SupaPM

AI-powered product management assistant with context-aware chat

Think "Cursor for Product Managers and Project Managers". SupaPM is your context-aware AI agent that comes with you in your browser to make everything you do quicker, easier and higher impact. It makes switching contexts easy, giving you the latest context about ongoing projects. It makes it easy to create your own automations and workflows that do the grunt work so you don't have to. Features: - Local-first - your data doesn't leave your browser - Use any tool-calling LLM you like. If a new one comes out, just switch over in seconds - Build your own shortcuts and automations - Create your own skills that your LLM can use to help you get your work donw - Give the LLM your own personal organisational context on how work should be done. What the product does, who's responsible for what, all so that your SupaPM agent can be as helpful as possible in getting your work done --- ## 0.0.15 Agent edits to Company Context, skills, and project descriptions are more reliable, you can see how much browser storage SupaPM is using in Settings, and chats no longer bleed into each other when you switch tabs while a response is still generating — start a new conversation or jump back without losing the in-flight reply. - Thinking-model reasoning is easier to follow while it streams: internal steps appear as collapsible headings (collapsed by default), the active step shimmers until the next one arrives, and you can expand any step to read the full text. - Chats are isolated while generating: switch tabs or open a new chat while another response is still streaming; each conversation keeps its own in-flight state and background tabs show a pulse indicator until the reply finishes. - Settings shows local storage usage for SupaPM data in this browser profile: total footprint, browser storage vs project database (IndexedDB), optional per-category breakdown, and browser quota when available. - `edit_document` find-and-replace tolerates minor whitespace differences (e.g. collapsed blank lines or indentation) so the agent can update Company Context, skills, and project descriptions without spurious "not found" errors. ## 0.0.14 This release fixes a crash loop that made the sidepanel unresponsive, ensures hyperlinks are preserved when you capture page context, and makes captured context much easier to understand and trust. Page, screen, and selection context now behaves like file attachments: it sits in the same staging area before you send, clears when the message goes, and shows up in your message history. Any context or attachment chip is clickable to inspect exactly what text was sent to the model — useful when page extraction produces noise or you want to verify a capture before a sensitive query. The sidepanel also opens noticeably faster: the AI model list is now served from a local cache on every reopen, so the panel is ready to use without waiting on a network round-trip. - Styling updates: Updated the font to use the same as the main website, improved tab management - Sidepanel no longer blocks on a network fetch when opening. The AI model catalog is persisted to `chrome.storage.local` after the first load and served from there on every subsequent open. If the cached catalog is older than 15 minutes a background refresh runs silently and updates the model list when it completes — the panel is always immediately usable. - Updated the model catalog with Gemini 3.5 Flash (stable, now the Google default — optimised for agentic and coding tasks), Gemini 3.1 Flash-Lite (stable), and GPT-5.4 nano (OpenAI's cheapest/fastest tier). Deprecated the preview aliases that those stable models replace, and fixed the OpenAI default model from the deprecated GPT-4o mini to GPT-5.4 mini. - Fixed an infinite injection loop: when the sidepanel mounted, it requested content script injection; Chrome then fired a `tabs.onUpdated` event with `status: 'complete'` for the same URL, which the background treated as a navigation change and notified the sidepanel, which requested injection again — looping endlessly. The background now only notifies the sidepanel when the tab ID or URL actually changes, so same-page `status: 'complete'` events are silently ignored. - `@selection` captures now serialise the selected DOM range to HTML and convert it to markdown, preserving links, headings, bold/italic, code, and list structure. Falls back to plain-text if the pa --- v0.0.13 This update renames "Org Knowledge" to "Company Context" throughout the extension, rebuilds the default document template with richer structure and table-based sections, rewrites the system prompt with explicit routing rules, and adds a Raw/Styled toggle directly in the markdown editor toolbar. It also makes chat context much more practical in day-to-day PM work: you can attach files or paste slide screenshots directly into chat for summarisation/extraction workflows, and you can stop an in-flight generation when the response is already good enough or heading the wrong way. - Renamed "Org Knowledge" to "Company Context" across the sidepanel nav, options nav, page titles, aria labels, skills, chat context labels, and all internal references. - Rebuilt the default Company Context document template with distinct sections: Company, Products, Customers, Strategy (including a baseline metrics subsection), Competitors, Pricing, Glossary, and Decisions -- followed by personal sections (My Team, My Stakeholders, My Notes) separated by a divider. - Glossary, My Team, and My Stakeholders sections now use markdown tables for better structure and scanability. - Added explicit load-routing hints to the system prompt so the agent knows which section to load for each question type (people/ownership: my-team/my-stakeholders, customers/ICP: customers, strategy/OKRs: strategy, pricing: pricing, competitors: competitors, products: products, unfamiliar terms: glossary, past decisions: decisions). - Rewrote the default system prompt with structured sections covering tone and style, company context usage, skills usage, proactive context capture, project defaults, and a clear list of things to avoid. - Added a Raw/Styled toggle button to the MarkdownEditor toolbar (far right, after Redo) so users can switch between the TipTap rich editor and a plain-text raw view in any markdown editor surface -- system prompt, skills, project descriptions. - Added file and image attachments in sidepanel chat: PMs can attach files from disk or paste images (for example slide screenshots with charts/text) and use them as direct chat context. - Added attachment-aware request building so text/code files are inlined as structured context blocks and image attachments are sent as model-safe image payloads (including pasted `data:` images), enabling summarise/extract workflows without manual copy-paste. - Added removable attachment chips in the composer plus attachment-only sends, so users can submit context artifacts even when there is little or no typed prompt text. - Added a Stop button during generation that aborts the in-flight response, preserves any useful partial output, and marks interrupted tool activity as cancelled for clearer chat trace visibility. - Simplified edit-event cards in chat so the header focuses on the meaning of the change (document + summary), while detailed diff metadata is de-emphasized behind expansion controls for faster sidepanel scanning. --- v0.0.12 This release helps PMs move faster in multi-step AI workflows by making follow-up chats more context-aware and easier to trust. You can quickly confirm what the model received for a turn, keep loaded skills and org knowledge reliably available across turns, and avoid duplicate or misplaced project links by checking where a URL is already referenced before creating new work. - Improved LLM auditability in debug mode by combining preflight payload capture (including the fully composed `systemPrompt` and `messagesForAgent`) with step-level request/response logging, making Inspect Turn easier to trust when validating model behavior. - Improved cross-turn memory reliability by carrying forward compact prior tool outcomes and preserving version-aware loaded context in `LOADED_CONTEXT_IN_THIS_CHAT` (including `updated_at` freshness checks), so follow-up turns retain loaded skills and org knowledge without replaying verbose tool JSON. - Fixed context window usage calculation to avoid double counting across a chat session: context percentage and context-attachment safety checks now use current-turn context input usage (plus active captured context), not cumulative lifetime session tokens. - Improved inline captured-context tokens in the chat composer to use readable page/screen labels (for example `[[screen:2026 TikTok...]]`) instead of opaque short IDs. - Improved edit-event cards in chat and Edit History with clearer no-border visual separation and more readable add/remove change highlighting. - Replaced the sidepanel header Home and Settings buttons with an options quick-nav dropdown that links directly to every options page (Chats, Skills, Prompts, Org Knowledge, Edit History, and Settings), now using DaisyUI's native dropdown behavior for more consistent open/close interactions and fewer custom-state edge cases. - Removed automatic sidepanel focus jumps (on open, send completion, loading completion, and session/context updates) so the project picker no longer briefly opens and accidental project selection is less likely when you refocus the panel. - Expanded the local `search_projects` agent tool with richer filtering and sorting (`sort_by`, `sort_order`, status, priority range, tags, recency windows like last 7 days, and stale-project queries), plus clearer parameter guidance so project triage requests resolve more reliably in chat. - Added a Skills page search box with instant local filtering by skill name/description and a clear no-results state, plus improved input surface contrast so the field is visually distinct in light theme. - Added strict duplicate-link protection to `add_project_link`: the tool now canonicalizes URLs and returns a clear error when the exact same link (including params) is already attached to a project. - Added a new `find_projects_by_link` tool so the assistant can answer “is this URL already linked anywhere?” and return all matching projects plus the matched link records, useful for workflows like Jira-ticket-to-project lookup. --- v0.0.11 This release gives PMs better control over both reasoning depth and captured page context. You can now tune how much thinking effort the assistant spends on harder work directly in chat, and add highlighted text to context explicitly with a keyboard shortcut instead of having context appear unexpectedly when the composer gains focus. This makes quote capture from customer calls, competitor pages, and docs more intentional and easier to trust. - Added model-level reasoning controls in the sidepanel chat model popup so you can set reasoning depth inline (low, medium, high) while selecting your model. - Added support flags in the model catalog (`supports_thinking`) so reasoning controls only appear when the currently selected model supports thinking, while still showing capability hints next to compatible models. - Added provider-specific reasoning settings at runtime for OpenAI, Anthropic, and Google models so the selected reasoning depth is actually applied to API calls. - Improved Google/Gemini reasoning behavior by requesting thought content explicitly (`includeThoughts`) so thinking output can be surfaced in the chat reasoning UI when the provider returns it. - Added persistent reasoning-level preference in extension AI settings so your selected depth carries across sessions instead of resetting each time. - Added a native highlighted-text capture shortcut (`Ctrl+Shift+Y` / `Command+Shift+Y`) that opens the sidepanel and adds the active selection to chat context in both Chrome and Firefox. - Replaced legacy focus-based selected-text auto-attach behavior with explicit shortcut capture, so highlighted text is only added when the user deliberately triggers it. - Added clearer in-panel capture feedback for shortcut flows, including no-selection and permission guidance, and deduped repeated captures of the same selection (`url + exact text`) in a chat. This release makes SupaPM easier to approve, trust, and use in day-to-day product work. Security and compliance teams can review the public CycloneDX SBOM before rollout, while PMs get clearer control over when page context is captured—explicit site permissions and removable grants in Settings—and a calmer capture flow: highlight a customer quote, competitor claim, roadmap note, or support detail on an approved site, then focus the sidepanel chat input to add it as context without a floating page button. The agent can chain document edits reliably on org knowledge, skills, and project descriptions (including right after install or a data reset), and the assistant can open a URL in the **current** tab when that matches “go to…” instead of always opening a new tab.