working-context

name: working-context description: Track active tasks and discovered issues per-branch, defer non-critical fixes without losing them

Skill: Working Context Management

Persistent working context across conversations using .copilot-tasks.md.

⚠️ CRITICAL — This Gets Overlooked Every Time

You MUST read .copilot-tasks.md at conversation start

Not optional. Not "when relevant." Every single conversation. Use read_file. Then acknowledge what you found:

• "Continuing: " if work in progress
• "You have N discovered issues" if deferred items exist
• "Ready for new work" if no active task

If the Active Task has a "Next step", verify it's still needed: "Last session, next step was X. Still needed?"

Failure tracking does NOT persist in your memory

You will forget patterns after a few prompts. That is why the Failure Tracking table exists. The table is the memory, not you. When a failure occurs:

1. Log to the table FIRST — before fixing, before retrying, before anything
2. Then address the issue

If you fix first and log later, you will forget to log. This has happened repeatedly. The sequence is: recognize → log → fix.

Staleness checks on read

When reading the file, also check:

Activation Context

• Starting a new conversation (ALWAYS)
• Detecting an issue while working on something else
• Developer says "add to TODO", "fix later", "note that"
• Developer asks "what was I working on?" or "what issues?"
• Switching branches, completing tasks, conversation reset

File Format

File: .copilot-tasks.md (repo root, gitignored, local only, never commit)

# Copilot Working Context

## General
Items not tied to a specific branch:
- [ ] Feature idea or future work

### Discovered Issues
- [ ] `path/to/file.py:123` - issue found during other work

### Completed (Recent)
- [x] 2026-01-28: Completed task description

## Failure Tracking
| Pattern | Count | Last seen | Context | Fix when promoted |
| ------- | ----- | --------- | ------- | ----------------- |

---

## Branch: feature/some-feature

### Active Task
What you're currently working on

- Next step: Specific next action

### Deferred (Post-Merge)
- [ ] Issue deferred from PR review

### Completed
- [x] Step that was finished

Key rules:

• General section for backlog + discovered issues not tied to a branch
• Branch sections for branch-specific work only
• No "Current Branch" header — you know what branch you're on from git

Updating Tasks

Starting work

Set Active Task with description and first next step.

Completing a step

Update "Next step" to the next action.

Completing a task

Move to "Completed (Recent)" with date, set Active Task to "None — ready for new work". After 3+ completed items accumulate, offer cleanup.

Discovering Issues During Work

1. Assess: Does it block the current Active Task?
2. If blocking: Explain why, ask before switching
3. If non-blocking: Ask — "I noticed X. Add to TODO or fix now?"
4. Always ask — never auto-switch without confirmation
5. Add to branch section or General > Discovered Issues as appropriate

Branch Switching

1. Save current branch context (Active Task + new Discovered Issues)
2. Load context for new branch
3. Mention active task and issue count on the new branch

PR Review Deferrals

When a review issue is valid but out of scope:

1. In the PR: Comment explaining deferral
2. In .copilot-tasks.md: Add to branch's ### Deferred (Post-Merge)
3. Do NOT resolve unaddressed review threads

After merge

1. Move deferred items to ## General > ### Discovered Issues with [from PR #N] prefix
2. Remove the branch section

Failure Tracking

When to log

• A command/approach failed due to a pattern (not a typo)
• User corrects you for a behavioral pattern — highest-signal failure

Logging sequence (NON-NEGOTIABLE)

1. Recognize failure
2. Log to table NOW — update .copilot-tasks.md
3. THEN fix/retry

Before logging

Check if pattern is already documented in *.instructions.md or relevant skill. If found, don't duplicate — it's already promoted.

Table format

| Pattern | Count | Last seen | Context | Fix when promoted |
| ------- | ----- | --------- | ------- | ----------------- |
| shell-escaping | 2 | 2026-02-02 | PowerShell quotes | Use --body-file |

One row per pattern. Increment count on recurrence. Update Last seen.

Promotion (count = 3)

1. Document fix in appropriate file (*.instructions.md, skill, or copilot-references.md)
2. Remove row from table
3. Announce: "Promoted [pattern] to [destination] after 3 occurrences"

Pruning

User-triggered only ("prune failures" / "reset failure tracking"):

• Prune: Remove rows older than 7 days
• Reset: Clear all rows

Conversation Reset & Context Window

Before summarization, ensure .copilot-tasks.md has:

• Current Active Task and next step
• All Discovered Issues
• Updated Failure Tracking

The file persists — context is not lost.

Minimize context window usage

.copilot-tasks.md is read at conversation start, consuming context window space. Keep it lean:

• Completed items: Prune after acknowledging (keep last 3-5 max)
• Branch sections: Delete after merge — stale branches waste tokens every conversation
• Failure Tracking: Promote at count 3, prune stale rows — don't let the table grow unbounded
• Discovered Issues: If >5, prompt user to prioritize or prune

The file should rarely exceed ~100 lines. If it does, it needs cleanup before adding more.

File Management

• Create immediately when this skill activates and file doesn't exist — don't ask
• Ensure .gitignore includes .copilot-tasks.md
• Clean up branch sections for deleted branches when user asks