Handoffs-tidy — retire handoffs that no longer point at live work
Keep ~/.claude/handoffs/ focused on live threads. Handoffs go stale three ways, and the
common one isn't supersede:
- Superseded — a newer handoff continues the same thread (same branch / topic / same-day re-wrap).
- Done — the work shipped: its PR merged, every referenced bead is closed, the branch landed, or
its Jira ticket is Done. These are usually never re-wrapped, so nothing supersedes them — they
just sit in the picker looking
🟢 liveuntil you notice the work is long gone. - Stale — abandoned: the PR was closed unmerged, or the branch is gone with no merge evidence.
- Old and signal-less — unknown branch liveness with no usable PR signal or resolvable beads. Age only earns an assisted question; it never makes the handoff an automatic archive candidate.
This command finds all four and archives the ones you confirm. A recent handoff is retained even when
finished or stale; superseded rows are the sole immediate exception. Nothing is deleted — archived
files move to ~/.claude/handoffs/archive/ and stay greppable.
Run it ad-hoc whenever the picker feels noisy, or right after a /wrap-up. It is the standalone
twin of /handoffs's opt-in archive step (§3b): same list.sh classification, same archive flow —
shared verbatim via REFERENCE.md — but with no full table, no picker, and no resume step. It only
ever offers candidates; it never touches a live or open-PR row.
Earlier versions only found superseded handoffs (it ran
list.shwith no flags and looked at one field). That's why it rarely found anything — supersede is the narrowest signal. From v0.2.0 it runs the full liveness pass, so finished-but-never-re-wrapped handoffs finally surface.
When to use
- After a
/wrap-upwhose handoff continued an earlier thread, or whose work has now shipped. - Periodically, when
/handoffsshows entries you've moved past — merged PRs, closed beads, dead branches. - Not needed if you keep only one live handoff per topic and archive as you go.
Instructions
MUST use the helper scripts. Never construct ad-hoc
ls/grep/git/gh/bdpipelines against~/.claude/handoffs/— they bypass the per-script permission allowlist and miss the repo-matching and liveness logic. Classification and the archive flow are specified once in~/.agents/skills/handoffs/REFERENCE.md; read it rather than re-deriving the rules here, including §Age-review for old rows with no usable signals.
1. Load the shared spec
Read ~/.agents/skills/handoffs/REFERENCE.md. It defines how to read list.sh's output, classify
each row, and run the archive flow — the same definitions /handoffs uses, so the two never drift.
2. Run the lister with liveness
~/.agents/skills/handoffs/scripts/list.sh --check-branches [--stale-days N]
Forward --stale-days N when the user supplied it; otherwise let the script match the age-review
floor to the recent window (3 days, or 4 on Tuesday). Shorter values are clamped to that window. The
threshold only controls the assisted age-review group and never changes archive-class.
--check-branches is what makes this command capable: it fills branch-state and (when gh is
present) pr-state, so the script can mark merged PRs, landed branches, and closed PRs. Bead-closure
(beads-done, keyed off the **Deliverable:** field when present) and supersede are computed
regardless. See REFERENCE §Run and §Fields for the flag semantics and the 22-field line format.
Degrades cleanly offline (REFERENCE §Run / the failure modes below).
Parse the ---HANDOFFS--- lines and the ---SUMMARY--- counts. For each current-repo row, derive its
Status (REFERENCE §Status) and archive-class (safe / keep / empty — REFERENCE §Fields).
Also note rows with needs-review=Y (step 5b) and needs-age-review=Y (step 5c) — both are
assisted judgement groups, never automatic archive candidates.
3. Resolve Jira-Done (optional)
Follow REFERENCE §Jira-Done for any current-repo row without another done/stale/supersede signal that names a Jira ticket. This catches handoffs whose only finished signal is a closed ticket (no merged PR, no closed beads). It's gated exactly as REFERENCE describes and degrades silently if the Jira MCP isn't configured — skip it freely if you want to stay network-light; PR/bead/branch/supersede classification still stands.
4. Present the candidates
If current_repo_superseded == 0 and current_repo_stale == 0 and §Jira-Done promoted no
older row to safe and no row has needs-review=Y and no row has needs-age-review=Y and no
member row is
archivable (step 5d's gate), report
nothing and stop at step 6 — the picker is already tidy:
_No archivable handoffs — remaining rows are live, unknown, or still inside the recent grace window._
Check step 5d's gate before stopping. In a multi-repo workspace the current repo can be perfectly tidy while its members hold many finished handoffs — member rows deliberately never feed
current_repo_stale. Stopping on the current-repo counts alone reports "nothing to tidy" while archivable handoffs sit one level away, which is precisely the blind spot this command exists to close. When only member candidates exist, skip to step 5d rather than stopping.
If there are auto-classified candidates, render them as a table, grouped by regret (REFERENCE
§Archive-flow defines the groups). When only an assisted flag is present, skip straight to step 5b
for needs-review or step 5c for needs-age-review.
## 🗂️ Archive candidates ({count})
| Archive? | Date | Slug | Branch | Status | Group |
|----------|------|------|--------|--------|-------|
- Status: the §Status glyph for the row.
- Group:
Superseded/Done/Stale. A row that is both superseded and otherwise archivable goes in Superseded (the safest reason to archive). Order the table Superseded → Done → Stale. - Archive?: pre-suggest
✅forsafe(Superseded, Done) and leave☐forkeep(Stale) — Stale rows may be the only record of an abandoned thread.
5. Confirm + archive
Run the archive flow exactly as REFERENCE §Archive-flow specifies: prompt with AskUserQuestion
(multiSelect, one option per candidate, grouped and described per that section), then archive the
selected filenames in one archive.sh call and parse ---ARCHIVED--- / ---SKIPPED---. Pre-check
safe candidates; leave keep unchecked. Surface every ---SKIPPED--- line verbatim with its
reason. Never offer a 🟢 live, 🟠 PR open, or unknown row. archive.sh only moves — never deletes.
✅ Archived {N} handoff(s) to `~/.claude/handoffs/archive/`.
5b. Trunk handoffs worth a look (assisted)
If any current-repo row has needs-review=Y, run the assisted prompt per REFERENCE §Trunk-review
— a separate, clearly-labelled prompt (not mixed into the step-5 groups) for legacy trunk-parked
handoffs the script couldn't auto-classify: partial bead closure (beads-progress like 1/4), no
**Deliverable:** marker, so it can't tell finished own-work from never-closing context. Present the
{beads-progress} and bead list per row and let the user decide; archive any they confirm via the
same archive.sh call. Skip entirely when no row is flagged. This prompt goes quiet on its own as
old handoffs age out and new ones carry **Deliverable:**.
5c. 🕰️ Old handoffs with no completion signal
If any current-repo row still has needs-age-review=Y after §Jira-Done promotion, run the assisted
prompt exactly as REFERENCE §Age-review specifies. Keep it separate from the safe/done/stale and
§Trunk-review groups. Age is not evidence of doneness: leave every option unselected, make skipping
trivial, and archive only filenames the user explicitly confirms. Skip when no row is flagged.
5d. 🧱 Workspace-member handoffs
Run the member archive flow exactly as REFERENCE §Archive-flow-members specifies, after the current-repo steps. In a multi-repo workspace, member repos are where finished handoffs accumulate — the root itself often holds one or two.
Read the member rows from ---WORKSPACE-MEMBER-HANDOFFS--- (24 fields: the usual 22 plus
{member-display}|{member-path}) and the per-repo totals from ---WORKSPACE-MEMBER-REPOS---.
Classify each with the same §Status / §Archive-glyph rules as current-repo rows — member rows are
already classified against their own repo when --check-branches was passed.
Key points from that section, so they aren't missed:
- Gate:
--check-brancheswas passed, and at least one member row isarchive-class=safe; recent non-superseded member rows have an empty class and are not offered. - Only
saferows are offered from here.keep-class member rows (PR closed unmerged, branch gone with no merge evidence) are higher-regret — name them and point atcd {path} && /handoffs-tidyrather than offering them. - One question per member repo, not one per candidate —
AskUserQuestioncaps at 4 options and a single repo can hold more candidates than that. - Archive every selection across repos in one
archive.shcall; it takes bare filenames and is repo-agnostic.
Skip the step entirely when the gate doesn't pass.
6. Done
If nothing was archivable or reviewable (steps 4–5d), or the user selected none, say so plainly and stop. This command never touches live work and never deletes — at worst it's a no-op.
Failure modes
list.sh/archive.sh/REFERENCE.mdmissing (handoffs skill not installed): say so plainly and stop — this command is a thin driver over the handoffs skill's scripts and shared spec.- No
~/.claude/handoffs/directory: nothing to tidy; say so and stop. - Not in a git repo: liveness needs a repo to classify against — say so and stop. (
cdinto the repo and re-run.) - Multi-repo workspace: member repos are covered by step 5d, but only their
saferows. Anything needing judgement (keep-class, trunk-review) still requirescding into that member and re-running. - Offline / remote unreachable:
branch-statedegrades to local-only (mergedstill detected against the local default tip; no falsegone). The Done/Stale groups just shrink. Don't retry. ghmissing, unauthenticated, or timed out:pr-statereportsunknownand classification falls back tobranch-state+beads-done. No error, no retry. Squash-merged branches may show as⚪ branch gone(keep?) rather than✅ merged(safe) unless a closed bead or Jira-Done marks them done.bdmissing or no.beads/:beads-doneis always empty — finished work then relies on PR/branch/supersede (and §Jira-Done) signals alone.- Jira MCP missing or errors: step 3 is skipped silently; PR/bead/supersede classification stands.