Eliteforge Task Progress Tracker

Task Progress Tracker

Goal

For each concrete requirement, keep one markdown task file under the selected task directory and update it as work progresses. The default directory is docs/tasks/. The task file is also the interruption recovery anchor: after every meaningful change, create a git commit and write the commit id into the related step note.

Use This Skill When

• The user wants the agent to make a plan before coding.
• The user wants TODOs, current status, or progress persisted in the repo.
• The user asks to keep updating task state while the agent is working.
• The user wants work to be resumable after interruption.

File Rules

• Default directory: docs/tasks/
• If the user or another active skill gives an explicit task directory, use that directory verbatim and do not also create a duplicate under docs/tasks/.
• Known override: when $eliteforge-google-stitch-onboarding is active, place its tracker files under docs/tasks/google-stitch-onboarding/.
• File name: <task-description>.md
• task-description should be a short kebab-case summary of the requirement.
• Prefer ASCII for task-description; if the request is in Chinese, convert it to a short English slug.
• Do not append fixed suffixes such as -status-process, -task, or -todo unless the user explicitly asks for them.
• If a matching task file already exists, update it instead of creating a duplicate.
• Never overwrite an unrelated task file.
• The task file must stay consistent with the real git checkpoint state of the work.

Allowed Status Values

• notrun
• running
• waiting_human_decision (等待人类决策: current thinking is complete, and work must pause until a human decision is provided)
• finished
• cancel

Git Checkpoint Rules

• Treat each meaningful change as a checkpoint. Before moving to the next change, create a git commit.
• Each step note must record the latest related commit id once a commit exists, using a short format such as commit: abc1234.
• If one step spans multiple commits, append them in order, for example commits: abc1234, def5678.
• Do not leave multiple unrelated edits uncommitted between task file updates.
• A task-file-only bookkeeping update that merely writes back an already-created commit id does not create a new step checkpoint by itself. Carry it in the next work commit, or in one final bookkeeping commit when the task stops.
• If the workspace is not a git repo, or a commit cannot be created, record the blocker in the current step note and update log, then tell the user before continuing more unrecoverable changes.
• When resuming interrupted work, read the latest matching task file, locate the latest recorded commit id in the relevant step note, and use that commit as the recovery baseline before making new changes.

Workflow

1. Before substantial work, create or refresh the task file with:

• requirement summary
• completion conditions
• step list
• initial status: notrun
• initial process: 0%
• step notes prepared to carry commit ids

2. When actual execution starts, immediately update:

• status: running
• process: 5%
• current_step to the first active step
• verify the workspace can create git commits for checkpoint recovery

3. After each meaningful milestone, update the same file:

• acceptance criteria checkboxes and any existing task checkboxes
• step statuses
• current_step
• process
• a short progress log entry
• the related commit id after the commit is created, with the write-back carried by the next work commit or a final bookkeeping commit

4. Before the final reply, update the file one last time:

• successful completion: status: finished, process: 100%, all meaningful changes committed, all completed acceptance criteria checked
• human decision required: status: waiting_human_decision, process kept at the current meaningful percent, and the exact decision needed recorded in the active step note and update log
• user stopped or abandoned task: status: cancel

5. If the task is interrupted, resume by:

• reading the latest matching task file under the selected task directory
• checking current_step, the latest step note, and the latest update log entry
• continuing from the most recent recorded commit id instead of guessing state

Progress Rules

• process must use percent format like 5%, 35%, 100%.
• Keep progress monotonic; do not decrease it unless the user explicitly changes scope.
• Prefer increments that reflect real milestones, not fake precision.
• Do not set 100% until implementation and validation are actually complete.

Required Template

Use this structure unless the repo already has a stronger convention:

# Task: <title>

## Requirement
<short requirement summary>

## Acceptance Criteria
- [ ] <condition 1>
- [ ] <condition 2>

## Overall Status
- status: notrun
- process: 0%
- current_step: not started

## Steps
| step | description | status | note |
| --- | --- | --- | --- |
| 1 | <step 1> | notrun | commit: pending |
| 2 | <step 2> | notrun | commit: pending |

## Update Log
| time | status | process | update |
| --- | --- | --- | --- |
| <timestamp> | notrun | 0% | task initialized |

Update Rules

• The step list is the TODO list. Do not keep a second disconnected TODO section elsewhere.
• Treat markdown checkboxes as live state, not static text. When a criterion or TODO is actually completed, change [ ] to [x] in the same update cycle.
• If work is only partially complete or not yet validated, keep the checkbox unchecked and explain the remaining gap in the step note or update log.
• If a previously completed item becomes invalid because of scope change or regression, revert it to [ ] and record why in the update log.
• Before replying with success, make sure checkbox state matches the real implementation and validation state. Do not leave completed items unchecked.
• current_step should always match one row in ## Steps, or completed.
• When a step reaches finished, its note should contain at least one commit id unless the user explicitly stopped before a checkpoint.
• If more commits are added to an existing step, update that same step note and the update log instead of leaving stale commit information.
• A bookkeeping-only task file edit that records an existing commit id does not require a new step entry; just make sure it is persisted in the next available commit.
• When scope changes, first update ## Requirement, ## Acceptance Criteria, and ## Steps, then continue execution.
• If blocked by technical failure, keep status: running and record the blocker in the step note and update log.
• If analysis is complete and the next action requires a human choice, set status: waiting_human_decision, record the exact decision needed, and stop further execution until the user decides.
• For interruption recovery, make sure the latest step note and update log entry tell the next agent what to continue and which commit to continue from.
• Keep entries concise and factual.

Response Behavior

• Mention the task file path in the working update or final response when helpful.
• If the user asks to continue an existing task, search the selected task directory first and resume from the latest matching file and its latest recorded commit id.
• If work happens outside a git repo or a commit could not be created, state that constraint explicitly in the response.