchestrator handed a specific package may invoke the selected package-local Eval Runner; this explicit handoff boundary does not weaken the repository prohibition or authorize CI/live evals.