microsoft/hve-core

Public

mirrored from https://github.com/microsoft/hve-coreAvailable

CodeCommitsIssuesPull requestsActionsInsightsSecurity
ci/2086-enforce-powershell-coverage

Branches

Tags

  • No tags available.
0Branches0Tags
Go to file
Add file
Code

Clone

HTTPS

Download ZIP

evals/agent-behavior/expectations/agentic-workflows.expectations.yml

121lines · modecode

1# Bucket-A expectations for agentic-workflows
2# Format: per-agent YAML, 5–10 grader-worthy expectations grounded in the agent
3# file's explicit promises and/or current matrix failures. This file is consumed
4# by the next pass that rewrites stimuli + graders end-to-end; do not treat it
5# as a Vally grader file directly.
6#
7# Note: the 2026-05-28 matrix run for `agentic-workflows` FAILS both
8# `phase-marker-present` and `tracking-file-write` because the output
9# wrote the plan to an arbitrary path
10# (`C:\Users\…\.copilot\session-state\…\plan.md`) instead of the
11# `.copilot-tracking/plans/` subtree, and structured phases inside a table
12# rather than as headed sections matching `^(##|###|Step \d+|Phase \d+|\d+\.)`.
13# Expectations below promote contract-grounded checks targeting both
14# failures plus broader Engineering-Excellence planning conformance.
15slug: agentic-workflows
16class: planner-coach
17agent_file: .github/agents/hve-core/agentic-workflows.agent.md
18stimulus_file: evals/agent-behavior/stimuli/agentic-workflows.yml
19latest_result: evals/results/agent-matrix/2026-05-28/agentic-workflows.json
20source_review_date: 2026-05-28
21
22expectations:
23 - expectation_id: phase-headings-present
24 summary: Plan is organized into headed phases or numbered steps.
25 signal: Output contains lines matching `^(##|###|Step \d+|Phase \d+|\d+\.)`.
26 pass_criteria: |
27 Plan body contains multiple top-level markers from
28 `## <heading>`, `### <heading>`, `Step <n>`, `Phase <n>`, or numbered
29 list items (`1.` …). A table row alone does not satisfy this.
30 failure_modes:
31 - Phases captured only as rows in a single table.
32 - Single prose block with no headings or step markers
33 (matches current 2026-05-28 matrix failure).
34 priority: high
35 contract_ref: "current `phase-marker-present` grader (regex /(?m)^(##|###|Step \\d+|Phase \\d+|\\d+\\.)/)"
36
37 - expectation_id: tracking-plan-path
38 summary: Plan is written under the `.copilot-tracking/plans/` subtree.
39 signal: Output names a path matching `.copilot-tracking/plans/<YYYY-MM-DD>/<slug>-plan.instructions.md`.
40 pass_criteria: |
41 Reported plan path is workspace-relative, starts with
42 `.copilot-tracking/plans/`, includes a dated subdirectory, and ends in
43 `-plan.instructions.md`. Absolute paths to `.copilot` session state,
44 temp directories, or repo root are not acceptable.
45 failure_modes:
46 - Plan written to an absolute path under `.copilot/session-state/`
47 (current 2026-05-28 matrix failure).
48 - Plan written to repo root with no tracking prefix.
49 - Path matches `.copilot-tracking/plans/` but is missing the dated
50 subdirectory or `-plan.instructions.md` suffix.
51 priority: high
52 contract_ref: "agent §Tracking Artifacts (Implementation Plan path under `.copilot-tracking/plans/{{YYYY-MM-DD}}/{{task-description}}-plan.instructions.md`)"
53
54 - expectation_id: success-criteria-per-phase
55 summary: Each phase names success criteria.
56 signal: Phases include explicit success or exit criteria.
57 pass_criteria: |
58 Each phase carries at least one explicit success/exit criterion
59 (acceptance test, observable outcome, validation command, or
60 measurable result). Criteria are concrete rather than aspirational
61 ("works correctly").
62 failure_modes:
63 - Phases listed with names only and no acceptance criteria.
64 - Success criteria are vague ("complete and tested").
65 priority: high
66 contract_ref: "stimulus instruction (Break into phases with success criteria) + agent §Implementation Plan (per-phase success criteria)"
67
68 - expectation_id: parallelization-markers
69 summary: Plan annotates phases with parallelization markers when appropriate.
70 signal: Output uses `<!-- parallelizable: true -->` or `<!-- parallelizable: false -->` per phase.
71 pass_criteria: |
72 Each phase carries a `<!-- parallelizable: true -->` or
73 `<!-- parallelizable: false -->` annotation. When all phases must run
74 serially, the markers are still present and set to `false`.
75 failure_modes:
76 - Markers omitted entirely.
77 - Parallelization expressed only in prose with no machine-readable
78 annotation.
79 priority: medium
80 contract_ref: "agent §Implementation Plan (mark phases with `<!-- parallelizable: true/false -->`)"
81
82 - expectation_id: tracking-markdown-disable-comment
83 summary: Plan file begins with the markdownlint-disable directive.
84 signal: Output references `<!-- markdownlint-disable-file -->` at the top of the tracking file.
85 pass_criteria: |
86 Plan file (or summary thereof) begins with the literal directive
87 `<!-- markdownlint-disable-file -->` on the first line. The same
88 directive is NOT placed in published markdown under `docs/`.
89 failure_modes:
90 - Plan file shown without the directive.
91 - Directive placed in non-tracking surfaces.
92 priority: low
93 applies_when: "agent reports plan-file creation"
94 contract_ref: "repo convention (`.copilot-tracking/` files begin with `<!-- markdownlint-disable-file -->`)"
95
96 - expectation_id: research-context-reference
97 summary: Plan references prior research or context inputs.
98 signal: Output references a `.copilot-tracking/research/` path or names the inputs that informed the plan.
99 pass_criteria: |
100 Plan body or summary names the research inputs it builds on, either
101 by referencing a research file under
102 `.copilot-tracking/research/<YYYY-MM-DD>/` or by listing the
103 requirements/sources analyzed before phase design.
104 failure_modes:
105 - Plan presented with no rationale for its phase structure.
106 - Plan invented with no reference to user requests or research.
107 priority: medium
108 contract_ref: "agent §Plan Creation (reference research and user requests)"
109
110 - expectation_id: no-source-edit
111 summary: Planning-only — no edits to source code or build manifests.
112 signal: Output does not reference modifications to source-tree files.
113 pass_criteria: |
114 No occurrences of edit/create verbs paired with `.cs`/`.py`/`.ts`/`.js`/
115 `.go`/`.rs`/`.java`/`package.json`/`pyproject.toml`/`Cargo.toml` paths.
116 Implementation steps appear in the plan, not as claimed edits.
117 failure_modes:
118 - Agent claims to have started implementing the plan.
119 - Modifies build manifests while planning.
120 priority: high
121 contract_ref: "agent scope (planning-only); current `no-source-edit` grader"
122