manage-state

name	manage-state
description	Creates, reads, and updates the per-workflow agent state file at .github/state/<TICKET-KEY>.md — a fast-access local mirror of workflow context that reduces GitHub API round-trips and enables richer resumption. Use when creating a new workflow, at phase transitions, or when resuming.
argument-hint	operation and ticket key (e.g., create KAN-12, update KAN-12)

Manage State

Creates, reads, and updates the per-workflow agent state file at .github/state/<TICKET-KEY>.md. The state file is the single source of truth for workflow context — the PR body is a rendered view produced from it at phase boundaries.

Schema reference: Block definitions and mutability rules are in .github/state/SCHEMA.md. The template is .github/state/workflow-state.tpl.md.

Content Depth Principle

The state file is the agent's working memory and the authoritative record of workflow context. The PR body is human-facing and deliberately summarized — it is rendered from state, not maintained in parallel. Write each block with enough detail that an agent resuming work in a new session can fully reconstruct context and continue from the current step without re-reading the JIRA ticket, re-running research, re-exploring the codebase, or re-deriving decisions.

Per-Block Depth Guide

Block	Depth	What to include	What to omit
UNDERSTANDING	Full reproduction of ticket requirements in the agent's own words	Every requirement and acceptance criterion from the ticket; edge cases and ambiguities identified during codebase exploration; key file paths and patterns discovered; existing conventions relevant to the task; build/test/lint commands confirmed	Verbatim copy-paste of the entire JIRA description (summarize in your own words instead)
RESEARCH	Full findings with rationale	Each option evaluated and why it was accepted or rejected; library versions, API patterns, algorithm trade-offs; links to docs or references consulted; final recommendation with explicit reasoning	Raw web-page dumps or full API reference pages
PLAN	Mirror the PR body task list, plus implementation notes per task	Task table with status; for each task: which files to create/modify, what pattern to follow, what tests to add, any gotchas discovered during planning; test strategy with exact commands; risk details with concrete mitigation steps	Vague tasks like "implement feature" — be specific enough that the task could be done without re-reading the ticket
IMPLEMENTATION	Running log of what was done and why	Every file created or modified (with brief note on what changed); every test added (with what it covers); any decisions made during implementation that deviated from the plan; command outputs if relevant (e.g., test results)	Full file contents or large diffs (reference the file path instead)
REVIEW	Structured findings and their resolution status	Risk level; each finding with severity, description, and resolution; what was fixed vs. deferred vs. accepted	Full code snippets from the review (summarize the issue and point to the file/line)

Rule of Thumb

After writing a block, ask: "If I lost my entire conversation history and only had this state file, could I continue from the current phase without re-doing any prior phase's work?" If the answer is no, add more detail.

When to Use

• Create: After creating the branch and before creating the draft PR — state file is created first, then create-pull-request renders the initial PR body from it
• Read: At workflow resume to restore context via the resume-workflow skill
• Update (phase transition): At each phase transition; then invoke update-pull-request to re-render the PR body from the updated state
• Update (task progress): After completing each task in Implement phase to keep PLAN block current

Procedure

Create a New State File

1. Copy the template:

   cp .github/state/workflow-state.tpl.md .github/state/<TICKET-KEY>.md
   

2. Populate each block per the Per-Block Depth Guide above. The data source for each block at creation:

   • META: workflow type, ticket key/URL, branch, timestamps (PR number/URL populated after create-pull-request returns)
   • PHASE: Implementing
   • UNDERSTANDING: jira-reader output + codebase exploration
   • RESEARCH: researcher output (leave empty if not invoked)
   • PLAN: planner-lite output + per-task implementation notes
   • PHASE_LOG: first row — "Branch created, state file initialized, entering implementation"

3. Write the populated content using single-quoted Python (never heredocs):

   python3 -c 'content = open(".github/state/workflow-state.tpl.md").read(); # ... populate content ...; open(".github/state/<TICKET-KEY>.md","w").write(content)'
   

4. Commit the state file alongside any other Phase 2 work:

   python3 ./.github/skills/git-operations/scripts/git_helper.py commit "chore(state): initialize workflow state [<TICKET-KEY>]"
   

Read / Parse State File

cat .github/state/<TICKET-KEY>.md

Parse STATE_BLOCK:*:BEGIN/END boundary markers to extract specific sections:

• UNDERSTANDING block → requirements, discovered file paths, conventions, build commands (primary context for resuming Phase 1 work)
• RESEARCH block → options evaluated, recommendation, references (prevents re-running research)
• META block → ticket, branch, PR number for API calls
• PHASE block → current phase for routing
• PLAN block → task table with statuses for progress assessment
• PHASE_LOG block → audit trail for context

If the state file does not exist (older workflow or first resume), fall back to fetching the PR body:

python3 ./.github/skills/create-pull-request/scripts/pr_helper.py fetch-body --pr-number <PR_NUMBER>

Update — Phase Transition

At each phase transition:

1. Update the PHASE block: replace the phase value

2. Update the META block: advance Updated At to current UTC timestamp

3. Update phase-specific MUTABLE blocks per the Per-Block Depth Guide:

   • Entering Reviewing: populate IMPLEMENTATION block
   • Entering Submitting: populate REVIEW block
   • Entering Ready: ensure all MUTABLE blocks are complete and consistent

4. Append a row to the PHASE_LOG block — do not append if the last row already has the same Phase value

5. Commit the updated state file:

   git add .github/state/<TICKET-KEY>.md
   # Include in the phase-transition commit
   

6. Then invoke update-pull-request to re-render the PR body from the updated state. The PR body reflects state; state is never derived from the PR body.

Update — Task Progress

During Phase 3/4, after completing each task:

1. In the PLAN block, update the task's Status column: pending → in-progress → done
2. Update the IMPLEMENTATION block per the depth guide (files modified, tests added, any plan deviations)
3. Stage the state file:

   git add .github/state/<TICKET-KEY>.md
   # Include in the same commit as the code change
   

Archive at Completion

When the pr-author finalizes the PR (moves to Ready), move the state file to the archive folder per the rules in .github/state/SCHEMA.md:

mkdir -p .github/state/archive
git mv .github/state/<TICKET-KEY>.md .github/state/archive/<TICKET-KEY>.md

Include this git mv in the finalization commit alongside the artifact registry update. If the PR is abandoned instead, delete the file with git rm rather than archiving it.

Important

• The state file lives in the working tree and must be committed to git — it is not a temp file
• Never store secrets or credentials in the state file
• If STATE_BLOCK boundary markers are missing or malformed, stop and report — do not proceed
• The state file is the source of truth. The PR body is always derived from it — never the reverse