Conversation
Installs direnv and exports .envrc into CLAUDE_ENV_FILE once at session start (static export/unset statements, not a live shell hook), then re-exports via a debounced PreToolUse(Bash) check whenever the .envrc fingerprint changes. Mirrors mise's install-mise.sh discipline around avoiding `direnv hook bash`'s PROMPT_COMMAND-style eval loop, and github-app's PreToolUse debounce pattern. Co-Authored-By: Claude Code (User Settings, in: /Users/nathan.heaps/src/nsheaps/ai-mktpl) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GdW3bEd5bo54U5eKm3njWe
Plugin Version PreviewPreview only — plugin versions and
|
![henry-nsheaps[bot]](https://avatars.githubusercontent.com/in/2371206?s=60&v=4){kind=small}
This comment was marked as outdated.
This comment was marked as outdated.
Sorry, something went wrong.
This comment was marked as outdated.
This comment was marked as outdated.
Sorry, something went wrong.
Wires in nsheaps/claude-utils' agent-plugin and agent-hook binaries where available, per the real (verified against a fresh build of packages/agent-plugin and packages/agent-hook, and a live end-to-end scratch-dir test) contract in that repo's README: - install-direnv.sh: tries `agent-plugin ensure-dependency direnv "direnv@latest"` (mise-based install; verified the mise registry maps `direnv` to aqua:direnv/direnv and `mise use direnv@latest` installs v2.37.1 successfully) before falling back to the existing direct GitHub-release download. Also routes log messages through `agent-plugin log-*` alongside shared-lib's hook-logging.sh. - direnv-check.sh: uses `agent-hook export-input` instead of hand-rolled jq parsing when agent-hook is on PATH, falling back to jq otherwise. Does NOT use agent-hook's allow/deny for the "nothing changed" case, per this repo's own hook-output-patterns.md (no opinion = no output, never an explicit allow). Both are treated as enhancements, not hard dependencies: agent-plugin's own settings file (.claude/settings.direnv.yaml) is separate from this plugin's existing plugins.settings.yaml convention, and no other ai-mktpl plugin depends on claude-utils yet. direnv.settings.yaml's autoInstall remains the single source of truth for whether to install; agent-plugin/mise is only an alternate mechanism. Debounce logic's TODO now links to the confirmed in-flight nsheaps/claude-utils#436 (agent-hook-throttle). Co-Authored-By: Claude Code (User Settings, in: /Users/nathan.heaps/src/nsheaps/ai-mktpl) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GdW3bEd5bo54U5eKm3njWe
There was a problem hiding this comment.
❌ Some changes need to be made — the two blocking findings from the last review are untouched, and the new commit adds a second code path rather than a fix
❌ Still open, re-reproduced on this revision: editing .envrc mid-session wipes every exported var and latches the fingerprint so it never retries
❌ Still open: PATH_add emits an absolute export PATH= that clobbers mise's and tool_ensure_path's PATH
⚠️ New: the agent-plugin install path logs an ERROR every session on the default configuration, for a non-error
⚠️ New: installToProject: true is silently ignored by the agent-plugin path, which does a global mise use -g
⚠️ Still open: .envrc secrets land at 0644 before chmod 600; DIRENV_* vars baked into every shell make SKILL.md's Verify steps lie
❔ New: the claude-utils adoption is dual-path with mandatory fallbacks and no tests — better as its own PR after the fixes
✅ The integration itself is correctly built: eval-safe parser, stderr-only logger, real upstream issue link — all verified against claude-utils source
🖱️ Click to expand for full details
.envrc mid-session wipes every exported var and latches the fingerprint so it never retries❌ Still open:
PATH_add emits an absolute export PATH= that clobbers mise's and tool_ensure_path's PATHagent-plugin install path logs an ERROR every session on the default configuration, for a non-errorinstallToProject: true is silently ignored by the agent-plugin path, which does a global mise use -g.envrc secrets land at 0644 before chmod 600; DIRENV_* vars baked into every shell make SKILL.md's Verify steps lie❔ New: the
claude-utils adoption is dual-path with mandatory fallbacks and no tests — better as its own PR after the fixes✅ The integration itself is correctly built:
eval-safe parser, stderr-only logger, real upstream issue link — all verified against claude-utils sourceState of the previous review
aaddf71 changed README.md, direnv-check.sh, install-direnv.sh and lib/direnv-export.sh, but none of the ten open threads were addressed. I re-ran the reproductions against direnv v2.37.1 on this exact revision rather than assuming they still held:
| Finding | Thread | Status on aaddf71 |
|---|---|---|
❌ Blocked .envrc wipes the env |
lib/direnv-export.sh:247 |
Re-reproduced. Editing an allowed .envrc → direnv export bash exits 1 with is blocked; lines 245–247 blank the output and line 259 overwrites the snapshot with just its header. |
| ❌ Re-export path never re-allows | direnv-check.sh |
Unchanged, and line 194 still commits the new fingerprint after the failed export, so it never retries. GitHub has marked this thread outdated because the file moved underneath it — the code it describes is still there. |
❌ PATH_add clobbers PATH |
lib/direnv-export.sh:259 |
Re-reproduced. PATH_add ./bin renders as export PATH=$'/tmp/dproj2/bin:/opt/other-plugin/bin:/usr/bin:/bin' — a whole-PATH absolute assignment, sourced last. |
chmod 600 after create |
lib/direnv-export.sh:261 |
Unchanged. |
DIRENV_* baked into every shell |
SKILL.md:110 |
Unchanged; the export output above still carries DIRENV_DIFF/WATCHES/DIR/FILE. |
❔ Docstring vs none sentinel |
lib/direnv-export.sh:199 |
Unchanged. |
❔ README overstates the mise parallel |
README.md:104 |
Unchanged. |
| ❔ No tests | lib/direnv-export.sh:1 |
Unchanged — and this commit added two more branches to the file. |
❔ PreToolUse table in plugins/CLAUDE.md |
hooks.json:26 |
Unanswered. |
❔ Hand-edited marketplace.json |
marketplace.json:322 |
Unanswered; entry still present at lines 339–358. |
One secondary detail worth repeating because it makes the first finding self-sustaining: direnv allow leaves .envrc's mtime and size identical (re-verified: 1786411988:19 before and after). So once the fingerprint is latched on a blocked file, the recovery SKILL.md documents cannot move it.
The new commit
I read claude-utils' source directly rather than the README, and the integration is built correctly:
agent-hook export-inputis genuinelyeval-safe.export-input.tswraps values in single quotes using the'\''close/escape/reopen sequence, and rebuilds variable names character-by-character from[A-Z0-9_]— rejecting anything that can't form an identifier — rather than trying to escape the left-hand side of an assignment. I have no injection finding againstdirenv-check.sh:81, and the file's own security commentary is accurate.agent-plugin's logger is stderr-only (log.ts: "Everything goes to stderr, never stdout"), so it can't corrupt hook JSON.AGENT_PLUGIN_NAMEis a documented equivalent of--plugin(plugin-name.ts).- The
TODO(agent-hook-throttle)now links nsheaps/claude-utils#436, which is a real, open PR in a public repo. That's a concrete improvement and closes one of my round-1 follow-ups.
The problems are at the seams, not in the parsing:
install-direnv.sh:137—ensure-dependency'sautoInstalldefaults tofalseupstream, and the decline branch logs aterrorlevel. Since theagent-pluginblock is only reached when direnv is already known to be absent, every user withclaude-utilsinstalled and no.claude/settings.direnv.yamlgets an unfilterableERROR [direnv] direnv is not installed…on every session start, immediately before the curl fallback succeeds. Same thread also covers the unredirected stdout on a function whose stdout is its return channel; one suggestion fixes both.direnv.settings.yaml:18—ensure-dependencyinstalls withmise use -g. That's a persistent rewrite of~/.config/mise/config.tomlaffecting every project on the machine, taken while this plugin's owninstallToProject: truesays project-local. Either gate the path on that setting or document the divergence.lib/direnv-export.sh:26— scope. Both new branches keep mandatory fallbacks, so nothing was replaced; theagent-hookbranch substitutes for two lines ofjqthat must stay anyway, and theagent-pluginbranch is declined-by-default upstream. Net: ~120 lines and two runtime forks added to a file that still has no tests, in a PR whose core loop has two unfixed ❌.direnv-check.sh:77— nit only: the payload travels as argv, so it's visible inps. Matches upstream's documented usage, so it's an upstream ask.
What's solid
Everything I praised last time still holds and isn't repeated here. New this round: the README.md section on the integration is unusually good — it states the fallback behaviour, names the settings-file divergence as the reason it's an enhancement, and links the upstream throttle PR. The failure shape on direnv-check.sh:77–80 (2>/dev/null || true plus a non-empty guard) degrades to jq cleanly for a malformed payload, an ARG_MAX overflow, or an older claude-utils. The commit message is precise about what was and wasn't verified.
Why 68% / 76% / 72% / 92%
- Quality 68 (was 72) — down because two ❌ in the core loop are now two revisions old while the diff grew, and the new code introduces a default-path
ERRORlog and a settings-vs-behaviour conflict. The scaffolding and documentation remain strong, which is what keeps it this high. - Security 76 (was 78) — the
chmodwindow is unchanged; new: a globalmiseconfig mutation the settings file doesn't disclose, and payload-in-argv. Offset upward by having verified theevalpath is safe rather than assuming. - Simplicity 72 (was 88) — the honest cost of two conditional integrations that can't remove the code they sit in front of.
- Confidence 92 (was 90) — up: I re-reproduced all three direnv behaviours against the shipped v2.37.1 binary on this revision, and read claude-utils'
log.ts,dependencies.ts,cli.ts,plugin-name.tsandexport-input.tsfor the exit-code, log-stream and escaping contracts instead of trusting the commit message. Residual 8%: still no live Claude Code session, soCLAUDE_ENV_FILEsource ordering across parallel SessionStart hooks stays reasoned rather than observed.
Recommended follow-ups (non-blocking):
- Raise the payload-on-stdin request with
nsheaps/claude-utilssoexport-inputdoesn't require argv, and consider asking upstream to log thedeclinedoutcome atdebug/inforather thanerror— a caller with its own fallback shouldn't have to swallow stderr to stay quiet. - Add checksum verification to
shared-lib'stool-install.shdownload helper somise,direnvand future plugins all get it — direnv publishesSHA256SUMSper release. (Repo-wide, not introduced here.) - Once nsheaps/claude-utils#436 ships, do the planned
should_check/record_checkswap and drop the duplicated debounce ingithub-app.
To improve future reviews on this repo I'd benefit from:
- A
mise run testtask wired toplugins/*/tests/*.test.sh, so plugin test coverage is a checkable convention rather than a per-plugin judgement call.
Footnotes
-
Workflow Run: https://github.com/nsheaps/ai-mktpl/actions/runs/31449463970/attempts/1 ↩
-
direnv v2.37.1 release — the binary used to re-reproduce the blocked-
.envrcexit code, thedirenv allowmtime behaviour, and thePATH_addexport shape ↩ -
nsheaps/claude-utils —
packages/agent-plugin/src/{log,dependencies,cli,plugin-name}.tsandpackages/agent-hook/src/export-input.ts, read for the log-stream, exit-code, plugin-name and shell-escaping contracts ↩
| # claude-utils or configured .claude/settings.direnv.yaml. | ||
| if [ -n "$(direnv_agent_plugin_bin)" ]; then | ||
| local via_agent_plugin | ||
| if via_agent_plugin="$(direnv_try_ensure_dependency_via_agent_plugin)" && [ -n "$via_agent_plugin" ]; then |
There was a problem hiding this comment.
ERROR on every SessionStart for something that is not an error — and ensure-dependency's stdout is unguarded inside a function whose stdout is the return channel.
I read claude-utils' actual source rather than relying on the README. Two things follow from it:
1. The decline is logged at error level. packages/agent-plugin/src/dependencies.ts defaults autoInstall to false (readAutoInstall returns false for undefined), and the decline branch is:
if (!deps.autoInstall) {
deps.log(
"error",
`${cli} is not installed. Install it with \`mise use -g ${misePackage}\`, or enable ` +
`automatic installation with \`agent-plugin settings autoInstall true\`.`,
);
return { status: "declined", cli, misePackage };
}Note the ordering in resolve_direnv_bin: this block is only reached when tool_is_available direnv was already false (line 118), so hasCommand("direnv") is false too and the decline branch is the one that runs. For any user who installs claude-utils but has never created .claude/settings.direnv.yaml — i.e. the default for everyone — every session start emits
ERROR [direnv] direnv is not installed. Install it with `mise use -g direnv@latest`, or enable automatic installation with `agent-plugin settings autoInstall true`.
…immediately followed by this plugin successfully installing direnv via the curl fallback. error is above the default info threshold, so it is never filtered. That is a scary, actionable-looking message telling the user to fix something that isn't broken, and it directly undercuts the "enhancement, not a hard dependency" framing the surrounding comment argues for.
2. $(direnv_try_ensure_dependency_via_agent_plugin) captures that function's stdout as the direnv path. agent-plugin's logger writes to stderr only today (log.ts: "Everything goes to stderr, never stdout"), so this is currently latent rather than broken — but the invocation on lib/direnv-export.sh:109 leaves stdout unredirected, so any future stdout from ensure-dependency becomes part of $via_agent_plugin and then part of $direnv_bin. Every other command in that function is careful about this (mise which direnv 2>/dev/null); this one isn't.
Both are fixed by capturing stderr and discarding stdout at the call site in lib/direnv-export.sh:
local ap_err
if ! ap_err="$(AGENT_PLUGIN_NAME="direnv" "$ap_bin" ensure-dependency direnv "direnv@latest" 2>&1 >/dev/null)"; then
# Declined (claude-utils' own autoInstall setting is unset/false — its
# default), no mise, or the mise install itself failed. All three are
# normal here: this plugin's own curl fallback handles every one of them,
# so agent-plugin's error-level decline message must NOT reach the session
# verbatim. Keep it as a debug breadcrumb instead.
[ -n "$ap_err" ] && direnv_log "debug" "agent-plugin ensure-dependency did not provide direnv: ${ap_err}"
return 1
fi(2>&1 >/dev/null — stderr to the capture, stdout to /dev/null — the order matters.)
Separately, and much smaller: the doc comment on direnv_log (lib/direnv-export.sh:71) says it routes through agent-plugin --plugin direnv log-<level>, but the code sets AGENT_PLUGIN_NAME=direnv instead. Both are supported (plugin-name.ts resolves --plugin then AGENT_PLUGIN_NAME), so this is a comment fix, not a behaviour one.
| # Install direnv to $CLAUDE_PROJECT_DIR/bin/.local. Only used when | ||
| # autoInstall is true. | ||
| installToProject: true |
There was a problem hiding this comment.
installToProject is silently ignored when the agent-plugin path wins, and that path mutates the user's global mise config.
This setting documents "Install direnv to $CLAUDE_PROJECT_DIR/bin/.local", and tool_resolve_install_dir in shared-lib honours it for the curl fallback. But agent-plugin ensure-dependency (tried first, at install-direnv.sh:137) installs via mise use -g — global. Its own help text is explicit:
ensure-dependency <cli> <mise-package>— Make sure<cli>is runnable, installing<mise-package>viamise use -g…
So on a machine with claude-utils + mise + .claude/settings.direnv.yaml: autoInstall: true, this plugin persistently rewrites ~/.config/mise/config.toml to pin direnv@latest globally — a change that outlives the session, affects every other project on the machine, and happens even though the user asked for a project-local install here. INSTALL_DIR ends up unused, and resolve_direnv_bin returns a path under mise's install root instead.
That's a bigger side effect than anything else this plugin does, and it isn't mentioned in README.md's claude-utils section, SPEC.md, or here. Two ways out, either is fine:
- Respect the setting — skip the
agent-pluginpath wheninstallToProjectistrue(the default), since a global install is by definition not a project-local one; or - Document it — say plainly in this comment and in the README that the
agent-pluginmechanism is a globalmise use -ginstall and ignoresinstallToProject.
If you keep the current ordering, this comment at minimum needs to stop claiming the install location:
| # Install direnv to $CLAUDE_PROJECT_DIR/bin/.local. Only used when | |
| # autoInstall is true. | |
| installToProject: true | |
| # Where the curl-based fallback installs direnv. When true, installs to | |
| # $CLAUDE_PROJECT_DIR/bin/.local. Only used when autoInstall is true. | |
| # | |
| # NOTE: this does NOT apply to the claude-utils `agent-plugin | |
| # ensure-dependency` path (tried first when `agent-plugin` is on PATH), | |
| # which installs via `mise use -g` — a GLOBAL install that rewrites | |
| # ~/.config/mise/config.toml and is not scoped to this project. | |
| installToProject: true |
Related: the log-noise and stdout-capture issues on the same agent-plugin call are in the sibling thread at #736 (comment).
| # --- claude-utils agent-plugin / agent-hook integration (enhancement-if-present) --- | ||
| # | ||
| # nsheaps/claude-utils ships two native binaries meant for exactly this kind | ||
| # of plugin (see that repo's README, "For plugin and hook authors"): | ||
| # agent-plugin — leveled logging + per-plugin settings + `ensure-dependency` | ||
| # agent-hook — parses a hook's JSON stdin into shell vars, emits decisions | ||
| # Installed via `brew install nsheaps/devsetup/claude-utils`. | ||
| # | ||
| # No other ai-mktpl plugin depends on these yet (mise/github-app/1pass predate | ||
| # them and use this repo's own shared-lib bash convention), and their settings | ||
| # story is genuinely different: agent-plugin reads a per-plugin autoInstall | ||
| # flag from `.claude/settings.direnv.yaml`, NOT ai-mktpl's own | ||
| # `plugins.settings.yaml` (read via shared-lib's plugin-config-read.sh) that | ||
| # this plugin's `direnv.settings.yaml` documents and every other ai-mktpl | ||
| # plugin's settings live in. Requiring users to configure autoInstall twice, | ||
| # in two different files, to get auto-install working would be a regression. | ||
| # | ||
| # So this plugin treats agent-plugin/agent-hook as an ENHANCEMENT, not a hard | ||
| # dependency: used opportunistically when on PATH, with ai-mktpl's own | ||
| # shared-lib-based logic remaining the source of truth and the fallback in | ||
| # every case. See install-direnv.sh and direnv-check.sh for where each is | ||
| # used. |
There was a problem hiding this comment.
❔ KISS/YAGNI: this integration doubles the number of code paths in a plugin that still has zero tests, and it does not change what any user gets today. Should it be its own PR, landed after the correctness fixes?
To be clear about what I verified rather than assumed — I read the upstream source, and the engineering here is sound on its own terms:
- ✅
agent-hook'sexport-inputreally iseval-safe.packages/agent-hook/src/export-input.tssingle-quotes values with the'\''close/escape/reopen sequence and rebuilds key names from[A-Z0-9_]rather than escaping them, rejecting anything that can't form an identifier. I have no injection finding against theevalondirenv-check.sh:81. - ✅
agent-plugin's logger is stderr-only (log.ts), so it can't corrupt a hook's JSON stdout. - ✅
AGENT_PLUGIN_NAMEis a documented equivalent of--plugin(plugin-name.ts). - ✅ The
TODO(agent-hook-throttle)now points at nsheaps/claude-utils#436, which is real and open — that's a genuine improvement over the bare TODO in the previous revision, and it closes out one of my round-1 follow-ups.
The concern is scope and cost, not correctness of the integration:
- Every path needs a full fallback anyway, by design ("enhancement, not a hard dependency"). So the jq parsing, the curl download, and the
hook_loglifecycle all have to stay and keep working. Nothing was removed; ~120 lines and two new runtime branches were added on top. - The
agent-hookbranch replaces two lines of jq.direnv-check.shneeds exactlytool_nameandtool_input.command. The jq that does that has to remain as the fallback regardless, so the branch buys no capability — it only adds a second way for the same two variables to get set, one of which is reachable only on machines that haveclaude-utilsinstalled. That's the branch least likely to be exercised and most likely to rot. - The
agent-plugininstall branch is declined-by-default upstream (autoInstalldefaults tofalseindependencies.ts), so for essentially every user it runs, logs, fails, and falls through — see the thread oninstall-direnv.sh:137for the error-level noise that produces, and the one ondirenv.settings.yaml:18for theinstallToProjectconflict. - There are still no tests, and this commit added the two branches whose interaction is hardest to reason about —
direnv_lognow fans out to two sinks, andresolve_direnv_binnow has four ways to produce a binary path.
Meanwhile the three findings from the previous review — all of which I re-reproduced against direnv v2.37.1 on this exact revision and none of which this commit touches — sit in the plugin's core loop. My suggestion is to land the correctness fixes plus a tests/direnv-export.test.sh first (this file is pure and stub-friendly, per the thread at #736 (comment)), and bring the claude-utils adoption back as a follow-up PR — at which point it can be reviewed on its own merits, and a test suite already exists to cover both branches of each fork.
Not a blocker on its own; the blockers are the unaddressed round-1 threads. But if the integration stays in this PR, both branches need the fixes in the sibling threads before merge.
| AGENT_HOOK_BIN="$(direnv_agent_hook_bin)" | ||
| AGENT_HOOK_EXPORTS="" | ||
| if [ -n "$AGENT_HOOK_BIN" ]; then | ||
| AGENT_HOOK_EXPORTS="$("$AGENT_HOOK_BIN" export-input "$INPUT" 2>/dev/null || true)" |
There was a problem hiding this comment.
Nit (non-blocking, and it matches upstream's own documented usage): passing the whole hook payload as an argv element puts it in the process table for the lifetime of the call, so any other local user's ps -ef sees the full Bash command the agent is about to run — which for this agent routinely includes tokens, gh invocations, and .env manipulation. The jq fallback below doesn't have this property (echo "$INPUT" | jq keeps it in a pipe).
export-input.ts's own doc block shows eval "$(agent-hook export-input "$HOOK_INPUT")", so you're following the contract as written — this is really an upstream request (accept the payload on stdin when no argument is given) rather than something to fix here. Worth raising on nsheaps/claude-utils alongside #436.
The 2>/dev/null || true + non-empty check on line 77–80 is a good failure shape, by the way: a malformed payload, an ARG_MAX overflow on a very large command, or a claude-utils version that doesn't have export-input all land silently in the jq branch rather than aborting the hook. ✅
1 participant
Summary
Adds a new
direnvplugin, mirroring themiseandgithub-appplugins' structure and conventions.install-direnv.sh): resolves a direnv binary, runsdirenv allowon the project's.envrc(and any git worktrees), then computes the.envrcdiff once viadirenv export bashand writes it as staticexport/unsetstatements to a runtime file sourced (once, idempotently) viaCLAUDE_ENV_FILE.direnv-check.sh, matcherBash): before each Bash call, checks (debounced, default 2s viacheckIntervalSeconds) whether the nearest.envrc's fingerprint (path + mtime + size) changed, and if so re-runs the export and rewrites the runtime file in place — so vars removed from.envrcare correctly dropped, not just left stale.direnv hook bash(thePROMPT_COMMAND-style shell hook) — same rationale documented at length inmise'sinstall-mise.sh: stderr/stdout pollution on every command, eval-injection risk, unnecessarycd/pushd/popdoverrides in a non-interactive agent shell.shared-lib(tool-install.sh,plugin-config-read.sh,hook-logging.sh,env-file.sh) exactly likemise/github-app/1pass.direnv.settings.yaml,README.md,SPEC.md,.release-it.js(copied verbatim from mise's pattern), and a manual-operationskills/direnv/SKILL.md..claude-plugin/marketplace.json(alphabetical position; CD bumps versions on merge per this repo'sversioning.md).On "claude-utils agent-plugin helpers"
Correction from an earlier version of this PR description: I initially reported that no such thing existed in
nsheaps/claude-utils, based on investigating a local checkout that turned out to be stale. That was wrong.packages/agent-plugin/andpackages/agent-hook/are real, documented in that repo's README ("For plugin and hook authors"), and this PR now integrates with both:agent-plugin—agent-plugin ensure-dependency <cli> <mise-package>installs a missing CLI viamise use -g <mise-package>, gated on that plugin's ownautoInstallsetting (read from.claude/settings.<plugin>.yaml, defaulting tofalse). Also provides leveled logging (agent-plugin log-info/log-warn/etc, threshold via$AGENT_PLUGIN_LOG_LEVEL) and per-plugin settings read/write.agent-hook—agent-hook export-inputparses a hook's JSON stdin into shell exports (HOOK_TOOL_NAME,HOOK_TOOL_INPUT_COMMAND, etc.) viaeval, replacing hand-rolledjqparsing. Also providesallow/deny/halt/warn-user/postfor emitting a hook's JSON decision.How I verified this for real (not just reading source): built both binaries from a clean
origin/maincheckout in an isolated worktree (bun install && bun run build), confirmedmise use direnv@latestresolves through the registry'saqua:direnv/direnvbackend and installs direnv v2.37.1, then raninstall-direnv.shanddirenv-check.shend-to-end in scratch project directories with the built binaries onPATH(and separately with them absent) to exercise every branch — see "Test plan" below.Where I landed: enhancement-if-present, not a hard dependency.
agent-plugin's settings file (.claude/settings.direnv.yaml) is a different file from this plugin's owndirenv.settings.yaml/plugins.settings.yamlconvention (the oneshared-lib'splugin-config-read.shreads, and what every other ai-mktpl plugin uses). Requiring both to be configured to get auto-install working would be a real regression, and no other ai-mktpl plugin (mise,github-app,1pass) depends on claude-utils yet — brew'sclaude-utilsformula on this machine is still 0.10.0 and doesn't even ship these binaries. So:direnv.settings.yaml'sautoInstallstays the single source of truth for whether to install at all.install-direnv.shtriesagent-plugin ensure-dependency direnv "direnv@latest"first whenagent-pluginis on PATH, and falls back to the direct GitHub-release download on any failure (agent-plugin absent, its own autoInstall declined, no mise, or the mise install failing) — zero behavior change for the 100% of users who don't have claude-utils installed today.agent-plugin log-*when present, always also through this repo'shook-logging.shlifecycle (SessionStart's summary output and failure diagnostics depend on it regardless).direnv-check.shusesagent-hook export-inputwhen present, falling back tojq. It never callsagent-hook allow/denyfor the "nothing changed" case — per this repo's ownhook-output-patterns.md, a PreToolUse hook with no opinion must emit nothing, not an explicitallow(which would bypass the permission system for every Bash call). That rule takes precedence over reaching for every agent-hook feature just because it exists.nsheaps/claude-utils#436(branchfeature/agent-hook-throttle), which generalizes this exact pattern into a reusable, sourceable helper. Swapping to it once merged is a one-function change (should_check/record_checkare isolated for exactly this).Full rationale (with more detail than fits here) is in the plugin's
README.mdunder "claude-utils integration".Test plan
plugin.json/hooks.json/marketplace.jsonJSON syntax anddirenv.settings.yamlYAML syntax validatedclaude plugin validate plugins/direnv/.claude-plugin/plugin.json— passesmise run validate-marketplace— passesshellcheckon all scripts — only info-levelSC1091/SC2317/SC2034findings, matching the same class of findings shellcheck reports on existing plugins (e.g.mise's install script)prettier --checkon all files — passesinstall-direnv.shend-to-end in a scratch project dir with a real.envrc, without claude-utils on PATH — confirmed the (unchanged) direct-download install path,direnv-envruntime file, andCLAUDE_ENV_FILEwiring are correctinstall-direnv.shend-to-end with locally-builtagent-plugin/agent-hookon PATH and direnv deliberately kept off PATH (isolatedMISE_DATA_DIR/MISE_CONFIG_DIRso nothing touched this machine's real mise state) — confirmedagent-plugin ensure-dependencyinstalls direnv viamise use -g direnv@latestfor real, resolves the binary viamise whichwhen it isn't yet on the current process'sPATH, and wires it intoCLAUDE_ENV_FILEcorrectlyagent-plugin's ownensure-dependencyto fail (untrusted mise config in the ambient directory) and confirmedinstall-direnv.shcorrectly fell through to the direct GitHub-release download rather than abortingdirenv-check.sh— confirmed: (a) no-op when nothing changed (silent, exit 0, no stdout), (b) re-exports and correctly drops a var removed from.envrc(overwrite, not append), (c) debounce window actually throttles a same-second re-check, (d)cd subdir &&prefix parsing correctly resolves a nested.envrc, (e) withagent-hookon PATH,agent-hook export-inputcorrectly parsesHOOK_TOOL_NAME/HOOK_TOOL_INPUT_COMMANDand drives the same directory-resolution and re-export logic, (f) withagent-hookabsent, thejqfallback still works identicallymise run check/mise run validatepipeline (localmise run setuphit an unrelatedaqua:cli/cliattestation/network failure installing this repo's own dev toolchain — pre-existing, unrelated to this change)claude-utilsbuild ofagent-plugin/agent-hook— the current brew formula (0.10.0) predates these binaries, so verification used binaries built directly fromorigin/mainof that repo in an isolated worktreeCo-Authored-By: Claude Code (User Settings, in: /Users/nathan.heaps/src/nsheaps/ai-mktpl) noreply@anthropic.com
https://claude.ai/code/session_01GdW3bEd5bo54U5eKm3njWe