Totem Shield

Version updated for https://github.com/mmnto-ai/totem to version @mmnto/totem@1.15.4.

  • This action is used across all versions by 0 repositories.

Action Type

This is a Composite action.

Go to the GitHub Marketplace to find the latest changes.

Action Summary

Totem is a CLI tool designed to provide a persistent memory and enforcement layer for AI coding agents, ensuring architectural integrity across repositories. It automates the process of converting plain-English lessons about coding mistakes into enforceable linting rules, which are applied locally and offline to prevent recurring errors during development workflows. By acting as a “tripwire,” Totem addresses challenges like context loss and redundant mistakes, enhancing collaboration between AI agents and human developers without relying on runtime large language models (LLMs).

What’s Changed

Patch Changes

  • d295439: 1.15.4 bundles two compile-worker prompt classifier improvements that surfaced from downstream consumer friction on mmnto-ai/liquid-city. Both close fidelity gaps between the lesson prose authors wrote and the compiled rule that shipped.

    Test-contract scope classifier (closes #1626)

    • New ### Test-Contract Scope Classifier (mmnto-ai/totem#1626) section on both COMPILER_SYSTEM_PROMPT and PIPELINE3_COMPILER_PROMPT. Teaches the compile-worker to recognize lessons whose hazard is behavior inside test files (assertion conventions, spy / mock contracts, test-fixture hygiene) and emit test-inclusive fileGlobs instead of the default !**/*.test.* exclusion.
    • Three positive signals classify a lesson as test-contract: the testing tag, test-framework calls in badExample/goodExample (describe(, it(, test(, expect(, vi.mock(, jest.mock(, beforeEach(, afterEach(, vi.spyOn(, jest.spyOn(), or lesson-body references to test-execution-specific behavior.
    • Broad test-inclusive glob set for test-contract rules: ["**/*.test.*", "**/*.spec.*", "**/tests/**/*.*", "**/__tests__/**/*.*"]. Narrow test-scoped globs (e.g., packages/e2e/**/*.spec.ts) are preserved when the lesson clearly targets them.
    • False-positive trap guard: the word “contract” alone does NOT classify a lesson as test-scoped. Lessons titled “Define strict API Data Contracts” or “Versioning contracts for REST endpoints” describe application-surface invariants. Classification requires the testing tag OR test-framework code in the examples alongside any keyword match.

    Downstream impact: Two liquid-city rules ("Normalize temp paths for cross-platform equality", "Spy on logger contracts in tests") were shipping with scopes that excluded tests and silently never fired. A follow-up chore cycle (totem compile --upgrade <hash> per rule) retriages existing corpus against the new prompt.

    Declared severity override (closes #1656)

    • New parseDeclaredSeverity(body: string) helper exported from @mmnto/totem. Parses **Severity:** error / Severity: warning prose declarations from a lesson body and returns a normalized 'error' | 'warning' | undefined. Tolerates common markdown and punctuation shapes: bold markers (**, *, _) on either side, backtick-wrapped values, trailing sentence punctuation (., ,, ;, :, !, ?), and combined shapes like **Severity: error**.. Strict enum equality follows the strip, so out-of-vocabulary tokens (info, critical) return undefined.
    • buildCompiledRule honors a new declaredSeverityOverride option on BuildCompiledRuleOptions. Post-LLM override wins over parsed.severity regardless of LLM emission. Marker fires in BuildRuleResult.severityOverride only when the override actually changed the outcome (declared value differs from emittedSeverity ?? 'warning'). Marker is threaded through rejection paths too, so telemetry captures prompt-drift even when the rule fails for other reasons.
    • New onSeverityOverride callback on CompileLessonCallbacks fires when the override changes the emitted severity. CLI compile.ts wires a writeSeverityOverrideTelemetry closure that appends records tagged type: 'severity-override' to .totem/temp/telemetry.jsonl via the cwd-aware totemDir (matches the mmnto-ai/totem#1645 pattern). Fire-and-forget; sink failures do not interfere with compile results.
    • New ### Declared Severity (mmnto-ai/totem#1656) directive section on both compile prompts instructs the LLM to honor prose-declared severity in its emitted JSON. Every Output Schema example and every concrete Lesson → Output few-shot example now carries "severity": "warning" (the default) to reduce drift at source.

    Downstream impact: Five liquid-city ADR-008 rules on PR 77 burned ~10 manual severity-edit commits across R2 + R3 rounds because the compile pipeline emitted "severity": "warning" despite lesson prose declaring Severity: error. The mechanical re-edit loop closes; the next totem lesson compile cycle on LC emits declared severity directly.

    Strategy submodule bump

    • .strategy submodule pointer advances from 113179c to 7892892b. Picks up strategy PR #125 (upstream-feedback items 015 + 016 from liquid-city session-17) and strategy PR #124 (upstream-feedback item 017 — three-layer language support gap addendum that documents the architectural surface of the pending Rust-support arc).