handon

name	handon
description	Use at the start of a session to orient to outstanding work — scans for HANDOFF.yaml (or HANDOFF.md) files, triages items by priority, and acts according to risk level without asking for approval on P1/P2 work.
model	haiku
effort	low
allowed-tools	["Read","Bash"]

handon — Session-Start Handoff Reader

Overview

Scan the current directory tree for handoff files, parse items by priority, and act:

Priority	Action
P0	Validate current state immediately. Report to user. Ask before touching anything.
P1	Execute autonomously. Stop only if scope expands or something unexpected happens.
P2	Delegate to subagents. Cap at 5 concurrent.

handon should use HANDOFF YAML plus handoff-db/SQLite only. Do not call doob here; valerie is the only skill that should touch doob or GitHub issue sync. Treat items as transient open-work context and log as durable completed-work history.

Steps

0. Load local config

Look for .ctx/handoff.<repo>.config.toml in the repo root (where <repo> is the directory name). If found, read it and make all [vars] entries available as session variables for the remainder of this skill's execution. If absent, continue — config is optional.

If the file does not exist but .ctx/handoff.<repo>.config.toml.example does, note it once: "No local config found. Copy .ctx/handoff.<repo>.config.toml.example to .ctx/handoff.<repo>.config.toml and fill in your values."

Do not error or stop if config is missing.

1. Find handoff file

handoff-detect          # returns path if exists, expected path + exit 2 if not

• Exit 0 → file exists, path printed — read it
• Exit 2 → file missing, expected path printed — offer to create via /atelier:handoff
• Exit 1 → not in a git repo — report and stop

If handoff-detect is not on PATH, fall back to globbing .ctx/ for HANDOFF.*.yaml.

If invoked from a workspace root (e.g. ~/dev) with no .git, sweep subdirs for HANDOFF.*.*.yaml files instead.

If only a legacy HANDOFF.md exists at repo root, read it as freeform. Do not convert unless asked.

2. Run preflight script if present

Before reading state and HANDOFF files manually, check for a .ctx/scripts/preflight.rs at the repo root:

test -f .ctx/scripts/preflight.rs && rxx .ctx/scripts/preflight.rs

If it exists and rxx is on PATH, run it. Its output covers branch, git status, recent history, tracked files, and full .ctx/HANDOFF* content — skip steps 2a and 3 entirely and proceed directly to step 4 using the displayed HANDOFF content.

If the script is absent or rxx is not on PATH, continue with the manual steps below.

2a. Read .ctx/HANDOFF...state.yaml

After locating the HANDOFF file, derive the state file path by appending .state before .yaml (e.g. HANDOFF.atelier.atelier.yaml → HANDOFF.atelier.atelier.state.yaml). Read it if it exists. Extract branch, build, tests, notes, touched_files and surface them in the triage header:

## Handoff Triage — <path/to/repo>

Branch: <branch> | Build: <build> | Tests: <tests>
<notes if non-null>
Recently touched: <touched_files as comma-separated list, omit if absent>

If the file is absent, omit the state line rather than guessing.

3. Pull latest state from SQLite

Before parsing the local file, check the local SQLite database for status overrides written outside this session (e.g. by another tool or a manual update):

handoff-db query --project <project> 2>/dev/null

For each row returned, if the SQLite status differs from the YAML status, prefer SQLite and update the in-memory copy before triaging. Do not write back to HANDOFF.yaml here — that happens in step 9.

If the script is not found or sqlite3 is not on PATH, skip and continue with the local file.

4. Review on wake

Before triaging by priority, scan all items for unreviewed human-edit entries — any extra entry with type: human-edit and no reviewed field (or reviewed is absent).

Surface these first, regardless of item priority:

## Review on Wake

- [id] "[title]" — human edited `<field>` → `<value>` on <date>
  <note if present>

Do not act on these items automatically. Present to user and wait for acknowledgement before proceeding to P0 triage. After the user acknowledges, note which items were reviewed — /atelier:handoff will stamp reviewed: <today> on those entries at session end.

5. Parse items

From HANDOFF.yaml: read items list directly. Filter to status: open or status: blocked. Apply any SQLite overrides from step 3 before triaging.

From HANDOFF.md: read the "Known Gaps", "Next Up", "Parked", or "Remaining Work" sections. Infer priority:

• P0: "broken", "fails", "blocked", "urgent", "security"
• P1: specific file + known fix mentioned
• P2: everything else that's safe

6. Triage P0 items

For each P0:

1. Run relevant validation (cargo check, git status, test run)
2. Report finding to user with current state
3. Ask for go/no-go before acting

Do not proceed to P1/P2 until all P0s are acknowledged by user.

7. Execute P1 items

Work through each open P1 without asking. Stop and surface to user when:

• Scope expands beyond what the item described
• Tests fail unexpectedly (not the known failure)
• More than 3 files need changing beyond what was described
• Any destructive operation would be needed

8. Delegate P2 items

Dispatch one subagent per P2 item (cap 5 concurrent). Each subagent must:

• Receive explicit --allowedTools list
• Verify git status is clean before starting
• Commit its own changes
• Report back with result

9. Report and update

After all work:

## Handoff Triage — <path/to/repo>

P0:
  - [id] [name] "[title]" → [state found] → [action / question]

P1:
  - [id] [name] "[title]" → done | blocked: <reason>

P2:
  - [id] [name] "[title]" → delegated | skipped: <reason>

Then update HANDOFF.yaml:

• Remove any item that was completed or closed upstream so items only carry open context
• Add log entry for this session (one-liner, prepend to list)
• Upsert remaining open/blocked items to SQLite via handoff-db.sh upsert (see handoff skill step 6)
• Commit: git add .ctx/HANDOFF.<project>.*.yaml && git commit -m "docs: update handoff"

Edge Cases

No handoff file found: Report "No HANDOFF.yaml found in <path>." Offer to create one via /atelier:handoff.

HANDOFF.md only: Read it, triage as normal, note at end: "Consider migrating to HANDOFF.yaml for structured triage."

All items done or parked: Report clean state, no action needed. Closed items should be pruned from items on the next write; log remains as history.

Blocked item: Do not attempt. Report the blocker to user verbatim from the description field.

Additional Resources

• references/triage-patterns.md — priority classification guide, P0 validation sequence, P1 autonomous execution rules, P2 subagent dispatch template, multi-repo sweep, human-edit review patterns, SQLite conflict resolution, blocked item handling