Local context engine for AI coding agents that ranks relevant repo files and builds compact task-focused context packs. Exposes context packing as MCP tools for Claude Code, Codex, Cursor, Windsurf, and CI workflows.
AgentPack defines 28 tools with consistent naming (verb_noun pattern: get_, create_, start_, pack_, route_, etc.) and complete input schemas. All tools have descriptions (avg ~80 chars, within baseline 34-392). However, descriptions are terse and lack WHEN/WHY context. Output schemas are not documented in the tool definitions, LLMs cannot predict return structure. Error handling is absent from descriptions. No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels (WRITE vs READ_ONLY). Parameter descriptions are present but generic ('Output format (auto, json, markdown, yaml)' repeated 28 times). No guidance on tool composition or dependencies (e.g., start_task → pack_context → get_context flow is implicit).
atomically claim, apply, and resume a handoff
package a structured report and complete Git-visible patch
show score breakdown + symbols for a specific file
relationship evidence receipt
read-only route with skill score reasons
read latest context pack; auto-refreshes when task.md changed
return selected-file delta since the previous pack
Output schemas not documented. Tool definitions show input schemas but no return type documentation. LLMs cannot predict response structure, forcing trial-and-error tool chaining.
Descriptions lack WHEN/WHY context. All 28 tools have descriptions under 100 chars with no guidance on when to call them, dependencies, or prerequisites. E.g., 'refresh' says 'refresh using the current task.md' but does not explain when task.md changes or what triggers a refresh.
No tool annotations (readOnlyHint/destructiveHint/idempotentHint). Tools are marked READ_ONLY or WRITE in metadata but not in MCP tool definitions. LLMs cannot infer safety/idempotency from names alone.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 63 | 2026-07-28+ | v2 |
bounded semantic graph neighbours
one graph node and its evidence
inspect one handoff by memorable name
get context for pull request analysis and review
return import-graph neighbours of a file
read one skill by name or path
read-only skill/rule inventory
token/saving stats for the latest pack
record structured proof for a learning session
return the next competency-backed learning queue
start a coached learning session
list pending or historical project handoffs
generate/refresh a context pack for a task
bounded semantic graph search
prove live MCP exposure and report server/CLI/tool status
refresh using the current task.md
release the current session's claim
read-only route: files + rules + skills + commands
shortest semantic relationship path
write task.md and return a fresh context pack
validate TOON syntax from content or a repo-relative path
Generic parameter descriptions repeated 28 times. 'Output format (auto, json, markdown, yaml)' appears identically in every tool. No guidance on when to use each format or what the default is.
No error handling guidance in descriptions. Tools like 'get_skill' accept 'name_or_path' but do not document what happens if the skill is not found, what alternatives exist, or how to recover.