v0.11.1

Fixed

• A tab no longer repeats a workspace whose name is the same as its directory's, spelled with a different separator. herdr's worktree manager labels a workspace after the branch, slashes and all, while the worktree directory the branch is checked out in gets those slashes flattened to hyphens. The two compares that ask whether a name is already on screen read them character for character, so feature/fh-10390-frame-tests and feature-fh-10390-frame-tests did not match, and the tab put the directory back in its label under a workspace already carrying it: FH-10390 › claude, the name twice. -, _, . and / now fold together in both compares, alongside the ASCII case fold that was already there. A separator is still required where the rule wanted one, so a tab in legacy-api under an api workspace still says where it is.

v0.11.0

Added

• Workspace labels can be shown under a shorter name than their directory carries. WORKSPACE_SUBSTITUTE_SETS is an ordered list of sed -E rewrites applied to the name herdr derives from the directory, so 's|^worktree-|wt-|' puts worktree-feature in the sidebar as wt-feature. The rewrite reaches the label and nothing else: the directory and the Git worktree keep their names, and a tab inside that directory still drops the workspace's name out of its own label rather than re-injecting the long spelling it was shortened to lose.

  Only a derived name is rewritten. A name you typed is left alone for good, the same promise the tab opt-out makes, with the same gap in it: herdr exposes no way to tell a typed name from the derivation it happens to match exactly.

  Numbering and rewriting are the two reasons the workspace pass runs, so a rewrite reaches a workspace whose numbering was never turned on. A rewrite is derived fresh every pass rather than built out of the label the last one wrote, so deleting the rules puts the derived names back at the next event by itself. With numbering off there is no next pass to do that, and the labels keep their last rewrite until clear.

  clear hands back the derived name too, where it used to strip the number off the rewrite and leave that standing. It is documented as the last step before uninstall, after which the plugin that could have restored the label is gone.

  Based on #15 by @engineersamuel.

v0.10.0

Added

• One command installs the plugin and the shell hook: curl -fsSL .../install.sh | bash. The hook is what makes a rename land the moment a command starts, and wiring it was a per-shell copy-paste out of the README, so anyone who read past step 1 got the numbering and none of the naming until they matched their shell to the right snippet. install.sh reads $SHELL, writes the zsh, bash, or fish snippet to that shell's startup file, and installs the plugin first when herdr does not already list it.

  A marker comment in the startup file is what makes a re-run change nothing, so the command doubles as the upgrade path. The snippet it writes globs herdr's managed plugin directory, whose name carries a version hash, and spells the home directory as $HOME so the startup file survives being carried to another machine. A shell argument (bash install.sh fish) wires a second shell and HAR_RC names a startup file other than the default, which is what a macOS login shell reading ~/.bash_profile needs. Run from a clone rather than curl'd, the hook points at the checkout and the plugin half is left alone, since herdr owns that copy.

  The two steps it will not take behind your back, turning off herdr's new-tab name prompt and installing the agent integrations, it prints when it finishes. The manual snippets are still in the README, folded away.

• A doctor action. It runs one real naming pass with tracing on and prints, for the current tab, the versions and paths in play, the tab's ownership record, the label it carries, and every decision that pass made about it. Until now every failure looked the same from outside: nothing happened. A tab that opted out after a hand rename, a placeholder deferred, a title refused, a lock held by a dead process, and a jq that was never installed all produced the same silence, and the issues filed against this plugin were mostly the reporter reasoning backwards from that silence. herdr plugin action invoke herdr-automatic-rename.doctor runs it, and the README has a Troubleshooting section keyed to what the tab bar shows.

  AR_TRACE=1 in herdr's environment writes the same decisions for every pass to trace.log in the state directory, at 0600 since a label can carry a task title.

Fixed

• Ownership records survive a pass that could not read everything. A tab whose record goes missing reads as renamed by hand on the next pass and opts out of naming for good, so every path that dropped one without cause was a permanent bug. A workspace whose tab list failed to read had every tab pruned. An empty workspace keep list pruned every workspace record. A store with one hand-edited key emptied every record after it. A jq that crashed read as a broken file and healed a good store to nothing, and on Debian's jq 1.6 a truncated file was never healed at all, since that release reports a parse error with a different status. Each of those now leaves the records alone, and the store is rewritten only when a record actually changed rather than on every event.

• The shell hook scrubs the workspace name it derives from $PWD. A directory name holding a control character reached herdr workspace rename raw. herdr handed the label back normalized, which read as a name somebody typed and opted the workspace out of directory tracking, with no reset to bring it back. The tab half already scrubbed for this and the workspace half now does the same.

• agent_panel_sort = "spaces" # or "priority" in config.toml reads as spaces. The comment used to be part of the value, and the word priority anywhere on the line stripped every agent number.

• SUBSTITUTE_SETS is documented as acting on the program name or command line, which is what it does. The comment said the final label, and PR #15 was written against that wording. The README says that untitled Claude Code tabs are named by reading the session transcript, and how to turn that off, and it writes MAX_TITLE_LEN's default as the derivation it is rather than the number it happens to be.

Changed

• A steady-state event spawns fewer processes. The state file is read once per pass rather than once per tab, herdr --version is asked once per process, the rename the plugin itself issues no longer buys a second full pass through the tab.renamed event it fires, and the wait for a closing tab backs off over about two seconds instead of polling sixty times.

• The test suite runs its files in parallel, about twenty seconds instead of forty-five. ./tests/run.sh --serial restores the old order. CI pins its actions by commit and checks the shellcheck download against a recorded digest, and markdownlint now refuses an em dash, which CONTRIBUTING.md already did.

v0.9.1

Fixed

• Workspace numbers follow a collapsed space again on herdr 0.9.0. That release moved the terminal UI into each client and the sidebar's collapse state with it, so the array the plugin read, collapsed_space_keys in session.json, is now written empty whatever the sidebar shows. Every collapsed space read as expanded: the hidden rows kept a number no keybind reaches, and every row below one carried a number that jumped somewhere else.

  Collapse is read from the file the client writes it to instead, client-shell/local-<hash>.json under herdr's state directory, whose name is the FNV-1a 64 of the client socket path herdr derives from the socket path it already exports to us. An older herdr writes no such file, which is what picks the source: the file that is there, not a version test, so numbering on herdr below 0.9.0 is unchanged. The agent panel's sort order moved the same way and is read the same way, which matters only to a herdr old enough to number agents at all.

  The click now reaches the plugin faster than it did. herdr wrote session.json on a five-second debounce and writes this file the instant a space is toggled, so the pass that runs after the click reads the new value rather than the old one. Two limits are new. The file is one per session socket rather than one per client, so two clients on one session share it and the last writer wins. A client narrow enough for herdr's mobile layout ignores collapse altogether, without writing that down anywhere.

• A Muse tab is named after the agent. herdr 0.9.0 detects Muse, which installs as muse-bin-<version> and never runs under a plain name, so a pane herdr had not yet detected reached the tab bar as muse-bin-0.1.0-R708.1 wearing the glyph for an unknown program. The versioned name is folded onto muse the way herdr folds it, a digit required after the prefix so an unrelated muse-binary keeps its own name. The fold runs ahead of the alias lookup, the program lists and the icon map, so one PROGRAM_ALIASES entry names Muse however it was installed. muse, muse-cli and muse-code are listed alongside the other agents.

v0.9.0

Added

• An agent's title is condensed into its keywords instead of shown with the tail cut off. MAX_TITLE_LEN takes the END off a title, which is where the words saying WHICH task this is tend to sit: "Investigate why the nightly ETL job drops rows" reached a tab as "Investigate why the nightly". TITLE_CONDENSE=1 drops a leading verb and the filler and joins what is left, so the same budget carries "nightly-ETL-job-drops-rows".

  It selects rather than generates. The words are the agent's own, in the order it wrote them, on the reasoning that it put the salient ones first. TITLE_LEAD_VERBS and TITLE_FILLER_WORDS are the two lists, measured against the last title Claude Code generated in each of 65 titled sessions rather than written from imagination, and TITLE_WORD_SEPARATOR and TITLE_CASE say what the surviving words are joined and cased with. A condensed label is charged to MAX_TITLE_LEN like any other, the icon glyph and its space reserved out of it first.

  Off by default, so a config that does not name it renders exactly what AGENT_TITLES rendered before. A title that condenses to nothing is left as the sentence, and so is one whose label would come out longer than the prose or wearing a leading "[12]", the shape a tab number has.

• The agent is shown alongside its task, cc:auth-flow where a tab read auth-flow. Which agent is on a task was not recoverable from an agent tab: every agent herdr detects draws the same robot glyph, deliberately, and a title then replaced the one place the program name appeared. That costs a session running one agent nothing, and in a session running three it is what tells two tabs apart.

  TITLE_STYLE=name_and_task asks for it and PROGRAM_ALIASES applies, since asking for the name is asking for the name you chose for it. The name and its colon are charged to MAX_TITLE_LEN, so the task gives up the characters rather than the tab growing. The prefix is all or nothing: it goes in only where the budget seats the name, its colon and MIN_TASK_LEN characters of task, and otherwise the name is what goes, because this asks for the task with the name added rather than the other way about. The glyph and its space are part of that budget where icons are on, and an alias carrying a space is never used as a prefix, truncation having cut such a name in half and left the tab reading a fragment of it.

  Off by default. A refused title is not prefixed either, "cc:cc" saying nothing twice.

Fixed

• The workspace name follows the shell's directory (#20). A cd emits no herdr event, so a workspace kept the name of the directory it was created in until something unrelated woke the plugin, while the tab beside it moved at the next prompt. The shell hook is the one thing that does fire on a cd, and it renamed only the tab.

  It now names the workspace from the same prompt, by the rules the reconcile uses: the directory is the shell's own $PWD, the base is the repository that directory belongs to or the directory itself, and a workspace somebody named by hand is numbered and nothing else. A quiet prompt costs one state read and no herdr call, since the base we own is recorded and only a cd that leaves the project has anything to ask. A workspace the plugin has not adopted is left to the next event, adopting one meaning a round-trip on every prompt.

  Only the tab the workspace is on moves it, since herdr tracks a workspace's directory through its active pane and a prompt drawn in a background tab is about somewhere else. Panes with no shell hook installed are unchanged, and so is the five-second wait for herdr to persist a brand-new workspace.

• One PROGRAM_ALIASES entry now names an agent however it was installed (#19). Two agents answer to two names, cursor-agent, which herdr calls cursor, and kiro-cli, which it calls kiro, and which of the two reaches naming says only how the agent was installed. A native install arrives as its own executable, an npm-fronted one as the kind herdr detected behind the runtime. Keyed by exact name, one config labelled the two panes differently: cu where the alias matched, cursor where it did not.

  An alias is now looked up under both spellings, the exact name first, so a config naming each separately still gets each. The pair is read off the agent list that already carries both spellings rather than out of a second table somebody would have to keep in step with it.

• Two herdr sessions no longer share one state store (#22). Every server numbers its tabs from w1:t1, so the plugin running under herdr --session work and the one under herdr --session home wrote the same keys into the same state.json, and each full pass pruned the other session's tabs as closed. A tab whose record is gone reads as renamed by hand on the next pass and opts out of naming until reset, which is how a machine running more than one session ended up with every tab frozen on whatever it was called when the other session last ran, and the lock they also shared dropped events on top of that.

  A named session now keeps its store under sessions/<name>/ inside the state directory, resolved the way the herdr CLI picks its server: from the session directory in the socket path herdr exports to plugin commands and pane environments alike, or from HERDR_SESSION when no socket path is set, so the herdr-invoked pass and the shell hooks resolve one file. The default session keeps the store where it always was.

  Upgrading costs nothing in the common case. A named session's store is created once from the owned records of the old shared one, so a tab the plugin was naming goes on being named. A record the shared store got wrong, which on a multi-session machine is most of them, ends where an empty store would put it: a tab whose label matches nothing the store owns is opted out, and the reset action or clearing the label hands it back as before. Nothing removes a session's store when the session is deleted, and the old shared file stays as the default session's; both are safe to delete by hand.

v0.8.0

Added

• A tab is named after where the work is, not only after what is running there. The label reads [N] <directory> › <branch> › <activity>, so five claude tabs across three checkouts stop reading alike. Each part drops out when it says nothing worth the width: the directory when it is your home directory, the filesystem root, or the name of the workspace the tab is already in, since herdr shows that above the tabs. TAB_CONTEXT=0 turns the whole half off, and every part keeps a budget of its own rather than sharing one total.

  A directory too long for MAX_CONTEXT_LEN is reduced rather than cut through the middle. Worktrees and branches are named the same way by the same people, so bugfix-proj-482-fix-rev-discrepancy reads as PROJ-482 where a plain cut leaves bugfix-proj-, which identifies nothing.

• The branch the pane's repository has checked out, read from the files under .git and never by running git. git rev-parse is a process where HEAD is one open, and this runs per named tab on every herdr event and again on every shell prompt. The repository's own default branch is left out, read from refs/remotes/origin/HEAD rather than from a list of names, so a team whose trunk is develop gets the same silence a main one does. A repository that records no default falls back to TRUNK_BRANCHES, because otherwise every tab of a local-only repo carries main alike.

  Worktrees and submodules are followed through their gitdir: pointer to their own HEAD and through commondir to the shared refs. A rebase keeps the branch it set aside, so a tab does not take a new hash on every step; any other detached HEAD shows the short hash, which is where commits get lost. A branch that repeats the directory or the workspace is dropped, since a worktree named after its branch would otherwise say one thing three times. SHOW_BRANCH=0 and MAX_BRANCH_LEN=0 both leave branches out.

• A pane running ssh is named after the machine it reached, prod-01 › ssh. The directory it was launched from is local and the branch checked out there would read as the remote machine's, so both are dropped. The destination is parsed rather than taken as the first word after ssh, which is as often an option's value as a host: clustered short options, attached values, the url form, and bracketed IPv6 addresses all resolve, the user and the port are dropped, and a -o setting whose value is a command is refused outright rather than guessed at, because a ProxyCommand parses as its own bastion.

• A coding agent that has not titled its terminal is named from its own session. Claude Code derives that title from what the user typed, so a session opened with a slash command and answered by the agent alone is never given one, and its tab read claude for as long as it ran. The transcript is read for the title the agent generated, or failing that the first prompt the user actually typed, which is what Claude Code's own session list shows for an untitled session.

  Only Claude Code is read, and only where herdr reports that pane's session, which herdr integration install claude is what sets up. Reads are bounded to the end of the file each answer needs. AGENT_TRANSCRIPT=0 leaves the file unread, and a transcript that stops carrying these fields yields nothing, so the tab is named as it was before this existed.

Fixed

• A numbered workspace goes on following its directory (#13). herdr names a workspace after identity_cwd, its own tracked directory, and the first workspace rename freezes that name for good: herdr keeps the directory current and never labels from it again. Numbering a workspace pinned it to whatever it was called when it opened, and no smaller rename would have helped, because there is no rename that leaves the derivation alive.

  The base now comes from identity_cwd, read out of the session.json the collapse rules already read, and resolves to the repository the directory belongs to, or to the directory itself outside a repo. A pane moving around inside a project is not a rename; a pane that leaves takes the new name. Two caveats come with the file. The value is in no API response, and herdr saves it on a 5-second debounce, so a cd lands on the label an event or two later rather than instantly. An older herdr without the field keeps the previous behavior.

  A workspace herdr has not written to session.json yet, which is every new one for up to five seconds, reads its directory off its panes instead. Without that, the numbering rename lands before the file exists, the stale label reads as a name somebody typed, and the workspace opts out of tracking for good.

  A cd still emits no event the plugin subscribes to, so the new name lands on whatever event arrives next, the same way a sidebar collapse does.

  A workspace named by hand, or by another plugin, keeps its name and is only ever numbered, which is the promise a tab already gets. The plugin tracks a label that is herdr's own derivation or its own last write, and nothing else. Renaming a workspace back to the directory name hands tracking back, since reset takes a tab and a workspace has no equivalent.

v0.7.3

Fixed

• Tab naming could stop for a session, or for half a minute at a time, and say nothing either way. Both come from the lock that keeps one reconcile pass running at a time. Its 30-second steal was three filesystem calls, so two passes that found the same abandoned lock could both come away holding it: each writes the whole state file, one overwrites the other's record of which tabs it named, and a tab whose record is gone reads as renamed by hand on the next pass and opts out for good, recoverable only through the reset action. Stealing is now a move, an immediate reservation of the lock path, and only then the question of whether what was moved was really abandoned. Against six processes racing one abandoned lock, that is the difference between two holders in roughly a third of bursts and somewhere between none and one in a hundred.

  The same rewrite turned up three ways to leave a lock that no process holds and none may steal until it ages out, which stops every event in between. A lock whose owner token could not be written was left behind rather than given back, and no race is needed to reach it: a umask of 222 does, and so do a full disk, a quota, and a read-only mount. The hand-back that returns a lock to a live holder truncated its token before reading it, and printed its own failure onto your prompt. And the move that starts a steal did not check its destination was free, where an existing one makes it nest the lock instead of renaming it, out of sight of everything that follows.

• A state.json that jq cannot read no longer ends tab naming for the session. Every write started from the file already on disk, so one that would not parse failed every write for as long as it sat there, and the reset action could not recover the tab either, because re-adopting one is another write. It renamed the tab once, recorded nothing, reported "Nothing to reset", and the next event went back to doing nothing. Missing and unreadable are now the same answer, and the check insists on exactly one JSON object: jq reads top-level values as a stream, so a file holding two of them parsed, and writers handed the pair updated each document and wrote both back, leaving it broken for good.

  Healing the file does not hand a tab its name back, and cannot. A tab whose record went with the file has a label matching nothing state knows, which is exactly what a hand rename looks like. What changes is that the opt-out is recorded, so reset works again.

Changed

• make lint can fail. Its recipe was command -v shellcheck && shellcheck ... || echo skipping, and a real warning took the || branch too, so the target printed "shellcheck not installed" and exited 0. make lint-md had the same shape, so a hard-wrapped line reported itself as a missing npx. CI now runs shellcheck as well, over the same file list, and its syntax check covers every bash file in the repo rather than three of them.

v0.7.2

Fixed

• An oh-my-pi tab showed the agent's spinner glyph and renamed itself on every status change. Most agents park a status glyph in front of the task and the existing strip takes it off; oh-my-pi puts its own brand there first and the glyph second: π ⠋ Fix the parser while it works, π > Fix the parser when the turn is yours, π ! ... when it wants you, π: ... with its title state off. The strip only takes a leading run of non-alphanumerics and jq reads the brand as a letter, so it stopped on character one and each state reached the tab as a different label (#12).

  TITLE_BRANDS names a brand per herdr agent kind (pi and omp by default, both the same program) and it comes off before the strip runs, at the very front only and only with a non-alphanumeric behind it, so πcalc rewrite keeps its first letter. Every state now reduces to the same label, so nothing renames on a status change. The reported case gets more than the glyph back: with no session title yet oh-my-pi labels itself after its directory, which the directory refusal already knew to hand to the program name. It just could not see it behind the brand.

v0.7.1

Fixed

• A Qwen Code or maki tab was missing from the lists keyed by program name. herdr 0.8.2 detects qwen as an agent kind, and Qwen Code installs from npm, so its pane fronts as node and takes the WRAPPER_PROGRAMS path that names a tab from herdr's detection: naming needed nothing. The two program lists did. With ICONS_ENABLED=1 a qwen tab drew the ? fallback instead of the agent glyph, and with SHOW_PROGRAM_ARGS=1 a natively installed qwen showed its whole command line. maki, a kind herdr added in 0.8.0, was missing from both the same way. Qwen's title needs no new rule: it parks a status glyph in front of the task (✳, ◐), which the existing strip takes off, and the no-task title Qwen Code is already refused as the agent kind plus code.

  The icon test reads its roster out of NAME_ONLY_PROGRAMS now rather than repeating it, so the next agent added to one list and not the other fails the suite instead of shipping. That is how maki went two herdr releases with the fallback on its tab.

Documentation

• The README covers herdr 0.8.2's ui.window_title, which is on by default at {hostname}: {workspace} and so already carries a workspace's [N] prefix. Adding {tab} puts the generated tab name in the outer terminal's title, a surface the plugin cannot write itself.
• The README notes that herdr 0.8.2 searches renamed single-tab labels, so the Session Navigator finds a tab by the name this plugin wrote.

v0.7.0

Names a tab running a coding agent after the task the agent reports -- Squash merge command instead of a fifth tab reading claude. herdr publishes the title on the pane object the reconcile already holds, so the label costs no extra herdr call. It follows the work on an event the plugin already subscribes to, so nothing polls for it.

A tab with more than one pane is named after the pane that matters. For the usual agent-and-shell split that is the agent, not the half you happen to be typing in. Several naming bugs go with it: a background split kept the name it had before the split, a tab whose rename herdr rejected was never named again, and a label carrying a backslash or a control character reached herdr mangled.

Upgrade note for anyone running a coding agent in a tab. AGENT_TITLES defaults to on, so those tabs stop reading claude and start reading the task. Set AGENT_TITLES=0 to keep the program name. MAX_TITLE_LEN (28) and TITLE_IGNORE tune what gets through.

Added

• A tab running a coding agent is named after the task the agent reports rather than after the agent (AGENT_TITLES, default on). Naming by foreground program has one case it cannot answer: five claude tabs all read claude, and the tab bar stops being a way to find anything. A coding agent already publishes the task as its terminal title and keeps it current as the work moves, so that is what the tab shows -- Squash merge command, Check PR 2169 relevance. herdr publishes the title on the pane object the reconcile already holds, so reading it costs no extra herdr call on any version, and a tab named this way needs no pane process-info at all, which leaves an agent-heavy session making fewer herdr round-trips than before. The label follows the work on pane.agent_status_changed, an event the plugin already subscribes to, so nothing polls for it.

  A title only replaces the program name when it says something. ar_title_clean refuses the three kinds that say nothing: one naming the agent instead of its work (the agent's own herdr kind, that kind followed by code, or a TITLE_IGNORE entry, which is what an agent titles a session it has no task for yet), one that is just the directory the pane sits in, and a bare number, which is herdr's own generated tab label handed back through the title. Every refusal falls through to the program name, PROGRAM_ALIASES and the WRAPPER_PROGRAMS unwrapping included. A path to that directory is refused the same way, since an agent titling itself ~/dev/api is still naming where it sits: only a title with no whitespace, though, because a description like Fix build for services/payments also ends in a directory name and is still the work. Every comparison is made against the title in the shape the tab would carry it, so a trailing space cannot walk a refused title past them.

  A title that survives replaces the program name outright, alias and all: AGENT_TITLES is the request for the task, and an alias shortening claude to cl is not a request to hide the work. The program still supplies the icon, and the label gets its own budget, MAX_TITLE_LEN (28) rather than MAX_NAME_LEN (20), cut at a word boundary when that leaves at least half of it: a title is a sentence, and 20 characters were sized for command names.

  Before any of that, the leading run of non-alphanumerics comes off. herdr keeps an ANSI-stripped copy of the title and stripped means exactly that: an agent parks a spinner glyph in front of its title while it works, and claude cycles four of them. Without the strip the label would flip between Task and <glyph> Task on every status event, and each flip is a rename. jq does that strip, and the case-folding for the comparisons, in one call, because its character classes know Unicode: a byte-wise strip under the C locale herdr may launch a plugin in would eat the first letter of a title that opens with a non-ASCII word. The function still takes strings and returns strings, which is what keeps it in naming.sh.

• A tab with more than one pane is named after the pane that matters. Naming it from its own focused pane, which is what the layouts fix below gives it, still read the wrong half of the usual agent split. An agent sits in one pane and a shell in the other, and focus is in the shell most of the time, so the tab advertised zsh while the interesting thing sat beside it. The pick is now, in order: the tab's own focused pane when herdr reports an agent in it, then any pane of the tab holding an agent that is working or blocked, then the focused pane. So a split stays about the agent while you read or type in the shell half, and an idle agent stops at the second rule on purpose -- a finished agent should not hold the name of a tab you have moved on to. The choice is made in the snapshot reshape and travels on the tab row as _name_pane, so the tab loop costs nothing extra for it. A herdr whose snapshot carries no layouts can answer neither the first rule nor the third, since both ask for the tab's own focused pane. The second asks only for an agent at work, so it still answers there, and everything it does not answer falls to the older inference (the sole pane of a single-pane tab, else the tab's own focused pane).

• The reset and clear actions report what they did as a herdr notification. Both are built for a keybinding, where the only feedback was the tab bar redrawing. A reset that found no tab to re-adopt, or that ran under NAME_TABS=0, looked exactly like one that worked, and clear on an already clean session looked like nothing at all. Each outcome now names itself. A re-adoption is reported only when both halves happened: the tab had opted out (read before its state is cleared) and it is named and owned again by the end of the pass. Reporting it from the first half alone told the user naming was back on even when the rename that followed failed, and a tab in that position opts itself straight back out on the next event.

  Both actions also wait for the lock instead of deferring to whoever holds it. Events can defer, because any pass computes the same names, but an action carries a request that lives in its own process, so handing the job over dropped it: a reset pressed during a burst of events did nothing at all, and reported nothing either, since the deferral exited before the notification. Best effort, so a herdr without notification show declines it and the action still does its work; the uninstall path (--clear) notifies through the same helper.

Fixed

• Running the test suite inside a herdr pane renamed that session's tabs. The hook tests source the real hooks, and a hook resolves the real engine next to itself, so with HERDR_TAB_ID inherited from the pane and nothing pointing the engine at the fake herdr, every hook call the tests made went to the live session. They run fully sandboxed now.

• A tab whose rename herdr rejected stopped being named at all. Ownership was recorded before the rename was issued, so a failed one left state claiming a base the tab did not carry; the next pass compared the two, read the mismatch as a name typed by hand, and opted the tab out of auto-naming for good -- recoverable only through the reset action. Ownership is now recorded for a name the tab actually carries: the label already matches, or the rename reported success. The shell-hook fast path has always ordered it this way.

• A background tab with more than one pane kept the name it had when it last had a single pane, so a split that started an agent still read nvim. Naming had no pane to read: none of a background tab's panes carries .focused, and the pane list is all the pass looked at. It now takes the tab's own focused_pane_id from the snapshot's layouts, which answers for every tab. Older herdr, whose snapshot has no layouts, keeps the previous behavior -- except for a tab with an agent at work in it, which the rule above names without asking for a layout.

• The focused tab was named from the GLOBALLY focused pane, which belongs to whichever client moved focus last. With a second client attached, or a remote attach, that pane can sit in a different tab. Per-tab layouts remove the guess, and the pane-list fallback now reads only the tab's own panes. A focused tab with several panes and no focused pane of its own -- which is what that same second client looks like on a herdr with no layouts -- keeps the label it has rather than being named after an arbitrary one of them.

• Where herdr named no foreground process group at all but still reported one process for the pane, no name was computed. Some Linux container and sandbox setups cannot expose a foreground group, which is what left tab naming doing nothing there. A single reported process is not a choice, so such a pane is now named after it. Two or more with no named group would be: herdr does not order that list and documents its own degraded detection as one where a background job can look like the foreground one, so those panes keep the label they have. A group herdr DID name whose process is absent from the list is a group racing its own exit, and still computes no name; so does a pane herdr reports no process for.

• A label carrying a control character reached herdr verbatim: argv can hold a newline or a tab, so SHOW_PROGRAM_ARGS=1 could put one in the tab bar. They are now replaced where they arrive, in the jq that reads pane process-info, and the label has its whitespace runs collapsed and its ends trimmed before truncation. A clean label -- nearly all of them -- costs no extra process.

  The two values that reader returns travel one per line now rather than as TSV. @tsv escaped a real tab into the printable two characters \t, which no scrub downstream can tell from text somebody typed, and it doubled every backslash in a command line on the way past.

• A tab carrying no label at all was never named again. Its row had an empty field ...