Changelog

[1.6.1] - 2026-08-10

Changed

• interface.shortDescription is now "Expert web search for agents", replacing "Live web search for AI agents".
• interface.longDescription opens on accuracy and cost, replacing "Nimble gives agents live access to the public web." with "Nimble gives agents governed, accurate access to the public web, at a lower token cost per answer." The rest of the paragraph is unchanged; it is now 639 characters.

Only .codex-plugin/plugin.json carries an interface block, so it is the only file whose copy changes; the rest of this release is the 1.6.1 version bump. The shared description remains byte-identical across all five plugin manifests and marketplace.json.

[1.6.0] - 2026-08-10

Changed

• Plugin listing copy is now consistent across every manifest. The same description is shared by all five plugin manifests and marketplace.json, and it now leads with live web search rather than data extraction. Trimmed from 397 to 313 characters.
• displayName is now Nimble, replacing Nimble Web Data Toolkit in .cursor-plugin/plugin.json and .codex-plugin/plugin.json. The plugin name is unchanged at nimble, so every skill identifier, npx skills add --skill <name> invocation, and MCP tool prefix keeps working.
• interface.shortDescription is now "Live web search for AI agents", and interface.longDescription is 591 characters, down from 1638. It describes what each capability returns rather than how it works internally.
• interface.defaultPrompt no longer uses internal product vocabulary. The starter prompts are phrased as questions a new user would actually ask.
• Categories reconciled. .codex-plugin/plugin.json moves from Data & Analytics to Developer Tools and marketplace.json from data to development; .cursor-plugin/plugin.json was already developer-tools.

[1.5.0] - 2026-08-09

Added

• Conformance with Agent Plugins v1.0.0, the vendor-neutral package format for Agent Skills and MCP servers used by VS Code, Cursor, GitHub Copilot, Kiro, and ChatGPT & Codex. A root plugin.json carries the portable manifest with the canonical $schema and only the permitted fields, and root mcp.json carries the portable MCP config. The skills/ tree already conformed. This is additive: no current consumer reads a root plugin.json, so the four per-client manifests keep working untouched.
• Portable-package CI gates in scripts/check-plugin-manifests.py. The manifest is checked for the exact $schema, the required fields, the closed field set, the name pattern and length, the restricted author sub-keys, and keyword and extension types. The MCP config is checked for the exact $schema, the two permitted top-level keys, and a transport from stdio / streamable-http / sse with its required field. scripts/tag-release.sh covers the new manifest, taking version parity to 22 references.

Changed

• Root mcp.json now declares native Streamable HTTP. It previously held the superseded mcp-remote stdio shim with a bearer token in argv, so a client reading the portable canonical path got the retired pattern instead of OAuth. .mcp.json is unchanged — it declares "type": "http", which the portable schema rejects, and it is the file Claude Code auto-registers as a Connector. The two describe one endpoint in mutually exclusive vocabularies and are kept separate deliberately; the portable schema sets additionalProperties: false at the root and on every server, so one file could not carry both.
• Corrected the platform-compatibility table. It listed root mcp.json as Cursor's MCP config. Cursor reads the user's own .cursor/mcp.json; nothing in this repository loads root mcp.json. README.md and CLAUDE.md now document the vocabulary split so the files are not merged later.

[1.4.0] - 2026-08-09

Added

• xAI/Grok plugin layer. New .grok-plugin/plugin.json declares the shared ./skills/ tree to Grok Build. Verified against xAI's own scripts/plugin_catalog.py: the plugin now indexes 15 skills, 2 agents, 1 command, and 1 HTTP MCP server. No Grok-only skill copies, no second repository — the marketplace is an index, so distribution is one catalog entry in xAI's repo pinning a commit from this one.
• mcpServers drift assertion in scripts/check-plugin-manifests.py. The Grok manifest declares its MCP server as an inline object rather than as a path to .mcp.json, which duplicates one server config. The checker now asserts the two copies agree — matching server names and identical config — and fails if the Grok manifest is changed to a path declaration. Verified against fixtures for URL drift, a server present in only one file, and the path-declaration case.

Changed

• scripts/tag-release.sh also asserts .grok-plugin/plugin.json. All five manifests, the README badge, and every skill's frontmatter must agree on the version, so a partial bump fails CI.
• README.md, CLAUDE.md, and CONTRIBUTING.md document Grok as a fourth packaging target, including why the Grok manifest declares MCP inline. xAI's indexer reads only the mcpServers key out of a .mcp.json file, and this repo's is a bare server map, so a path declaration would index zero servers and the Grok listing would claim the plugin ships no MCP server. Declaring inline is also what keeps .mcp.json byte-identical, preserving Claude Code's Connector registration over native HTTP with OAuth.

[1.3.0] - 2026-08-06

Added

• OpenAI/Codex plugin layer. New .codex-plugin/plugin.json declares skills as the single ./skills/ string path (an array is rejected with plugin_skills_path_wrong_type), mcpServers pointing at the existing root .mcp.json, and an interface block carrying the listing metadata: display name, short and long descriptions, developer name, category, capabilities, brand colors, three starter prompts, and the required logo and composerIcon assets. The long description and capabilities separate the four data capabilities explicitly — Search, Extract, Extraction Templates, and Web Search Agents — alongside Map and Crawl. The shared skill tree, the hosted MCP configuration, and all Claude, Cursor, and CLI behavior are unchanged; no OpenAI-only skill copies were added.
• assets/nimble-logo.png at the repository root, referenced by interface.logo and interface.composerIcon. Square 200×200 PNG, satisfying the square, minimum 48×48, maximum 4096×4096, and 5 MiB asset gates.
• scripts/check-plugin-manifests.py. Validates the manifest fields the structure check and a JSON parse both pass straight over: skills declared as a string rather than an array, the required logo and composerIcon assets and their dimensions read from the file header, brand-color contrast against both the light and dark surfaces, the category enum, and the display-name, description, and starter-prompt length limits. Each check is named after the error code OpenAI would raise, so a CI failure maps onto the published submission-error reference directly. Standard library only — no new dependency.
• scripts/check-plugin-structure.sh and a Plugin packaging CI workflow. Asserts every SKILL.md is an immediate child of skills/, that no SKILL.md exists under any references/ directory, that each skill's frontmatter name matches its directory, that <plugin>:<skill> stays within 64 characters, that frontmatter is present and closed, that descriptions stay within 1024 characters, and that marketplace.json enumerates exactly the skills that exist. Runtime discovery is recursive on both Claude Code and Codex, so these violations work locally while blocking OpenAI submission and silently emptying the xAI index — a CI gate is the only thing that surfaces them.

Changed

• Reference documents under skills/nimble-web-expert/references/ are now named reference.md instead of SKILL.md. The seven affected documents cover Search, Extract, Extraction Templates, Web Search Agents, Map, Crawl, and Tasks. A .codex-plugin/plugin.json is read as a Legacy-format manifest whose loader scans for SKILL.md recursively, so these progressive-disclosure documents registered as seven extra skills in the model-visible catalog — 22 instead of 15. Every pointer was updated across CLAUDE.md, _shared/nimble-playbook.md and its synced copies, nimble-web-expert's SKILL.md, README.md, rules/nimble-web-expert.mdc, and references/batch-patterns.md. Directory names are unchanged, so no skill identifier moves.
• scripts/tag-release.sh also asserts .codex-plugin/plugin.json. All four manifests, the README badge, and every skill's frontmatter must agree, so a partial bump fails CI instead of shipping manifests that disagree about the version.
• Trimmed the brand-mention-monitor and launch-monitor descriptions to 1006 and 1005 characters, from 1102 and 1072. Both exceeded the 1024-character cap and would have been rejected with skill_description_too_long. Only descriptive prose was removed — every trigger phrase and every negative trigger is intact.

[1.2.0] - 2026-08-05

Changed

• Flattened the skills tree. All 15 skills are now direct children of skills/ instead of being grouped into vertical subdirectories (business-research/, data-platforms/, healthcare/, human-resources/, marketing/, productivity/, seo/, web-search-tools/). Skill directory names are unchanged, so every skill identifier (nimble:competitor-intel, nimble:meeting-prep, …) and every npx skills add --skill <name> invocation keeps working exactly as before. Both plugin platforms require the flat layout: OpenAI/Codex rejects nested skill directories at submission (skill_manifest_nested, skill_manifest_missing), and xAI indexes only direct children of skills/. Runtime discovery in Claude Code and Codex is recursive, so the previous layout worked locally while being absent from both catalogs.
• Verticals are now recorded as metadata.category in each skill's SKILL.md frontmatter rather than as directory names. The README skills table keeps the same categories and links to each skill directly.
• .claude-plugin/plugin.json skills is now a single ./skills/ entry instead of eight vertical paths; .claude-plugin/marketplace.json lists the 15 flat skill paths. .cursor-plugin/plugin.json was already pointing at ./skills/ and is unchanged.
• scripts/sync-shared.sh globs skills/*/references instead of skills/*/*/references.
• CONTRIBUTING.md now instructs contributors to create skills directly under skills/ and set metadata.category, and to register them in marketplace.json only — the plugin manifests no longer need a per-skill path.
• CLAUDE.md documents the flat-tree requirement and both platforms' constraints so the layout can't silently regress.

Fixed

• Broken relative link in skills/nimble-web-expert/README.md — ../../README.md resolved to a nonexistent skills/README.md at the previous nesting depth and now resolves to the root README.

[1.1.0] - 2026-07-29

Added

• Three Web Search Agent run modes, taught and routed. The canonical WSA reference and the shared playbook now lead with a mode decision table: named create-or-reuse (agents run --agent-name <name>, no agent ID — the default for skills, since a repeated name resolves to the same web_search_agent_id), explicit agent (agents:runs create --agent-id <id>), and caller-anonymous (agents run with neither, which still returns a generated web_search_agent_id). Documented the routing trap behind it: agents:runs create requires --agent-id and ignores --agent-name, so Modes 1 and 3 must go through nimble agents run.
• use_case locking semantics. research / enrichment / dataset_building are documented exactly. use_case is stored when the agent is created (including on a Mode 1 first call), accepted as a no-op when it matches an existing agent, and rejected when it differs — it is not a per-run override. Also documented the two server-side rules that come with dataset_building: an output_schema is required, and effort must be high or above.
• Run-level skill override. Against an existing agent, --skill applies to that run only and leaves stored config untouched; on the call that creates the agent, --skill and --use-case become its stored configuration instead.
• Verified live-progress documentation. On the CLI, --enable-events plus agents:runs stream-events emits task_run.state / task_run.progress_msg.* / task_run.progress_stats events and closes on its own at a terminal state (--max-items <n> closes it after n events). The stream never carries the output — agents:runs result is still required. On production MCP the documented approach is bounded status polling, verified separately rather than assumed from the CLI behavior.
• Error table for the run contract — 409 "Run still active", 404 run/agent mismatch, the four distinct 422 shapes (locked use_case, invalid enum, missing dataset_building schema, effort too low), the sources shape rejection, and the Required flag "agent-id" not set routing mistake.

Added

• Routing eval for nimble-web-expert (scripts/run-routing-eval.py + evals/nimble-web-expert-routing.json). Ten cases assert which capability the routing guidance selects for a given prompt, and whether the researched-report/quick-scan fork is offered. The runner reads the routing text out of SKILL.md at run time, so the eval and the doc can't drift; it makes no Nimble calls, so it needs no API key and spends no credits. Currently 10/10, unanimous over three runs per case. It earned its keep immediately: it caught that dataset_building's high+ effort floor was documented only in the on-demand reference, leaving the fork rule undecidable at routing time — now stated inline.

Changed

• nimble-web-expert routing split into two gates. Gate A asks where the data lives (URL → extract, site + item → Extraction Template, sitemap → map, section → crawl); anything with no location signal — and any named site with no template — falls through to Gate B, which asks what the user wants back. nimble search returns raw material to skim; a Web Search Agent returns a finished, cited answer, and the prompt's deliverable noun decides which. Previously the routing table keyed only on input shape, so a request like "create a report on X" carried no distinguishing signal and defaulted to search — the flat table has been replaced, the core-principles bullet no longer contradicts it, a lexical trigger table was added, and a guardrail now forbids answering a synthesis deliverable with raw search results.
• The expensive fork is now offered rather than silently taken. When Gate B lands on a Web Search Agent at high effort or above, the skill offers one AskUserQuestion — researched report (minutes, cited) vs quick scan (nimble search, seconds) — instead of picking for the user. Below high it just runs. Web Search Agent runs must also announce their expected duration and narrate progress, since on MCP progress comes from bounded polling and an un-narrated multi-minute run reads as a hang. Added "Research & report" to the ambiguous-request options, which previously offered no path to a Web Search Agent at all.
• CLI prerequisite raised to 1.2.0, the first release exposing the complete run contract. Updated in the playbook transport table, the onboarding install/upgrade flow, and the nimble-web-expert description.
• Corrected the use_case enum. The stale value data_enrichment is replaced with the released enrichment everywhere.
• Documented the real --sources shape. allow / block are arrays of ordered source groups ({title, domains, order}); prioritize / avoid are plain guidance strings. The previous docs implied a single uniform shape, which the API rejects.
• Separated enrichment input from output shape. --input-data carries the rows you already have; --output-schema describes the shape of the answer. Enriching several rows needs an array schema — an object schema returns a single object. Fields carried in from input_data come back as confidence: "pre_existing" with no citations and must not be presented as sourced findings.
• Discovery is client-side filtering everywhere. agents list and agents:templates list take no server-side search term, matching the already-documented extract:templates list behavior.
• Transport differences are now documented per capability rather than assumed parity. Modes 1 and 2, use_case, skill, sources, output_schema, input_data, and effort work the same on both transports. Three CLI-path capabilities have a documented MCP approach instead: enable_events/stream-events → bounded status polling; previous_interaction_id → restate prior context in input; and Mode 3 → pass an agent_name, since nimble_agents_run requires agent_id or agent_name. Added nimble_agents_get to the nimble-web-expert MCP allow-list so an agent's stored use_case can be checked before a run.
• Attribution unchanged — --client-source nimble-agent-skills on every CLI call; production MCP still exposes no integration-specific attribution parameter, so the managed-transport exception is retained.

Verified

Every command, parameter, and error documented above was exercised against live services — nothing was inferred from SDK types or --help output.

On CLI 1.2.0: all three run modes; same-name reuse returning the same agent ID; a run-level skill override leaving stored config untouched; a locked-use_case rejection; a structured enrichment run with input_data + output_schema + sources; event streaming to self-termination, plus a capped stream (--max-items) confirmed not to cancel its run; terminal result retrieval with trust/citations; and the not-ready (409) and validation-error (422) paths, including the sources group title requirement.

On production MCP: Modes 1 and 2, status polling, and result retrieval with trust metadata. Mode 3 was probed too: nimble_agents_run takes agent_id or agent_name, so the transport table routes Mode 3 to a named agent on MCP rather than claiming blanket parity.

[1.0.0] - 2026-07-22

Changed

• Migrated the plugin to Nimble's current web-data surface (CLI 1.1.0+ / production MCP). The taxonomy is now Search · Extract · Extraction Templates · Web Search Agents · Map · Crawl, using Nimble's own product names. nimble-web-expert was reworked around this taxonomy and the Ask-Nimble tool-selection model (including the "no template for a site → use a Web Search Agent" rule and the Web Search Agent reuse-priority chain). Added references/nimble-extract-templates/ (existing-only discover/inspect/run) and reworked references/nimble-agents/ into the Web Search Agents / Agent API V2 lifecycle (agents / agents:templates / agents:runs) with the agent-authoring schema and per-claim trust/citation metadata.
• Repointed every skill off the removed singular nimble agent command group. CLI 1.1.0 replaced it with extract:templates * (structured site scrapers — the old amazon_pdp/google_maps/yelp/NPI-style agents) and agents:runs * (open-ended research). Business, healthcare, marketing, HR, and SEO skills now discover via extract:templates list + client-side filter (the --search flag is gone) and run via extract:templates run. The response envelope is unchanged (data.parsing).
• Request attribution is now the stable integration tag nimble-agent-skills, carried on the CLI path via --client-source / CLIENT_SOURCE. MCP requests are attributed at the transport level, so the CLI path is used when per-integration attribution matters.

Removed

• nimble-agent-builder skill — removed entirely. Building/publishing new templates or agents is out of scope for the plugin; use existing Extraction Templates and Web Search Agents, or the Nimble app to author new ones. Removed from both plugin manifests and the marketplace.

[0.25.0] - 2026-06-24

Added

• launch-monitor skill in marketing/ — monitors press, social, developer communities, and competitor channels around a product launch, tracking sentiment, flagging mischaracterizations, surfacing competitor responses, and recommending actions in real time. Delivers a Response War Room dashboard with a live signal feed, mischaracterization tracker, competitor response panel, and sentiment velocity chart. Ships with references/template.html and references/sources.md.

[0.24.0] - 2026-06-24

Added

• brand-mention-monitor skill in marketing/ — scans Reddit, X, LinkedIn, Instagram, TikTok, YouTube, blogs, news, and review platforms for brand mentions, scores each on reach/velocity/sentiment/risk-topic match, and buckets into Crisis / Watch / Engage / Log with routing and response windows. Ships with references/template.html (interactive triage console) and references/sources.md (per-platform query templates).

[0.23.0] - 2026-06-14

Changed

• nimble-databricks-data-products — SQL-native agent discovery. Replaced the nimble agent get CLI introspection with the new nimble_agent_describe() Unity Catalog SQL function, so the full discovery path runs through SQL (nimble_agent_list() → nimble_agent_describe() → nimble_agent_run()) with no nimble CLI dependency. references/nimble-agents.md §2 now reads inputs via nimble_agent_describe('<agent>') and splits input discovery (from describe) from output-field discovery (the §2.5 run-probe). SKILL.md updates the "discover, don't assume" golden rule and Phase 2, and drops Bash(nimble:*) from allowed-tools (it existed only for nimble agent get). This removes the skill's only hard nimble-CLI dependency for discovery — the Phase 0 unblock for packaging it into Databricks Genie Code (SQL/notebook-native, no shell). Requires the nimble_agent_describe UDTF from the Nimble cookbook.

[0.22.0] - 2026-06-12

Added

• nimble-databricks-data-products skill in a new data-platforms/ vertical — turns a one-line brief (e.g. "pricing comparison on dog products from Amazon and Walmart in Databricks") into working Databricks data products (Delta tables, an AI/BI dashboard, and/or a deployed app), end to end — equally for production data products or quick demos. It runs Phase 0 preflight (auth, warehouse, integration gate, writable-schema confirm), discovers the right Nimble agents at runtime via the nimble_integration Unity Catalog functions (never hardcoded), ingests live web data into Delta tables using a control-table + correlated LATERAL nimble_agent_run pattern (one set-based, expandable INSERT — not per-keyword files), and produces an AI/BI dashboard and/or a deployed Databricks App, branded "Powered by Nimble". Bundles reference cookbooks (preflight, agent discovery/ingest, Lakeview dashboard, AppKit app, branding, official-databricks-*-skill delegation map, integration-install fallback), helper scripts (build_dashboard.py for one-shot dashboard creation, ingest.sh for async statement fan-out), and the Nimble logo asset. Encodes hard-won gotchas: defensive numeric casts for currency-string prices, per-agent localization, probe-before-fan-out, reconcile-against-control-table verification, Lakeview JSON pitfalls, and AppKit numeric-string / light-mode fixes. Registered in .claude-plugin/plugin.json, .cursor-plugin/plugin.json, and .claude-plugin/marketplace.json.

[0.21.3] - 2026-05-27

Fixed

• Cowork / claude.ai connector detection & hallucinated auth flows — in those hosts the plugin is often installed while its MCP connector isn't connected. The previous guidance was reactive ("if the first call errors, then guide"), so agents fired a data call, got back an OAuth authorization URL, and improvised a fabricated recovery flow ("paste the full URL from your address bar back here and I'll complete the connection"). The fix makes connection verification a proactive preflight (one read-only nimble_agents_list probe before any work), corrects the connect path everywhere to Customize → Connectors → Nimble → Connect → browser login (superseding the 0.21.1 Personal plugins → … → Add to your team wording), adds an inline sign-up note for users without a Nimble account, and adds a hard anti-hallucination rule (present any "Authorize" link verbatim and stop; never invent a completion step; never claim tools "will activate" then call them in the same turn). Touches _shared/ canonical sources, all 11 synced business-skill references/ directories, both core skills' SKILL.md + rules/setup.md, and README.md.
• Missing probe-tool permission — added mcp__plugin_nimble_nimble__nimble_agents_list to nimble-agent-builder's allowed-tools so it can actually run the mandated preflight probe in MCP-only hosts.
• Dangling reference — both core SKILL.md files pointed connector guidance at a references/profile-and-onboarding.md that doesn't exist in those skills; guidance is now self-contained and points to rules/setup.md.

[0.21.1] - 2026-05-13

Fixed

• Cowork onboarding — when the plugin is installed but the connector isn't activated yet (typical Cowork / claude.ai first-run state), the skills were silently falling back to WebFetch instead of guiding the user to activate. The fix adds a dedicated "plugin installed but connector not activated" state to the transport-selection table in _shared/nimble-playbook.md, a matching sub-section in _shared/profile-and-onboarding.md, and an explicit hard-stop rule in both core SKILL.md files. When this state is detected, the skill now surfaces verbatim instructions: Customize → Personal plugins → Nimble → Connectors → Add to your team, then stops. WebFetch / WebSearch / curl substitution is explicitly forbidden. Synced into all 11 business-skill references/ directories via scripts/sync-shared.sh.

[0.21.0] - 2026-05-13

Changed

• MCP server ID renamed to nimble (was nimble-mcp-server) across .mcp.json, mcp.json (Cursor), README, both core-skill SKILL.md allowed-tools whitelists, the claude mcp list health check, and all CLI→MCP mapping tables in nimble-agent-builder. New tool prefix: mcp__plugin_nimble_nimble__*. Aligns with the lowercase server-ID convention used by every Anthropic-official plugin (linear, github, asana, etc.).
• Onboarding rewritten around the plugin install path. Every install instruction now leads with /plugin install nimble for any Claude product (Claude Code, Claude Cowork, claude.ai) — the plugin's .mcp.json auto-registers as a Connector in Customize → Personal plugins → Nimble → Connectors, and OAuth handles auth on first tool call. Updated _shared/nimble-playbook.md (Transport Selection block), _shared/profile-and-onboarding.md (prerequisite-check flow), README.md, nimble-agent-builder/README.md, both rules/setup.md files, and both core skills' Prerequisites sections. Synced into all 11 business-skill references/ folders via scripts/sync-shared.sh.
• Transport selection is now explicit. Skills pick CLI or MCP at session start (nimble --version → CLI; claude mcp list | grep nimble → plugin MCP) and stick with it for the session.
• CLI package name corrected in onboarding docs: @nimbleway/cli → @nimble-way/nimble-cli (the actual published package).

[0.20.0] - 2026-05-13

Changed

• Plugin MCP config — migrated .mcp.json from the mcp-remote stdio shim with bearer-token header to native HTTP transport with OAuth (type: "http", url, no headers). Matches the canonical pattern in anthropics/claude-plugins-official (Linear, Asana). Removes the npx -y mcp-remote@latest cold-start and the ${NIMBLE_API_KEY} header — Claude Code now drives the full OAuth 2.1 + PKCE + Dynamic Client Registration flow against mcp.nimbleway.com automatically. Root-level mcp.json preserved for Cursor compatibility.

[0.19.0] - 2026-04-20

Added

• Human Resources vertical (skills/human-resources/) — new vertical for HR and recruiting workflows; scoped to cover future skills like comp-analysis, interview-prep, and onboarding
• talent-sourcing — distribution step (Step 8) added; skill now offers Notion/Slack report delivery via connector detection pattern from memory-and-distribution.md

Changed

• Renamed skills/talent/ → skills/human-resources/ — "talent" was ambiguous; human-resources is the standard term and the right scope for the vertical
• Updated .claude-plugin/plugin.json and marketplace.json to reference ./skills/human-resources/

[0.18.0] - 2026-04-13

Added

• seo-intel — single-skill SEO intelligence toolkit covering the full lifecycle via 7 router-dispatched workflows: keyword research, rank tracking, technical site audit, content gap analysis, competitor keyword reverse-engineering, AI visibility measurement (ChatGPT, Perplexity, Google AI, Gemini, Grok), and GitHub repository SEO. One install, one entry point — intent detection routes to the right workflow, and workflows chain naturally across a session.

Changed

• _shared/ — removed ai-platform-profiles.md (only SEO-specific) and output-quality.md (marginal value, overlaps playbook) from the shared sync set

[0.17.0] - 2026-04-07

Added

• Two-tier wiki index — global index.md (one line per directory) + per-directory {dir}/index.md catalogs for scalable entity lookup
• Chronological activity log (log.md) — append-only timestamped record of skill runs and findings, rotated at 90 days
• Cross-entity references — Obsidian-compatible [[path/entity]] wiki links between related entities (people → companies, competitors → competitors)
• Ad-hoc insights — "save this" / "remember that" files insights into relevant entity pages with [ad-hoc] tags
• Cross-entity synthesis pages (synthesis/) — dynamically created when patterns emerge across 3+ entities, with YAML source tracking for deterministic staleness detection
• Research backlog (backlog.md) — tracks knowledge gaps and unanswered questions across skill runs

Changed

• All 9 business skills now run 5 preflight calls (added index load) and reference wiki update patterns in their save steps
• competitor-intel — new Step 7.5 generates synthesis pages using nimble-analyst agent; save step adds cross-references and appends to backlog
• meeting-prep — preflight follows cross-references from person files to load related competitor/company context
• company-deep-dive — save step adds cross-references for discovered people and related competitors
• Bootstrapping now creates synthesis/ directory alongside existing entity directories

[0.16.0] - 2026-04-06

Changed

• All 5 business skills now discover WSAs dynamically at runtime via nimble agent list --search instead of hardcoding agent names
• local-places — replaced 10 hardcoded WSA names across Steps 4-7 with a new WSA Discovery step (Step 4) that discovers, classifies, and caches WSAs for all phases
• competitor-intel, company-deep-dive, meeting-prep, competitor-positioning — added slim WSA discovery step before main execution, prioritizing search/extract/crawl/map WSAs
• wsa-pipeline.md reference converted from static WSA inventory to discovery strategy document
• Added sibling skill suggestions to follow-up sections across all business skills
• Fixed weakness language in competitor-positioning battlecard template

[0.15.1] - 2026-04-06

Changed

• All 5 business skills now reference Scaled Execution pattern from nimble-playbook.md with call estimation guidance
• Added explicit 500 retry and timeout handling to error sections of company-deep-dive, meeting-prep, competitor-positioning, local-places, and competitor-intel

[0.15.0] - 2026-04-05

Added

• healthcare-providers-verify skill — cross-references provider data against authoritative sources (NPI registry, state licensing boards) for accuracy verification

[0.14.0] - 2026-04-05

Added

• healthcare-providers-enrich skill — enriches extracted provider records with reviews, credentials, and contact info from multiple web sources

[0.13.0] - 2026-04-05

Added

• healthcare-providers-extract skill — first skill in the healthcare/ vertical; extracts structured practitioner directories from any web source using dynamic WSA discovery

Changed

• Fixed --shared-inputs YAML syntax in shared playbook
• Strengthened entity dedup normalization rules

[0.12.1] - 2026-04-05

Added

• Audit mode for market-finder skill — validates a reference list against live web data, scores market presence, and produces gap analysis

[0.12.0] - 2026-04-05

Added

• market-finder skill in business-research/ — discovers all businesses of a given type in any geography using Nimble WSAs
• 6 vertical presets (Healthcare, SaaS, Restaurants, Legal, Auto/Home, Custom) with dynamic WSA discovery
• SaaS treated as first-class vertical with two-pass discovery and funding verification
• Shared "Scaled Execution" pattern added to _shared/nimble-playbook.md (tiered: individual → batch → multi-batch → confirmation gate)

[0.11.0] - 2026-04-05

Added

• local-places skill in productivity/ — discovers, enriches, and scores local businesses using Nimble WSAs (Google Maps, Yelp, Facebook, Instagram, DoorDash, Uber Eats)
• Skill-specific WSA pipeline reference with category detection, location disambiguation, and interactive Leaflet.js map generation

[0.10.2] - 2026-04-03

Added

• CONTRIBUTING.md with contributor guidelines
• Shared patterns and conventions for new skill development

Changed

• Rewrote README with category-level skill table
• Fixed sync script for shared reference distribution

[0.10.1] - 2026-04-02

Added

• CLAUDE.md with repo context, skill authoring rules, and conventions

Changed

• Added agent CLI commands (nimble agent list, nimble agent get, nimble agent run) to all business skill playbooks
• Added MCP fallback table and fixed agent get flag syntax
• Updated plugin descriptions across both manifests

[0.10.0] - 2026-03-30

Changed

• Grouped skills into vertical directories: business-research/, healthcare/, marketing/, productivity/, web-search-tools/
• Updated plugin.json and marketplace.json for grouped directory structure
• Standardized author metadata across all skills

[0.9.0] - 2026-03-26

Added

• Business skills foundation — shared references (_shared/nimble-playbook.md, profile-and-onboarding.md, memory-and-distribution.md), custom sub-agents (nimble-researcher, nimble-analyst)
• competitor-intel skill — ongoing competitor monitoring with signal classification and delta detection
• company-deep-dive skill — 360-degree company research from web sources
• meeting-prep skill — attendee research and meeting briefings with calendar detection
• competitor-positioning skill — marketing-focused positioning analysis with before/after change tracking

Changed

• Added signal date validation to filter stale events from reports
• Added value positioning section to meeting-prep briefings
• Enforced DRY across all business skills: replaced site: operator with --include-domain, standardized placeholder names, added same-day detection

[0.8.0] - 2026-03-08

Changed

• nimble-agents skill renamed to nimble-agent-builder — clearer name that reflects its purpose (build, discover, and run structured-data agents)
  • Folder: skills/nimble-agents/ → skills/nimble-agent-builder/
  • YAML name: field updated from nimble-agents to nimble-agent-builder
• nimble-web-expert skill — major structural overhaul (v2.0.0)
  • Rewritten as thin hub (~430 lines) with 12 load-on-demand reference files under references/
  • References reorganised into subfolders: nimble-agents/, nimble-crawl/, nimble-extract/, nimble-map/, nimble-search/
  • Added YAML argument-hint, allowed-tools (9 tools), license, and metadata fields
  • Added $ARGUMENTS variable at top of skill body
  • Added Core principles section (10 hard rules replacing prose CRITICAL BEHAVIOR block)
  • Added Response shapes table (all command/flag combinations with output shape and access pattern)
  • Added Final response format (Step 4 summary table + attribution)
  • Added Guardrails section (11 NEVER/hard rules consolidated at bottom)
  • Added run_in_background=False rule for all Task agents
  • Added Hard 429 rule and hard retry limit (max 2 on error)
  • Added AskUserQuestion format constraints: header ≤12 chars, label 1–5 words, (Recommended) first
  • All reference files gained YAML frontmatter (name, description)
  • Playwright added as free Tier 6 alternative to browser-use
  • Nimble Docs MCP section added (claude mcp add nimble-docs)
• Version bumped to 0.8.0 across all plugin configs
• README.md updated with new skill name and directory structure

[0.7.0] - 2026-02-28

Added

• nimble-web-expert skill — extract-first scraping expert replacing nimble-web-tools
  • Lean SKILL.md (~500 lines) covering extract, search, map, crawl, parallelization, and example workflows
  • 5 reference files: parsing-schema, browser-actions, network-capture, search-focus-modes, error-handling
  • 2 rules files: nimble-web-expert.mdc (routing), output.md (security)
  • Render escalation tiers (1-5): static → render → stealth → browser actions → network capture
  • Geo targeting, parser schemas, XHR mode for public APIs

Removed

• nimble-web-tools skill (fully replaced by nimble-web-expert)

Changed

• Version bumped to 0.7.0 across all plugin configs
• README.md updated with new skill name, directory structure, and examples
• rules/nimble-tools.mdc updated to reference nimble-web-expert

[0.6.1] - 2026-02-24

Changed

• nimble-agents skill — comprehensive rewrite for MCP reliability and best practices
  • Fixed allowed-tools prefix (mcp__plugin_nimble_nimble-mcp-server__ format)
  • Task agents now use run_in_background=False to preserve MCP access (#13254)
  • Added MCP tool registry blocks to all Task prompt templates
  • Enforced nimble_web_search (MCP) as only search method — banned WebSearch, WebFetch, curl
  • Description rewritten to third-person with specific trigger phrases
  • Step 3 condensed; detailed content moved to references/generate-update-and-publish.md
  • Added anti-hallucination guardrails for subagent prompts
• Version bumped to 0.6.1 across all plugin configs
• Deduplicated google_search caveat in error-recovery.md

[0.5.0] - 2026-02-23

Added

• nimble-web-tools skill — replaces nimble-web-search with full Nimble CLI wrapper
  • nimble search — web search with 8 focus modes
  • nimble extract — extract content from any URL (JS rendering, geolocation, parsing)
  • nimble map — discover URLs and sitemaps on a website
  • nimble crawl — bulk crawl website sections with depth/path control

Changed

• Skills now use Nimble CLI (@nimble-way/nimble-cli) instead of curl-based wrapper scripts
• Version bumped to 0.5.0 across all plugin configs
• rules/nimble-tools.mdc updated to reference nimble-web-tools skill
• README.md updated with CLI installation and new skill documentation

Removed

• nimble-web-search skill (replaced by nimble-web-tools)
• scripts/search.sh and scripts/validate-query.sh curl wrapper scripts
• examples/ and references/ directories from old web-search skill

[0.4.0] - 2026-02-18

Added

• nimble-agents skill — find, generate, and run agents for structured data from any website
• .cursor-plugin/plugin.json — Cursor IDE plugin support
• .mcp.json / mcp.json — MCP server configuration for Claude Code and Cursor
• rules/nimble-tools.mdc — Cursor rule for preferring Nimble tools
• Multi-platform support: Claude Code, Cursor, and Vercel Agent Skills CLI

Changed

• Plugin renamed from nimble-web to nimble (unified plugin)
• Version bumped to 0.4.0 across all skills and config files
• .claude-plugin/plugin.json updated with new name, description, and keywords
• .claude-plugin/marketplace.json updated to reflect unified plugin
• .gitignore updated to include .cursor/, .claude/, *.bak
• README.md rewritten to cover all installation channels

[0.1.0] - 2025-01-01

Added

• Initial release with nimble-web-search skill
• 8 focus modes: general, coding, news, academic, shopping, social, geo, location
• AI-powered answer generation
• Agent Skills standard compatibility