Fully automatic context preservation for Claude Code - Works out of the box with zero configuration. Never lose valuable work again!
c0ntextKeeper provides three well-named tools with complete schemas and reasonable descriptions. All tools follow verb_noun naming (fetch_, search_, get_), have non-empty descriptions (71-171 chars, within the 10-1024 baseline), and comprehensive input parameter definitions with types, descriptions, and constraints. However, tool descriptions lack WHEN to use context (distinguishing between tools), dependency hints, and guidance on multi-step workflows. Parameters are well-constrained (enums for scope/sortBy/type, numeric ranges with min/max), but output schemas are not documented in the source code, a critical gap per pattern:tool. Error handling is not visible in the schema definitions. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite being READ_ONLY tools per the rubric. The tools correctly accept natural identifiers (query strings) and avoid exposing secrets, but the lack of documented return structures and error recovery guidance limits LLM reasoning and downstream tool chaining.
Fetch relevant archived context for the current task. Returns previously extracted problems, solutions, and implementations.
Retrieve recurring patterns from archived sessions. Identifies common solutions and approaches.
Search through archived context with advanced filters. Find specific implementations, errors, or decisions.
Output schemas not documented. Pattern:tool requires documenting return types. Without documented output schemas, LLMs cannot plan downstream tool calls or extract the right data for chaining. fetch_context, search_archive, and get_patterns all lack visible return type documentation in the source code.
Tool descriptions lack 'WHEN to use' context. Descriptions state WHAT each tool does (e.g. 'Fetch relevant archived context') but not WHEN to select it over similar tools or WHAT prerequisites exist. For example, fetch_context vs search_archive distinction is unclear, both search archived content, but the description does not explain when to use fetch_context (natural language, relevance scoring) vs search_archive (advanced filters, date ranges, file patterns).
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). All three tools are READ_ONLY but have no explicit annotation. Pattern:tool requires idempotency declarations. Without annotations, LLMs cannot determine if it is safe to retry or cache results.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | C | 62 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 0 | - | v1 |
No error handling or recovery guidance visible. Pattern:recovery-guide and pattern:error-classification require error responses to tell the LLM what to do next. If search fails or query is malformed, the agent has no guidance on recovery steps.
Parameter 'limit' uses generic description without explaining enforcement. For search_archive and get_patterns, 'limit' description says 'Maximum results (default: 10, max: 100)', this is acceptable. However, fetch_context default is 5 and max is 100, creating inconsistency. Description should clarify 'default 5 results, up to 100' explicitly.
No documented pagination or result capping in descriptions. Pattern:paginated-result and mxe:enforce-result-limits recommend explicitly stating in descriptions that results are capped and offering pagination. Returning 100 results at once risks context window exhaustion.
Parameter 'minRelevance' on fetch_context uses range 0-1 with default 0.3, but description does not explain what relevance score represents or how the LLM should tune it. Should clarify 'higher values return only closely matching contexts; lower values broaden the search'.
dateRange parameter on search_archive lacks example or validation hint. Description says 'Date range for filtering results' but does not explain what happens if 'from' > 'to', whether both are required, or how open ranges (only 'from', no 'to') behave. Per review:param-validation-rules, format and constraint hints should be in the description.
scope parameter on fetch_context lacks dependency hints. Description says "Search scope (default: project)" but does not explain what 'session', 'project', and 'global' mean or how to decide between them. Should clarify: 'session = current session only, project = all sessions in this project, global = all projects'.