chore(ci): weekly renovation 2026-08-23 (#3880)

## Summary

Weekly CI renovation. Rust stable moved to 1.98.0 on 2026-08-20, so MSRV
and the development toolchain go to 1.97 (latest stable − 1); one
cargo-install pin had drifted. Everything else is already current.

- **MSRV + toolchain: 1.96 → 1.97** — `Cargo.toml`,
`tests/helpers/wt-perf/Cargo.toml`, `rust-toolchain.toml` (`1.97.0`),
plus the `flake.lock` refresh the channel bump requires.
- **`cargo-llvm-cov`: 0.8.7 → 0.9.0** — `coverage.yaml`. MSRV 1.87, well
under the pinned toolchain.

Toolchain compatibility for every other pinned crate against 1.97.0:
`cargo-affected` 0.4.0 (1.94), `cargo-insta` 1.48.0 (1.66.0),
`cargo-nextest` 0.9.143 (1.91), `cargo-llvm-cov` 0.9.0 (1.87),
`cargo-msrv` 0.19.3 (1.91.1), `lychee` 0.24.2 (1.88.0), `worktrunk`
0.74.0 (1.96), `cargo-udeps` 0.1.61 (unspecified). All build under
1.97.0.

- **Pinned nightly: `nightly-2026-03-01` → `nightly-2026-08-01`** —
`nightly.yaml` ×2. Not drift: the MSRV bump *broke* these two jobs, see
below.
- **The skill's `flake.lock` recipe** — the in-session Nix install it
pointed at can never run. The setup step that fixes that lives in #3891,
not here; see below.

The remaining commits are fallout from the same sweep: `cargo-llvm-cov`
0.9.0 changes the shape of the `--show-missing-lines` column
`tests/CLAUDE.md` sends you to during a codecov investigation, and the
weekly statusline cache-check has been reporting two duplicates that are
artifacts of its own `-vv` flag.

## The MSRV bump broke two nightly jobs; that is fixed here

`check-unused-dependencies` and `minimal-versions` each pin their own
toolchain, and `nightly-2026-03-01` is `1.96.0-nightly` — so raising
`rust-version` to 1.97 made cargo refuse the workspace outright, before
either job ran its actual check:

```
error: rustc 1.96.0-nightly is not supported by the following packages:
  worktrunk@0.74.0 requires rustc 1.97
```

Both are now on `nightly-2026-08-01` (`1.99.0-nightly`), which clears
the new MSRV with headroom rather than landing on it. Reproduced locally
in both directions: `cargo +nightly-2026-03-01 check` fails with the
error above, `cargo +nightly-2026-08-01 check` passes. The coupling is
easy to miss because neither failure mentions udeps or minimal versions,
so both lines now say so, and the weekly MSRV file table gained a row
for the pins. The rule on that row is a date, not a version: bump only
when the pinned date is more than three months old. Nightly runs two
channels ahead of stable and MSRV tracks stable − 1, so a pin that
recent is still a release above the new MSRV — which makes staleness
checkable from the pinned date alone, where reading it off the rustc
version means fetching that day's `channel-rust-nightly.toml`.
`nightly-2026-03-01` was five and a half months old, which is exactly
what the rule catches.

## The `flake.lock` bump was computed by hand — and the recipe that
forced that is fixed here

`rust-toolchain.toml` moving to `1.97.0` forces a lock refresh: the
locked `rust-overlay` (`4a408e1f`, 2026-06-02) carries stable manifests
only up to `1.96.0`, so `nix flake check` cannot resolve the new
channel. `rust-overlay` is bumped to `f60c1b57` (2026-08-23), which has
`1.97.0.nix`. `nixpkgs` and `crane` are left alone — only `rust-overlay`
has to know about the new channel, and a full `nix flake update` is a
wider change than this PR needs.

This session could not run `nix flake update`: the weekly runner has no
Nix, and the agent runs as a sandbox user with no sudo, so the installer
the skill pointed at stops at `sudo: a password is required`. The lock
entry was therefore computed directly from the NAR serialisation, and
the method was validated first by recomputing the **existing**
`rust-overlay` entry from its own revision and getting
`sha256-7mDa7OBAaf7MU6ZovT9ENfD62kH911SsSazBb4KTDF0=` back byte-for-byte
— the value already in `flake.lock`. The same script then produced the
new entry.

That should not have been the answer, and it no longer is. #3891 has
merged (`2960bfb`), so `tend-setup` installs Nix and next week's refresh
runs `nix flake update rust-overlay` for real. This branch briefly
carried its own `cachix/install-nix-action@v31` step in the same file;
`adbeb76` dropped it in favour of #3891's, which asserts the multi-user
install branch directly and doesn't couple to `github.workflow ==
'tend-weekly'`, a gate that stops firing if the generated workflow is
renamed. `.github/actions/tend-setup/action.yaml` on this branch is
byte-identical to `main`.

What stays here is the skill side: the in-session install recipe and the
NAR-serialisation fallback are both deleted, because the Determinate
installer they rest on cannot run in the sandbox. Both PRs rewrote that
section and converged on nearly the same wording, so they merged clean
while both were open — but #3891's squash collapsed its five commits
into one, and its last two had moved the "if `nix` isn't on the PATH"
line and folded the flake-input rationale into the code block. That
turned the section into the merge's one conflict, resolved by taking
`main`'s wording wholesale; the section is now byte-identical to
`main`'s, and this branch's own additions to that file are untouched by
it.

That is a strong check but not the real one. `nightly`'s `nix-flake` job
runs `nix flake check` and fires on this PR (the paths filter matches
`rust-toolchain.toml`, `Cargo.toml`, and `flake.lock`), so it is the
gate that actually proves the lock evaluates. **It has now run on this
head and passed**, so the computed lock resolves the 1.97.0 channel and
the flake's checks build under it. Flagging the provenance anyway, since
a hand-written lock entry deserves a second look on review.

<details><summary>How the hash was computed</summary>

Nix serialises the unpacked flake input as a NAR and reports SHA-256 of
that stream in SRI form. The format is length-prefixed strings padded to
8 bytes, with directory entries in bytewise-sorted order; a ~90-line
Python script reproduces it from the GitHub tarball with the top-level
`<repo>-<rev>` directory stripped, which is the same tree Nix hashes.

```
rust-overlay 4a408e1fc99ad517b4cb402fd0de7464f40c05e1 (currently locked)
  computed  sha256-7mDa7OBAaf7MU6ZovT9ENfD62kH911SsSazBb4KTDF0=
  in lock   sha256-7mDa7OBAaf7MU6ZovT9ENfD62kH911SsSazBb4KTDF0=   ✓

rust-overlay f60c1b57ff805a46b5175c76fc981fb4f81efbcc (new)
  computed  sha256-r4LDUF+zmJnkftvCVkCrUhSJazsf6EVJF+V2l4/MYbI=
  lastModified 1787454509 (committer date 2026-08-23T03:08:29Z)
```

</details>

## Already up to date

- `cargo-affected` 0.4.0, `cargo-insta` 1.48.0, `cargo-nextest` 0.9.143,
`cargo-msrv` 0.19.3, `cargo-udeps` 0.1.61, `lychee` 0.24.2, `worktrunk`
0.74.0 (matches the newest crates.io release)
- Nushell 0.115.0 — all four call sites, and `dev/codex.sh`
- Codex Cloud pins in `dev/codex.sh`: pre-commit 4.6.2, cargo-insta
1.48.0, cargo-nextest 0.9.143, PowerShell 7.6.5; `setup-web`'s own
Nushell and PowerShell pins agree
- Runner images `ubuntu-24.04` and `windows-2022` (the latter
deliberately pinned — actions/runner-images#12677)

## Found, not bumped: `macos-15` is no longer what `macos-latest`
resolves to

`ci.yaml`'s header comment says `ubuntu-24.04, macos-15` are "pinned to
current -latest equivalents". That is still true of Ubuntu but no longer
of macOS: `macos-latest` now resolves to **macOS 26 arm64**, and
`macos-15` has become the older of the two GA images — which, under the
runner-images deprecation policy ("deprecation of the oldest image label
begins once the newest OS image label has been released to GA"), is the
one that starts aging out next. `macos-14` already carries the
deprecated badge.

Not bumped here, because it is an OS upgrade rather than a version pin,
and this repo drives PTY and shell-integration snapshots on macOS that a
new image and Xcode can move. There is also nothing this session can do
to check it — the only macOS available to it is the `test (macos)` job
on a PR, and a red macOS matrix would have blocked the MSRV bump
alongside it. Flagging rather than guessing at the timing: whether to
take macOS 26 now or wait for the deprecation announcement is a call
about this repo's snapshot surface. Happy to open it as its own PR if
you want the signal.

## CI

Fully green on `7940a97`, the first merge of `main`: all 33 non-skipped
checks, including `msrv`, all three `test` legs, `full-tests` on every
platform, `nix-flake`, `minimal-versions`, `check-unused-dependencies`,
`check-docs`, `link-check`, `feature-powerset`, and `codecov/patch`.
That head is where the hand-written lock entry and the nightly pins were
proven against the current `main`.

`main` has since moved four commits (#3886, #3893, #3895, #3897) and is
merged in again at `d8d1553`, alongside the nightly-pin rule above.
#3895 touches `nightly.yaml`, so the pins were re-checked after that
merge: still exactly two sites, `minimal-versions` and
`check-unused-dependencies`, both on `nightly-2026-08-01`. Locally
`pre-commit run --all-files` and `test_docs_are_in_sync` pass on this
head.

## Weekly checks that produced no change

- **LLM model pins** — no drift. `gpt-5.6-luna` is still how OpenAI's
models page positions the "fast and affordable … lowest cost in the
family" 5.6 variant; `gpt-5.3-codex-spark` is a Pro-only research
preview, not a default. `haiku` / `claude-haiku-4.5` is still the
smallest current Anthropic model.
- **Statusline cache-check** — clean. `command_count` was 27, which
includes the three calls the `-vv` diagnostic collector makes after the
render, so the render itself is ~24 against a ~29 baseline measured
before the recipe used `-vv`. The two duplicates the report flagged (`gh
--version`, `git worktree list --porcelain`) are that same collector,
not the render. Both the baseline arithmetic and the trap are now
written down in the skill.
- **Agent app integration surfaces** — nothing that changes what
Worktrunk consumes. Claude Code 2.1.239 fixed `metadata.pluginRoot`
resolution, but `.claude-plugin/marketplace.json` names an explicit
`./plugins/worktrunk` source rather than a bare name, so it is
unaffected; the `WorktreeCreate`/`WorktreeRemove` hooks and the
statusline stdin schema are unchanged (`workspace.git_worktree` and
`workspace.repo` exist but predate this window by several hundred
releases). Codex, Gemini CLI, and OpenCode shipped only routine
releases.
- **README month** — already "August 2026".
- **Dependency PRs** — none open.

<details><summary>Verification</summary>

- `rustup` resolves the new channel: `cargo 1.97.0 (c980f4866
2026-06-30)`, `rustc 1.97.0 (2d8144b78 2026-07-07)`, `clippy 0.1.97`.
- `cargo clippy --all-targets --all-features -- -D warnings` passes on
1.97.0 — the check that matters most for a toolchain bump, since it
compiles every target and feature under the new rustc and holds the new
clippy to zero warnings. Doctests passed too.
- Unit tests pass on 1.97.0: 979 + 11 + 1, zero failures.
- Integration tests: 2000 passed, 1 failed —
`test_copy_ignored_preserves_file_executable_permissions`, which expects
`0644` and got `0664`. That is this sandbox's `umask 0002`, not the
toolchain; the `test (linux|macos|windows)` matrix is the real gate.
- `cargo run -- hook pre-merge --yes` is not a clean local signal on
this runner: it ends on `pre-commit: exit status: 127` because
pre-commit isn't installed in the tend sandbox, unrelated to the bump.
- `cargo msrv verify` was not run locally (it rebuilds the graph per
candidate); `ci.yaml`'s `msrv` job covers it.
- Every version above was read from its upstream source of truth: `cargo
info` / crates.io for the cargo tools,
`static.rust-lang.org/dist/channel-rust-stable.toml` for Rust stable
(1.98.0, 2026-08-20), the `nushell/nushell` and `PowerShell/PowerShell`
release APIs, PyPI for pre-commit, `actions/runner-images` README for
the image labels, and the vendors' own model pages for the LLM pins.
- `rust-overlay` at the currently-locked revision was confirmed to stop
at `1.96.0.nix`, and at the new revision to contain `1.97.0.nix` — this
is why the lock has to move at all.

</details>



> _This was written by Claude Code on behalf of max-sixty_

---------

Co-authored-by: worktrunk-bot <254187624+worktrunk-bot@users.noreply.github.com>
Co-authored-by: Maximilian Roos <m@maxroos.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

.claude/skills/running-tend/SKILL.md

@@ -266,6 +266,11 @@ Files to update:
 | `Cargo.toml` | `rust-version` | `"1.93"` |
 | `tests/helpers/wt-perf/Cargo.toml` | `rust-version` | `"1.93"` |
 | `rust-toolchain.toml` | `channel` | `"1.93.0"` |
+| `.github/workflows/nightly.yaml` | `rustup override set nightly-<date>`, twice (`minimal-versions`, `check-unused-dependencies`) | a nightly from the last few weeks |
+
+Bump the nightly pins only when the pinned date is more than three months old.
+Cargo refuses a workspace whose `rust-version` exceeds the toolchain, so an
+MSRV bump past a stale pin fails both jobs before they check anything.
 
 `flake.nix` reads the channel from `rust-toolchain.toml`, so no separate bump
 is needed. After updating the toolchain, refresh `flake.lock` so the locked
@@ -339,9 +344,26 @@ Triage each duplicate:
   open an issue or fix it. Common shapes: `merge_base("main", "<sha>")` vs
   `merge_base("main", "branch")` keying separately;
   `worktree_at(cwd)` vs `worktree_at(porcelain_path)` not canonicalizing.
+- **Measurement artifact** — `-vv` is what produces the trace, and the
+  diagnostic collector that writes the bundle runs its own `git --version`,
+  `gh --version`, and `git worktree list --porcelain` (`src/diagnostic.rs`)
+  *after* the render, into the same trace. So the report reliably pairs those
+  with the render's calls on an otherwise clean run. Only a duplicate *within*
+  the render is a finding; the collector's calls are the trailing block in
+  `"$(git rev-parse --git-common-dir)/wt/logs/trace.log"`.
 
 Baseline: ~29 git subprocesses per render on a clean tree; a jump above
-~32 warrants investigation.
+~32 warrants investigation. `command_count` in the same JSON report is the
+number to read, minus the collector's three calls: the baseline was measured
+before the recipe used `-vv`, so it counts the render alone, and three is the
+whole width of the ~32 threshold.
+
+Discounting by hand is a stopgap — the durable fix is to give the collector's
+subprocesses a reserved context in `Profile::from_entries` and exclude it from
+both `command_count` and the duplicate bucketing, so the number means the
+render again. When that lands, delete this paragraph, the **Measurement
+artifact** bullet, and the discount clause in the paragraph above — the
+`~29`/`~32` baseline itself stays.
 
 ## Weekly Maintenance: LLM Model Names in Docs
 

.github/workflows/coverage.yaml

@@ -35,7 +35,7 @@ jobs:
     - uses: baptiste0928/cargo-install@v3
       with:
         crate: cargo-llvm-cov
-        version: "=0.8.7"
+        version: "=0.9.0"
 
     - uses: baptiste0928/cargo-install@v3
       with:

.github/workflows/nightly.yaml

@@ -288,8 +288,9 @@ jobs:
     - name: 📂 Checkout code
       uses: actions/checkout@v7
 
-    # cargo-udeps requires nightly; update date periodically
-    - run: rustup override set nightly-2026-03-01
+    # cargo-udeps requires nightly. Keep at or above `rust-version` in
+    # `Cargo.toml`; see the note on `minimal-versions`' identical pin.
+    - run: rustup override set nightly-2026-08-01
 
     - name: 💰 Cache
       uses: Swatinem/rust-cache@v2
@@ -324,7 +325,12 @@ jobs:
     - name: 📂 Checkout code
       uses: actions/checkout@v7
 
-    - run: rustup override set nightly-2026-03-01
+    # Pinned so `-Z direct-minimal-versions` resolves reproducibly. Must stay
+    # at or above `rust-version` in `Cargo.toml`: cargo refuses a workspace
+    # whose MSRV exceeds the toolchain, so a stale pin fails the job outright
+    # rather than reporting on dependency floors. Bumped by the weekly MSRV
+    # pass, with headroom so that isn't every week.
+    - run: rustup override set nightly-2026-08-01
 
     - name: 💰 Cache
       uses: Swatinem/rust-cache@v2

Cargo.toml

@@ -50,7 +50,7 @@ filterset = "test(/readme_sync/)"
 name = "worktrunk"
 version = "0.74.0"
 edition = "2024"
-rust-version = "1.96"
+rust-version = "1.97"
 build = "build.rs"
 repository = "https://github.com/max-sixty/worktrunk"
 homepage = "https://worktrunk.dev"

flake.lock

@@ -64,11 +64,11 @@
         ]
       },
       "locked": {
-        "lastModified": 1780370483,
-        "narHash": "sha256-7mDa7OBAaf7MU6ZovT9ENfD62kH911SsSazBb4KTDF0=",
+        "lastModified": 1787454509,
+        "narHash": "sha256-r4LDUF+zmJnkftvCVkCrUhSJazsf6EVJF+V2l4/MYbI=",
         "owner": "oxalica",
         "repo": "rust-overlay",
-        "rev": "4a408e1fc99ad517b4cb402fd0de7464f40c05e1",
+        "rev": "f60c1b57ff805a46b5175c76fc981fb4f81efbcc",
         "type": "github"
       },
       "original": {

rust-toolchain.toml

@@ -1,3 +1,3 @@
 [toolchain]
-channel = "1.96.0"
+channel = "1.97.0"
 components = ["rustfmt", "clippy"]

tests/CLAUDE.md

@@ -48,9 +48,11 @@ When `codecov/patch` fails, investigate before declaring ready (the merge gate i
 
 ```bash
 task coverage
-cargo llvm-cov report --show-missing-lines | grep <file>   # authoritative miss list; matches codecov line-for-line
+cargo llvm-cov report --show-missing-lines | grep <file>   # authoritative miss list; same lines codecov counts
 ```
 
+Since cargo-llvm-cov 0.9.0 that column collapses consecutive misses into ranges (`12-18`, not `12, 13, …`), so expand a range before comparing it against a codecov line list.
+
 For each uncovered function, either write a test (integration tests via `assert_cmd_snapshot!` do capture subprocess coverage) or document why it's intentionally untested.
 
 **Querying codecov directly** serves two cases the local report can't: disputing a posted check, and running in CI, where `task coverage` isn't installed. Prefer measuring everywhere else.

tests/helpers/wt-perf/Cargo.toml

@@ -7,7 +7,7 @@
 name = "wt-perf"
 version = "0.1.0"
 edition = "2024"
-rust-version = "1.96"
+rust-version = "1.97"
 publish = false
 description = "Performance testing and tracing tools for worktrunk"