此前截断分散在多层且互相冲突:Pi bash 保留尾部 50KB,RTK 兜底截断又只留 开头 12000 字符(砍掉报错与完整输出路径),pi-chrome 截开头 30K,web_fetch 各 provider 不一致。新扩展统一负责"新结果给模型看多少": - capture.ts 最先加载暂存原文,index.ts 最后加载(system-prompt 之前) - 超过 24K 字符时原文落盘 <agentDir>/my-pi/pi-overflow/(700/600、原子写), 优先取 bash fullOutputPath 完整文件;模型看到 头 + 中段信号行 + 尾 的有界 视图,行号基于落盘原文、结尾永远可见 - overflow_query:按 id 正则搜索、行区间、jsonPath 只读查询,id 校验防穿越 - 收编:RTK truncate 默认关闭、删除 pi-chrome truncateText、web_fetch 默认 每页 20000(上限 100000)、web_search max_results 限定 1–20 - pi-compaction 遮蔽带 details.overflow.id 的结果时在占位符提示 overflow_query - 登记 tools.jsonc / routing / 权限 allow / tool-search 常驻与分组 Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
8.7 KiB
pi-tool-search
Bounded, model-assisted tool-group loading for Pi. This source is maintained directly in the my-pi bundle from upstream v0.3.6; see UPSTREAM.md.
Why
A name plus a truncated first sentence can misrepresent a tool, while an add-only active set eventually restores every full schema in a long session. This maintained version instead builds validated retrieval metadata from complete tool definitions, presents compact group cards, activates whole workflow groups, and caps the dynamic active set with LRU eviction.
Generated metadata never replaces executable schemas. Exact names, parameter types, required fields, and source definitions remain authoritative in Pi. The repository provides curated metadata for its own tools; a model may generate only summaries, usage guidance, keywords, and assignments for additional unrecognized tools.
Catalog lifecycle
session_startkeepstool_search, Pi core tools, and configuredalwaysEnablednames active. PowerShell is filtered out on non-Windows hosts; on Windows it is pinned beside Bash as the same local-shell capability class.- Hidden tool definitions are hashed with the grouping constraints.
- Every tool exposed by the standard
my-pibundle is matched against the checked-in catalog inextensions/bundle-groups.ts. If all tools are recognized, no model call and no user cache are needed. - User
groupOverridestake priority over checked-in assignments. Optional or unavailable bundle tools are simply filtered out of their predefined groups. - Additional user or third-party tools receive deterministic prefix/source groups immediately. A matching private model-enriched cache at
<agent-dir>/my-pi/tool-search/catalog-v1.jsonis reused when present. - Only when unrecognized tools exist and the call does not name an exact existing group may
tool_searchask the current authenticated model to enrich the complete catalog. Exact group activation always uses the current deterministic catalog immediately. Validation requires all checked-in bundle tools to remain in their predefined groups before a0600cache is accepted. - Missing authentication, invalid JSON, timeout, cancellation, changed bundle assignments, or cache errors leave the checked-in plus deterministic hybrid catalog usable.
A standard bundle installation therefore sends no tool definitions to a model during catalog setup. Complete definitions are sent to the selected provider only when extra unrecognized tools require enrichment during an unknown-group or query lookup; exact existing group loads never wait for that nested call. Any nested-call usage is attached to the tool_search result.
Run /tool-search-status to inspect the current catalog source, pinned count, dynamic capacity, LRU order, and latest eviction reason. Run /tool-search-rebuild to remove model enrichment; standard bundle tools immediately return to the checked-in catalog and extra tools can be enriched lazily on the next search.
Precomputed my-pi groups
| Group | Bundle tools |
|---|---|
filesystem-navigation |
ls, multi_grep |
ssh-connection |
ssh_connect |
ssh-remote-files |
ssh_read, ssh_write, ssh_edit |
ssh-remote-search |
ssh_find, ssh_grep |
ssh-remote-shell |
ssh_cd, ssh_bash |
code-intelligence |
CodeGraph and all LSP tools |
web-search |
web_search, web_fetch (pi-web-search unified tools) |
context-execution |
ctx_execute, ctx_execute_file, ctx_batch_execute |
context-knowledge |
Context indexing, searching, and web-doc ingestion |
context-observability |
Context stats, doctor, and Insight |
context-administration |
Context upgrade and destructive purge |
memory-recall |
Hermes memory/session search and condensed-output recovery |
memory-management |
Hermes memory writes and procedural skill management |
subagents |
Launch agents, collect background results, and steer running agents |
user-interaction |
ask_user_question model-initiated TUI/RPC questions |
chrome-navigation |
Chrome bridge setup, tabs, navigation, snapshots/find/inspect, waits, and screenshots |
chrome-interaction |
Chrome click/type/fill/key/hover/drag/tap/scroll input |
chrome-debugging |
Chrome page evaluation, console/network diagnostics, response retrieval, and file upload |
Pinned tools are removed from these dynamic groups at runtime. The defaults pin codegraph_explore, lsp_diagnostics, the three high-frequency Context Mode execution tools, and all six Hermes memory tools, so none of them consume dynamic group or tool capacity.
Loading and eviction
Call tool_search with an exact catalog group id when possible:
{ "group": "web-search" }
A workflow can atomically activate up to five exact groups in one call. The complete request must fit the configured group and tool limits; otherwise no active state is changed:
{
"groups": ["ssh-connection", "ssh-remote-shell", "ssh-remote-files", "ssh-remote-search"]
}
A natural-language query uses weighted group, alias, and exact-tool-name metadata when the group is unclear. Weak or closely tied matches return ranked candidates without activating an arbitrary group. Loading exposes the original full schemas on Pi's next model request.
Pinned tools do not consume the dynamic budget. Dynamic groups update their recency when loaded or when a member starts executing. Loading beyond maxActiveGroups or maxDynamicTools atomically evicts least-recently-used non-requested groups before activation.
The first loads are purely additive and retain Pi's native deferred-loading propagation. An eviction necessarily removes and adds schemas in one transition, so Pi uses its safe fallback rather than treating that transition as purely additive.
Configuration
Pi reads this configuration directly from the installed bundle's config/tool-search.json — the single source of truth, refreshed by pi update. A missing, unreadable, or invalid file falls back to the built-in BUNDLE_TOOL_SEARCH_DEFAULTS; the extension never writes the file:
{
"alwaysEnabled": [
"codegraph_explore",
"lsp_diagnostics",
"ctx_execute",
"ctx_execute_file",
"ctx_batch_execute",
"memory_search",
"session_search",
"memory_add",
"memory_replace",
"memory_remove",
"skill_manage",
"context_tree_query",
"overflow_query"
],
"showToolSearchFooterStatus": false,
"maxActiveGroups": 5,
"maxToolsPerGroup": 8,
"maxDynamicTools": 28,
"groupOverrides": {}
}
| Key | Bundle default | Description |
|---|---|---|
alwaysEnabled |
CodeGraph explore, LSP diagnostics, Context execution trio, all six Hermes memory tools, context_tree_query, overflow_query |
Exact names pinned alongside Pi core tools |
showToolSearchFooterStatus |
false |
Show active tool/group counts |
maxActiveGroups |
5 |
Maximum simultaneously active dynamic groups |
maxToolsPerGroup |
8 |
Maximum size accepted from model-generated groups |
maxDynamicTools |
28 |
Hard cap across active dynamic group tools |
groupOverrides |
{} |
Exact tool lists that override checked-in, generated, or fallback assignments |
Example override:
{
"groupOverrides": {
"web-search": ["web_search", "web_fetch"],
"lsp": ["lsp_definition", "lsp_references", "lsp_hover", "lsp_symbols", "lsp_diagnostics"]
}
}
An explicit override larger than maxDynamicTools remains visible but cannot be activated until the limit or override is corrected. To customize, edit config/tool-search.json in the bundle repository and update the bundle — pi update refreshes the installed copy.
Current host integration
This version directly targets @earendil-works/pi-coding-agent, current typebox, ModelRegistry.complete(), and ExtensionAPI.setActiveTools(). It does not load or alias @mariozechner/pi-coding-agent, install another Pi runtime, rewrite provider payloads, proxy tool execution, or enqueue hidden sendMessage loops.
Repository layout
extensions/index.ts— Pi lifecycle, model call, activation, and group LRUextensions/catalog.ts— hashing, strict catalog validation, fallback grouping, cache, and rankingextensions/bundle-groups.ts— checked-in authoritative groups for tools exposed bymy-piextensions/config.ts— bundle-deployed config path and config parsingtest/— catalog, config, model-generation, and active-set testsdocs/dynamic-tool-loading.md— current Pi propagation and replacement behaviorUPSTREAM.md— imported snapshot and local divergence
Development
npm install
npm run typecheck
npm run test
npm run build
The root bundle loads ./pi-tool-search/extensions/index.ts directly and references this package through file:./pi-tool-search.