[feature] MCP external-analysis surface: recording-level write-back for external AI analysts - #229
Merged
Merged
Conversation
…recording_analysis + list_recordings since/analyzed + finding author/quotes 外部 AI 分析迴圈的寫回面： - update_recording_meta：改任意已存錄音的 meetingType / meetingContext / speakerNames（per-key 合併，空字串移除），修正誤判的會議類型與 me/them 對映 - set_recording_analysis：以 id 全量替換 findings / actionItems、寫 brief、標 analyzed；經前端 normalize 後落盤 + 雲端同步，載入中的 replay 即時反映 - list_recordings 加 since（createdAt >= ms）過濾；summary 卡帶 analyzed 旗標，外部分析者可便宜輪詢新錄音 - TimelineEvent 加 author（外部 pass 署名）與 quotes（原本 schema 有宣告但型別缺欄位，寫入被靜默丟棄的 bug 一併修掉） - get_recording 回傳補 analyzed / companyId / threadId
|
This was referenced Aug 14, 2026
YJack0000
added a commit
that referenced
this pull request
Aug 17, 2026
The site and the README were still selling three subsystems that shipped their last release weeks ago — the live intelligence board, the accounts mini-CRM, and live voice translation — while the thing that actually grew, the MCP server, got one code block and three bullet points. Both are rewritten around where the app converged in v0.22.14–v0.22.18: record, transcribe, analyze, search, and an MCP surface wide enough that an external agent can own the library, including the analysis. website/index.html - Hero, metadata, OG/Twitter cards and JSON-LD rewritten around the record + MCP story instead of real-time coaching. Pre-flight is gone from the workflow steps; meeting context replaces the meeting-type picker. - The MCP section grows from one code block to all 48 tools in five groups, each name checked against src-tauri/src/mcp.rs, plus the auto-analyze-off handover flow that #230 and #229 built. - The iPhone section gets the real App Store link (id6795031201). Android is described honestly as built and syncing but not on Play yet. - Footer regrouped into Product / Project / Legal & support; account deletion moves out of the top-level nav into Legal (the /account-deletion/ URL is unchanged — Play requires it to stay reachable). - Pathors is de-emphasised: the nav link and the "A Pathors AI open source project" line are gone, the JSON-LD Organization graph collapses to a single author field, and the open-source section explains the origin without leaning on the brand. website/{privacy,support,account-deletion} - English-only, per the repository convention. lang was zh-Hant with an "English summary" appended; now there is one language and no duplication. - Support gains three desktop entries that did not exist before: a failed system-audio tap, an MCP client that cannot connect, and analysis not running because auto-analyze is off. README.md - Drops the live-translation section, the Gemini-key install step and the virtual-mic build step. Adds the MCP server as a first-class section with the tool table and the external-analyst flow. showcase-accounts.png is deleted: it pictures the accounts mini-CRM removed in #245 and nothing referenced it any more. Note that the remaining desktop screenshots predate the removals — hero still shows the intel board — and need a reshoot in a follow-up.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What this changes
Adds the MCP write-back surface that lets an external AI analyst (e.g. a Claude Code session) own the analysis of saved recordings while Parley stays the capture/transcribe/replay layer:
update_recording_meta— patch a saved recording's frame by id:meetingType(scenario id),meetingContext, andspeakerNames(merged per key; empty string removes a custom name). This is the repair path for a misclassified meeting (e.g. an office hour analyzed as a sales call) or a wrong me/them speaker mapping.set_recording_analysis— replace whole analysis fields on a saved recording by id:findings,actionItems,brief(markdown debrief), and theanalyzedflag. Omitted fields stay untouched; payloads are normalized by the frontend before hitting disk; the open replay syncs immediately; cloud push rides the existing lifecycle.list_recordingsgains asincefilter (createdAt >= epoch ms) and the summary cards now mirror the entry'sanalyzedflag, so an external analyst can poll cheaply for new, not-yet-analyzed recordings.TimelineEventgainsauthor(which analyst wrote a marker — external passes stamp e.g. "claude", Parley's own pass omits it) andquotes.quoteswas already advertised in the MCP finding schema but missing from the TS type, so quotes sent by MCP clients were silently dropped bynormalizeFinding— fixed.get_recordingadditionally returnsanalyzed,companyId,threadId.Both new tools ride the existing
call_frontendRPC so writes go through the frontend's read-modify-write + cloud-sync path (same asrename_recording), never raw file writes behind the app's back.Why
The external-AI analysis loop (transcribe in Parley → analyze in an AI app over MCP → write conclusions back) currently dead-ends at write-back: findings tools only touch the loaded session, and meeting type / speaker mapping / brief / action items are read-only over MCP. In practice Parley's built-in pass can misread a meeting (a mentoring office hour classified as
sales, the user mapped tothem-1) with no way to correct it from outside.How it was verified
bunx tsc --noEmitpassesbunx vitest runpasses (330 tests)cargo check+cargo clippyclean insrc-tauribun run tauri dev) and exercised the changezh-TWandeninsrc/i18n/messages.ts(no user-facing strings — MCP surface only)End-to-end exercise against a real recording (rename + meta fix + external findings/brief write on an existing office-hour recording) planned right after this lands in a release build.