Agent handoff
Portable evidence-based handoffs for AI agents
npx -y skills add sonu07-star/agent-handoff-skill --skill agent-handoffAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 24 days oldThe repository was created 24 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Create, validate, and resume structured handoff documents that let another AI agent continue work without hidden conversation context. Use when pausing or transferring unfinished work, approaching a context limit, switching agents or sessions, reporting a blocker, recovering after interruption, resuming from HANDOFF.md, or auditing a handoff for evidence, actionable next steps, secret leakage, and unsupported completion claims.
SKILL.md
6.1 KB, ~1.1k tokens by cl100k_base, as published. Nobody here has run it
Agent Handoff
Create a compact, evidence-based continuation packet. Make the handoff useful to an agent that can access the artifacts but cannot see the current conversation.
Preserve the contract
- Treat repository state, files, command output, and external records as evidence. Do not treat memory as evidence.
- Separate verified facts from inferences and unknowns.
- Record the smallest exact next action that can restart progress.
- Preserve user intent, scope, constraints, and outstanding commitments.
- Never copy secrets, credentials, session tokens, private keys, or unnecessary personal data into a handoff.
- Do not label work complete without verification evidence.
- Use
None — <reason>orUnknown — <how to resolve>instead of leaving a required field blank.
Choose an operation
- Create or update: Capture work so another agent or session can continue it.
- Validate: Audit an existing handoff for missing, vague, unsafe, or unsupported claims.
- Resume: Reconstruct the task from a handoff and compare it with live state before acting.
Create or update a handoff
1. Reconstruct live state
Inspect the artifacts directly. Collect only information relevant to continuation:
- Original goal and measurable success criteria
- Authorized scope and explicit constraints
- Workspace, repository, branch, revision, and relevant external identifiers
- Modified, staged, untracked, committed, or published artifacts
- Work completed and work still outstanding
- Decisions, rationale, rejected approaches, and reversal conditions
- Verification commands and their exact summarized results
- Blockers, risks, failed attempts, and unresolved unknowns
- The first action the next agent should take
For a Git workspace, inspect at least the current branch, revision, working-tree status, and relevant diff. Do not run commands that mutate state merely to prepare a handoff.
2. Create the document
Prefer HANDOFF.md in the active project root unless the user or repository specifies another location. Seed the required structure with:
python scripts/create_handoff.py \
--output HANDOFF.md \
--goal "Ship the requested change with passing checks" \
--next-action "Inspect the failing authentication test"
Resolve script paths relative to this skill directory, not the user's current directory. The generator records Git context when available and refuses to overwrite a file unless --force is supplied.
3. Replace every placeholder
Write for a cold reader:
- Use paths, URLs, issue numbers, branch names, and commands instead of phrases such as "the file" or "the test."
- Summarize results; do not paste large logs or diffs. Point to durable artifacts when detail matters.
- Mark each file or artifact with its state and why it matters.
- Record failed attempts only when they prevent repeated work or reveal a constraint.
- Make the first next action atomic and executable. Put later actions in dependency order.
- Include a retrieval method for external state without embedding credentials.
Read references/handoff-format.md when a field is unclear, when adapting the format, or when reviewing status and evidence semantics.
4. Validate before transfer
Run:
python scripts/validate_handoff.py HANDOFF.md --strict
Fix every error. Fix warnings or explain why the exception is accurate in the handoff. Re-run validation after edits.
5. Transfer explicitly
Tell the recipient where the handoff is and state its status. Do not claim that another agent has received or resumed it unless that transfer was verified.
Validate an existing handoff
Run the validator first, then inspect semantic accuracy against live state:
python scripts/validate_handoff.py path/to/HANDOFF.md --strict
Check that:
- The recorded branch, revision, file states, and external identifiers still exist.
- Verification results name a command or durable artifact and distinguish pass, fail, and not run.
- A
completehandoff has checked success criteria and successful evidence. - Unknowns include a concrete resolution method.
- The next action remains valid after any state drift.
- No secret or sensitive data is present.
Report discrepancies; do not silently rewrite the source of truth unless the user asked for an update.
Resume from a handoff
- Read the entire handoff and any directly referenced artifact required for the first action.
- Verify the workspace, branch, revision, working-tree state, and relevant external state.
- Compare live state with the snapshot. Classify differences as expected progress, harmless drift, or a conflict.
- Stop and surface a conflict when continuing could overwrite work, exceed authorization, or act on stale external state.
- Otherwise execute the first valid next action and continue toward the recorded success criteria.
- Update the handoff after material progress if another transfer is likely. Preserve prior decisions that still constrain the work.
Do not repeat completed work merely because the original conversation is unavailable. Do not assume the handoff grants authority for publishing, deletion, deployment, messaging, purchases, or other consequential actions.
Status vocabulary
Use exactly one status:
in-progress: Work can continue and an actionable next step exists.blocked: Progress requires named user input, permission, dependency, or external state.ready-for-review: Implementation is done and evidence is recorded; review or approval remains.complete: Every success criterion is met and supported by verification evidence.
Prefer ready-for-review over complete when a human decision remains.
What ships with it: 4 files
17.7 KB alongside SKILL.md, 2 of them executable
agents/
- openai.yaml216 B
references/
- handoff-format.md3.2 KB
scripts/
- create_handoff.pyruns5.2 KB
- validate_handoff.pyruns9.1 KB
Gives 0 of the 12 instructions most agent orchestration skills give in ~1.1k tokens
Counted across 742 of the 995 authors here whose files we hold, read 2026-08-07
- Reference existing artifacts by path or URLin 53 of 742, across 25 files
- Run the full test suite after integrating changesin 51 of 742, across 19 files
- Dispatch one agent per independent problem domainin 50 of 742, across 17 files
- Verify fixes do not conflictin 45 of 742, across 13 files
- Include a suggested skills section in the documentin 45 of 742, across 17 files
- Redact sensitive informationin 41 of 742, across 11 files
- Save to the temporary directory of the operating systemin 39 of 742, across 10 files
- Tailor the document to user-provided focus argumentsin 39 of 742, across 9 files
- Spot check agent changes for systematic errorsin 34 of 742, across 7 files
- Write a handoff document summarising the current conversationin 31 of 742, across 6 files
- Assign each agent a specific scopein 23 of 742, across 8 files
- Provide specific scope and clear goalin 23 of 742, across 5 files
Said here and by no other author read
- separate verified facts from inferences and unknowns
- record the smallest exact next action
- never copy secrets into a handoff
- inspect artifacts directly to reconstruct live state
- use explicit paths and identifiers instead of vague phrases
- run the validation script before transfer
Grouped from the skills themselves: near-identical wordings counted once, and counted by distinct author, so one author publishing three of these counts once. Length counted with cl100k_base; the agent that loads this file may tokenize it differently.