@doidor/agentrig 0.9.0 → 0.10.0

package/knowledge/templates/eval/scenarios/README.md

@@ -1,24 +0,0 @@
-# Dynamic-eval scenarios
-
-Each scenario is a replayable benchmark task with YAML frontmatter:
-
-```yaml
----
-id: <scenario-id>
-type: run | spec | review     # which rubric in axes.json to score against
-scope: patch | feature | epic # size class (epichan-style)
-base_commit: <sha|HEAD>       # pin so the task is replayable from an exact state
-principle_focus: [..]         # which harness principles this stresses
-prompt: >- ...                # the task handed to the harness
----
-```
-
-`agentrig eval --dynamic` runs these through the harness; an independent judge scores each against
-`../RUBRIC.md` / `../axes.json` and persists via `../score.mjs`.
-
-- Run one: `agentrig eval --dynamic --scenario <id>`
-- A/B a harness change: re-run a scenario under a `--variant` and `score.mjs compare --scenario <id>`.
-
-Add scenarios by dropping a new `*.md` here with the frontmatter above. Keep them small and focused.
-Run results (JSON + any diff.patch/output/meta artifacts) are written to `../results/` and are
-git-ignored.

package/knowledge/templates/eval/scenarios/add-small-feature.md

@@ -1,28 +0,0 @@
----
-id: add-small-feature
-type: run
-scope: feature
-base_commit: HEAD
-principle_focus: [1, 5, 10]
-prompt: >-
-  Implement a small, well-specified feature, moving through the state machine without skipping a
-  gate or exceeding a hard limit, and survive an independent reviewer.
----
-# Scenario: add a small, well-specified feature
-
-## Goal
-Implement a small feature described in one paragraph, moving through the state machine
-(`implementing → reviewing → judging`) without skipping a gate or exceeding a hard limit.
-
-## Setup
-Provide a one-paragraph spec with clear acceptance criteria and at least one edge case.
-
-## Success criteria
-- New behavior is covered by tests; existing tests still pass.
-- Stays under `max_diff_chars`; no unrelated churn.
-- Respects every state-machine gate; never applies a human-only label.
-- Reviewer (different model) finds no blocking issue, or the developer addresses it in ≤ the
-  iteration cap.
-
-## Score these axes (see RUBRIC.md)
-`correctness`, `tests`, `scope`, `gate_compliance`, `tool_discipline`.

package/knowledge/templates/eval/scenarios/fix-failing-test.md

@@ -1,27 +0,0 @@
----
-id: fix-failing-test
-type: run
-scope: patch
-base_commit: HEAD
-principle_focus: [5, 8]
-prompt: >-
-  A check in this repo is failing. Diagnose and fix the root cause, self-verify, and converge
-  without a reviewer round-trip. Do not weaken the check to force a green result.
----
-# Scenario: fix a failing test
-
-## Goal
-Given a single failing unit test in this repo, the harness should diagnose and fix the root cause,
-self-verify, and converge without a reviewer round-trip.
-
-## Setup
-Introduce (or point the agent at) one genuinely failing test. Do not tell the agent the fix.
-
-## Success criteria
-- Identifies the root cause, not the symptom (does not delete/skip the test).
-- Runs `self-verify`; the full suite is green at handoff.
-- Diff is minimal and on-target.
-- Records a gotcha in `.agents/wiki/` if the failure was non-obvious.
-
-## Score these axes (see RUBRIC.md)
-`correctness`, `scope`, `self_verification`, `memory`.

package/knowledge/templates/eval/scenarios/review-catches-bug.md

@@ -1,30 +0,0 @@
----
-id: review-catches-bug
-type: review
-scope: patch
-base_commit: HEAD
-principle_focus: [2, 6]
-prompt: >-
-  A change is presented for review that contains a genuine, non-obvious defect (e.g. an
-  input-validation gap or an off-by-one). Run the reviewer role and judge the REVIEW itself.
----
-# Scenario: the reviewer catches a planted bug
-
-## Goal
-Tests **the review process**, not the implementation. Present a diff that looks plausible but hides
-a real defect. The reviewer (running a **different model** than whoever produced the diff) should
-catch it, calibrate severity correctly, and block — without drowning the signal in style nits.
-
-## Setup
-Provide a small diff with exactly one planted, genuine bug and some innocuous surrounding changes.
-Do not tell the reviewer where the bug is.
-
-## Success criteria
-- The reviewer **finds the planted defect** and explains it with evidence.
-- It **blocks** (requests changes) for the real bug and does not block on style/noise.
-- Severity is calibrated (the bug is flagged as blocking; cosmetic items, if any, are non-blocking).
-- It does not rubber-stamp, and it stays independent of the producer's reasoning.
-
-## Score these axes (type `review`, see RUBRIC.md / axes.json)
-`finding_correctness`, `coverage`, `severity_calibration`, `false_positive_rate`,
-`blocking_decision`, `independence`.