eq-layer 0.2.4__tar.gz → 0.2.6__tar.gz

eq_layer-0.2.6/.codex-plugin/plugin.json

@@ -0,0 +1,6 @@
+{
+  "name": "eq-layer",
+  "version": "0.2.6",
+  "description": "Inspectable EQ-Layer response-control skill.",
+  "skills": "./skills/"
+}

eq_layer-0.2.6/.github/ISSUE_TEMPLATE/bug_report.yml

@@ -0,0 +1,115 @@
+name: Bug report
+description: A reproducible defect in the library, CLI, packaging, or documented behaviour
+title: "[bug] "
+labels: ["bug"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Use the Case template when the disagreement is about which policy should
+        apply, or when two policies seem equally right. Use this form for a
+        reproducible defect: a crash, traceback, CLI or packaging failure, or
+        output that contradicts the documentation. Policy changes go through an
+        RFC; see CONTRIBUTING.md.
+
+        For a suspected vulnerability, use the private reporting route in
+        SECURITY.md. Do not open a public issue with exploit details.
+
+  - type: textarea
+    id: expected-actual
+    attributes:
+      label: What happened, and what did you expect?
+      description: Quote the exact output and name the documented behaviour it contradicts.
+      placeholder: |
+        `Selector().select(state, intent)` raised `AttributeError` for a state
+        that the policy table documents as valid.
+    validations:
+      required: true
+
+  - type: textarea
+    id: reproduction
+    attributes:
+      label: Minimal reproduction
+      description: The smallest transcript, Python snippet, or CLI invocation that shows the problem.
+      placeholder: |
+        1. Install with `python -m pip install -e ".[dev]"`
+        2. Run the snippet or `eq-layer route` command below.
+        3. Observe the selection or error.
+    validations:
+      required: true
+
+  - type: textarea
+    id: output
+    attributes:
+      label: Exact output, errors, and exit code
+      description: Include the traceback and exit code, not only the final message.
+      render: shell
+    validations:
+      required: true
+
+  - type: input
+    id: version
+    attributes:
+      label: Version and install source
+      description: Which version are you running, and was it installed from PyPI, npm, or a local checkout?
+      placeholder: "0.2.5, local checkout"
+    validations:
+      required: true
+
+  - type: dropdown
+    id: surface
+    attributes:
+      label: Which surface is affected?
+      multiple: true
+      options:
+        - "Python library (`eq_layer`)"
+        - "CLI (`eq-layer`)"
+        - "npm installer or CLI bridge"
+        - "Agent Skill (`skills/eq-layer/`)"
+        - "Evaluation harness (`eval/`)"
+        - "Install, build, or packaging"
+        - "Documentation"
+        - "Not sure"
+    validations:
+      required: true
+
+  - type: input
+    id: environment
+    attributes:
+      label: Environment
+      description: Operating system, Python version, and Node version if the npm package is involved.
+      placeholder: "Windows 11, CPython 3.12, Node.js 24"
+    validations:
+      required: true
+
+  - type: textarea
+    id: state
+    attributes:
+      label: Relevant control state
+      description: If routing or selection is involved, paste the policy name, intent belief, affect state, or repair state. Redact private transcript text.
+    validations:
+      required: false
+
+  - type: textarea
+    id: additional
+    attributes:
+      label: Additional context
+      description: Logs, related issues, or a proposed cause. Redact credentials and private data.
+    validations:
+      required: false
+
+  - type: checkboxes
+    id: checks
+    attributes:
+      label: Before submitting
+      options:
+        - label: I searched the open and closed issues for this problem.
+          required: true
+        - label: I read SECURITY.md and this is a defect, not a vulnerability.
+          required: true
+        - label: I removed credentials, private transcripts, and local paths from what I pasted.
+          required: true
+        - label: This is not a policy proposal; policy proposals belong in an RFC.
+          required: true
+        - label: I can reproduce this on the current version or on `main`.
+          required: true

eq_layer-0.2.6/.github/ISSUE_TEMPLATE/case.yml

@@ -0,0 +1,50 @@
+name: Case
+description: Add a transcript where two policies seem equally right, or where the current behaviour is clearly wrong
+labels: [case, "good first issue"]
+body:
+  - type: textarea
+    id: transcript
+    attributes:
+      label: Transcript
+      description: Alternate user/assistant turns. Plain JSON, same shape as eval/cases.jsonl
+      placeholder: |
+        [
+          {"role": "user", "content": "..."},
+          {"role": "assistant", "content": "..."},
+          {"role": "user", "content": "..."}
+        ]
+    validations:
+      required: true
+
+  - type: textarea
+    id: expectations
+    attributes:
+      label: What should happen, and why
+      description: If you are not sure which policy is right, say so. That ambiguity is the interesting part.
+    validations:
+      required: true
+
+  - type: input
+    id: concepts
+    attributes:
+      label: expected_concepts
+      description: Comma-separated words that a good response must contain
+    validations:
+      required: false
+
+  - type: dropdown
+    id: ambiguity
+    attributes:
+      label: Which policy is it, if any?
+      options:
+        - unclear — this is why I filed it
+        - mirror_specific
+        - validate_then_redirect
+        - direct_no_padding
+        - ask_one_question
+        - deflate_tension
+        - repair_hold_position
+        - hold
+        - boundary
+    validations:
+      required: true

eq_layer-0.2.6/.github/ISSUE_TEMPLATE/config.yml

@@ -0,0 +1,11 @@
+blank_issues_enabled: false
+contact_links:
+  - name: Security vulnerability
+    url: https://github.com/Furox-Art/eq-layer/security/advisories/new
+    about: Private reporting route described in SECURITY.md. Do not open a public issue for a suspected vulnerability.
+  - name: Contributing and the RFC process
+    url: https://github.com/Furox-Art/eq-layer/blob/main/CONTRIBUTING.md
+    about: Start with a case. Policy changes go through an RFC before a pull request.
+  - name: Code of conduct
+    url: https://github.com/Furox-Art/eq-layer/blob/main/CODE_OF_CONDUCT.md
+    about: Report concerns to the maintainers as described in the Code of Conduct.

eq_layer-0.2.6/.github/ISSUE_TEMPLATE/feature_request.yml

@@ -0,0 +1,103 @@
+name: Feature request
+description: Propose a capability, metric, integration, or documentation change
+title: "[feature] "
+labels: ["enhancement"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Read CONTRIBUTING.md first. Cases are the main entry point: a transcript
+        where the current behaviour is wrong is more useful than a policy idea
+        on its own. Policy changes go through an RFC before a pull request.
+
+        A policy that cannot lose cannot be evaluated, so say where the
+        proposal would be the wrong move.
+
+  - type: textarea
+    id: problem
+    attributes:
+      label: What cannot be done today
+      description: Describe the situation, not the solution. Which transcript, case id, or workflow shows the gap?
+    validations:
+      required: true
+
+  - type: textarea
+    id: proposal
+    attributes:
+      label: Proposed change
+      description: The interface you have in mind, such as a policy, precondition, metric, CLI command, or packaging change.
+    validations:
+      required: true
+
+  - type: dropdown
+    id: category
+    attributes:
+      label: Area
+      options:
+        - New or changed policy (RFC required)
+        - Selector precondition or tie-breaking
+        - Scorer or evaluation metric
+        - Affect, intent, repair, or interaction model
+        - Steering or response auditor
+        - CLI, npm package, or Agent Skill packaging
+        - Evaluation harness or dataset
+        - Documentation
+        - Performance
+    validations:
+      required: true
+
+  - type: dropdown
+    id: routing
+    attributes:
+      label: Does this change policy routing?
+      options:
+        - "Yes, and an RFC is linked below"
+        - "Yes, but I need help drafting the RFC"
+        - "No"
+        - "Not sure"
+    validations:
+      required: true
+
+  - type: input
+    id: rfc
+    attributes:
+      label: RFC or case link
+      description: Link to `.github/rfcs/NNN-short-title.md`, an existing case, or the related issue.
+    validations:
+      required: false
+
+  - type: textarea
+    id: counter-case
+    attributes:
+      label: Where would this be the wrong move?
+      description: Name a transcript or case where this proposal should lose. Required for policy changes.
+    validations:
+      required: false
+
+  - type: textarea
+    id: evidence
+    attributes:
+      label: How would this be evaluated?
+      description: Name the case, metric, or harness result that would show the change works, including negative results.
+    validations:
+      required: false
+
+  - type: checkboxes
+    id: integrity
+    attributes:
+      label: How this keeps the evaluation honest
+      description: Tick what applies. If none apply, explain why in the proposal or evidence section.
+      options:
+        - label: I did not weaken a score or adjust a case to make it pass.
+        - label: A policy addition includes at least one case where the policy loses.
+        - label: Evidence levels (verified, direct_learned, derived, structural, fallback) remain accurate.
+        - label: Documentation claims stay within the measured result and do not imply an EQ benchmark that was not run.
+        - label: Not applicable.
+
+  - type: textarea
+    id: contribution
+    attributes:
+      label: Willing to implement it
+      description: Say what part you can take on. Cases, counter-cases, and harness improvements are all useful.
+    validations:
+      required: false

eq_layer-0.2.6/.github/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,69 @@
+<!--
+Thanks for contributing. Read CONTRIBUTING.md first.
+
+Cases are the main entry point. Policy changes need an RFC in `.github/rfcs/`
+before a pull request. Do not weaken a score to make a case pass, and do not
+include credentials or private transcript data.
+-->
+
+## What this changes
+
+<!-- One paragraph. What is different after this merges, and for whom. -->
+
+Closes #
+
+## Type
+
+- [ ] Case or eval addition
+- [ ] Policy change (RFC linked below)
+- [ ] Bug fix in the layer or CLI
+- [ ] npm package or Agent Skill packaging
+- [ ] Evaluation harness or metric
+- [ ] Documentation
+- [ ] CI, build, or release process
+
+RFC: <!-- link to .github/rfcs/NNN-short-title.md, or "not applicable" -->
+
+## Cases and evidence
+
+<!--
+What proves the change is correct? Paste commands and real output. For policy
+changes, name at least one case where the new policy is the wrong move. For
+metric changes, include the before/after result and state what it does and does
+not establish.
+-->
+
+```
+```
+
+## Score and claim integrity
+
+- [ ] I did not weaken a score or adjust a case to make it pass.
+- [ ] A policy addition includes at least one case where the policy loses.
+- [ ] Evidence levels (`verified`, `direct_learned`, `derived`, `structural`, `fallback`) remain accurate.
+- [ ] Documentation claims stay within the measured result; no EQ or benchmark claim is implied beyond the data.
+- [ ] Not applicable.
+
+## Checks run
+
+Paste results, including exit codes.
+
+- [ ] `python -m pip install -e ".[dev]" "pytest-cov>=5,<8"`
+- [ ] `python eval/run.py`
+- [ ] `node tools/crosscheck.js`
+- [ ] `python -m pytest -q -rs tests/ --cov=eq_layer --cov-branch --cov-report=term-missing --cov-fail-under=100`
+- [ ] `python -m compileall -q eq_layer eval`
+- [ ] `npm test` (if `npm/`, `package.json`, or `skills/` changed)
+- [ ] `python -m build` and `python -m twine check dist/*` (if packaging changed)
+- [ ] Not applicable
+
+```
+```
+
+## Review notes
+
+<!--
+Anything a reviewer should look at first, anything you deliberately left out,
+and anything you are unsure about. Saying "I am not sure this is right" is more
+useful here than a confident guess.
+-->

eq_layer-0.2.6/.github/RFC_TEMPLATE.md

@@ -0,0 +1,38 @@
+# RFC: <short title>
+
+## Status
+
+Proposed | Accepted | Rejected | Superseded
+
+## Problem
+
+What failure does this address? Name a concrete transcript or case id, not a category.
+
+"Empathy is too generic" is not a problem statement. "In case-002 the model capitulates entirely after correction" is.
+
+## Proposal
+
+What changes — new policy, changed precondition, new scorer metric.
+
+For policies:
+
+- **Name:**
+- **Register:**
+- **Preconditions:** which affect conditions must hold
+- **Avoid:** failure modes this policy is meant to prevent
+
+## Distinguishing cases
+
+Which existing cases does this *not* cover? A policy with no distinguishing case is a policy nobody can disagree about, which usually means it is too vague to be testable.
+
+Point at specific case ids.
+
+## Counter-cases
+
+Where is this policy the wrong move? A policy that cannot lose is a policy that cannot be evaluated.
+
+## Alternatives considered
+
+What else could resolve this? Including "do nothing" if that is defensible.
+
+## Open questions

eq_layer-0.2.6/.github/dependabot.yml

@@ -0,0 +1,37 @@
+version: 2
+updates:
+  - package-ecosystem: github-actions
+    directory: /
+    open-pull-requests-limit: 2
+    schedule:
+      interval: weekly
+    commit-message:
+      prefix: "ci"
+    labels:
+      - dependencies
+    groups:
+      routine-actions:
+        applies-to: version-updates
+        patterns:
+          - "*"
+        update-types:
+          - minor
+          - patch
+
+  - package-ecosystem: pip
+    directory: /
+    open-pull-requests-limit: 2
+    schedule:
+      interval: weekly
+    commit-message:
+      prefix: "deps"
+    labels:
+      - dependencies
+    groups:
+      routine-python:
+        applies-to: version-updates
+        patterns:
+          - "*"
+        update-types:
+          - minor
+          - patch

eq_layer-0.2.6/.github/mypy-baseline.json

@@ -0,0 +1,23 @@
+{
+  "eq_layer/ab_eval.py": 19,
+  "eq_layer/actions.py": 1,
+  "eq_layer/affect.py": 11,
+  "eq_layer/auditor.py": 4,
+  "eq_layer/cli.py": 5,
+  "eq_layer/coarse_discourse.py": 2,
+  "eq_layer/dbdc3.py": 6,
+  "eq_layer/dialogbank_iso.py": 2,
+  "eq_layer/dialogue_reference.py": 4,
+  "eq_layer/intent.py": 2,
+  "eq_layer/interaction.py": 6,
+  "eq_layer/policies.py": 1,
+  "eq_layer/preferences.py": 3,
+  "eq_layer/response_experiment.py": 23,
+  "eq_layer/steer.py": 4,
+  "eq_layer/tracker.py": 7,
+  "eq_layer/trained_affect.py": 10,
+  "eq_layer/trained_breakdown.py": 6,
+  "eq_layer/trained_disagreement.py": 6,
+  "eq_layer/trained_intent.py": 11,
+  "eq_layer/trained_subtext.py": 18
+}

eq_layer-0.2.6/.github/scripts/mypy_ratchet.py

@@ -0,0 +1,236 @@
+"""Strict mypy ratchet for eq-layer.
+
+Runs ``mypy --strict`` over ``eq_layer`` and compares the per-file error count
+with the committed baseline in ``.github/mypy-baseline.json``.
+
+The gate fails when a file has more errors than its baseline, or when a file
+has errors but no baseline entry at all. Errors may only move down: a file
+whose count falls below its baseline is reported as fixed and passes. The
+baseline itself is regenerated from a fresh measurement with::
+
+    python .github/scripts/mypy_ratchet.py --write-baseline
+
+Exit codes: 0 when the ratchet holds (or the baseline was written), 1 when it
+does not.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import re
+import subprocess
+import sys
+import tempfile
+from collections import Counter
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[2]
+SOURCE = "eq_layer"
+BASELINE = ROOT / ".github" / "mypy-baseline.json"
+MYPY_ARGS = (
+    "--strict",
+    "--ignore-missing-imports",
+    "--no-error-summary",
+    "--show-error-codes",
+)
+ERROR_LINE = re.compile(
+    r"^(?P<path>.+?):(?P<line>\d+)(?::\d+)?: error: (?P<message>.*)$"
+)
+ANSI_ESCAPE = re.compile(r"\x1b\[[0-9;]*m")
+
+
+def _normalize_path(raw: str) -> str:
+    """Return a repo-relative path with forward slashes."""
+    path = Path(raw.replace("\\", "/"))
+    if path.is_absolute():
+        try:
+            path = path.resolve().relative_to(ROOT)
+        except ValueError:
+            return path.as_posix()
+    return path.as_posix()
+
+
+def _run_mypy() -> tuple[int, Counter[str], list[str], list[str]]:
+    """Run mypy and return (exit code, per-file counts, unattributed, output)."""
+    env = os.environ.copy()
+    # Keep generated cache files out of the working tree; mypy honours this
+    # environment variable in place of a --cache-dir argument.
+    with tempfile.TemporaryDirectory(prefix="mypy-ratchet-") as cache_dir:
+        env["MYPY_CACHE_DIR"] = cache_dir
+        proc = subprocess.run(
+            [sys.executable, "-m", "mypy", *MYPY_ARGS, SOURCE],
+            cwd=str(ROOT),
+            capture_output=True,
+            text=True,
+            encoding="utf-8",
+            errors="replace",
+            env=env,
+            check=False,
+        )
+
+    output = proc.stdout.splitlines() + proc.stderr.splitlines()
+    counts: Counter[str] = Counter()
+    unattributed: list[str] = []
+    for raw_line in output:
+        line = ANSI_ESCAPE.sub("", raw_line).strip()
+        match = ERROR_LINE.match(line)
+        if match:
+            counts[_normalize_path(match.group("path"))] += 1
+        elif ": error: " in line:
+            unattributed.append(line)
+    return proc.returncode, counts, unattributed, output
+
+
+def _measure() -> Counter[str] | None:
+    """Measure the tree, or print a failure and return None if mypy broke."""
+    returncode, counts, unattributed, output = _run_mypy()
+    problems: list[str] = []
+    if returncode not in (0, 1):
+        problems.append(f"mypy exited with status {returncode} (expected 0 or 1)")
+    if unattributed:
+        problems.append("mypy emitted errors that could not be attributed to a file")
+    if returncode == 1 and not counts and not unattributed:
+        problems.append("mypy reported failure but no file-scoped errors were parsed")
+    if problems:
+        print("FAIL: " + "; ".join(problems), file=sys.stderr)
+        for line in output[-25:]:
+            print(f"  {line}", file=sys.stderr)
+        return None
+    return counts
+
+
+def _load_baseline() -> dict[str, int]:
+    """Read the baseline, rejecting malformed or ambiguous entries."""
+    relative = BASELINE.relative_to(ROOT)
+    if not BASELINE.exists():
+        raise SystemExit(
+            f"FAIL: {relative} is missing; create it with --write-baseline"
+        )
+    try:
+        raw = json.loads(BASELINE.read_text(encoding="utf-8"))
+    except (OSError, json.JSONDecodeError) as exc:
+        raise SystemExit(f"FAIL: cannot read {relative}: {exc}") from None
+    if not isinstance(raw, dict):
+        raise SystemExit(f"FAIL: {relative} must contain a JSON object")
+
+    baseline: dict[str, int] = {}
+    for key, value in raw.items():
+        if not isinstance(key, str) or not isinstance(value, int) or isinstance(value, bool):
+            raise SystemExit(
+                f"FAIL: {relative} entry {key!r}: count must be a non-negative integer"
+            )
+        if value < 0:
+            raise SystemExit(
+                f"FAIL: {relative} entry {key!r}: count must be a non-negative integer"
+            )
+        path = _normalize_path(key)
+        if path in baseline:
+            raise SystemExit(f"FAIL: {relative} lists {path!r} more than once")
+        baseline[path] = value
+    return baseline
+
+
+def _write_baseline() -> int:
+    """Replace the baseline with the current measurement."""
+    counts = _measure()
+    if counts is None:
+        return 1
+    payload = {path: counts[path] for path in sorted(counts) if counts[path] > 0}
+    BASELINE.parent.mkdir(parents=True, exist_ok=True)
+    text = json.dumps(payload, indent=2) + "\n"
+    BASELINE.write_text(text, encoding="utf-8", newline="\n")
+
+    relative = BASELINE.relative_to(ROOT)
+    total = sum(payload.values())
+    print(f"wrote {relative}: {total} error(s) in {len(payload)} file(s)")
+    for path, count in payload.items():
+        print(f"  {count:>5}  {path}")
+    return 0
+
+
+def _print_diff(counts: Counter[str], baseline: dict[str, int]) -> None:
+    """Print one row per file whose count differs from the baseline."""
+    paths = sorted(set(counts) | set(baseline))
+    rows = [
+        (path, baseline.get(path, 0), counts.get(path, 0))
+        for path in paths
+        if baseline.get(path, 0) != counts.get(path, 0)
+    ]
+    if not rows:
+        print("per-file diff: no changes")
+        return
+
+    width = max(len("FILE"), *(len(path) for path, _, _ in rows))
+    print(f"{'FILE':<{width}}  {'BASE':>6}  {'NOW':>6}  {'DELTA':>5}  STATUS")
+    for path, base, now in rows:
+        delta = now - base
+        if path not in baseline:
+            status = "NEW"
+        elif delta > 0:
+            status = "REGRESSED"
+        else:
+            status = "FIXED"
+        print(f"{path:<{width}}  {base:>6}  {now:>6}  {delta:>+5}  {status}")
+
+
+def _check() -> int:
+    """Compare a fresh measurement with the baseline."""
+    baseline = _load_baseline()
+    counts = _measure()
+    if counts is None:
+        return 1
+
+    total_now = sum(counts.values())
+    total_base = sum(baseline.values())
+    relative = BASELINE.relative_to(ROOT)
+    print(f"mypy --strict {SOURCE}: {total_now} error(s) in {len(counts)} file(s)")
+    print(f"baseline {relative}: {total_base} error(s) in {len(baseline)} file(s)")
+    _print_diff(counts, baseline)
+
+    regressions: list[tuple[str, int | None, int]] = []
+    for path in sorted(set(counts) | set(baseline)):
+        base = baseline.get(path)
+        now = counts.get(path, 0)
+        if now > 0 and base is None:
+            regressions.append((path, None, now))
+        elif base is not None and now > base:
+            regressions.append((path, base, now))
+
+    if regressions:
+        sys.stdout.flush()
+        print(
+            f"FAIL: {len(regressions)} file(s) exceed the mypy baseline",
+            file=sys.stderr,
+        )
+        for path, base, now in regressions:
+            if base is None:
+                print(f"  - {path}: no baseline entry, {now} error(s)", file=sys.stderr)
+            else:
+                print(f"  - {path}: baseline {base}, now {now} (+{now - base})", file=sys.stderr)
+        print(
+            "Fix the new errors, or review and accept the backlog with --write-baseline.",
+            file=sys.stderr,
+        )
+        return 1
+
+    print("PASS: no file exceeds the mypy baseline")
+    return 0
+
+
+def main(argv: list[str] | None = None) -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "--write-baseline",
+        action="store_true",
+        help="replace .github/mypy-baseline.json with the current counts",
+    )
+    args = parser.parse_args(argv)
+    if args.write_baseline:
+        return _write_baseline()
+    return _check()
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())