--- name: orphan-conhost-cwd-pins-worktree description: "the \"worktree that will not delete\" is whatever process holds the crate root as its CWD — usually a surviving descendant, but a --headless ConPTY conhost can be the sole holder (measured 2026-08-02), so check the cwd not the image name; write-inside works, rename/delete blocked; find the holder with the PEB cwd reader, then read its identity before killing" metadata: node_type: memory type: reference originSessionId: 0241438f-ee37-4144-979e-b5499de58a8e modified: 2026-08-02T13:16:04.974Z --- **Mechanism (corrected 2026-08-02 by measurement — the earlier conhost attribution in this file was WRONG and is deleted, not annotated).** A `git worktree remove` / `rmdir` that fails on `\crates\` is a **live process whose CURRENT DIRECTORY is that dir**. Cargo runs test binaries with cwd = the package root, so every child a test spawns inherits `\crates\\` — and so does every **grandchild**. Any of them that outlives the run pins the dir. The crate in the path is whatever crate the test ran under; nothing is special about spt-daemon. **A conhost is USUALLY not the pin — but "never" is false.** `CREATE_NO_WINDOW` mints exactly one `conhost.exe` per console child, and *those* conhosts take an **OS-assigned** cwd: box-wide census = 43 readable, all `C:\WINDOWS\` or `C:\WINDOWS\system32\`, zero under any worktree; force-killing such a child leaves its orphan conhost still reading `C:\WINDOWS\`, and on a scratch-dir positive control killing the conhost left the dir BLOCKED while killing the real holder — an orphaned **grandchild** (`PING.exe`, cwd inherited through `cmd`) — deleted it first try. So for the CREATE_NO_WINDOW class, killing conhosts is not the cure and "no conhosts" is not the all-clear. **Falsified 2026-08-02 (todlando, BAROMETER W4 teardown): a `--headless` ConPTY conhost carries an INHERITED cwd and pins that dir BY ITSELF.** Wild specimen: `.worktrees/barometer-w4` refused removal at `crates\spt`; the PEB sweep row was `46556 conhost …\barometer-w4\crates\spt\` (positive control: my own shells read the repo root). Killing a stray `du.exe` first left it blocked; killing the conhost freed it. *Retracted from my own first argument:* I claimed the chain was tight because the conhost had **zero children**. hertz was right that this proves nothing — **a ConPTY client is a child of whoever called `CreatePseudoConsole`, not of the conhost**, so killing the conhost tears down the pty and can take a living client with it. Never argue holder-identity from a conhost's child list. What replaces it — two controlled rigs (`pin-experiment.ps1`): 1. **Inheritance:** launch `conhost.exe --headless` from a probe dir ≠ system32 → its PEB reads the probe dir. Not OS-assigned. (That rig launched the binary directly with a working directory.) The `CreatePseudoConsole` path — the one the wild specimen took — was then covered separately by hertz: mint a pseudoconsole from a process whose cwd is a chosen dir and the new `conhost.exe --headless …` reads THAT dir, with the pin's own control beside it — write-inside **OK**, delete **BLOCKED** while it lives, **DELETED** after `ClosePseudoConsole`. Reading a path and blocking a delete are different claims; both are now held. 2. **Sole-holder, by KILL ORDER:** conhost cwd = probe dir, client deliberately `cd`'d to `C:\`. Both alive → blocked; **kill the CLIENT only → still blocked**; kill the conhost → deletes. Step 2 is causal, so pty-teardown-killed-a-client cannot explain it. **A census that looked like confirmation and had zero power:** all 8 live `--headless` conhosts read `C:\WINDOWS\system32\` — but their creator (`spt.exe`) sits there too, and that value is exactly what OS-assignment predicts. Uniform rows, both hypotheses fit ([[uniform-table-blindness-sweep]]). **Elevation is NOT the discriminator, and neither is access mask:** `elevated=False` and my reader still opened 49/54 conhosts (all 8 headless). hertz's "unelevated opens NONE" turned out to be a **parse bug in his caller**, not an access denial — his finder reads 46/50 unelevated, same mask as mine (`PROCESS_QUERY_INFORMATION|PROCESS_VM_READ`). No elevation caveat went into the doc. I had guessed *requested access rights*; that guess was also wrong. The durable lesson is only that the premise was worth checking rather than accepting — [[ruling-rests-on-a-premise]]. **Landed:** hertz's amend is H2 tip `2ab3e92` (population-scoped strike-and-amend), and he closed my method caveat by minting a `--headless` conhost through `CreatePseudoConsole` from a chosen cwd: its PEB reads that cwd, and the pin has its own control (write-inside OK, delete BLOCKED while the pseudoconsole lives, deleted after `ClosePseudoConsole`). Both launch paths are in the entry. Scope: the CREATE_NO_WINDOW half of the census stands, surviving descendants stay the common case. Only the word **"never"** dies, for the `--headless` population. Check the cwd, never the image name. **Signature, no tool needed:** a cwd pin allows writes INSIDE the dir but blocks rename/delete (content locks do the opposite), and persists (a scanner does not). **Finder** (`.github/ci/find-cwd-holders.ps1`): landed in hertz's `d058d34` — **on `golden/barometer`, NOT on main** (verified 2026-08-02; `git ls-files` on main finds nothing, which is why I rebuilt the PEB reader from scratch rather than reusing it). Read-only PEB read, `-Match worktrees`. It reports its readable/not-openable **denominator** — an empty result is "no pin I could see", never "no pin". **It answers WHO pins, never whether it is SAFE TO KILL** — see [[worktree-pin-holder-triage]]: orphan (parent dead, headless) → kill; a live agent's own shell → `cd` out (pwsh and Bash hold SEPARATE cwds, so clearing one leaves the other). **Two fixes refused on evidence:** flipping the rigs' `no_window()` to `DETACHED_PROCESS` removes the conhost (1 → 0) but not the pin, and under that flag **powershell does not execute at all** while `cmd`'s **piped stdout is lost** — which would make `job_escape_e2e::conhost_children_of` (`unwrap_or(0)`) turn the `REQ-HAZARD-WMI-DAEMON-WINDOW` guard **vacuously green**. Nor can the helper hand children a neutral cwd: `bind_cwd_project_e2e` / `resume_template_e2e` assert on the **inherited** cwd. Prevention is rig **teardown completeness** (the #99 class, extended to grandchildren). Written up as KNOWN-HAZARDS **5.18** `[REQ-HAZARD-TREE-CWD-PIN]` + a 5.17 amendment. See [[stale-breadcrumb-tree-kill-class]], [[worktree-target-junction]], [[reap-subtree-not-session-dir]], [[live-agents-lock-target-debug-spt-exe]], [[stacked-defects-mask-each-other]].