sauron-cli 1.2.0 → 1.3.0

package/OpenSpec-main/openspec/changes/archive/2026-02-17-project-config/design.md

@@ -0,0 +1,665 @@
+# Design: Project Config
+
+## Context
+
+OpenSpec currently has a fixed schema resolution order:
+1. `--schema` CLI flag
+2. `.openspec.yaml` in change directory
+3. Hardcoded default: `"spec-driven"`
+
+This forces users who want project-level customization to fork entire schemas, even for simple additions like injecting tech stack context or adding artifact-specific rules.
+
+The proposal introduces `openspec/config.yaml` as a lightweight customization layer that sits between preset schemas and full forking. It allows teams to:
+- Set a default schema
+- Inject project context into all artifacts
+- Add per-artifact rules
+
+**Constraints:**
+- Must not break existing changes that lack config
+- Must maintain clean separation between "configure" (this) and "fork" (project-local-schemas)
+- Config is project-level only (no global/user-level config)
+
+**Key stakeholders:**
+- OpenSpec users who need light customization without forking
+- Teams sharing workflow conventions via committed config
+
+## Goals / Non-Goals
+
+**Goals:**
+- Load and parse `openspec/config.yaml` using Zod schema
+- Use config's `schema` field as default in schema resolution
+- Inject `context` into all artifact instructions
+- Inject `rules` into matching artifact instructions only
+- Gracefully handle missing or invalid config (fallback to defaults)
+
+**Non-Goals:**
+- Structural changes to schemas (`skip`, `add`, inheritance) - those belong in fork path
+- File references for context (`context: ./file.md`) - start with strings
+- Global user-level config (XDG dirs, etc.)
+- Config management commands (`openspec config init`) - manual creation for now
+- Migration from old setups (no existing config to migrate from)
+
+## Decisions
+
+### 1. Config File Format: YAML vs JSON
+
+**Decision:** Use YAML (`.yaml` extension, support `.yml` alias)
+
+**Rationale:**
+- YAML supports multi-line strings naturally (`context: |`)
+- More readable for documentation-heavy content
+- Consistent with `.openspec.yaml` used in changes
+- Easy to parse with existing `yaml` library
+
+**Alternatives considered:**
+- JSON: More strict, but poor multi-line string UX
+- TOML: Less familiar to most users
+
+### 2. Config Location: Project Root vs openspec/ Directory
+
+**Decision:** `./openspec/config.yaml` (inside openspec directory)
+
+**Rationale:**
+- Co-located with `openspec/schemas/` (project-local-schemas)
+- Keeps project root clean
+- Natural namespace for OpenSpec configuration
+- Mirrors structure used by other tools (e.g., `.github/`)
+
+**Alternatives considered:**
+- `./openspec.config.yaml` in root: Pollutes root, less clear ownership
+- XDG config directories: Out of scope, no global config yet
+
+### 3. Context Injection: XML Tags vs Markdown Sections
+
+**Decision:** Use XML-style tags `<context>` and `<rules>`
+
+**Rationale:**
+- Clear delimiters that don't conflict with Markdown
+- Agents can easily parse structure
+- Matches existing patterns in the codebase for special sections
+
+**Example:**
+```xml
+<context>
+Tech stack: TypeScript, React
+</context>
+
+<rules>
+- Include rollback plan
+</rules>
+
+<template>
+## Summary
+...
+</template>
+```
+
+**Alternatives considered:**
+- Markdown headers: Conflicts with template content
+- Comments: Less visible to agents
+
+### 4. Schema Resolution: Insert Position
+
+**Decision:** Config's `schema` field goes between change metadata and hardcoded default
+
+**New resolution order:**
+1. `--schema` CLI flag (explicit override)
+2. `.openspec.yaml` in change directory (change-specific binding)
+3. **`openspec/config.yaml` schema field** (NEW - project default)
+4. `"spec-driven"` (hardcoded fallback)
+
+**Rationale:**
+- Preserves CLI and change-level overrides (most specific wins)
+- Makes config act as a "project default"
+- Backwards compatible (no existing configs to conflict with)
+
+### 5. Rules Validation: Strict vs Permissive
+
+**Decision:** Warn on unknown artifact IDs, don't error
+
+**Rationale:**
+- Future-proof: If schema adds new artifacts, old configs don't break
+- Dev experience: Typos show warnings, but don't halt workflow
+- User can fix incrementally
+
+**Example:**
+```yaml
+rules:
+  proposal: [...]
+  testplan: [...]  # Schema doesn't have this artifact → WARN, not ERROR
+```
+
+### 6. Error Handling: Config Parse Failures
+
+**Decision:** Log warning and fall back to defaults (don't halt commands)
+
+**Rationale:**
+- Syntax errors in config shouldn't break all of OpenSpec
+- User can fix config incrementally
+- Commands remain usable during config development
+
+**Warning message:**
+```
+⚠️  Failed to parse openspec/config.yaml: [error details]
+    Falling back to default schema (spec-driven)
+```
+
+## Implementation Plan
+
+### Phase 1: Core Types and Loading
+
+**File: `src/core/project-config.ts` (NEW)**
+
+```typescript
+import { z } from 'zod';
+import { readFileSync, existsSync } from 'fs';
+import { parse as parseYaml } from 'yaml';
+import { findProjectRoot } from '../utils/path-utils';
+
+/**
+ * Zod schema for project configuration.
+ *
+ * Purpose:
+ * 1. Documentation - clearly defines the config file structure
+ * 2. Type safety - TypeScript infers ProjectConfig type from schema
+ * 3. Runtime validation - uses safeParse() for resilient field-by-field validation
+ *
+ * Why Zod over manual validation:
+ * - Helps understand OpenSpec's data interfaces at a glance
+ * - Single source of truth for type and validation
+ * - Consistent with other OpenSpec schemas
+ */
+export const ProjectConfigSchema = z.object({
+  schema: z.string().min(1).describe('The workflow schema to use (e.g., "spec-driven", "tdd")'),
+  context: z.string().optional().describe('Project context injected into all artifact instructions'),
+  rules: z.record(
+    z.string(),
+    z.array(z.string())
+  ).optional().describe('Per-artifact rules, keyed by artifact ID'),
+});
+
+export type ProjectConfig = z.infer<typeof ProjectConfigSchema>;
+
+const MAX_CONTEXT_SIZE = 50 * 1024; // 50KB hard limit
+
+/**
+ * Read and parse openspec/config.yaml from project root.
+ * Uses resilient parsing - validates each field independently using Zod safeParse.
+ * Returns null if file doesn't exist.
+ * Returns partial config if some fields are invalid (with warnings).
+ */
+export function readProjectConfig(): ProjectConfig | null {
+  const projectRoot = findProjectRoot();
+
+  // Try both .yaml and .yml, prefer .yaml
+  let configPath = path.join(projectRoot, 'openspec', 'config.yaml');
+  if (!existsSync(configPath)) {
+    configPath = path.join(projectRoot, 'openspec', 'config.yml');
+    if (!existsSync(configPath)) {
+      return null; // No config is OK
+    }
+  }
+
+  try {
+    const content = readFileSync(configPath, 'utf-8');
+    const raw = parseYaml(content);
+
+    if (!raw || typeof raw !== 'object') {
+      console.warn(`⚠️  openspec/config.yaml is not a valid YAML object`);
+      return null;
+    }
+
+    const config: Partial<ProjectConfig> = {};
+
+    // Parse schema field using Zod
+    const schemaField = z.string().min(1);
+    const schemaResult = schemaField.safeParse(raw.schema);
+    if (schemaResult.success) {
+      config.schema = schemaResult.data;
+    } else if (raw.schema !== undefined) {
+      console.warn(`⚠️  Invalid 'schema' field in config (must be non-empty string)`);
+    }
+
+    // Parse context field with size limit
+    if (raw.context !== undefined) {
+      const contextField = z.string();
+      const contextResult = contextField.safeParse(raw.context);
+
+      if (contextResult.success) {
+        const contextSize = Buffer.byteLength(contextResult.data, 'utf-8');
+        if (contextSize > MAX_CONTEXT_SIZE) {
+          console.warn(
+            `⚠️  Context too large (${(contextSize / 1024).toFixed(1)}KB, limit: ${MAX_CONTEXT_SIZE / 1024}KB)`
+          );
+          console.warn(`   Ignoring context field`);
+        } else {
+          config.context = contextResult.data;
+        }
+      } else {
+        console.warn(`⚠️  Invalid 'context' field in config (must be string)`);
+      }
+    }
+
+    // Parse rules field using Zod
+    if (raw.rules !== undefined) {
+      const rulesField = z.record(z.string(), z.array(z.string()));
+
+      // First check if it's an object structure
+      if (typeof raw.rules === 'object' && !Array.isArray(raw.rules)) {
+        const parsedRules: Record<string, string[]> = {};
+        let hasValidRules = false;
+
+        for (const [artifactId, rules] of Object.entries(raw.rules)) {
+          const rulesArrayResult = z.array(z.string()).safeParse(rules);
+
+          if (rulesArrayResult.success) {
+            // Filter out empty strings
+            const validRules = rulesArrayResult.data.filter(r => r.length > 0);
+            if (validRules.length > 0) {
+              parsedRules[artifactId] = validRules;
+              hasValidRules = true;
+            }
+            if (validRules.length < rulesArrayResult.data.length) {
+              console.warn(
+                `⚠️  Some rules for '${artifactId}' are empty strings, ignoring them`
+              );
+            }
+          } else {
+            console.warn(
+              `⚠️  Rules for '${artifactId}' must be an array of strings, ignoring this artifact's rules`
+            );
+          }
+        }
+
+        if (hasValidRules) {
+          config.rules = parsedRules;
+        }
+      } else {
+        console.warn(`⚠️  Invalid 'rules' field in config (must be object)`);
+      }
+    }
+
+    // Return partial config even if some fields failed
+    return Object.keys(config).length > 0 ? (config as ProjectConfig) : null;
+
+  } catch (error) {
+    console.warn(`⚠️  Failed to parse openspec/config.yaml:`, error);
+    return null;
+  }
+}
+
+/**
+ * Validate artifact IDs in rules against a schema's artifacts.
+ * Called during instruction loading (when schema is known).
+ * Returns warnings for unknown artifact IDs.
+ */
+export function validateConfigRules(
+  rules: Record<string, string[]>,
+  validArtifactIds: Set<string>,
+  schemaName: string
+): string[] {
+  const warnings: string[] = [];
+
+  for (const artifactId of Object.keys(rules)) {
+    if (!validArtifactIds.has(artifactId)) {
+      const validIds = Array.from(validArtifactIds).sort().join(', ');
+      warnings.push(
+        `Unknown artifact ID in rules: "${artifactId}". ` +
+        `Valid IDs for schema "${schemaName}": ${validIds}`
+      );
+    }
+  }
+
+  return warnings;
+}
+
+/**
+ * Suggest valid schema names when user provides invalid schema.
+ * Uses fuzzy matching to find similar names.
+ */
+export function suggestSchemas(
+  invalidSchemaName: string,
+  availableSchemas: { name: string; isBuiltIn: boolean }[]
+): string {
+  // Simple fuzzy match: Levenshtein distance
+  function levenshtein(a: string, b: string): number {
+    const matrix: number[][] = [];
+    for (let i = 0; i <= b.length; i++) {
+      matrix[i] = [i];
+    }
+    for (let j = 0; j <= a.length; j++) {
+      matrix[0][j] = j;
+    }
+    for (let i = 1; i <= b.length; i++) {
+      for (let j = 1; j <= a.length; j++) {
+        if (b.charAt(i - 1) === a.charAt(j - 1)) {
+          matrix[i][j] = matrix[i - 1][j - 1];
+        } else {
+          matrix[i][j] = Math.min(
+            matrix[i - 1][j - 1] + 1,
+            matrix[i][j - 1] + 1,
+            matrix[i - 1][j] + 1
+          );
+        }
+      }
+    }
+    return matrix[b.length][a.length];
+  }
+
+  // Find closest matches (distance <= 3)
+  const suggestions = availableSchemas
+    .map(s => ({ ...s, distance: levenshtein(invalidSchemaName, s.name) }))
+    .filter(s => s.distance <= 3)
+    .sort((a, b) => a.distance - b.distance)
+    .slice(0, 3);
+
+  const builtIn = availableSchemas.filter(s => s.isBuiltIn).map(s => s.name);
+  const projectLocal = availableSchemas.filter(s => !s.isBuiltIn).map(s => s.name);
+
+  let message = `❌ Schema '${invalidSchemaName}' not found in openspec/config.yaml\n\n`;
+
+  if (suggestions.length > 0) {
+    message += `Did you mean one of these?\n`;
+    suggestions.forEach(s => {
+      const type = s.isBuiltIn ? 'built-in' : 'project-local';
+      message += `  - ${s.name} (${type})\n`;
+    });
+    message += '\n';
+  }
+
+  message += `Available schemas:\n`;
+  if (builtIn.length > 0) {
+    message += `  Built-in: ${builtIn.join(', ')}\n`;
+  }
+  if (projectLocal.length > 0) {
+    message += `  Project-local: ${projectLocal.join(', ')}\n`;
+  } else {
+    message += `  Project-local: (none found)\n`;
+  }
+
+  message += `\nFix: Edit openspec/config.yaml and change 'schema: ${invalidSchemaName}' to a valid schema name`;
+
+  return message;
+}
+```
+
+### Phase 2: Schema Resolution
+
+**File: `src/utils/change-metadata.ts`**
+
+Update `resolveSchemaForChange()` to check config:
+
+```typescript
+export function resolveSchemaForChange(
+  changeName: string,
+  cliSchema?: string
+): string {
+  // 1. CLI flag wins
+  if (cliSchema) {
+    return cliSchema;
+  }
+
+  // 2. Change metadata (.openspec.yaml)
+  const metadata = readChangeMetadata(changeName);
+  if (metadata?.schema) {
+    return metadata.schema;
+  }
+
+  // 3. Project config (NEW)
+  const projectConfig = readProjectConfig();
+  if (projectConfig?.schema) {
+    return projectConfig.schema;
+  }
+
+  // 4. Hardcoded default
+  return 'spec-driven';
+}
+```
+
+**File: `src/utils/change-utils.ts`**
+
+Update `createNewChange()` to use config schema:
+
+```typescript
+export function createNewChange(
+  changeName: string,
+  schema?: string
+): void {
+  // Use schema from config if not specified
+  const resolvedSchema = schema ?? readProjectConfig()?.schema ?? 'spec-driven';
+
+  // ... rest of change creation logic
+}
+```
+
+### Phase 3: Instruction Injection and Validation
+
+**File: `src/core/artifact-graph/instruction-loader.ts`**
+
+Update `loadInstructions()` to inject context, rules, and validate artifact IDs:
+
+```typescript
+// Session-level cache for validation warnings (avoid repeating same warnings)
+const shownWarnings = new Set<string>();
+
+export function loadInstructions(
+  changeName: string,
+  artifactId: string
+): InstructionOutput {
+  const projectConfig = readProjectConfig();
+
+  // Load base instructions from schema
+  const baseInstructions = loadSchemaInstructions(changeName, artifactId);
+  const schema = getSchemaForChange(changeName); // Assumes we have schema loaded
+
+  // Validate rules artifact IDs (only once per session)
+  if (projectConfig?.rules) {
+    const validArtifactIds = new Set(schema.artifacts.map(a => a.id));
+    const warnings = validateConfigRules(
+      projectConfig.rules,
+      validArtifactIds,
+      schema.name
+    );
+
+    // Show each unique warning only once per session
+    for (const warning of warnings) {
+      if (!shownWarnings.has(warning)) {
+        console.warn(`⚠️  ${warning}`);
+        shownWarnings.add(warning);
+      }
+    }
+  }
+
+  // Build enriched instruction with XML sections
+  let enrichedInstruction = '';
+
+  // Add context (all artifacts)
+  if (projectConfig?.context) {
+    enrichedInstruction += `<context>\n${projectConfig.context}\n</context>\n\n`;
+  }
+
+  // Add rules (only for matching artifact)
+  const rulesForArtifact = projectConfig?.rules?.[artifactId];
+  if (rulesForArtifact && rulesForArtifact.length > 0) {
+    enrichedInstruction += `<rules>\n`;
+    for (const rule of rulesForArtifact) {
+      enrichedInstruction += `- ${rule}\n`;
+    }
+    enrichedInstruction += `</rules>\n\n`;
+  }
+
+  // Add original template
+  enrichedInstruction += `<template>\n${baseInstructions.template}\n</template>`;
+
+  return {
+    ...baseInstructions,
+    instruction: enrichedInstruction,
+  };
+}
+```
+
+**Note on validation timing:** Rules are validated lazily during instruction loading (not at config load time) because:
+1. Schema isn't known at config load time (circular dependency)
+2. Warnings shown when user actually uses the feature (better UX)
+3. Validation warnings cached per session to avoid spam
+
+### Phase 4: Performance and Caching
+
+**Why config is read multiple times:**
+
+```typescript
+// Example: "openspec instructions proposal --change my-feature"
+
+// 1. Schema resolution (to know which schema to use)
+resolveSchemaForChange('my-feature')
+  → readProjectConfig()  // Read #1
+
+// 2. Instruction loading (to inject context and rules)
+loadInstructions('my-feature', 'proposal')
+  → readProjectConfig()  // Read #2
+
+// Result: Config read twice per command
+// More complex commands may read 3-5 times
+```
+
+**Performance Strategy:**
+
+V1 approach: No caching, read config fresh each time
+- Simpler implementation
+- No cache invalidation complexity
+- Acceptable if config reads are fast enough
+
+**Benchmark targets:**
+- Typical config (1KB context, 5 artifact rules): **< 10ms** per read (imperceptible even 5x)
+- Large config (50KB context limit): **< 50ms** per read (acceptable for rare case)
+
+**If benchmarks fail:** Add simple caching:
+
+```typescript
+// Simple in-memory cache with no invalidation
+let cachedConfig: { mtime: number; config: ProjectConfig | null } | null = null;
+
+export function readProjectConfig(): ProjectConfig | null {
+  const projectRoot = findProjectRoot();
+  const configPath = path.join(projectRoot, 'openspec', 'config.yaml');
+
+  if (!existsSync(configPath)) {
+    return null;
+  }
+
+  const stats = statSync(configPath);
+  const mtime = stats.mtimeMs;
+
+  // Return cached config if file hasn't changed
+  if (cachedConfig && cachedConfig.mtime === mtime) {
+    return cachedConfig.config;
+  }
+
+  // Read and parse config
+  const config = parseConfigFile(configPath); // Extracted logic
+
+  // Cache result
+  cachedConfig = { mtime, config };
+  return config;
+}
+```
+
+**Performance testing task:** Add to Phase 6 (Testing)
+- Measure typical config read time (1KB context)
+- Measure large config read time (50KB context limit)
+- Measure repeated reads within single command
+- Document results, add caching only if needed
+
+## Data Flow
+
+```
+┌──────────────────────────────────────────────────────────────┐
+│                                                              │
+│  User runs: openspec instructions proposal --change foo     │
+│                                                              │
+└────────────────────────────┬─────────────────────────────────┘
+                             │
+                             ▼
+┌──────────────────────────────────────────────────────────────┐
+│  resolveSchemaForChange("foo")                               │
+│                                                              │
+│  1. Check CLI flag ✗                                         │
+│  2. Check .openspec.yaml ✗                                   │
+│  3. Check openspec/config.yaml ✓ → "spec-driven"             │
+│                                                              │
+└────────────────────────────┬─────────────────────────────────┘
+                             │
+                             ▼
+┌──────────────────────────────────────────────────────────────┐
+│  loadInstructions("foo", "proposal")                         │
+│                                                              │
+│  1. Load spec-driven/artifacts/proposal.yaml                 │
+│  2. Read openspec/config.yaml                                │
+│  3. Build enriched instruction:                              │
+│     - <context>...</context>                                 │
+│     - <rules>...</rules>  (if rules.proposal exists)         │
+│     - <template>...</template>                               │
+│                                                              │
+└────────────────────────────┬─────────────────────────────────┘
+                             │
+                             ▼
+┌──────────────────────────────────────────────────────────────┐
+│  Return InstructionOutput with enriched content              │
+│                                                              │
+│  Agent sees project context + rules + schema template        │
+│                                                              │
+└──────────────────────────────────────────────────────────────┘
+```
+
+## Risks / Trade-offs
+
+**[Risk]** Config typos silently ignored (e.g., wrong artifact ID in rules)
+→ **Mitigation:** Validate and warn on unknown artifact IDs during config load. Don't error to allow forward compatibility.
+
+**[Risk]** Context grows too large, pollutes all artifact instructions
+→ **Mitigation:** Document recommended size (< 500 chars). If this becomes an issue, add per-artifact context override later.
+
+**[Risk]** YAML parsing errors break OpenSpec commands
+→ **Mitigation:** Catch parse errors, log warning, fall back to defaults. Commands remain functional.
+
+**[Risk]** Config cached incorrectly across commands
+→ **Mitigation:** Read config fresh on each `readProjectConfig()` call. No caching layer for v1 (simplicity over perf).
+
+**[Trade-off]** Context is injected into ALL artifacts
+→ **Benefit:** Consistent project knowledge across workflow
+→ **Cost:** Can't scope context to specific artifacts (yet)
+→ **Future:** Add `context: { global: "...", proposal: "..." }` if needed
+
+**[Trade-off]** Rules use artifact IDs, not human names
+→ **Benefit:** Stable identifiers (IDs don't change)
+→ **Cost:** User needs to know artifact IDs from schema
+→ **Mitigation:** Document common artifact IDs, show in `openspec status` output
+
+## Migration Plan
+
+**No migration needed** - this is a new feature with no existing state.
+
+**Rollout steps:**
+1. Deploy with config loading behind feature flag (optional, for safety)
+2. Test with internal project (this repo)
+3. Document in README with examples
+4. Remove feature flag if used
+
+**Rollback strategy:**
+- Config is additive only (doesn't break existing changes)
+- If bugs found, config parsing can be disabled with env var
+- Users can delete config file to restore old behavior
+
+## Open Questions
+
+**Q: Should context support file references (`context: ./CONTEXT.md`)?**
+**A (deferred):** Start with string-only. Add file reference later if users request it. Keeps v1 simple.
+
+**Q: Should we support `.yml` alias in addition to `.yaml`?**
+**A:** Yes, check both extensions. Prefer `.yaml` in docs, but accept `.yml` for users who prefer it.
+
+**Q: What if config's schema field references a non-existent schema?**
+**A:** Schema resolution will fail downstream. Show error when trying to load schema, suggest valid schema names.
+
+**Q: Should rules be validated against the resolved schema's artifact IDs?**
+**A:** Yes, validate and warn, but don't halt. This allows forward compatibility if schema evolves.