Spec Index

Spec Memory Index

Use this skill to turn one-off discoveries into durable project memory and to retrieve that memory when a later task would benefit from it.

This skill is about memory with reuse value, not generic note taking. Capture decisions, setup knowledge, root causes, fix patterns, and constraints in a form that future planning and correction work can actually use.

Bundled resources

• references/template.md defines the canonical Markdown entry shape, YAML frontmatter, category field, and trigger tag fields.
• scripts/index.py provides deterministic add, best-match search, audit, and rebuild commands for Markdown knowledge bases.

Prefer the script for creating, searching, auditing, and rebuilding the index when the local environment can run Python. If the script cannot run, follow the same template manually and explain why the deterministic helper was not used.

Repository preference override

Before retrieving or saving memory, look for a SPEC.md in the target repository root.

If it exists, treat it as the repository-specific preference layer for:

• the knowledge-base location
• capture categories and storage expectations
• environment and configuration gotchas that are worth remembering
• cross-skill workflow conventions

If it is missing, use the built-in defaults for this skill series instead of blocking on the missing file.

If it is partial, use the fields it defines and fall back to defaults for the rest.

Built-in defaults

When no repository root SPEC.md is available, use these defaults:

• knowledge-base root: docs/knowledge-base/
• knowledge-base index: docs/knowledge-base/index.md
• entry template: references/template.md
• indexing helper: scripts/index.py

Search aliases

The search helper ships with no built-in synonym table so it stays project-neutral. Search-term synonyms are a repository-specific concern, so they live with the knowledge base rather than in the script.

To widen retrieval with project vocabulary, add an optional search-aliases.json in the knowledge-base root. Map each canonical term to a list of synonyms; when a query hits any of a group's terms, the search expands to include the rest. Terms may be in any language.

{
  "auth": ["authentication", "login", "session", "token", "登录", "鉴权"],
  "http-404": ["404", "not-found", "route-missing", "找不到"]
}

If the file is absent or malformed, search runs without alias expansion instead of failing.

Core responsibilities

• Save durable knowledge in a normalized, retrieval-friendly form.
• Retrieve relevant past knowledge using concrete keywords and context clues.
• Investigate, update, and reorganize existing knowledge-base entries when memory quality matters.
• Reduce repeated debugging and repeated planning by surfacing what is already known.
• Feed high-value context back into spec-plan and spec-crlp.

Operating modes

Choose one mode from the user's intent:

• Retrieve mode: Use when the user wants to know whether the project already has related decisions, root causes, setup notes, or fix patterns.
• Capture mode: Use when the user wants to save a new reusable finding, decision, setup rule, or root cause.
• Investigate mode: Use when the user wants to inspect, audit, clean up, reorganize, update, deduplicate, or validate existing knowledge-base entries.

If the user asks to "整理", "调查", "audit", "clean up", "review existing knowledge", "update old notes", "merge duplicates", or "fix tags", use investigate mode.

What belongs in the index

• Architecture or product decisions and the reason behind them.
• Requirement constraints and non-obvious boundaries.
• Setup steps, environment prerequisites, and configuration gotchas.
• Confirmed root causes and the fixes that resolved them.
• Reusable implementation patterns, migration rules, or troubleshooting heuristics.
• Known pitfalls, edge cases, and lessons learned from previous work.

Entry categories

Use one primary category per entry:

• decision: durable architecture, product, workflow, or tooling decision
• constraint: requirement boundary or non-obvious long-term limitation
• setup: environment, dependency, installation, or configuration prerequisite
• root-cause: confirmed cause of an observed failure
• fix-pattern: reusable correction strategy for a recurring class of bugs
• implementation-pattern: reusable implementation approach
• pitfall: known trap, anti-pattern, or edge case to avoid
• validation: lint, test, build, CI, or verification rule

If an entry seems to fit multiple categories, choose the category that best explains why future work should retrieve it.

Trigger tags

Trigger tags are retrieval hooks, not documentation prose.

Build tags from signals that future users, logs, or code searches are likely to contain:

• type:<category>
• phase:planning, phase:correction, phase:setup, or phase:retrieval
• domain:<area>
• component:<component-name>
• framework:<framework-name>
• tool:<tool-name>
• symptom:<observed-behavior>
• error:<normalized-error-name>
• file:<normalized-file-name>
• env:<environment-name>

Prefer lowercase, hyphenated tags. Keep them specific enough to retrieve the entry later, but avoid one-off tags that no future query is likely to use.

What does not belong in the index

• Raw chatter with no future reuse value.
• Large unstructured logs when a concise summary would do.
• Temporary guesses that have not been validated.
• Duplicate entries that say the same thing in different words.

Retrieval workflow

1. Identify what question you are trying to answer.
2. Build search terms from the problem domain, component name, symptom, tool, environment, and error wording.
3. Use scripts/index.py search against the knowledge-base root when Python is available.
4. If the script is unavailable, retrieve the most relevant matches manually from the location defined in the repository root SPEC.md when present, otherwise use the built-in default location.
5. Rank them by closeness to the current task.
6. Return concise takeaways, why they matter, and what action they suggest.
7. If nothing relevant is found, say so plainly instead of pretending there is memory that does not exist.

Use scripts/index.py search --explain when the reason for a match matters. The search helper uses weighted best-match scoring across trigger tags, titles, frontmatter fields, summaries, Applies When, root-cause sections, resolutions, and validation text. Treat search results as candidates to inspect, not as proof.

Capture workflow

1. Decide whether the new information is likely to help with future work.
2. Choose exactly one primary category.
3. Generate trigger tags from phase, domain, component, tool, framework, symptom, error, file, and environment signals.
4. Normalize the content using references/template.md.
5. Use scripts/index.py add when Python is available so the entry and index are updated deterministically. The helper automatically adds the type:<category> trigger tag if it is missing.
6. Preserve the context that explains when the entry applies.
7. Include the evidence, resolution, or decision rationale.
8. Avoid creating a second entry when an update to an existing one would be cleaner.

Investigate workflow

Use this workflow when maintaining an existing knowledge base:

1. Resolve the knowledge-base root from the repository root SPEC.md when present, otherwise use the built-in default location.
2. Read index.md and scan all memory entries under the knowledge-base root.
3. Check for broken index links, entries missing from the index, duplicate or overlapping entries, weak trigger tags, missing metadata, stale setup notes, conflicting rules, and entries without validation context.
4. Group findings by action: keep, update, merge, split, archive, or needs human review.
5. Prefer safe updates first: add missing trigger tags, fix category metadata, improve summaries, fill Applies When, and add validation context.
6. Ask before destructive changes such as deleting entries, overwriting substantial content, or merging entries where information could be lost.
7. Run scripts/index.py audit when Python is available to get a deterministic structural report.
8. After manual edits, run scripts/index.py rebuild when Python is available.
9. Produce an investigation report summarizing what changed, what was left untouched, and what still needs review.

Investigation checks

Check each entry for:

• exactly one primary type
• meaningful trigger_tags
• a clear Summary
• an Applies When section that explains retrieval conditions
• a Decision Or Root Cause section with validated knowledge, not speculation
• a Resolution Or Rule section that is actionable
• a Validation section that explains how the knowledge was confirmed or should be checked
• related files or systems when they help future retrieval

Check the index for:

• links to missing files
• entries that exist on disk but are missing from index.md
• duplicate titles or near-duplicate summaries
• categories that are too broad or inconsistent
• stale entries that should be updated or archived

Preferred investigation report shape

Prefer this structure after investigating existing memory:

# Knowledge Base Investigation
## Scope
## Summary
## Entries Reviewed
## Changes Applied
## Recommended Updates
## Duplicates Or Conflicts
## Broken Or Missing Index Links
## Needs Human Review
## Rebuild Status

Preferred entry shape

Prefer the bundled references/template.md when saving or presenting memory. The entry should include this shape:

# Memory Entry
## Title
## Context
## Applies When
## Decision Or Root Cause
## Resolution Or Rule
## Validation
## Tags
## Related Files Or Systems

Preferred retrieval shape

Prefer a structure like this when returning matches:

# Relevant Memory
## Match
## Why It Matters
## Key Takeaway
## Recommended Next Step

Storage guidance

Use the storage location defined in the repository root SPEC.md when present.

If no repository root SPEC.md is available, default to docs/knowledge-base/ and docs/knowledge-base/index.md.

When adding or changing entries, use scripts/index.py rebuild after manual edits to keep index.md synchronized.

If the default storage location is clearly unsuitable for the task or the repository is read-only, present the normalized memory entry in the response and explain that persistent storage could not be updated.

Handoffs

• Accept decisions, assumptions, and constraints from spec-plan when they are worth preserving.
• Accept confirmed root causes, setup fixes, and troubleshooting patterns from spec-crlp.
• Return relevant memory to spec-plan when planning a new task with similar constraints.
• Return relevant memory to spec-crlp when debugging a failure that resembles a known issue.
• Recommend creating or updating a repository root SPEC.md when repeated memory entries reveal a repository-specific convention that should become explicit.

Example requests

• "Save this root cause so we do not lose it after this session."
• "Have we solved something like this migration issue before?"
• "Index the setup steps and env gotchas from today's work so future tasks can retrieve them."
• "Audit the knowledge base and fix weak trigger tags."
• "Merge duplicate root-cause entries and rebuild the index."
• "Investigate existing setup notes and mark anything stale."