Add Cowork Task Route skill

submissions/cowork-task-route/README.md

@@ -0,0 +1,23 @@
+# Cowork Task Route
+
+Cowork Task Route gives users a clear, ordered explanation of how Cowork completed
+a task: which skills, agents, and tools were invoked, what each step did, and what
+artifacts or outcomes were produced.
+
+## How to use it
+
+Ask Cowork questions like:
+
+- "Show me the task route."
+- "Which skills and tools did you use?"
+- "How did you complete that task?"
+- "Trace what Cowork did."
+
+The skill responds with a concise route and a Task Details table covering the
+model label, tool count, observable context size, and runtime environment.
+
+## Good to know
+
+This skill reports only what actually happened in the current session. It does
+not fabricate missing steps, rerun actions, expose sensitive payloads, or provide
+a raw debug transcript.

submissions/cowork-task-route/SKILL.md

@@ -0,0 +1,161 @@
+---
+name: cowork-task-route
+description: |
+  Reports the task route: the skills, subagents, and tools Cowork invoked to
+  complete the current task, as an ordered plain-language trail. Use this skill
+  when the user asks what was used to do something, asks to show the task route,
+  asks what was invoked to complete a task, asks which skills and tools were used,
+  asks how Cowork completed the task, asks to trace what Cowork did, asks to break
+  down the steps taken, or asks to show the work. Reconstructs the invocation
+  trail from the current session's real actions, groups it by skills, agents, and
+  tools, and presents any produced artifacts plus a task-details summary table.
+  Do not use this skill to debug a failed session, render the full raw transcript,
+  answer general how-to questions, or explain a task Cowork has not actually
+  performed in this session.
+cowork:
+  category: productivity
+  icon: Flowchart
+---
+
+# Cowork Task Route
+
+Show the user, in plain language, what Cowork actually invoked: which skills,
+subagents, and tools were used to complete the task they are asking about. This is
+a transparency skill: it traces the routing decisions of the current session and
+presents them as a clean, ordered route the user can follow.
+
+## When to use
+
+Use this skill when the user wants visibility into how a task was carried out:
+
+- "What did you use to do that?"
+- "Show me the task route."
+- "Which skills and tools did you use?"
+- "How did you complete that task? Walk me through the steps."
+- "Trace what Cowork did."
+- "Show your work."
+- "Break down the steps you took."
+
+## When not to use
+
+- Debugging a failed session or wanting the full raw transcript. This skill gives
+  a concise route, not a byte-level trace.
+- General how-to questions, such as "How would you build a deck?" That is advice
+  about a hypothetical task, not a trace of real work. Answer directly instead.
+- A task Cowork never performed in this session. Do not invent a route. If there
+  is no real invocation history for the referenced task, say so.
+- Explaining why a result is what it is. Content questions are about the output,
+  not the routing.
+
+## Scope: what "current task" means
+
+Default to the most recent completed user request in this session. If the session
+contains several distinct tasks and the reference is ambiguous, such as "that" or
+"the last thing," briefly confirm which task the user means. List candidate tasks
+by their user-facing descriptions and do not guess across unrelated tasks.
+
+## Workflow
+
+1. Identify the target task. Pin down which request the user means, using the most
+   recent completed request by default. Note its user-facing goal in one line.
+
+2. Reconstruct the invocation trail from the current session in the order things
+   happened. Gather three layers:
+   - Skills invoked via the Skill tool.
+   - Subagents spawned via the Agent tool, including the sub-task each was given.
+   - Tools called, grouped by purpose: searched email, read a file, created a
+     calendar event, generated an image, wrote an output file, posted to Teams,
+     ran a script, or similar.
+
+3. Classify each step so the route reads cleanly:
+   - What capability was used: skill, agent, or direct tool.
+   - What it did, in business language.
+   - What it produced or changed, naming artifacts exactly when applicable.
+
+4. Compile the task details for the summary table. Source each value from real
+   session context. Never guess:
+   - Model used: always report the assistant as "Microsoft Copilot (powered by
+     Microsoft AI technology)." Never expose or imply an underlying third-party
+     model name, vendor, or model ID.
+   - Tools invoked: count the tool calls actually made while completing this task,
+     including Skill and Agent invocations. Give the total and a short by-purpose
+     breakdown if useful.
+   - Context size: report the best measure you can genuinely observe, such as the
+     number of turns or messages exchanged, or an explicitly labelled approximate
+     token figure if one is available. If no reliable figure exists, write
+     "not precisely measurable."
+   - Runtime: describe the environment in business terms, such as
+     "Microsoft 365 Copilot (Cowork) - secure cloud workspace." Never surface
+     session IDs, raw file paths, internal hostnames, or other low-level
+     identifiers.
+
+5. Present the route. Lead with a one-line summary of the task, then the ordered
+   steps, then produced artifacts or outcomes, and finish with the Task Details
+   table. Offer to expand into granular tool-level detail if the user asks for it.
+
+## Output format
+
+Default to a compact ordered list in chat:
+
+```markdown
+Task: <one-line description of what the user asked for>
+
+Route:
+1. <Skill/Agent/Tool> - <what it did, plain language> -> <artifact/result, if any>
+2. ...
+3. ...
+
+Produced: <artifacts/outcomes, each named exactly>
+```
+
+- Name skills and subagents plainly because they are user-facing named
+  capabilities, such as "Used the pptx skill" or "Ran the deep-research agent."
+- Describe tool activity in business language, such as "searched your email,"
+  "checked your calendar," or "saved a file for you." Do not show raw tool
+  identifiers unless the user explicitly asks for tool-level names.
+- For a route with four or more ordered steps or three or more distinct
+  capabilities, render it as a numbered table with columns for Step, Capability,
+  What it did, and Result. For a short route of one to three steps, plain text is
+  better.
+- Keep it scannable. This is a map of the work, not a re-explanation of the
+  result.
+
+## Task Details
+
+Always include a Task Details table with exactly these four rows, in this order:
+
+| Detail | Value |
+|--------|-------|
+| Model used | Microsoft Copilot (powered by Microsoft AI technology) |
+| Tools invoked | `<total count>` (`<optional by-purpose breakdown>`) |
+| Context size | `<turns/messages exchanged, or approximate tokens if available, else "not precisely measurable">` |
+| Runtime | `<business-language environment, such as "Microsoft 365 Copilot (Cowork) - secure cloud workspace">` |
+
+- Render the Task Details as a Markdown table by default.
+- Fill every cell from real session context. If a value genuinely cannot be
+  determined, write what can be observed and mark the rest as unavailable.
+- Never leave a cell blank and never fabricate a value.
+
+## Guardrails
+
+- Only report what actually happened. Reconstruct the route from real session
+  actions: skill invocations, agent spawns, tool calls, and the task tracker.
+  Never fabricate steps, tools, or artifacts to make the route look complete. If
+  part of the trail is unclear, say which part cannot be reconstructed.
+- If the referenced task was not performed in this session, or the session has no
+  prior actions, tell the user plainly: "I don't have a record of carrying out
+  that task in this session," and offer to run it now.
+- Honor privacy and identity rules. Describe tool activity without leaking
+  sensitive payloads such as message bodies, private event titles, credentials,
+  folder paths, or session IDs. Present tool activity by purpose. Surface raw tool
+  names only if the user explicitly asks for that level of detail.
+- The "Model used" cell is fixed by policy. It is always "Microsoft Copilot
+  (powered by Microsoft AI technology)" and never a third-party model name,
+  vendor, version, or model ID.
+- Never fabricate metrics. Tool count, context size, and runtime must come from
+  what can actually be observed in the session. If context size is not
+  measurable, say so plainly rather than estimating a precise-looking number.
+- Do not duplicate side effects. This skill only reports; it never re-runs the
+  actions it is describing.
+- Verify artifacts before naming them as delivered. If you claim a file was
+  produced, confirm it exists before presenting it in the Produced line.

submissions/cowork-task-route/metadata.json

@@ -0,0 +1,10 @@
+{
+  "name": "Cowork Task Route",
+  "description": "Show the skills, agents, and tools Cowork used to complete the current task as a concise ordered route.",
+  "platforms": ["Cowork"],
+  "tags": ["cowork", "productivity", "transparency"],
+  "author": "tanaveed",
+  "version": "1.0.0",
+  "createdAt": "2026-07-27",
+  "updatedAt": "2026-07-27"
+}