@meridiona/meridian-darwin-arm64 1.0.0

package/services/skills/activity/task-classifier/SKILL.md

@@ -0,0 +1,205 @@
+---
+name: task-classifier
+description: Classify a user work session against open Jira tickets — pick the best match (or null) using session evidence, and infer dimension tags.
+version: 2.0.0
+metadata:
+  meridian:
+    tags: [classifier, task-linker]
+---
+
+# Session Task Classifier
+
+You are Meridian's AI classifier. Your job is to classify work sessions captured from the user's screen and match them to open Jira tickets when appropriate and return response in strcutured json output.
+
+## Purpose
+
+The task classifier sits at the center of Meridian's workflow understanding:
+
+1. **Screen frames** → **app sessions** (Rust daemon combines frames by app into sessions)
+2. **Sessions** → **task classification** (you classify each session)
+3. **Classification outcome** dictates downstream usage:
+   - Sessions marked as **overhead** → completely discarded. Never surfaced in the UI, never used for inference, never used to create tasks. Treat as throwaway. Examples:  music players, system settings, idle browsing, etc.
+   - Sessions marked as **untracked** → retained and used downstream. Fed into workload analysis, task inference, and new-task creation pipelines. These are real work signals the user performed that just didn't match an open ticket. Examples: standup meetings, config housekeeping, repo exploration, general research.
+   - Sessions with **task matches** → linked to Jira tickets, routing=auto for time-tracking and progress.
+
+## Classification Decision Tree
+
+For each session, you must decide:
+
+### 1. Is this overhead?
+If the session is **idle, music, system settings,or clearly personal/unrelated activity** → return:
+```json
+{"task_key": null, "confidence": 0.0, "session_type": "overhead", "routing": "skip"}
+```
+**overhead is a hard discard.** These sessions are thrown away — never surfaced, never used for inference, never create tasks. When in doubt between overhead and untracked, ask: *"Would a manager care that this happened?"* If no, it's overhead.
+
+### 2. Is this work-related?
+If the session shows **any real work signal** (coding, research, meetings, writing, debugging, reviewing, learning) but **no Jira candidate matches** → mark as **untracked** and return:
+```json
+{"task_key": null, "confidence": 0.6-0.8, "session_type": "untracked", "routing": "queue"}
+```
+**untracked sessions are kept and used downstream** — for workload analysis, capacity reporting, and automatic new-task creation. Mark dimensions to capture *what* the work was. Examples that must be `untracked` (not `overhead`): standups, retros, code reviews on untracked PRs, config/infra housekeeping, general repo exploration, internal tool usage.
+
+### 3. Can it map to an open Jira ticket?
+If the session evidence **directly or contextually matches** an open ticket → return:
+```json
+{"task_key": "KEY-123", "confidence": 0.50-0.90, "session_type": "task", "routing": "auto"}
+```
+Cite the evidence (window title, OCR snippet, context from previous sessions) and infer activity dimensions.
+
+## Your inputs
+
+The user message contains:
+
+- **SESSION** — app, category (with confidence), duration, top window titles, and counts of OCR/audio captures.
+- **CANDIDATE TICKETS** — all open Jira tickets. These are the only tickets you may choose from.
+- **RECENT SESSIONS** (previous 5) — context to help disambiguate. Example: *"User was on KAN-42 (coding) 5 minutes ago, then Slack, now back in VS Code."* → likely same task, even if Slack doesn't directly match KAN-42.
+
+## Available capabilities
+
+**Database access** — You can query the meridian database for verification or additional context if needed:
+```
+sqlite3 "~/.meridian/meridian.db" "<SQL>"
+```
+
+Available tables:
+- `app_sessions` — all captured work sessions (id, app_name, duration_s, session_text, task_key, task_routing, etc.)
+- `pm_tasks` — open Jira tickets (task_key, title, description_text, issue_type, status, epic_title, sprint_name)
+
+Use database queries sparingly — session data and candidate tickets are already provided in the message. Only query if you need to verify a detail or look up historical context not included in the current inputs.
+
+## Your job
+
+Pick **exactly one** of the candidate `task_key` values, OR return `null` if **none** fit the session.
+
+Use **context from previous sessions** to make smarter decisions:
+- If the current session is **generic** (e.g., Slack) but follows/precedes work on a specific ticket, consider linking it to that task.
+- If sessions alternate (coding → Slack → coding), treat them as potentially the **same task** if separated by only a few minutes.
+- Overhead (system settings, music, etc) should always be `null` regardless of context.
+
+## Output format
+
+Reply with ONE valid JSON object — no preamble, no markdown fences, no follow-up text:
+
+```json
+{
+  "task_key": "KAN-86",
+  "confidence": 0.85,
+  "session_type": "task",
+  "reasoning": "Editing run_watcher.py with KAN-86 ticket open in adjacent tab; matches the migration task described.",
+  "dimensions": {"activity": ["coding"], "intent": ["implementation"], "tool": ["vscode"]},
+  "session_summary": "Opened run_watcher.py in VS Code and rewrote the inotify polling loop to use the new ETLConfig path; introduced a 250 ms debounce window and removed the obsolete `_last_tick` global. Ran `cargo check` twice — second attempt clean. Reviewed migration 023 in DBeaver to confirm the pm_sync_state schema matches what the watcher expects. Briefly read the openobserve docs tab for OTLP retry semantics before deciding to defer the retry change to a follow-up. No tests written this session — they are queued behind the watcher refactor."
+}
+```
+
+### Field rules
+- `task_key` — must be one of the supplied candidates, or `null`. Never invent a key.
+- `confidence` — see Scoring heuristics section for exact ranges per outcome type.
+- `session_type` — `"task"` links to Jira; `"overhead"` is thrown away; `"untracked"` is kept for workload analysis.
+- `reasoning` — must cite specific window titles, OCR snippets, or context clues.
+- `dimensions` — omit keys with no evidence; return `{}` if no clear signals.
+- `session_summary` — see the dedicated section below. This is the SINGLE most important field for downstream PM updates.
+
+## session_summary — THE PM-update payload
+
+This field is what the PM-update workflow consumes to write Jira worklog comments. It REPLACES the raw OCR text downstream — the PM agent will not see the original session_text unless it explicitly asks. So **every SDLC-relevant detail you observe must be captured here, written so a tech-lead reading it understands exactly what happened.**
+
+### Length
+
+**Adaptive to the content.** Match depth to evidence:
+
+| Session shape | Target length |
+|---|---|
+| 12-second session, one terminal command, screen mostly static | ~5-10 sentences, factual |
+| 1-3 minute session, single file edited, no errors | ~10-20 sentences |
+| Content-rich session: multiple files, tests, errors, decisions, research | ~25-80 sentences |
+| Long session (5+ min) with a clear narrative arc — debug → fix → verify | up to ~80 sentences |
+
+Do not pad. If a session was genuinely uninteresting, write the truth in a few sentences. If a session was rich, give the full picture.
+
+### Voice
+
+- Past tense, third person ("edited", "ran", "reviewed", "decided") — never "I" or "you"
+- Concrete, file-name-specific — `"edited services/agents/run_task_linker_mlx.py"`, not `"worked on the classifier"`
+- Cite exact commands, error messages, function names when visible
+- No marketing language ("successfully implemented", "made great progress")
+- No speculation about future work ("will need to…", "next step is…")
+
+### What MUST be in the summary (SDLC checklist)
+
+Capture every category the session shows evidence of:
+
+1. **Files / paths / modules touched** — full paths or recognisable basenames
+2. **Commands / scripts / queries run** — and their outcome (success, error message, exit code)
+3. **Errors hit** — exception names, stack-trace snippets, failing assertions
+4. **Tests** — written, run, passed, failed; specific test names
+5. **Technical decisions** — choice made, alternative considered, reason ("picked process-tree over osascript because…")
+6. **Schema / DB changes** — migrations applied, columns added, queries run in DBeaver
+7. **Commits / branches / PRs** — visible in terminal output or VS Code source control panel
+8. **Blockers / open questions** — explicit "stuck on…", unanswered prompts, missing deps
+9. **External research** — docs read, Stack Overflow visited, Claude/ChatGPT consulted (and which question)
+10. **Validations** — manual smoke tests, UI inspections, screenshot reviews, log tailing
+11. **Design discussions** — architecture sketches, ADR notes, conversations with Claude/Codex about approach
+12. **Time-on-subtask hints** — when the session has clearly distinct phases ("first 2 min reading docs, then 6 min editing")
+
+### What NEVER goes in the summary
+
+- "Worked on KAN-XXX" (vague — say what *specifically*)
+- "Successfully implemented X" (drop "successfully"; just say what was done)
+- "Will continue tomorrow" (no speculation)
+- "User seems frustrated" (no interpretation of mood)
+- "This is important because…" (no editorialising)
+- Repeated content from `reasoning` — `reasoning` is for *why* the session matches the ticket; `session_summary` is for *what happened*
+
+### Examples
+
+**Good — short trivial session (~6 sentences):**
+> Ran `git status` and `git diff` in the meridian repo terminal to review pending edits. Output showed three modified files in `services/agents/pm_update/`. No commands executed beyond the status review. No errors. No tests. The session ended when focus shifted to the VS Code window.
+
+**Good — content-rich session (~30-80 sentences):**
+> Edited `services/agents/pm_update/workflow.py` to remove the chunked-summariser heavy-path; deleted the `_is_heavy_bundle` evaluator, the `Parallel(*chunk_summarisers)` block, and the `merge_chunks` step. Workflow now linear: collect → synthesise → ground → route. Then removed `build_chunk_agent` and `build_merger_agent` from `agents.py` along with their `_CHUNK_NAME` / `_MERGER_NAME` constants. Cleaned up `config.py` — dropped `PM_UPDATE_CHUNK_MAX_TOKENS`, `PM_UPDATE_MERGE_MAX_TOKENS`, `PM_UPDATE_CHUNK_PARALLELISM`, `PM_UPDATE_KNOWLEDGE_DIR`. Decision: removed the heavy path entirely instead of just disabling parallelism, because the 9B model's 262K context fits all bundles comfortably. Ran `python -c "from agents.pm_update.workflow import build_workflow; print([s.name for s in build_workflow().steps])"` to verify; output was `['collect','synthesise','ground','route']`. No new tests written. Confirmed via `rm -rf __pycache__` then re-import that no stale references remain. Briefly considered keeping the chunk/merger factories dormant for future use but chose to delete since the 9B context window makes them strictly unnecessary.
+
+**Bad — too vague:**
+> Made changes to the workflow file. Removed some unused code. Cleaned things up. Ran the workflow and it worked.
+
+**Bad — speculative + marketing:**
+> Successfully refactored the workflow to be more efficient. The new linear design will be much faster. Next steps include adding the worklog poster and testing end-to-end on Jira.
+
+## Using Context from Previous Sessions
+
+You have access to **the previous 5 sessions** to disambiguate the current session:
+
+**Example: Coding → Communication about same work → Coding**
+- Session 1 (5 min ago): VS Code, editing KAN-42 implementation → task_key: KAN-42, confidence: 0.90
+- Session 2 (3 min ago): Slack, discussing PR review for KAN-42 → **if related to same work**, task_key: KAN-42, confidence: 0.75 (work mention + prior context)
+- Session 3 (now): VS Code, editing same file → task_key: KAN-42, confidence: 0.85 (context continuity)
+
+**Decision:** If Session 2 (Slack) content shows it's about the same work (discussing the work or searching about it), classify it to **KAN-42** using context from Session 1. If Slack is generic work discussion with no connection to the prior task, return `null` with `session_type: "untracked"`.
+
+Example reasoning for Session 2 (if task-related): `"Slack discusses PR review for KAN-42 implementation mentioned in prior VS Code session; linked via work context."`
+
+## Scoring heuristics
+
+**When task_key is not null (matched to a ticket):**
+- **Task key + work alignment**  — `confidence ≥ 0.90`, `session_type: "task"`
+- **Work description alignment**  — `0.75–0.85`, `session_type: "task"`
+- **Context continuity**  — `0.75–0.85`, `session_type: "task"`
+- **Generic project-level match**  — `0.50–0.65`, `session_type: "task"`
+- **Task key only**  — `0.60–0.75`, `session_type: "task"` (lower than key+alignment because work intent unclear)
+
+## Task mapping
+
+- **Clear overhead signals** (music, SIM browsing, system popups, idle) — `confidence: 0.0–0.2`, `session_type: "overhead"`, `routing: "skip"` → **discarded**
+- **Work activity, no matching ticket** (coding, meetings, reviews, research, config) — `confidence: 0.6–0.8`, `session_type: "untracked"`, `routing: "queue"` → **retained for inference and task creation**
+- **Ambiguous — leans work** (unclear but some work signal present) — `session_type: "untracked"` → **default to untracked, not overhead, when uncertain**
+
+**Decision rule:** Always verify work matches ticket *intent*, not just visible metadata. If equally plausible, pick the ticket whose description best aligns with what the user is *actually doing*.
+
+## Hard rules
+
+- Output JSON only. No fences, no thinking-out-loud before or after the JSON.
+- `task_key` MUST be one of the supplied candidates, or `null`. Never invent a key.
+- Cite specific window titles, OCR snippets, OR context clues (e.g., *"returning to same task after brief Slack"*) in your reasoning.
+- Don't speculate about tickets not in the candidate list.
+- Overhead and breaks should always be `null`, regardless of any other signals.
+- When two candidates seem equally plausible, pick the one whose description more directly matches what the session evidence shows the user *actually doing*.

package/services/tests/conftest.py

@@ -0,0 +1,14 @@
+"""conftest.py — stubs out hermes and observability imports so tests run
+without hermes or an LLM installed."""
+import sys
+from unittest.mock import MagicMock
+
+# Stub modules that require hermes or external services at import time.
+# These must be inserted before run_task_linker is imported.
+_hermes_setup = MagicMock()
+_hermes_setup.ensure_hermes_importable = lambda: None
+sys.modules["agents._hermes_setup"] = _hermes_setup
+
+_observability = MagicMock()
+_observability.setup = lambda *a, **kw: MagicMock()
+sys.modules["agents.observability"] = _observability

package/services/tests/evals/.dataset.json

@@ -0,0 +1,65 @@
+[
+  {
+    "input": "RECENT WORK CONTEXT:\n  03:31  Google Chrome   <1min    → [pending]  [research]\n  03:31  Terminal        <1min    → [pending]  [coding]\n  03:31  Google Chrome   <1min    → [overhead]  [meeting]\n  03:31  Code            <1min    → KAN-107  [coding]\n  03:32  Google Chrome   <1min    → [pending]  [research]\n\nSESSION:\napp: Google Chrome\ntime: 03:32–03:33  (1min)\ncategory: meeting (confidence 0.9)\ntop windows:\n  • teams.cloud.microsoft (×8)\nscreen content [hybrid]:\n[03:32:14]\nOpen office app launcher\nCollapse app bar\nSettings and more\nYour profile, status In a call\nActivity (⌃ ⇧ 1)\nChat (⌃ ⇧ 2)\nCalendar (⌃ ⇧ 3)\nCalls (⌃ ⇧ 4)\nOneDrive (⌃ ⇧ 5)\nCopilot (⌃ ⇧ 6)\nView more apps\nApps\nNo Microphone\nNo Microphone No microphone was found, plug one in or enjoy just listening in. Close\nNo microphone was found, plug one in or enjoy just listening in.\nMyself video, Adithya Harish, Muted, Has context menu\nAdithya Harish\nMore options\nChethan M N, muted, Context menu is available\nChethan M N\nSuyash Diwan, Context menu is available\nSuyash Diwan\nSiddhi Jhade, Context menu is available\nSiddhi Jhade\nAnshul Modh, Context menu is available\nAnshul Modh\nChat | Banavo DevOps StandUp | Microsoft Teams - Audio playing - High memory usage - 1.1 GB\nMute tab\nClose\nNew Tab\nOpen Gemini in Chrome\nChromeFileEditViewHistoryBookmarks ProfilesTabWindowHelp• • •Ti Chat | Banavo DevOps Sta 4) xteams.cloud.microsoft01:28вAdithya HarishChethan MN XSJSiddhi JhadeChat85PeopleRaiseReactNo MicrophoneNo microphone was found, plug one in orenjoy just listening in.XCNSuyash DiwanAnshul ModhNotes*+QWed May 20 9:02AM• TerminalAsk GeminiWi-FiThis one_5gBluetoothAirDropEveryoneNot Playing10FocusDisplaySound-()Edit Controls\nSiddhi Jhade, muted, Context menu is available\n→ii Chat | Banavo DevOps Sta 4) Xteams.cloud.microsoftDQ Press T G to go right to a chat or channelO € 01:31No MicrophoneNo microphone was found, plug one in orenjoy just listening in.CNAdithya Harish & ...Chethan MN &SJSiddhi JhadedAnshul ModhChat85PeopleRaiseReactSuyash DiwanHODSMoreballlelaSD* Ask GeminiCa1Share\nContent shared by Suyash Diwan\nZoom is reset to 100%\nChat | Banavo DevOps StandUp | Microsoft Teams - High memory usage - 1.1 GB\nii Chat | Banavo DevOps Standl X→teams.cloud.microsoftDQ Type \"with:\" to narrow your search to a personO € 01:46• Andromeda 360, Inc.XIC Accounts IAWS access pomtalmlxE shople-processed-datalake -S ×+ >teams.cloud.microsoft*No MicrophoneNo microphone was found, plug one in orenjoy just listening in.88Q Search (Ctrl+Alt+E)Your status is set to do not disturb. You'llonlyget notifications for urgent messages and from your priority contacts.•E 01:42€ g5ChatPeopleRaiseReactSDCNChat85PeopleRaisem Daily Standup T xViewNotesAppsMoreSuyash DiwanChethan MN &SiddhiJhade &Adithya Harish 2324°CSuyash Diwanstly sunnyteams.cloud.microsoft is sharing your screen.Stop sharingQ SearchHideReact• *ViewNotes• Ask GeminiCameraMicSJ0ENG• *AppsMoreWorkM All BookmarksChange settingsStop shar\n  … (8608 more chars)\n\nCANDIDATE TICKETS:\n1. KAN-67  [Feature · Epic: PM Write-back]\n   title: Jira Worklog Write-back\n   description: Automatically submit time spent (from  app_sessions ) as Jira worklogs for sessions that have been confidently mapped to tasks. Includes a human-in-the-loop approval queue for uncertain mappings, and a full undo/correction capability.\\n\\n A…\n\n2. KAN-65  [Feature · Epic: UI Intelligence Dashboard]\n   title: Session-Task Linkage Display\n   description: Show developers which PM task each session was mapped to, along with confidence and routing status. This makes the intelligence engine visible and auditable, and gives users confidence in the system before write-back is enabled.\\n\\n Accepta…\n\n3. KAN-64  [Feature · Epic: Session Intelligence Engine(Session to task categorizer)]\n   title: Session-to-Task Mapping Engine\n   description: Design and implement the pipeline that takes a classified session and produces a confident match to a PM task (or a deliberate skip). The mapping engine runs inline after each ETL cycle and writes results to  ticket_links .\\n\\n Why this mat…\n\n4. KAN-109  [Task]\n   title: LLM benchmarking and prompt evaluation framework for Session-Task classifier\n   description: Set up an offline evaluation harness (DeepEval + local LLm) to compare model quality and prompt variants for the session→task classification pipeline. Goal:  Make model-swap decisions data-driven rather than intuition-based. The results fee…\n\n5. KAN-134  [Task · Epic: Redesign product UI and messaging]\n   title: UI Branding & Messaging Standardization\n   description: Audit and standardize product terminology; create a style guide; implement consistent labeling and messaging across all UI surfaces; and add contextual help and tooltips.\n\n6. KAN-135  [Task · Epic: Redesign product UI and messaging]\n   title: UI Component Redesign & Implementation\n   description: TaskBadge visual design (color palette, states, animations), dashboard layout redesign (widget hierarchy, spacing, collapsible sections), session timeline visualization, and classification confidence feedback UI",
+    "expected_output": "{\"task_key\": \"none\", \"session_type\": \"untracked\", \"reasoning\": \"Banavo DevOps standup meeting on Microsoft Teams — team coordination session. Work-related activity that does not map to a specific technical ticket but is retained for workload analysis and potential task inference.\"}",
+    "additional_metadata": {
+      "session_id": 1792,
+      "app_name": "Google Chrome",
+      "task_method": "hermes_aiagent"
+    }
+  },
+  {
+    "input": "RECENT WORK CONTEXT:\n  05:27  DBeaver         <1min    → [pending]  [research]\n  05:27  Terminal        <1min    → [pending]  [coding]\n  05:29  Code            <1min    → [overhead]  [coding]\n  05:29  Google Chrome   <1min    → KAN-67  [code_review]\n  05:30  Code            <1min    → [overhead]  [coding]\n\nSESSION:\napp: Google Chrome\ntime: 05:30–05:33  (3min)\ncategory: code_review (confidence 0.9)\ntop windows:\n  • github.com (×95)\nscreen content [hybrid]:\n[05:30:20]\nSkip to content\nOpen menu\nHomepage ( g then d )\nMeridiona\nmeridian\nSearch or jump to…\nType\nto search\nCreate new...\nAll issues\nAll pull requests\nAll repositories\nYou have unread notifications ( g then n )\nOpen user navigation menu\nRepository navigation\nCode\nIssues\nPull requests  (2)\nPull requests\nActions\nProjects\nSecurity and quality\nInsights\nSettings\nFeat/production cleanup #12 Edit title\nFeat/production cleanup\nEdit title\nMerged\nadityaharishch\nmerged 8 commits into\nmain\nfrom\nfeat/production-cleanup\nCopy head branch name to clipboard\nnow\nLines changed: 1757 additions & 967 deletions\nConversation  (0)\nConversation\nCommits  (8)\nCommits\nChecks  (1)\nChecks\nFiles changed  (24)\nFiles changed\n@adityaharishch\nShow options\nadityaharishch commented 1 minute ago\ncommented\n1 minute ago\nNo description provided.\nAdd or remove reactions\nAditya Harish\nand others\nadded\ncommits\n3 minutes ago\n@claude\nrefactor(task-linker): split db.rs into read/write modules and extrac…\nCommit message body\n2a93563\nfeat(task-linker): add startup preflight check for classification stack\nb95d324\nfix(config): disable jira updater by default, skip gracefully if modu…\n7a94a4a\nrefactor(services): consolidate DB connections and harden input valid…\nc6d7685\ntest(services): add parser and run_task_linker test suites\na97dc67\ntest(task-linker): replace stub smoke tests with real hermes integrat…\nf7a8fc2\nfeat(scripts): add setup-services.sh for co-dev onboarding\ne316c83\ndocs: update README onboarding flow and TESTING.md task-linker section\nPull requests: Meridiona/meridian\nFilters\nis:pr is:open\nLabels 9\nLabels\nMilestones 0\nMilestones\nNew pull request\nSelect all issues\n1 Open\n11 Closed\nAuthor\nLabel\nReviews\nAssignee\nSort\nPull requests list\nfeat(intelligence): dynamic local LLM selection for task classifier (Apple Silicon)\n1 / 1 checks OK\n#11 opened\n2 days ago\nAkarsh-Hegde\n5 tasks\nProTip!\nExclude everything labeled\nbug\nwith\n-label:bug\nFooter\nGitHub Homepage\n© 2026 GitHub, Inc.\nFooter navigation\nTerms\nPrivacy\nSecurity\nStatus\nCommunity\nDocs\nContact\nManage cookies\nDo not share my personal information\nhttps://github.com/Meridiona/meridian/pull/11\nUsage · Settings\nClose\nPull requests · Meridiona/meridian - Memory usage - 459 MB\nNew Tab\nOpen Gemini in Chrome\nMeridiona/meridian\nMay 18\nfeat(intelligence): dynamic local LLM selection for task classifier (Apple Silicon) #11\n#11\nOpen\nSummary Introduces\nselect_model_for_hermes()\nllm_selector.py\n— discovers the best…\nKAN-97-dynamic-llm-choice-and-download-logi\n  … (48184 more chars)\n\nCANDIDATE TICKETS:\n1. KAN-67  [Feature · Epic: PM Write-back]\n   title: Jira Worklog Write-back\n   description: Automatically submit time spent (from  app_sessions ) as Jira worklogs for sessions that have been confidently mapped to tasks. Includes a human-in-the-loop approval queue for uncertain mappings, and a full undo/correction capability.\\n\\n A…\n\n2. KAN-65  [Feature · Epic: UI Intelligence Dashboard]\n   title: Session-Task Linkage Display\n   description: Show developers which PM task each session was mapped to, along with confidence and routing status. This makes the intelligence engine visible and auditable, and gives users confidence in the system before write-back is enabled.\\n\\n Accepta…\n\n3. KAN-64  [Feature · Epic: Session Intelligence Engine(Session to task categorizer)]\n   title: Session-to-Task Mapping Engine\n   description: Design and implement the pipeline that takes a classified session and produces a confident match to a PM task (or a deliberate skip). The mapping engine runs inline after each ETL cycle and writes results to  ticket_links .\\n\\n Why this mat…\n\n4. KAN-109  [Task]\n   title: LLM benchmarking and prompt evaluation framework for Session-Task classifier\n   description: Set up an offline evaluation harness (DeepEval + local LLm) to compare model quality and prompt variants for the session→task classification pipeline. Goal:  Make model-swap decisions data-driven rather than intuition-based. The results fee…\n\n5. KAN-134  [Task · Epic: Redesign product UI and messaging]\n   title: UI Branding & Messaging Standardization\n   description: Audit and standardize product terminology; create a style guide; implement consistent labeling and messaging across all UI surfaces; and add contextual help and tooltips.\n\n6. KAN-135  [Task · Epic: Redesign product UI and messaging]\n   title: UI Component Redesign & Implementation\n   description: TaskBadge visual design (color palette, states, animations), dashboard layout redesign (widget hierarchy, spacing, collapsible sections), session timeline visualization, and classification confidence feedback UI",
+    "expected_output": "{\"task_key\": \"KAN-107\", \"session_type\": \"task\", \"reasoning\": \"User is reviewing PR #11 on GitHub, which explicitly mentions 'dynamic local LLM selection for task classifier' and references 'KAN-97' (not in candidates), but the content align strongly with the 'agent reasoning to the session to task mapping' described in KAN-107. This is a continuation of the code review activity seen in the previous session (KAN-67 was likely a misclassification or shift, but the screen content here is specifically about the classifier's intelligence layer).\"}",
+    "additional_metadata": {
+      "session_id": 1961,
+      "app_name": "Google Chrome",
+      "task_method": "hermes_aiagent"
+    }
+  },
+  {
+    "input": "RECENT WORK CONTEXT:\n  05:35  Code            <1min    → [pending]  [coding]\n  05:35  DBeaver         2min     → KAN-76  [coding]\n  05:37  Terminal        <1min    → [pending]  [deployment_devops]\n  05:37  Code            2min     → KAN-107  [coding]\n  05:40  Google Chrome   <1min    → KAN-59  [code_review]\n\nSESSION:\napp: Code\ntime: 05:41–05:42  (1min)\ncategory: coding (confidence 0.9)\ntop windows:\n  • .gitignore — meridian — Modified (×11)\n  • .mcp.json — meridian — Modified (×3)\n  • pyproject.toml — meridian — Modified (×2)\n  • .env — meridian (×2)\nscreen content [accessibility]:\n[05:41:13]\nExplorer (⇧⌘E)\nSearch (⇧⌘F)\nSource Control (⌃⇧G) - 64 pending changes\nRun and Debug (⇧⌘D)\nExtensions (⇧⌘X) - 1 requires restart\nClaude Code\nEXPLORER\nExplorer Section: meridian\nMERIDIAN\n.agents\n.cargo\n.claude\n.claude-flow\n.githooks\n.github\n.swarm\nnode_modules\npackages\nscripts\nservices\n.hermes\n.pytest_cache\n.venv\nagents\n__pycache__\ntests\n__init__.py\n_parser.py\n_prompts.py\nconfig.py\nobservability.py\nREADME.md\nrun_task_linker.py\nserver.py\nhermes-config\nmeridian_agents.egg-info\nskills\n.env\n.env.example\n.gitignore\npyproject.toml\nrequirements.txt\nsrc\nswift\ntarget\nOutline Section\nOUTLINE\nTimeline Section\nTIMELINE\nconfig.yaml\nREADME.md, preview\n# meridian — hermes gateway secrets template\n# Copy to services/.hermes/.env and fill in values.\n# NEVER commit .hermes/.env — it is gitignored.\n# ── Model provider ────────────────────────────────────────────────────────────\n# ── API server ────────────────────────────────────────────────────────────────\n# Bearer token Rust uses to call the gateway. Pick any string.\nAPI_SERVER_ENABLED=true\n[ENV_SECRET]\n# ── Optional overrides ────────────────────────────────────────────────────────\n# API_SERVER_PORT=8642\n# API_SERVER_HOST=[IP_ADDRESS]\n# OLLAMA_MODEL=gemma4:31b        # override model\n# OLLAMA_HOST=https://ollama.com/v1\nProblems (⇧⌘M)\nPROBLEMS\nOutput (⇧⌘U)\nOUTPUT\nDebug Console (⇧⌘Y)\nDEBUG CONSOLE\nTerminal (⌃`)\nTERMINAL\nPorts\nPORTS\nremote\nmeridian (Git) - feat/production-cleanup*, Checkout Branch/Tag...\nfeat/production-cleanup*\nmeridian (Git) - Synchronize Changes\nNo Problems\nNotifications\nSign In\nDotenv\nEditor Language Status: $(copilot) No inline suggestion available, Inline suggestions\nUTF-8\nSpaces: 4\nLn 6, Col 1\nScreen Reader Optimized\nexpanded\nCommand Succeeded\n# Generated by Cargo\ndebug\n/target\nmeridian.db\nmeridian.db-shm\nmeridian.db-wal\n/node_modules\n# These are backup files generated by rustfmt\n**/*.rs.bk\n# MSVC Windows builds of rustc generate these, which store debugging information\n*.pdb\n# Generated by cargo mutants\n# Contains mutation testing data\n**/mutants.out*/\n# RustRover\n#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can\n#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore\n#  and can be added to the global gitignore or merged into this file.  For a more nuclear\n#  option (not recommended) you can uncomment the following to ignore the entire idea folder.\n#.idea/\n# Cargo local config — ignore cache but track config.toml\n.cargo/*\n\n  … (3037 more chars)\n\nCANDIDATE TICKETS:\n1. KAN-67  [Feature · Epic: PM Write-back]\n   title: Jira Worklog Write-back\n   description: Automatically submit time spent (from  app_sessions ) as Jira worklogs for sessions that have been confidently mapped to tasks. Includes a human-in-the-loop approval queue for uncertain mappings, and a full undo/correction capability.\\n\\n A…\n\n2. KAN-65  [Feature · Epic: UI Intelligence Dashboard]\n   title: Session-Task Linkage Display\n   description: Show developers which PM task each session was mapped to, along with confidence and routing status. This makes the intelligence engine visible and auditable, and gives users confidence in the system before write-back is enabled.\\n\\n Accepta…\n\n3. KAN-64  [Feature · Epic: Session Intelligence Engine(Session to task categorizer)]\n   title: Session-to-Task Mapping Engine\n   description: Design and implement the pipeline that takes a classified session and produces a confident match to a PM task (or a deliberate skip). The mapping engine runs inline after each ETL cycle and writes results to  ticket_links .\\n\\n Why this mat…\n\n4. KAN-109  [Task]\n   title: LLM benchmarking and prompt evaluation framework for Session-Task classifier\n   description: Set up an offline evaluation harness (DeepEval + local LLm) to compare model quality and prompt variants for the session→task classification pipeline. Goal:  Make model-swap decisions data-driven rather than intuition-based. The results fee…\n\n5. KAN-134  [Task · Epic: Redesign product UI and messaging]\n   title: UI Branding & Messaging Standardization\n   description: Audit and standardize product terminology; create a style guide; implement consistent labeling and messaging across all UI surfaces; and add contextual help and tooltips.\n\n6. KAN-135  [Task · Epic: Redesign product UI and messaging]\n   title: UI Component Redesign & Implementation\n   description: TaskBadge visual design (color palette, states, animations), dashboard layout redesign (widget hierarchy, spacing, collapsible sections), session timeline visualization, and classification confidence feedback UI",
+    "expected_output": "{\"task_key\": \"none\", \"session_type\": \"untracked\", \"reasoning\": \"The user is editing project configuration files (.gitignore, .mcp.json, pyproject.toml) and .env secrets. While this is clearly work-related 'coding' activity, it doesn't align with any specific candidate ticket's goal (like the mapping engine or UI dashboard); it appears to be general environment setup or project maintenance.\"}",
+    "additional_metadata": {
+      "session_id": 1972,
+      "app_name": "Code",
+      "task_method": "hermes_aiagent"
+    }
+  },
+  {
+    "input": "SESSION:\napp: Google Chrome\ntime: 08:46–08:49  (3min)\ncategory: idle_personal (confidence 0.9)\ntop windows:\n  • google.com (×36)\n  • airtel.in (×5)\nscreen content [unknown]:\n[08:46:15]\nPlease wait ...\nHold on!\nWe are loading the page for you\nUsage · Settings\nClose\nmeridian/services/skills/activity/task-classifier/SKILL.md at feat/production-cleanup · Meridiona/meridian - Memory usage - 165 MB\nMeridian\n(147) Season 1 Suit Clips To Play Over A Coffee Break - YouTube - Memory usage - 353 MB\nport esim jio to airtel - Google Search\nPostpaid SIM Card: Get Unlimited 4G & 5G with New Postpaid Connection\nSearch results - [EMAIL] - Gmail - Memory usage - 419 MB\nairtel new sim cost - Google Search\nminimum recharge to keep jio sim active - Google Search\nNew Tab\nOpen Gemini in Chrome\n• • • Usage • Settings x | meridian/services/s x |® Meridianx • (147) Season 1 Suit X G port esim jio to airte X© airtel.in/new-conne xM Search results - adi Xdlc arte new sım costxaaci minimum recharge XSI-+>CAairtel.in/new-connection/postpaid-sim/new-sim/?price=549&utm_source=ps_pfx_postpaid_google&utm_campaign=PFX_Postpaid_Brand_NonCore_Exact_SFCM_Port_549_loss_aversion_port%20to%2Oairtel&utm_m.** Ask GeminiPlease wait ...Hold on!We are loading the page for you\nINDIVIDUAL\nBUSINESS\nINVESTORS\nExplore Airtel\nAirtel Black\nPrepaid\nPostpaid\nWi-Fi\nDTH\nBank\nAirtel Finance\nHelp and Support\nAccount\nHome\nGet Airtel Postpaid Connection\nEnjoy a stronger network with Postpaid\nstep 1 Pick a connection type New Connection change connection type\nstep 2 Add contact details\nstep 3 Add delivery address to view plans\nSuperfast speeds in crowds with Fast Lane Tech step 4 Select a priority postpaid plan Unlimited data + top OTTs\nSubmit Postpaid SIM Request\nQuick Access\nAirtel Black Bill Payment\nPrepaid Recharge\nPostpaid Bill Payment\nDTH Recharges\nBroadband Payment\nExciting Offers\nPersonal Loans\nAirtel Payments Bank\nAirtel Thanks App\nFree Sim Delivery\nTelecom Tariff\nRights Issue\nQIP & FCCB Issue\nTerms & Conditions\nAirtel Blogs\nHelp At Hand\nFind a store\nCheck Coverage\nOwnership Transfer\nManage Your account\nISP Near Me\nWynk Music\nBroadband FAQs\nDTH FAQs\nAirtel Xstream FAQs\nPostpaid FAQs\nStores near me\nDTH Consumer Corner\nDo not disturb\nAbout Airtel\nBrand\nCareer\nPartner\nBusiness\nPress Release\nInvestor\nInvestor Support\nTelecom Charter\nNew Connections\nBuy DTH New Connection\nBuy Prepaid New Connection\nBuy Postpaid New Connection\nBuy Fiber New Connection\nBuy Airtel Black Plan\nBuy Xstream Box\nSwitch to Airtel\nInternational Roaming\nGet the app\nFollow us on\nBACK TO TOP\n[08:46:48]\nWhy miss out on superfast speeds\nin crowds with\nPriority Postpaid\nJust place a request and we will get in touch \n  … (32130 more chars)\n\nCANDIDATE TICKETS:\n1. KAN-67  [Feature · Epic: PM Write-back]\n   title: Jira Worklog Write-back\n   description: Automatically submit time spent (from  app_sessions ) as Jira worklogs for sessions that have been confidently mapped to tasks. Includes a human-in-the-loop approval queue for uncertain mappings, and a full undo/correction capability.\\n\\n A…\n\n2. KAN-65  [Feature · Epic: UI Intelligence Dashboard]\n   title: Session-Task Linkage Display\n   description: Show developers which PM task each session was mapped to, along with confidence and routing status. This makes the intelligence engine visible and auditable, and gives users confidence in the system before write-back is enabled.\\n\\n Accepta…\n\n3. KAN-64  [Feature · Epic: Session Intelligence Engine(Session to task categorizer)]\n   title: Session-to-Task Mapping Engine\n   description: Design and implement the pipeline that takes a classified session and produces a confident match to a PM task (or a deliberate skip). The mapping engine runs inline after each ETL cycle and writes results to  ticket_links .\\n\\n Why this mat…\n\n4. KAN-109  [Task]\n   title: LLM benchmarking and prompt evaluation framework for Session-Task classifier\n   description: Set up an offline evaluation harness (DeepEval + local LLm) to compare model quality and prompt variants for the session→task classification pipeline. Goal:  Make model-swap decisions data-driven rather than intuition-based. The results fee…\n\n5. KAN-134  [Task · Epic: Redesign product UI and messaging]\n   title: UI Branding & Messaging Standardization\n   description: Audit and standardize product terminology; create a style guide; implement consistent labeling and messaging across all UI surfaces; and add contextual help and tooltips.\n\n6. KAN-135  [Task · Epic: Redesign product UI and messaging]\n   title: UI Component Redesign & Implementation\n   description: TaskBadge visual design (color palette, states, animations), dashboard layout redesign (widget hierarchy, spacing, collapsible sections), session timeline visualization, and classification confidence feedback UI",
+    "expected_output": "{\"task_key\": \"none\", \"session_type\": \"overhead\", \"reasoning\": \"The session consists of personal activity including browsing for SIM cards on airtel.in and watching YouTube ('Season 1 Suit Clips'), as evidenced by window titles and screen content.\"}",
+    "additional_metadata": {
+      "session_id": 2181,
+      "app_name": "Google Chrome",
+      "task_method": "hermes_aiagent"
+    }
+  },
+  {
+    "input": "RECENT WORK CONTEXT:\n  09:20  Google Chrome   <1min    → [pending]  [research]\n  09:21  Google Chrome   <1min    → KAN-109  [code_review]\n  09:21  Google Chrome   <1min    → [pending]  [research]\n  09:21  Code            <1min    → KAN-109  [coding]\n  09:22  Terminal        <1min    → KAN-109  [coding]\n\nSESSION:\napp: Code\ntime: 09:22–09:35  (12min)\ncategory: coding (confidence 0.9)\ntop windows:\n  • SKILL.md — meridian (×27)\n  • test_config.py — meridian — Deleted (×3)\n  • test_task_classifier_agent.py — meridian — Deleted (×2)\n  • test_parser.py — meridian (×1)\nscreen content [unknown]:\n[09:22:25]\nExplorer (⇧⌘E)\nSearch (⇧⌘F)\nSource Control (⌃⇧G)\nRun and Debug (⇧⌘D)\nExtensions (⇧⌘X) - 1 requires restart\nClaude Code\nEXPLORER\nExplorer Section: meridian\nMERIDIAN\n.cargo\n.claude\n.claude-flow\n.githooks\n.github\npackages\nscripts\nservices\n.hermes\n.pytest_cache\n.venv\nagents\nhermes-config\nmeridian_agents.egg-info\nskills\nactivity\ntask-classifier\nSKILL.md\ntests\n.env\n.env.example\n.gitignore\npyproject.toml\nREADME.md\nrequirements.txt\nsrc\nbin\netl\nintelligence\ncategory_llm\nproviders\nsession_categorizer\ntask_linker\ndb_write.rs\ndb.rs\nmod.rs\ncategory_settler.rs\nmigrations\nOutline Section\nOUTLINE\nTimeline Section\nTIMELINE\nrun_task_linker.py\nscoring.rs\n_prompts.py\nname: task-classifier\ndescription: Classify a user work session against open Jira tickets — pick the best match (or null) using session evidence, and infer dimension tags.\nversion: 2.0.0\nmetadata:\nmeridian:\ntags: [classifier, task-linker]\n# Session Task Classifier\nYou are Meridian's AI classifier. Your job is to classify work sessions captured from the user's screen and match them to open Jira tickets when appropriate.\n## Purpose\nThe task classifier sits at the center of Meridian's workflow understanding:\n1. **Screen frames** → **app sessions** (Rust daemon combines frames by app into sessions)\n2. **Sessions** → **task classification** (you classify each session)\n3. **Classification outcome** dictates downstream usage:\n- Sessions marked as **overhead** → idle/system/unrelated activity, truly overhead, not tracked, routing=skip\n- Sessions marked as **untracked** → work-related but no matching ticket, useful for potential new task creation, routing=queue\n- Sessions with **task matches** → linked to Jira tickets, routing=auto for tracking\n## Classification Decision Tree\nFor each session, you must decide:\n### 1. Is this overhead?\nIf the session is **idle, system settings, app chrome, random browsing, or unrelated activity** → return:\n```json\n{\"task_key\": null, \"confidence\": 0.0, \"session_type\": \"overhead\", \"routing\": \"skip\"}\nThese sessions should never force a Jira link.\n### 2. Is this work-related?\nIf the session shows **clear work signals** (coding, writing, research) but **no Jira candidates match** → mark as **untracked** and return:\n{\"task_key\": null, \"confidence\": 0.6-0.8, \"session_type\": \"untracked\", \"routing\": \"queue\"}\nMark dimensions to show *what* the work was (activity, intent, tool, etc.). High confidence here means we're sure it's work; it's just not mapped to a tracked task. The session is us\n  … (11729 more chars)\n\nCANDIDATE TICKETS:\n1. KAN-67  [Feature · Epic: PM Write-back]\n   title: Jira Worklog Write-back\n   description: Automatically submit time spent (from  app_sessions ) as Jira worklogs for sessions that have been confidently mapped to tasks. Includes a human-in-the-loop approval queue for uncertain mappings, and a full undo/correction capability.\\n\\n A…\n\n2. KAN-65  [Feature · Epic: UI Intelligence Dashboard]\n   title: Session-Task Linkage Display\n   description: Show developers which PM task each session was mapped to, along with confidence and routing status. This makes the intelligence engine visible and auditable, and gives users confidence in the system before write-back is enabled.\\n\\n Accepta…\n\n3. KAN-64  [Feature · Epic: Session Intelligence Engine(Session to task categorizer)]\n   title: Session-to-Task Mapping Engine\n   description: Design and implement the pipeline that takes a classified session and produces a confident match to a PM task (or a deliberate skip). The mapping engine runs inline after each ETL cycle and writes results to  ticket_links .\\n\\n Why this mat…\n\n4. KAN-109  [Task]\n   title: LLM benchmarking and prompt evaluation framework for Session-Task classifier\n   description: Set up an offline evaluation harness (DeepEval + local LLm) to compare model quality and prompt variants for the session→task classification pipeline. Goal:  Make model-swap decisions data-driven rather than intuition-based. The results fee…\n\n5. KAN-134  [Task · Epic: Redesign product UI and messaging]\n   title: UI Branding & Messaging Standardization\n   description: Audit and standardize product terminology; create a style guide; implement consistent labeling and messaging across all UI surfaces; and add contextual help and tooltips.\n\n6. KAN-135  [Task · Epic: Redesign product UI and messaging]\n   title: UI Component Redesign & Implementation\n   description: TaskBadge visual design (color palette, states, animations), dashboard layout redesign (widget hierarchy, spacing, collapsible sections), session timeline visualization, and classification confidence feedback UI",
+    "expected_output": "{\"task_key\": \"KAN-109\", \"session_type\": \"task\", \"reasoning\": \"User is editing 'SKILL.md' for the 'task-classifier' agent in VS Code, which directly aligns with KAN-109 (LLM benchmarking and prompt evaluation framework for Session-Task classifier). Recent history also shows continuous work on KAN-109 in Code and Terminal.\"}",
+    "additional_metadata": {
+      "session_id": 2276,
+      "app_name": "Code",
+      "task_method": "hermes_aiagent"
+    }
+  },
+  {
+    "input": "RECENT WORK CONTEXT:\n  11:14  DBeaver         <1min    → [pending]  [research]\n  11:14  Google Chrome   <1min    → [pending]  [research]\n  11:14  Google Chrome   <1min    → [pending]  [planning]\n  11:14  Google Chrome   <1min    → KAN-109  [planning]\n  11:15  Code            1min     → KAN-109  [coding]\n\nSESSION:\napp: DBeaver\ntime: 11:16–11:20  (3min)\ncategory: deployment_devops (confidence 0.9)\ntop windows:\n  • DBeaver 26.0.4 - app_sessions (×11)\n  • DBeaver 26.0.4 - pm_tasks (×1)\n  • Save result set edit (×1)\n  • Exit DBeaver (×1)\nscreen content [unknown]:\n[11:16:32]\nIST\npm_tasks\nmeridian.db\nframes\napp_sessions\ndb.sqlite\nactivity_context\nagent_cursor\nFiles - General\nDatabase Navigator\nProjects\nDBeaver 26.0.4 - pm_tasks\n$= / sQL =:@ Commit =L Rollback - T=:@ Database Navigator X G Projectsriter oolecis ilaoles. oroceoures. users. erc.ov namne›db.sqlitemeridian.dbE Tables› _sqlxmigrations› Eactive_session› activity_context› Eagent_cursor› Eagent_runs› Eapp_sessions› Econtext_graph_nodes› E dispatch_queue› Eetl_cursor› Eetl_runs> Egaps›Ejira_update_log> Epm_task_embeddings> Epm_tasks> Esession_dimensions› E session_embeddings› Esession_summaries> • Views› •Indexes› •SequencesTable Triggers› I Data TypesAutoDmeridiand-8Tframesmeridian.dbE app_sessionsmeridian.dbE PropertiesEg Data6o6 DiagramE Show SQL I{3 Enter a SQL expression to filter results (use Ctrl+Space)E GridA-Z statusnce layer inside the Meridian Rust daem: To Doas todo → inprogressTo DoT Textme spent (from app_sessions ) as Jira i To Doh PM task each session was mapped to,To Do5line scores a session, persist the result To Doa pipeline insrc/intelligence/ usingtheTo Doselect the best approach for mapping: To Dothe pipeline that takes a classified sessi To Dong time-tracking data back to PM tools.To Do-task mappings produced by the intellig To Doract a submitted worklog or move it to a To Do121314routing = 'queue' in the Ul for develope To Doing = 'auto' that have not yet been logg To Doer-facing text in the Meridian dashboarc To Dooups sessions by their linked PMtask ar To Doit page (/sessions) to show the PM tast To Do1718copy, and empty-state messaging inth To Dongineering work for Meridiona's core pla In Progressdb.sqliteEactivity_contextE agent_cursor#pm_tasksXA2 status_categoryTOOOtodotodotodotodotodoTOOOtodotodotodotodotodotodotodotodotodoin_progressAZ issue_typeEpicTaskFeatureFeatureSubtaskSubtaskSubtaskFeatureEpicEpicSubtaskSubtaskSubtaskSubtaskSubtaskSubtaskFeatureEpicAZ project_keyKANKANKANKANKANKANKANKANKANKANKANKANKANKANKANKANKANKANAZ urlz https://meridiona.atlassian.net/browse/KAN-59@ https://meridiona.atlassian.net/browse/KAN-107rhttps://meridiona.atlassian.net/browse/KAN-67e https://meridiona.atlassian.net/browse/KAN-65https://meridiona.atlassian.net/browse/KAN-76https://meridiona.atlassian.net/browse/KAN-75https://meridiona.atlassian.net/browse/KAN-74@ https://meridiona.atlassian.net/browse/KAN-64rhttps://meridiona.atlassian.net/browse/KAN-61• https://meridiona.atlassian.net/browse/KAN-60https://meridiona.atlassian.net/browse/KAN-82http\n  … (8111 more chars)\n\nCANDIDATE TICKETS:\n1. KAN-67  [Feature · Epic: PM Write-back]\n   title: Jira Worklog Write-back\n   description: Automatically submit time spent (from  app_sessions ) as Jira worklogs for sessions that have been confidently mapped to tasks. Includes a human-in-the-loop approval queue for uncertain mappings, and a full undo/correction capability.\\n\\n A…\n\n2. KAN-65  [Feature · Epic: UI Intelligence Dashboard]\n   title: Session-Task Linkage Display\n   description: Show developers which PM task each session was mapped to, along with confidence and routing status. This makes the intelligence engine visible and auditable, and gives users confidence in the system before write-back is enabled.\\n\\n Accepta…\n\n3. KAN-64  [Feature · Epic: Session Intelligence Engine(Session to task categorizer)]\n   title: Session-to-Task Mapping Engine\n   description: Design and implement the pipeline that takes a classified session and produces a confident match to a PM task (or a deliberate skip). The mapping engine runs inline after each ETL cycle and writes results to  ticket_links .\\n\\n Why this mat…\n\n4. KAN-109  [Task]\n   title: LLM benchmarking and prompt evaluation framework for Session-Task classifier\n   description: Set up an offline evaluation harness (DeepEval + local LLm) to compare model quality and prompt variants for the session→task classification pipeline. Goal:  Make model-swap decisions data-driven rather than intuition-based. The results fee…\n\n5. KAN-134  [Task · Epic: Redesign product UI and messaging]\n   title: UI Branding & Messaging Standardization\n   description: Audit and standardize product terminology; create a style guide; implement consistent labeling and messaging across all UI surfaces; and add contextual help and tooltips.\n\n6. KAN-135  [Task · Epic: Redesign product UI and messaging]\n   title: UI Component Redesign & Implementation\n   description: TaskBadge visual design (color palette, states, animations), dashboard layout redesign (widget hierarchy, spacing, collapsible sections), session timeline visualization, and classification confidence feedback UI",
+    "expected_output": "{\"task_key\": \"KAN-107\", \"session_type\": \"task\", \"reasoning\": \"User is using DBeaver to inspect the 'pm_tasks' and 'app_sessions' tables in meridian.db. The screen content explicitly shows a checklist item 'todo → inprogress' and mentions mapping sessions to PM tasks, which directly aligns with KAN-107's title ('Add agent reasoning to the session to task mapping') and description ('mark the task as todo → inprogress').\"}",
+    "additional_metadata": {
+      "session_id": 2354,
+      "app_name": "DBeaver",
+      "task_method": "hermes_aiagent"
+    }
+  },
+  {
+    "input": "RECENT WORK CONTEXT:\n  13:09  Google Chrome   1min     → KAN-109  [meeting]\n  13:10  Google Chrome   <1min    → [overhead]  [communication]\n  13:11  Google Chrome   <1min    → [pending]  [research]\n  13:11  Google Chrome   <1min    → [pending]  [research]\n  13:11  Google Chrome   1min     → [overhead]  [idle_personal]\n\nSESSION:\napp: Google Chrome\ntime: 13:12–13:13  (1min)\ncategory: code_review (confidence 0.9)\ntop windows:\n  • github.com (×4)\nscreen content [unknown]:\n[13:12:27]\nSkip to content\nOpen menu\nHomepage ( g then d )\nMeridiona\nmeridian\nSearch or jump to…\nType\nto search\nCreate new...\nAll issues\nAll pull requests\nAll repositories\nYou have unread notifications ( g then n )\nOpen user navigation menu\nRepository navigation\nCode\nIssues\nPull requests  (1)\nPull requests\nActions\nProjects\nSecurity and quality\nInsights\nSettings\nPrivate\nEdit Pins\nWatch: Participating in Meridiona/meridian\nWatch\nFork 0\nFork\nSee your forks of this repository\nStar this repository (0)\nStar\nAdd this repository to a list\nMeridiona/meridian\nmain branch\nmain\n20 Branches\nBranches\n0 Tags\nTags\nGo to file\nAdd file\nFolders and files\nName\nLast commit message\nLast commit date\nLatest commit\nShow 2 authors\npeople\nand\nAditya Harish\ndocs(task-classifier): restructure scoring heuristics with section he…\nOpen commit details\nfailure\nCommit 4f7d3a2\n4f7d3a2\n5 hours ago\nHistory\n261 Commits\n.cargo, (Directory)\n.cargo\nci: add GitHub Actions CI pipeline with fmt, clippy, test, build\n2 weeks ago\n.claude/skills, (Directory)\n.claude/\nskills\nfeat(intelligence): add backfill binaries for session categories and …\n5 days ago\n.githooks, (Directory)\n.githooks\nchore(hooks): parallelize pre-push checks and fix hooksPath\nlast week\n.github/workflows, (Directory)\n.github/\nworkflows\nfix(ui): restore correct page.tsx and remove stale ocr_samples refs\npackages/meridian-mcp, (Directory)\npackages/\nmeridian-mcp\nrefactor(schema): merge ticket_links into app_sessions\nscripts, (Directory)\nscripts\nfeat(scripts): add setup-services.sh for co-dev onboarding\n7 hours ago\nservices, (Directory)\nservices\nsrc, (Directory)\nsrc\nfix(task-linker): add reasoning and session_type to classification de…\nswift, (Directory)\nswift\nfeat(etl): filter OCR noise from session_text at storage time\ntests, (Directory)\ntests\ntest(task-linker): replace stub smoke tests with real hermes integrat…\nui, (Directory)\n.claudeignore, (File)\n.claudeignore\nfeat(etl): split browser sessions by domain, add cat_smoke diagnostic…\n.env.example, (File)\n.env.example\nchore: add .env.example with all supported environment variables\n.gitignore, (File)\n.gitignore\nchore: ignore .claude/settings.json local configuration\n.mcp.json, (File)\n.mcp.json\nfix(mcp): use local node path instead of npx until package is publish…\n.release-please-manifest.json, (File)\n.release-please-manifest.json\nci: add release-please for automated semantic versioning and tagging\nCLAUDE.md, (File)\nCLAUDE.md\ndocs(meridian-agents): cover the tagger pipeline end-to-end\nCarg\n  … (670 more chars)\n\nCANDIDATE TICKETS:\n1. KAN-67  [Feature · Epic: PM Write-back]\n   title: Jira Worklog Write-back\n   description: Automatically submit time spent (from  app_sessions ) as Jira worklogs for sessions that have been confidently mapped to tasks. Includes a human-in-the-loop approval queue for uncertain mappings, and a full undo/correction capability.\\n\\n A…\n\n2. KAN-65  [Feature · Epic: UI Intelligence Dashboard]\n   title: Session-Task Linkage Display\n   description: Show developers which PM task each session was mapped to, along with confidence and routing status. This makes the intelligence engine visible and auditable, and gives users confidence in the system before write-back is enabled.\\n\\n Accepta…\n\n3. KAN-64  [Feature · Epic: Session Intelligence Engine(Session to task categorizer)]\n   title: Session-to-Task Mapping Engine\n   description: Design and implement the pipeline that takes a classified session and produces a confident match to a PM task (or a deliberate skip). The mapping engine runs inline after each ETL cycle and writes results to  ticket_links .\\n\\n Why this mat…\n\n4. KAN-109  [Task]\n   title: LLM benchmarking and prompt evaluation framework for Session-Task classifier\n   description: Set up an offline evaluation harness (DeepEval + local LLm) to compare model quality and prompt variants for the session→task classification pipeline. Goal:  Make model-swap decisions data-driven rather than intuition-based. The results fee…\n\n5. KAN-134  [Task · Epic: Redesign product UI and messaging]\n   title: UI Branding & Messaging Standardization\n   description: Audit and standardize product terminology; create a style guide; implement consistent labeling and messaging across all UI surfaces; and add contextual help and tooltips.\n\n6. KAN-135  [Task · Epic: Redesign product UI and messaging]\n   title: UI Component Redesign & Implementation\n   description: TaskBadge visual design (color palette, states, animations), dashboard layout redesign (widget hierarchy, spacing, collapsible sections), session timeline visualization, and classification confidence feedback UI",
+    "expected_output": "{\"task_key\": \"none\", \"session_type\": \"untracked\", \"reasoning\": \"User is browsing the main repository page of Meridiona/meridian on GitHub. While this is work-related (exploring the codebase/repo), it does not specifically align with the benchmarking in KAN-109 or the reasoning UI implementation in KAN-107.\"}",
+    "additional_metadata": {
+      "session_id": 2514,
+      "app_name": "Google Chrome",
+      "task_method": "hermes_aiagent"
+    }
+  }
+]