Pkfire

pkfire — authoring Taskfile.pkl

pkf is a typed task runner: tasks declare their inputs, outputs, and dependencies in Pkl, and a content-addressed cache decides what actually runs. This skill is for whoever is editing the Taskfile.pkl.

When to invoke this skill

• A new task needs to be added or an existing task changed.
• A task is "always re-running" or "always hitting cache when it shouldn't" — that is a hashing question, not a runner bug.
• A team wants to share cache hits across machines / CI (remote cache).
• Migrating from justfile or Taskfile.yml.

Do not invoke for unrelated build-system questions (Make, Bazel, Gradle, etc.) or for editing Pkl that does not amend pkfire's schema.

Mental model

A Taskfile is a Pkl module that amends the pkfire schema and declares one local Task per unit of work, then lists them in tasks { ... }. The runner builds a DAG over deps, executes only the tasks whose action key changed, and restores the cached outputs of every other task from a CAS.

Action key = BLAKE3 over (cmd, shell, shellFlags, sorted env, sorted tools, sorted input file digests, the Pkl module's canonical form, plus any resolved CLI params and tail args when the task is the invocation target). Two invocations with the same key are guaranteed to produce the same outputs (assuming honest inputs declarations).

Above the pure run-it-once DAG, pkfire grows three orthogonal layers:

• service: true + pkf up — long-running processes (databases, dev servers) with process-group cleanup, readiness probes, and reuse-or-spawn semantics.
• acceptsArgs + params { ... } — typed CLI input (pkf run task --bump=minor -- file1 file2) folded into the action key so different invocations cache as different entries.
• pkf hooks install + pkf affected — git-aware orchestration: convention-based hook installation and changed-files-driven plan computation.

Each layer is opt-in via a schema field or a CLI subcommand; the base "one cached task per change" model is unchanged.

The two non-obvious rules

1. Every authored task must appear in tasks { ... }. Pkl's amends semantics forbid adding new top-level fields, so reflection-based auto-collection cannot work. A local foo = new Task { ... } that is not listed in tasks is dead code from the runner's perspective.
2. deps are real Task references, not strings. Write deps { build }, not deps { "build" }. A typo in a name is then a Pkl evaluation error, not a runtime DAG error.

Schema cheat sheet

class Task {
  name: String                                    // required, regex-checked
  cmd: String(length > 0)? = null                 // null = deps-only umbrella task
  shell: String = "bash"
  shellFlags: List<String> = List("-c")           // args before cmd, e.g. bash -c or node -e
  inputs: Listing<String>  = new {}               // glob; missing files silently ignored
  outputs: Listing<String> = new {}               // restored on cache hit
  deps: Listing<Task>      = new {}               // direct references
  env: Mapping<String, String>   = new {}         // contributes to action key
  tools: Mapping<String, String> = new {}         // declared toolchain versions
  cache: Boolean = true                           // false = always run
  workdir: String?     = null                     // relative to Taskfile dir
  description: String? = null
  visibility: "public"|"internal" = "public"      // hidden from list/graph unless --all
  quiet: Boolean = false                          // suppress pkfire per-task diagnostic lines
  service: Boolean = false                        // long-running, supervised by `pkf up`
  shutdownTimeoutSeconds: Int = 5                 // SIGTERM grace before SIGKILL
  services: Listing<Task> = new {}                // services to bring up while this task runs
  readyPort: Int = 0                              // TCP port to probe; doubles as a reuse detector
  readyCmd: String = ""                           // shell snippet that exits 0 when ready
  readyTimeoutSeconds: Int = 30                   // wait budget for the probe after spawning
  inheritEnv: Boolean = true                      // false = hermetic (PATH/HOME/LANG/... only)
  acceptsArgs: Boolean = false                    // forwards `pkf run task -- a b` as $@
  params: Listing<Param> = new {}                 // typed --name=value flags (see below)
}

class Param {
  name: String                     // lower-case; exposed to cmd as $NAME (uppercased)
  type: "string"|"enum"|"int"|"bool"
  choices: Listing<String>         // enum only
  default: String?                 // null = required; "10" for int, "true"/"false" for bool
  description: String?
}

class WorkflowTest {
  name: String
  changed: Listing<String>         // repo-relative files to simulate
  tasks: Listing<String>           // expected affected run plan, in order
  direct: Listing<String> = new {} // optional: direct input matches
}

Authoring template (always start from this)

amends "package://pkg.pkl-lang.org/github.com/mizchi/pkfire/pkfire@0.11.0#/Taskfile.pkl"

local sources: Listing<String> = new {
  // file globs your build reads from
}

local build: Task = new {
  name = "build"
  cmd = "go build -o bin/app ./cmd/app"
  inputs = sources
  outputs { "bin/app" }
}

local test: Task = new {
  name = "test"
  cmd = "go test ./..."
  inputs = sources
  deps { build }      // direct Task reference
}

tasks { build; test }

The package://... URI is already version-pinned, so the same checkout reproduces across machines and CI. To upgrade, bump the @<version> segment — Pkl re-resolves and re-fetches on the next run, then caches the new package locally.

Recipes (copy-paste starters)

Project layout

Three layouts cover the realistic spectrum of Taskfile sizes. Adopt them in this order — upgrade only when the previous layout hurts, not before.

A. Single Taskfile (default; up to ~30 tasks)

project/
└── Taskfile.pkl

Pkl's local function + for keep even matrix-heavy single files readable; see recipe 02. Stay here unless one of B / C is solving a real pain.

B. Split + import (single entry point, definitions distributed)

project/
├── Taskfile.pkl              # entry: amends pkfire schema, imports + spreads
├── shared/
│   ├── sources.pkl           # `sources: Listing<String>`, `toolchain: Mapping<...>`
│   └── env.pkl
└── tasks/
    ├── build.pkl             # `import "../shared/..."`; exports `tasks: Listing<Task>`
    └── test.pkl

The root Taskfile.pkl imports each fragment and spreads its tasks Listing:

amends "package://pkg.pkl-lang.org/github.com/mizchi/pkfire/pkfire@0.11.0#/Taskfile.pkl"
import "tasks/build.pkl" as bt
import "tasks/test.pkl" as tt

tasks { ...bt.tasks; ...tt.tasks }

Use this when one file is fine for the runner but humans want smaller files. Pkfire still consumes a single pkf run -f Taskfile.pkl <task> invocation, and Task references work across files because import brings real values across the boundary.

C. Hierarchical amends ("Know Your Place" pattern)

project/
├── Taskfile.pkl              # root: shared sources/tools, project-wide tasks (lint, format)
└── services/
    ├── api/
    │   └── Taskfile.pkl      # amends "../../Taskfile.pkl"; adds build:api, test:api
    └── web/
        └── Taskfile.pkl      # amends "../../Taskfile.pkl"; adds build:web, test:web

Inspired by Know Your Place from the Pkl team: the directory tree itself encodes structure. Each leaf Taskfile amends the root and tasks { ...new local tasks } appends to the inherited Listing. Run a service in isolation:

pkf run -f services/api/Taskfile.pkl ci    # only api's subgraph
pkf run -f Taskfile.pkl lint               # project-wide root tasks

Why this beats (B) at scale:

• A team owning services/api/ only edits services/api/Taskfile.pkl; cross-team reviews stay scoped.
• pkl:reflect can derive the leaf's identity (e.g. api, web) from its file path — see the upstream blog post for a findRootModule helper that walks the amends chain.
• The root file stays the canonical place for shared sources, tools, and policy tasks, so changing the lint command propagates everywhere by editing one file.

Trade-offs:

• One pkf run invocation only sees one Taskfile. There is no "umbrella ci that runs every leaf" without a small wrapper script (or a deliberately authored services/Taskfile.pkl that imports the leaves).
• Leaf authors must remember the amends URI is relative, and changing the root file's location breaks every leaf at once. Pin the root path, or have leaves amend the pkfire package URI directly if they don't actually need anything from a project-root Taskfile.

Picking between B and C

Discovery (walk-up)

pkf discovers Taskfile.pkl the way git finds .git/: when -f / --file is not specified, it walks up from the current working directory and uses the nearest ancestor that has one. Layout C benefits the most from this:

cd services/api/internal
pkf run ci      # finds services/api/Taskfile.pkl automatically
cd ../../..
pkf run lint    # finds the project-root Taskfile.pkl

An explicit -f opts out of walk-up — useful when the path is relative to the invocation directory rather than to the Taskfile.

Long-running services (pkf up)

Tasks marked service = true are long-running processes — dev servers, databases, watchers — that pkfire supervises rather than caches. pkf up <task> starts every service in the target's subgraph, blocks until Ctrl+C, then sends SIGTERM to each service's process group (so bash -c "node server.js" does not leak its node child) and escalates to SIGKILL after shutdownTimeoutSeconds (default 5).

local db: Task = new {
  name = "db"
  cmd = "exec postgres -D ./data"
  service = true
  shutdownTimeoutSeconds = 15
}

local api: Task = new {
  name = "api"
  // `until pg_isready ...` is a hand-rolled readiness wait. v1 of
  // `pkf up` starts every service simultaneously; dependent services
  // must retry until upstream is ready.
  cmd = """
    until pg_isready -q; do sleep 0.2; done
    exec node server.js
    """
  service = true
  deps { db }
}

tasks { db; api }

pkf up api               # starts db + api, blocks until Ctrl+C
pkf up --watch api       # also restarts both on a source save

Rules of thumb:

• Use exec in the cmd so the real binary replaces the shell and receives signals directly — gives one fewer pid in ps and clearer logs. The runner sets the cmd's process group regardless, so non-exec cmds still get reaped.
• pkf run rejects services in v1. The runner happily blocks forever, but pkf up is the supervisor that knows about the service lifecycle. Use pkf run for tasks that terminate.
• Non-service tasks may not depend on services. A build shouldn't deps { db } — that means "wait for db to finish", which a service never does. The reverse (db as a service that needs migrate to run first) is fine.
• Cache is implicitly disabled for service tasks — you never want to "skip" starting a server because its inputs haven't changed.

Tests that need live servers (services { ... } on the task)

For one-shot commands that need backing services for the duration of their cmd (e2e tests, smoke scripts, migration checks), declare them via services { ... } on the body task instead of running a separate pkf up:

local e2e: Task = new {
  name = "e2e"
  cmd = """
    until curl -fsS http://localhost:3000/health >/dev/null; do sleep 0.2; done
    pnpm exec playwright test
    """
  inputs { "tests/**/*.ts" }
  cache = false
  services { api }   // api in turn declares services { db }
}

pkf run e2e brings up api (and recursively api's own services like db), runs the test cmd, then stops everything in reverse order — same SIGTERM-grace-SIGKILL flow as pkf up.

services { ... } differs from deps { ... } along the "finishing" axis: deps { build } waits for build to exit successfully before this task starts; services { db } waits for db to start (a service is never expected to exit) and keeps it running for the duration. The two compose — the same task can have both.

Recipe 09 has the full picture, including a Drizzle migration check that uses just services { db } without api.

Reuse vs spawn (readyPort / readyCmd)

A service with a readiness probe is reused when the probe already passes:

local db = new Task {
  name = "db"
  cmd = "exec postgres -D ./data"
  service = true
  readyPort = 5432
  readyTimeoutSeconds = 15
}

When pkf run e2e (or any task with services { db }) starts, pkfire dials localhost:5432 once. If the dial succeeds, db is already up — typically because pkf up dev is running in another shell, or you started postgres yourself. pkfire logs reusing existing service "db", skips the spawn, and skips the teardown so the existing process keeps running after this run finishes.

If the dial fails, pkfire spawns the cmd and then polls the probe every 250ms for up to readyTimeoutSeconds before letting dependent services or the body task proceed. Without a probe the runner would race — the body would pnpm exec playwright against a server that hasn't bound its port yet.

readyCmd covers the cases TCP can't: readyCmd = "pg_isready -h localhost" (a port may be open before postgres has finished crash-recovery), or redis-cli ping, or any exit-0-when-ready shell snippet. readyPort and readyCmd compose — set both and both must pass.

Runtime args (acceptsArgs and params)

Tasks can take command-line input. Two shapes:

• acceptsArgs = true — positional tail args after -- land on $1, $2, ... and $@ (always quote: "$@"). The just analogue is just run *ARGS.
• params { Param ... } — typed named flags. The caller passes pkf run <task> --<name>=<value>; inside cmd the value is available as $NAME (uppercased). Four types:
  • "string": any value.
  • "enum": must match one of choices.
  • "int": parsable as a signed decimal integer.
  • "bool": --flag alone = true, --flag=true|false explicit. Bool params never consume the next token as a value, so --watch --port=80 is unambiguous. default = "..." makes the flag optional (the string itself is validated against type); omitting default makes it required.

local bumpVersion = new Task {
  name = "bump-version"
  cmd = "npm version $BUMP --no-git-tag-version"
  cache = false
  params {
    new { name = "bump"; type = "enum"; choices { "patch"; "minor"; "major" }; default = "patch" }
  }
}

local script = new Task {
  name = "script"
  cmd = "node \"$@\""
  acceptsArgs = true
  cache = false
}

pkf run bump-version --bump=minor      # validated, sets $BUMP
pkf run script -- src/main.ts arg1     # $@ = "src/main.ts arg1"
pkf run script --lang=ts -- src/main.ts arg1   # combined

Both forms participate in the action key when cache = true, so different invocations produce different cache entries — convenient for codegen-with-flags but typically you want cache = false for generic command wrappers (vitest filters, single-file scripts).

The CLI rejects unknown flags and missing required params before the cmd runs, so a typo like --bump=mahor fails fast with a clear message.

Environment inheritance (inheritEnv)

Default: inheritEnv = true. The cmd sees every var in pkfire's own environment (same semantics as just/make/npm run), so SSH_AUTH_SOCK, GPG_AGENT_INFO, your locale, editor preferences, and developer tokens all pass through without ceremony.

Switch to inheritEnv = false for hermetic tasks (release builds, reproducibility-sensitive CI steps). In that mode only a small allowlist passes through (PATH, HOME, USER, LOGNAME, SHELL, TERM, TMPDIR, LANG, LC_ALL, LC_CTYPE, TZ) — the pre-0.4 behavior.

In both modes the action key only hashes the env { ... } declared in Pkl, never the host environment. So a NODE_ENV shift in your shell never silently invalidates cache; if you want a host var to affect caching, copy it into env explicitly:

env { ["NODE_ENV"] = read("env:NODE_ENV") }

Publishing a task library on top of pkfire

Several patterns recur when a project grows beyond one Taskfile and becomes a library of reusable tasks — shared across repos, or published as its own Pkl package on pkg.pkl-lang.org for other projects to consume. The worked example is kawaz/pkf-tasks; the patterns below are extracted from its design records (DR-0001 … DR-0004).

Package layout

my-task-lib/
├── PklProject              # name, version, dependencies { ["pkfire"] {...} }
├── PklProject.deps.json    # generated by `pkl project resolve`
└── tasks/                  # one .pkl module per "topic"
    ├── vcs/
    │   ├── iface.pkl       # abstract module
    │   ├── jj.pkl          # extends iface.pkl
    │   ├── git.pkl         # extends iface.pkl
    │   └── auto.pkl        # extends iface.pkl — runtime dispatch
    └── docs/translations.pkl

Consumers import:

amends "package://pkg.pkl-lang.org/github.com/mizchi/pkfire/pkfire@0.11.0#/Taskfile.pkl"
import "package://pkg.pkl-lang.org/github.com/<you>/my-task-lib/my-task-lib@0.1.0#/vcs/auto.pkl" as vcs

tasks { ...vcs.allTasks }

Polymorphic modules: abstract module + extends

When the same logical task has multiple implementations (jj vs git, docker vs podman, npm vs pnpm), declare an abstract module with abstract task properties, then extends it from each implementation. Use extends, not amends — amends against an abstract module fails with Cannot instantiate abstract class.

// vcs/iface.pkl
@ModuleInfo { minPklVersion = "0.31.0" }
abstract module com.example.vcs.iface
import "@pkfire/Taskfile.pkl"

abstract commit: Taskfile.Task
abstract push: Taskfile.Task

allTasks: Listing<Taskfile.Task> = new { commit; push }

// vcs/jj.pkl
module com.example.vcs.jj
extends "iface.pkl"
import "@pkfire/Taskfile.pkl"

commit = new Taskfile.Task { name = "vcs:commit"; cmd = "jj describe ..." }
push   = new Taskfile.Task { name = "vcs:push";   cmd = "jj git push" }

The abstract property list acts as the interface contract: every implementation must fill in every abstract entry, enforced at Pkl evaluation time.

Field-level override: (base) { cmd = ...; description = ... }

To inherit one task's structure (params, env, cache, shell, ...) and only override specific fields, use Pkl's object-amends syntax:

// auto.pkl picks one impl as the base and overrides cmd:
commit = (jj.commit) {
  description = "vcs commit (auto-detect: jj or git)"
  cmd = autoCmd(jj.commit.cmd, git.commit.cmd)
}

(jj.commit) { ... } returns a new Task with the listed fields replaced and everything else (params, env, cache, ...) inherited. Cleaner than rebuilding a Task from scratch when only the cmd differs.

Runtime dispatch (Pkl can't see CWD)

Pkl evaluation is intentionally hermetic — it cannot read the working directory's filesystem state. read("file:.jj") is a syntax error (file URIs must be absolute), and absolute paths can't be portable across consumers. So "detect what the user is running on" must happen at task execution time, inside the cmd:

// auto.pkl
local function autoCmd(jjCmd: String, gitCmd: String): String = """
  if jj root >/dev/null 2>&1; then
  \(indent(jjCmd))
  elif git rev-parse --git-dir >/dev/null 2>&1; then
  \(indent(gitCmd))
  else
    echo "vcs: no jj or git repository found" >&2; exit 1
  fi
  """

Use the underlying tool's own ancestor-search (jj root, git rev-parse --git-dir, npm prefix, cargo locate-project) instead of [ -d .jj ] style cwd-only checks — the latter break when invoked from a sub-directory of the workspace.

Conventions worth following

• name = "<scope>:<action>" — namespace tasks by domain (vcs:commit, docs:check-translations, lint:pkl). The schema's name regex allows : and /, so lint:rust/fmt is also fine. Avoids collision with the consumer's own task names.
• allTasks: Listing<Task> at module level — let consumers do tasks { ...mod.allTasks } instead of listing each task by name. Keeps the consumer's Taskfile small and lets you add new tasks in the library without every consumer updating their import line.
• Avoid read("env:X") for ergonomics-only env vars. It fails with "resource not found" when X is unset, breaking eval in environments (CI containers, sandboxes) that don't have your developer-machine env. Rely on inheritEnv = true (the default) for SSH_AUTH_SOCK, GPG_AGENT_INFO, locale, etc. Use read() only when the value should be part of the action key (a release channel, a git SHA, a NODE_ENV you want to cache against).
• Pin pkfire as a dependency, not via amends to a URL. Library authors should declare pkfire in their PklProject.dependencies and reference it as @pkfire/Taskfile.pkl — that way the consumer gets a single Pkl package graph and pkl project resolve keeps things consistent.

Releasing the library

Mirror pkfire's tag scheme: <name>@<version> as the release tag, because packageZipUrl in PklProject expects that shape. Once v0.1.0 of your package ships, the major-only pin (@<name>@1) becomes resolvable by Pkl's package resolver. Stay exact-pinned during 0.0.x.

CI cache for Pkl downloads

Pkl caches resolved packages under ~/.pkl/cache. Add this to any workflow that runs pkl project resolve or evaluates remote Pkl packages — first-run network cost stays in the cache instead of hitting pkg.pkl-lang.org on every CI run:

- uses: actions/cache@v4
  with:
    path: ~/.pkl/cache
    key: pkl-cache-${{ hashFiles('**/PklProject.deps.json') }}
    restore-keys: pkl-cache-
- uses: mizchi/pkfire@v0.11.0
- run: pkl project resolve

Git hooks (pkf hooks install)

pkf hooks install is a thin opt-in convenience: any task whose name matches a git client-side hook event (pre-commit, pre-push, commit-msg, prepare-commit-msg, post-commit, post-checkout, post-merge, pre-rebase, post-rewrite) becomes installable. The install step writes .git/hooks/<event> as a 3-line shim:

#!/bin/sh
# managed by pkf hooks install
exec pkf run pre-commit -- "$@"

The trailing -- "$@" forwards git's per-hook args. commit-msg receives the message-file path as $1 (your task should declare acceptsArgs = true to receive it inside cmd); pre-push gets the remote name and URL; etc. See git help githooks for the per- event arg list.

pkf hooks install         # write shims for every matching task
pkf hooks install --force # also replace hand-written or other-tool hooks
pkf hooks list            # report which events have tasks + which are installed
pkf hooks uninstall       # remove only pkfire-managed shims

The shim is marked with a comment so uninstall and reinstall are idempotent and will not touch hooks installed by other tools (husky, lefthook, hk, etc.) unless you pass --force.

Auto-install on cd (direnv)

pkf hooks install is silent on no-op: when every matching shim is already present and bit-identical to what pkfire would write, the command produces no output. That makes it safe to drop into .envrc so the hooks always get re-installed (and re-aligned to the current Taskfile) the first time you cd into the repo on a fresh checkout:

# .envrc
if command -v pkf >/dev/null; then
  pkf hooks install
fi

The first direnv reload after the Taskfile changes writes a line to stdout. Every subsequent reload is quiet. File writes go through tempfile + rename so a concurrent reload (two terminals hitting cd at the same moment) won't tear the shim.

Rules of thumb

• cache = false on every hook task. Hooks fire on a tree that shifts every commit; the action key (which is keyed on declared inputs) is not a meaningful cache for "did the staged set change". Set cache = false explicitly. pkfire warns when an installable hook task has cache = true.
• Don't reuse the same task name for the hook target and the underlying gate. Declare a thin pre-commit task that deps { lint; typecheck } instead of inlining everything — it keeps the hooks layer (config: which event → which task) separable from the gate logic.
• Staged-file scoping happens in cmd, not in inputs. pkfire does not currently inject a $STAGED_FILES env or filter inputs to the staged subset. If you want lint-only-staged-files behavior, write git diff --cached --name-only inside the cmd and feed the result to your linter. Recipe 13 shows the pattern.

Relationship to hk / lefthook / pre-commit

hk (by the mise author) is a purpose-built git-hook manager configured in Pkl. It has features pkfire deliberately doesn't: staged-file globs with {{files}} templating, separate check vs fix commands, exclusive locks, file-set scoping, fail-fast. If you need those, use hk and have its steps shell out to pkf run <task> for the actual work — the two compose cleanly.

pkf hooks install exists for the case where the project already runs everything through pkfire and a single pkf run pre-commit is enough; adding a second config tool just to dispatch a hook is overhead. Same logic for lefthook and pre-commit — pkfire's built-in is the dependency-light path; the dedicated managers are the feature-rich path.

Common pitfalls

• A non-local object property cannot have a type annotation — you wrote field: Task = ... inside an amends. Use a local binding and add it to tasks { ... }.
• Self-shadowing names cause stack overflow — a function with (p, deps: Listing<Task>): Task = new { deps = deps } recurses forever because deps resolves to the field on the value being built. Rename the parameter (taskDeps).
• new {} inside Listing<Task> is ambiguous — write new Listing<Task> { taskA; taskB } to disambiguate.
• outputs parents are not auto-created — your cmd must mkdir -p bin && ... if it writes to bin/app and bin/ does not exist yet (some tools, like go build -o, do this themselves).
• Inputs with literal paths that do not exist are silently ignored — useful tolerance for optional files (e.g. go.sum in a no-deps Go project), surprising when you misspelled a real file. Use pkf run --print-hash <task> to see what actually contributed.

Selective execution

Real workflows rarely "run every task". A few patterns cover the common selection questions:

pkf run                        # no args → the task named `default` (errors clearly when absent)
pkf run -- a b c               # no task name + tail args → `default` receives "$@"
pkf run a b c                  # multi-target: topological union of all subgraphs
pkf run --refresh build        # skip cache lookup but re-store result (re-baseline)
pkf run --no-cache test        # skip lookup AND store for this run

Monorepo CI: pkf affected --since=<ref>. Runs only the tasks whose declared inputs glob matches a file in the asymmetric diff <since>...HEAD, plus their transitive dependents in the deps DAG. The unaffected deps that come along in the plan hit cache, so the actual work is minimal.

pkf affected --since=origin/main           # everything affected by the PR
pkf affected --since=origin/main test:unit # filter to specific targets (exact names)
pkf affected --since=HEAD~1 --dry-run      # preview without running
pkf affected --files src/main.go --explain --dry-run  # check file -> task matching
pkf affected --check                       # run workflowTests from Taskfile.pkl

Default --since chain: origin/main → origin/master → HEAD~1. Tasks with empty inputs are never affected by file changes (they explicitly declared no file dependencies). Tasks with cache = false are still subject to the same affected-set test — they don't get auto-pulled in just because they're always-run.

Use workflowTests { ... } next to the tasks when authoring a new Taskfile or changing inputs / deps; it is the Taskfile-level equivalent of a unit test for the affected workflow.

CLI tools

Inspection (read-only)

pkf list                       # public tasks, one per line + description
pkf list --unsorted            # Taskfile declaration order
pkf list --all                 # include visibility = "internal" tasks
pkf list --color=always        # force ANSI color (auto, always, never)
pkf list -v                    # cmd preview + deps + cache status
pkf list --long --all          # compact audit table: visibility, cache, quiet, deps, io, shell
pkf list --json                # machine-readable; for editor / CI tooling
pkf graph                      # Graphviz DOT
pkf graph --format mermaid     # GitHub-renderable
pkf graph --format tree        # terminal-readable dependency tree
pkf graph --target build       # only the subgraph rooted at `build`
pkf affected --files PATH --explain --dry-run  # show matching inputs + affected closure
pkf affected --check           # verify workflowTests
pkf lint                       # dead local tasks + suspicious task definitions
pkf lint --json                # structured findings for editor / CI tooling
pkf lint --fix --dry-run       # preview safe edits; currently cache=false for outputs-without-inputs
pkf doctor                     # diagnose pkf PATH / pkl / cache / remote / Taskfile setup
pkf doctor --json              # structured environment checks
pkf doctor --fix --dry-run     # preview replacing a stale pkf on PATH

pkf lint --fix is deliberately narrow. It only edits a task when the safe default is unambiguous: a cacheable task declares outputs but no inputs, so pkfire adds cache = false. Findings such as dead local tasks, no-op tasks, and services without readiness probes remain suggestions because automatic edits could change intent.

pkf doctor --fix can replace the pkf binary found on PATH with the currently running binary, backing up the old file first. Always start with --dry-run and inspect the path before applying it.

Execution preview

pkf run --dry-run build        # table: per-task hit/will-run/uncached + action key + cmd
pkf run --print-hash build     # raw action keys for the subgraph
pkf run --no-cache --dry-run build   # "what would --refresh re-run?" introspection

The --dry-run table classifies each task as hit (cache will short-circuit), will run (cacheable but no entry), uncached (cache = false), or service (skipped by pkf run — preview only). With --no-cache / --refresh, the lookup is inert so every cacheable task shows will run. Remote cache is NOT consulted during dry-run to keep it fast.

Execution

pkf run build                  # run + cache as usual
pkf run --watch build          # re-run on input change (Ctrl+C to stop)
pkf run -j 8 build             # cap parallelism (default: NumCPU)
pkf run --timing build         # add per-task duration breakdown at the end
pkf up dev                     # supervise every `service: true` task in the subgraph
pkf up --watch dev             # plus restart-on-change

After every non-watch run, pkfire prints a single summary line:

[pkf] done: 6 tasks · 3 hit · 3 ran · 2 uncached (11.3s wall, 12.0s CPU)

--timing follows with a table sorted by duration descending — pinpoints the slow task without a profiler.

Maintenance / hooks

pkf format                     # pkl format -w (defaults to Taskfile's dir)
pkf format --check pkl examples skills  # exit 11 on unformatted (CI-friendly)
pkf hooks install              # write .git/hooks/<event> shims for hook-named tasks
pkf hooks list                 # show which events are wired
pkf hooks uninstall            # remove only pkfire-managed shims
pkf init                       # write a starter Taskfile.pkl

pkf hooks install is silent on no-op, so it's safe in .envrc — see the Git hooks section.

Remote cache (optional)

export PKFIRE_REMOTE_CACHE=https://pkfire-cache.<account>.workers.dev
export PKFIRE_REMOTE_TOKEN=<bearer token>

Local cache stays the source of truth; the remote is consulted on a local miss and warmed on a remote hit. See examples/remote-cache-worker/ in the upstream repo for a Cloudflare Worker reference backend that GCs entries older than CACHE_TTL_DAYS (default 7) on a daily cron.

Running pkfire in GitHub Actions

jobs:
  ci:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v5
      - uses: mizchi/pkfire@v0.11.0      # or @v0 to track latest 0.x
      - run: pkf run ci

The mizchi/pkfire action is a setup-only composite that installs the matching pkf binary plus the Pkl CLI on the runner (linux/darwin × amd64/arm64) and adds them to PATH. After it runs, the rest of the workflow uses pkf directly — no extra tooling steps.

Do not write uses: mizchi/pkfire@pkfire@0.11.0. GHA's parser treats the @ in the ref as a syntax break and fails the whole workflow file, with no jobs run and a generic "workflow file issue" error. Pkl release tags happen to be pkfire@<ver> (because the package URI requires it), so the Release workflow also pushes v<ver> + a floating v<major> tag at the same commit — use those for uses:. Or SHA-pin: uses: mizchi/pkfire@<sha> # v0.5.0.

To share cache hits across CI runs and developer machines, point pkf at a remote cache via env:

      - uses: mizchi/pkfire@v0.11.0
      - run: pkf run ci
        env:
          PKFIRE_REMOTE_CACHE: ${{ vars.PKFIRE_REMOTE_CACHE }}
          PKFIRE_REMOTE_TOKEN: ${{ secrets.PKFIRE_REMOTE_TOKEN }}

Inputs:

Source: mizchi/pkfire — distributed by TomeVault.