Add /continue reopen mode to recover /exit-closed herdr tabs by gering · Pull Request #22 · gering/claude-plugins

gering · 2026-07-02T12:13:15Z

Summary

A kickoff tab runs Claude as its root pane, so a bare /exit (even one only meant to restart Claude Code) closes the whole tab — the worktree and resumable session persist, but you lose your place.
Adds a reopen path: /continue <task> from the main session reopens the task's herdr tab at its worktree and resumes the existing Claude session (claude -c), then focuses it.
Reopened tabs are hardened against /exit: Claude runs inside a shell pane, so a later /exit drops back to the shell and the tab survives.

Changes

scripts/herdr-launch.sh — split into two subcommands sharing one precondition/parse/output contract: launch (kickoff, argv-exec, root-pane, unchanged) and resume (tab create + pane run "claude -c" in a shell pane, then tab focus). Same pane=/tab=/moved= contract.
skills/continue/SKILL.md — branch by location: inside a worktree = in-session resume (unchanged); main session + <task> = reopen via the helper; outside herdr = manual cd <worktree> && claude -c block. Description updated for the new mode.
skills/kickoff/SKILL.md — call the helper's new launch subcommand.
Design decisions: keep kickoff argv-exec (race-freedom is verified, and /close's teardown depends on the root-pane model); resume uses claude -c (worktree cwd is a 1:1 proxy for the task session, so no session-id stash is needed). A race-free prevention variant is documented as deferred pending live herdr verification.
Docs: bump work-system 1.5.1 → 1.6.0; update root + plugin READMEs and the herdr knowledge entry (+ index).

Readiness

✅ README freshness — root + plugin README updated
✅ Version bump — 1.6.0 (plugin.json + marketplace.json in sync)
➖ Changelog — N/A (no changelog in repo)
✅ Knowledge — herdr entry + index updated
✅ Structure check — 0 errors (bash -n, JSON, version sync, refs, description budget)
✅ Rebase — 0 commits behind main

Test plan

Inside herdr: /exit a kickoff tab, then /continue <task> from the main session → tab reopens at the worktree, session resumes, tab is focused.
In the reopened tab, /exit again → drops to the shell, tab survives.
Outside herdr: /continue <task> prints the manual cd <worktree> && claude -c block.
Inside a worktree: /continue still does in-session resume unchanged.

🤖 Generated with Claude Code

A kickoff tab runs Claude as its root pane, so a bare /exit (even to just restart Claude Code) closes the whole tab. Recover it with `/continue <task>` from the main session. - herdr-launch.sh: split into `launch` (kickoff, argv-exec, unchanged) and `resume` (tab create + `pane run "claude -c"` inside a SHELL pane, then focus). Same pane=/tab=/moved= contract. Because resume runs Claude in a shell pane, a later /exit drops back to the shell and the tab survives. - continue/SKILL.md: branch by location — worktree = in-session resume (unchanged); main session + <task> = reopen via the helper; outside herdr = manual `cd <worktree> && claude -c` block. Update description. - kickoff/SKILL.md: call the helper's new `launch` subcommand. - Decisions: keep kickoff argv (race-freedom + /close teardown depend on root-pane Claude); resume uses `claude -c` (worktree cwd is 1:1 with the task session, so no session-id stash needed). - Bump work-system 1.5.1 -> 1.6.0; update READMEs + knowledge entry. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AQaQpf2Y8hn2aUV6SuLfFF

- Guard against a duplicate session: resume now reuses herdr-teardown.sh's worktree-tab (cwd→tab) lookup and, if a tab is already open at the worktree, just focuses it (reused=yes) instead of starting a second `claude -c` on the same working tree. - Report resumed=no when the `pane run "claude -c"` send fails, so the skill never claims a resume that didn't happen; SKILL branches on reused/resumed and softens the success wording (adopted worktrees may have no session to resume). - Route `/continue <other-task>` from inside a worktree to reopen mode for that task instead of silently ignoring the argument. - Parse the tab-create response in one python3 pass (was two spawns). - Restore the "pick up where I left off" trigger phrase; tighten the description to stay within budget. - Trim the over-long knowledge index line back to a terse pointer. - Document the resume-vs-/close teardown asymmetry (shell-launched Claude may not be a registered agent; /close falls back to its manual-close line) as a known, gracefully-degrading limitation. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AQaQpf2Y8hn2aUV6SuLfFF

Address xhigh review findings on the duplicate-session guard added last round: - Retry the worktree-tab lookup (~3x): it returns empty for both "no tab" and a transiently empty `herdr pane list` (e.g. just after a herdr restart), so a single empty read failed open and could still duplicate. - Don't assert a live resume on reuse: a cwd match can't tell a live Claude from a bare shell that survived a prior /exit, so the reuse branch emits `resumed=` (empty) and the skill tells the user to run `claude -c` if the focused tab is just a shell — no false "resumed". - Parse the tab-create response pipe-delimited (`<pane>|<tab>`) so an empty pane id can't be mis-read as the tab id (would have targeted a tab id as a pane id). - Emit `focused=<yes|no>`; skill no longer claims a focus that failed, and reorders its branches so an empty `tab=` with resumed=yes still gets the manual-focus hint. - Document the workspace-scoped lookup as an accepted limitation (a same-worktree tab in another herdr workspace is out of scope), consistent with /kickoff and /close. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AQaQpf2Y8hn2aUV6SuLfFF

Address xhigh round-2 findings; the duplicate-session guard is redesigned to fail closed instead of open: - Add `herdr-teardown.sh worktree-tab-state`: a tri-state cwd->tab lookup returning <tab>|none|unverified. resume now creates only on `none` (populated pane list, no match), FAILS CLOSED on `unverified` (herdr unreachable or an empty/repopulating pane list) by exiting 1 so the skill shows the manual block — a plain empty-return guard fails open and could still duplicate. This also drops the retry-with-sleep, so the common no-tab path pays no latency. - Search all workspaces (empty workspace arg) so a still-live tab for the same worktree in a different herdr workspace is found; the new tab is still created in $HERDR_WORKSPACE_ID. - Fix a dispatch regression: inside a worktree, an argument that doesn't resolve to another task's worktree (typo/alias) now resumes the current task in-session instead of refusing with "no worktree — /kickoff". - Header said "two keys", lists three (reused/resumed/focused). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AQaQpf2Y8hn2aUV6SuLfFF

Round-3 review returned only PLAUSIBLE findings; two were already- documented, gracefully-degrading tradeoffs (the shell-startup keystroke race and reopened tabs not being registered agents). Addressed: - Remove the duplicated norm()/cwd-match logic: worktree-tab-state now reuses the canonical extract_tab matcher plus a trivial extract_panes_populated check, so path/realpath matching is defined once and the reopen guard and /close's worktree-tab lookup can't drift. Guard an empty target as unverified (fail closed). - Word the reopen success message as "Sent `claude -c`", not "is running" — the helper delivers the keystrokes but can't confirm Claude came up (shell-startup race); tell the user to re-run it if they land on a bare shell. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AQaQpf2Y8hn2aUV6SuLfFF

Round-4 findings centered on the mode-selection prose (which had drifted across edits). Make it mechanical: - Normalize the argument through task-status.sh before comparing/looking up the worktree, so a `task/`-prefixed or aliased name matches the prefix-stripped worktree dir name instead of misrouting to in-session resume of the wrong task. - Compare the resolved arg name to the current worktree's task name first, so `/continue <current-task>` from its own worktree resumes in place rather than no-op-focusing the tab you're already in. - Mode-aware usage string: resume no longer advertises [session-name] (only launch takes one). - Knowledge: resume is called from continue's reopen path (main session, OR a different task's name from inside a worktree) — not "main-session only"; document the exact-cwd match gap (a shell wandered into a subdir isn't matched) as an accepted limit of the shared matcher. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AQaQpf2Y8hn2aUV6SuLfFF

Final review-loop round. Fixes the four surviving guard-path findings: - unverified now surfaces as `blocked=unverified` (exit 0), and the skill tells the user to CHECK herdr for an already-open tab before reopening by hand. Previously the fail-closed exit fell through to the plain manual block ("cd && claude -c") with no cue, which could recreate the duplicate the guard exists to prevent. - Collapse worktree-tab-state into ONE python pass (was two: populated check + extract_tab), removing the double parse. - The single tri-state extractor now fails closed (`unverified`) on a malformed/errored pane list AND when a pane sits in a SUBDIRECTORY of the worktree (a tab may have wandered in) — both previously read as `none` and could duplicate. norm() mirrors extract_tab (kept in sync by comment) since /close's exact-match matcher must not inherit the subtree/tri-state behavior. Unit-tested the extractor across exact/none/subtree/empty/malformed/ sibling cwd cases. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AQaQpf2Y8hn2aUV6SuLfFF

Round 6 verified the round-5 fixes and caught a regression they introduced. Corrections: - Revert the subtree-cwd matching: it was deterministically failing closed whenever ANY unrelated pane sat in a worktree subdir (e.g. a shell in <worktree>/logs), permanently blocking auto-reopen. Back to exact-match; the wandered-own-tab gap it targeted is re-documented as an accepted narrow limitation. - Fail closed on a tab pane whose cwd is empty/unreadable: it was being skipped and falling through to `none` (fail open -> possible duplicate claude -c). Now yields `unverified`. - blocked outcome now prints a lone `blocked=unverified`, matching the header contract (was three keys). - Factor norm() into a shared prelude string concatenated into both extract_tab and extract_tab_state, so path normalization is defined once (the two can no longer drift). - Document that reopening a different task from inside a worktree can place its tab in the current workspace, which a later /close may not locate (falls back to its manual-close line; no data loss). Unit-tested both extractors across exact/none/subtree/empty/null/ non-string/no-tab cwd cases. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AQaQpf2Y8hn2aUV6SuLfFF

Round 7 (verification of round 6). Its top finding was self-inflicted: the round-6 subtree revert changed the code but left comments claiming subdir -> unverified. Fixes: - Sweep the stale subtree=fail-closed comments (herdr-launch header + teardown) to match the code: subdir cwd -> `none` (accepted narrow gap), documented as such. - extract_tab_state: check cwd BEFORE tab_id, so a worktree-cwd pane momentarily missing its tab_id fails closed (`unverified`) instead of being skipped -> `none` (a fail-open duplicate risk). - resume sends `cd <worktree> && claude -c` (shell-quoted): the pane is created with --cwd, but an rc that drifts the cwd (direnv/zoxide) would make `claude -c` resume the wrong task's session. Re-anchor first. - Factor the herdr-down/empty-list guard chain into a `pane_query` helper shared by tab_status() and worktree-tab-state (were duplicated). - /continue reopen step 1 reuses the ARG_TASK the dispatcher already resolved instead of re-running task-status.sh. - Correct the "all workspaces ... globally unique" overclaim: the search spans the current herdr SERVER only (a separate server is invisible); document that and the unreadable-cwd over-block as accepted fail-safe. Re-unit-tested both extractors (8 + 3 cases); structure check green. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AQaQpf2Y8hn2aUV6SuLfFF

- continue/SKILL.md: correct the blocked=unverified cause list — a subdir pane now yields `none` (creates), not `unverified` (a stale claim the R7 sweep missed in the skill, matching the exact-match/subtree-revert design) - continue/SKILL.md: the resumed=yes template no longer tells a never-hosted worktree to re-run `claude -c` (which errors); start fresh `claude` instead - continue/SKILL.md: on reuse focused=no, cue reopening by hand if the tab closed between snapshot and focus (no auto-create — that risks a duplicate) - continue/SKILL.md: defer MAIN_REPO off the hot in-session-resume path; resolve it only in the linked+switch branch that needs it - herdr-launch.sh: close an orphaned tab (parsed tab id, no pane id) before bailing, so a schema-drift response can't leak a blank tab on every resume - Knowledge: note the unscoped-pane-list-spans-workspaces assumption + the orphan-close; bash -n + structure check green Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AQaQpf2Y8hn2aUV6SuLfFF

Second dogfooding round at the deepest profile (/swarm:review --max --fix: per-lens split live — gate pruned style, 10 xhigh finders + codex gpt-5.6-sol@xhigh + grok-4.5 + composer; 24 findings, 10 consensus). Fixes the agreed findings: - Never auto-accept an all-untagged consensus cluster: no tagged lens backs it, so it is verified like a solo. Verify/auto-accept now derive from ONE needsVerify predicate (structural exactly-once partition); kind vote in clearer every-form (#1, consensus critical; #21, #22) - Design verifier sees the finding's recommendation — the proposal the applicability rubric actually tests — and carries an escape hatch: a genuine defect mis-filed under a design lens is not refuted away (#12, #3); "solo" dropped from verifier prompts (#4) - Untagged findings from multi-lens cluster finders fall back to 'unspecified' (safe defect bucket), not lenses[0] (#7); merge-agent free-text lens validated, majority-member fallback (#9) - Improvement invitation scoped to design finder units — defect-lens finders stay defect-only (#8); merge prompt clusters by issue, not only defect (#13); schema descriptions generalized for design findings (#14) - LENS_BRIEF startup assertion (#6); gate prompt interpolates LENS_CLUSTERS.design (#19); new test_lens_sync.py guards all lens mirrors: SKILL HDR prompt, LENS_BRIEF, pr-post DESIGN_LENSES (#5) - Workflow assigns stable finding num (defects first, shared sequence); presenter/pr-post render it verbatim (#20) - pr-post.py: design lens is the backup kind signal when the handoff drops kind (explicit defect still wins); single-pass partition (#17, #24) - Balance: REFUTED is its own segment (refuted ⊄ solo since design consensus can be refuted) (#2); LOCKED design-table column precedence in --loop rounds clarified (#11); "verifies solos" doc sweep across manifests/README/knowledge (#4); cluster failure-isolation trade-off documented (#10); knowledge index line trimmed (#15) Declined: #16 (JS test harness for sandbox code; lens-sync test covers the drift class), #18 (gate-fail under --max runs all lenses — the documented never-silently-narrower degrade), #23 (micro-opt vs readability). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0126cGxsBYxqgEFH4AcpriNk

gering and others added 10 commits July 2, 2026 14:06

gering merged commit 8a7e1a5 into main Jul 2, 2026
1 check passed

gering deleted the task/continue-reopen-herdr branch July 2, 2026 20:45

Navigation Menu

Add /continue reopen mode to recover /exit-closed herdr tabs - #22

Add /continue reopen mode to recover /exit-closed herdr tabs#22
gering merged 10 commits into
mainfrom
task/continue-reopen-herdr

gering commented Jul 2, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

gering commented Jul 2, 2026

Summary

Changes

Readiness

Test plan

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant