pi-dev-tools

Published to npm as @nredd/pi-dev-tools (public). From 0.5.1, each version is cut by the monorepo Publish workflow from a signed @nredd/pi-dev-tools@<version> tag, with npm provenance and a GitHub release carrying the tarball, checksums and manifest (docs/releases.md); earlier versions were published manually without provenance. 0.6.0 moves resource_snapshot onto the published @nredd/pi-resource-sampler 0.1.1 (breaking structuredContent, see Resource awareness). 0.5.1 adds createDevTools(options) (injectable ssh, py, cargo tool options; the default export is unchanged) and fixes process output that dropped a short line sharing a chunk with an oversized one. Use the monorepo development commands. References below to standalone repositories, chezmoi pins, and installed-runtime harness gates describe the legacy deployment, not verification of this package.

Standalone use

The packed tarball is self-contained: Pi host modules (@earendil-works/pi-*, typebox) are * peers supplied by Pi, there are no runtime dependencies and no sibling plugin is required. Verified by installing the tarball alone and running it in exact stock Pi 1.3.0 (see the repository's docs/package-contracts.md). Optional fork integrations (subagents, Plan mode) are absent in that setup; the exercised behavior does not need them.

Typed pi tools for the workflows agents otherwise drive through raw Bash: git, ssh, Python (uv, ruff, ty, pytest) and Rust (cargo). Each tool spawns the real binary with an argv array, maps exit codes to named states, bounds output, and returns structuredContent alongside the compact text the model reads.

Tool groups land one per PR. Current state is in the table below.

Conventions every tool follows

• No shell. spawn(cmd, argv), stdin closed or fed a fixed script, child in its own process group, SIGTERM then SIGKILL 2s later on timeout or abort, hard resolve 1s after that so a grandchild holding the pipes cannot hang the caller. macOS and Linux only
• When a command settles, including a normal leader exit, whatever is left of its process group is SIGKILLed, so a background descendant cannot outlive the tool call. Only the owned group is signalled (never a bare pid). SSH is the exception: the local ssh is the group, so remote processes it started can keep running after the local client exits
• Output is capped at pi's DEFAULT_MAX_BYTES / DEFAULT_MAX_LINES, truncation notice included. When cut, the full text is written to $TMPDIR/pi-dev-tools-*/<label>-N.txt and the path is named in the result
• run caps captured stdout/stderr independently. Its stderr line listener drops only individual oversized lines, through their newline or EOF; valid following lines are delivered even when coalesced into the same chunk or after the capture cap is reached. Listener errors are swallowed
• Exit codes become named states (violations, no-tests, tool-error, ...), never a bare "exit 1"
• annotations.readOnlyHint is set honestly. The pi-plan-mode fork admits annotated read-only tools in Plan mode; everything else is denied there
• outputSchema + structuredContent on every tool, isError: true on red with the data intact
• Default transcript renderer, so every call settles as one call · summary row

Resource awareness

resource_snapshot({process_limit?, interval_ms?}) is an always-read-only observer for a parent or subagent to call before it knowingly starts a large test or compile. A before_agent_start system-prompt hook carries this guidance in parent and child sessions even while the tool is deferred: discover it with the available tool-search facility first, then use the result to choose targeted scope, workers, or timing. It never forces activation, schedules work, locks commands, or changes concurrency.

• Collection is the published @nredd/pi-resource-sampler, pinned to exactly 0.1.1 as a runtime dependency (installed from the registry, never the workspace copy; a test asserts the resolved module). npm would link the same-versioned packages/resource-sampler workspace, so the root lockfile carries an explicit nested packages/dev-tools/node_modules/@nredd/pi-resource-sampler registry entry with the published integrity. npm ci/npm install keep it; npm dedupe or regenerating the lockfile drops it, and that test then fails. Each call creates a private sampler, samples twice interval_ms apart (default 500, 100..5000), then disposes it, which awaits any helper reaping. No background polling or state survives the call.
• Reports logical CPU count, 1/5/15-minute load (null on Windows), aggregate host interval CPU cpu.utilization (0..100) and interval.{requestedMs,measuredMs}, total/free memory, Linux MemAvailable, and optional macOS memory_pressure -Q free percentage or Linux PSI some avg10 from the sampler. Missing optional metrics remain null, never invented. Sampler 0.1.1 copies os.cpus() counters when sampled, so Bun's lazy counters no longer produce a false 0 or 100.
• Returns at most 50 process entries, 12 by default: PID, parent PID, interval CPU, RSS bytes and executable (see privacy below). Interval CPU is a typed metric (available with a value where 100 means one logical core, otherwise warming or unavailable with a reason), never a zero stand-in. Linux measures it; Darwin cannot (one-second ps start precision cannot exclude PID reuse), so it is unavailable there. ranking is interval-cpu (measured entries first, then RSS) or rss. No ps lifetime or decaying average is reported or relabelled.
• coverage passes through the sampler's bounded scan (complete, enumerated, sampled, skipped, bytesRead, finite reasons); truncated is !coverage.complete. The scan is not atomic: a process exiting mid-scan marks it partial (complete: false) even when nothing else was cut, and short-lived work may be missed.
• executable is the sampler's kernel executable basename: Darwin ucomm (kinfo p_comm, written at exec from the executable file name, cut to 16 bytes), Linux the basename of readlink /proc/PID/exe. Never argv, process title, Linux comm (node-MainThread), environment or user. Sampler 0.1.0 reported Darwin ps -c comm, which is argv[0] and leaked titles such as npm exec rea-agents@6.5.0 mcp; 0.1.1 fixes that. "" means unknown (Linux processes of other users without ptrace access, unreadable or over-budget links) and is never filled in; text output prints (unknown). Whoever names, links or copies the executable file chooses the name, so it stays untrusted display text, bounded and control-stripped. Cancellation and sampler failures reject the call rather than returning a partial snapshot, including a cancellation that lands after the second sample while the optional memory probes run. The optional Darwin pressure probe gets the same signal, and failed, aborted, timed-out or truncated pressure probes stay null. A disposal failure, rejected or thrown synchronously, never hides the sample, probe or cancellation error that caused it; after a successful snapshot it rejects the call.
• Parent PIDs let a caller see real OS process trees, but OS process metadata has no proven link to a Pi parent, session, or agent. Use actual subagent/task information separately; do not claim process-to-agent attribution from this tool.
• macOS and Linux process metadata are supported. Other platforms return an explicit unavailable reason while retaining the portable memory fields.

API change from 0.5.x (structuredContent): entries drop cpuPercent (ps average) and elapsed, which the sampler does not measure, and gain parentPid plus the cpu metric. Additive: interval, cpu.utilization, processes.ranking, processes.coverage, and the interval_ms parameter. cpuSemantics, truncated, limit and the memory fields keep their names. The resourceSnapshot(limit, options, signal, intervalMs?) seams are now createSampler, readFile and run; the per-OS-primitive seams are gone.

Release: removing structuredContent fields is breaking, so this ships only as @nredd/pi-dev-tools 0.6.0, never a 0.5.x patch (0.5.2 was never published). It is cut by the signed-tag publish.yml flow in releases. The pin is the published sampler 0.1.1 (registry integrity sha512-rPwZxqqG...wisdwg==), whose process-name privacy and Bun CPU fixes this API depends on; 0.1.0 must never be pinned again.

git

Reads (git_status, git_diff, git_log, git_show) are annotated read-only and run with --no-pager --no-color --literal-pathspecs, GIT_TERMINAL_PROMPT=0, GIT_EDITOR=true, LC_ALL=C, timeout_s default 60. Refs (base, ref, rev) may not start with - and go after --end-of-options. git_status parses porcelain v2 (-z, renames and copies carry the old path, a gone upstream reads gone). git_diff leads with the stat and caps the patch; git_log is sha date author: subject, max 200; git_show peels tags to the commit. Failures name a state (not a git repository, unknown revision, path not tracked, no upstream, push rejected) and return { error } as structured content.

git_commit {subject, why?, changes?, paths, push?=true}:

• paths is required; the repo root in any spelling, .. segments, pathspec magic (:/) and flags are rejected. The commit is git commit -- <paths>, so it contains exactly those files; anything else already staged stays in the index. A rename needs both the old and the new path; a deletion (unstaged or already staged) is named like any other path. Nothing to commit is its own state
• The subject must be a capitalized imperative, <= 72 chars, no trailing period, no conventional-commit prefix, no em or en dash. An area: prefix is fine
• Message is subject, optional why, Changes: bullets, then Co-Authored-By: Claude <noreply@anthropic.com>, passed on stdin with -F -
• Never --no-verify, --amend or --no-gpg-sign. Hook output is relayed on failure, and files were modified by this hook is flagged as "re-stage and call again". Nothing retries
• Push is git push <remote> HEAD where <remote> is branch.<n>.pushRemote, else remote.pushDefault, else branch.<n>.remote, else the only remote (with -u). Several remotes and none configured fails closed at stage: "push"; it never defaults to origin. A branch.<n>.remote of . (local tracking) is not a push remote and counts as unconfigured and ambiguous. Only an unset config key (exit 1) is tolerated, and all three keys are read eagerly, so an unreadable one blocks the push even when a higher-precedence key is set; a broken config or git remote fails closed with git's stderr. git push <remote> HEAD pushes to the same-named branch on <remote>. A push failure returns stage: "push" with the sha so the caller fixes the push instead of committing twice. signed is true for %G? in G, U, X, Y

Bash guard (tool_call on bash, always on): blocks --no-verify (and -n on commit, where it means --no-verify; push -n is a dry run and passes), plain --force / -f / +ref pushes including clustered -fu (not --force-with-lease), and --no-gpg-sign or -c commit.gpgsign=false. Quoted strings are ignored and git must be the command word of its segment or substitution, so echo 'git push --force' passes and $(git commit -n) does not. Known limits, by design (Plan mode is the policy, this is a seatbelt): /usr/bin/git, sudo git, command git, xargs git, and bash -c '...' bodies are not inspected.

git worktrees

git_worktree_list is annotated read-only (Plan mode admits it); add, remove and prune are not. All run the real git with the same env and flags as the other git tools, take cwd and timeout_s (default 60), and return { error } on failure with a named state in the text. Mutations (add, remove, prune) take two locks keyed on the git COMMON dir: withLock in-process, then withFileLock across processes, so concurrent pi panes cannot race on .git/worktrees either. The file lock is an atomic mkdir of <common dir>/pi-dev-tools-worktree.lock holding owner.json {pid, startedAt, token}, always released in finally (token-checked, never throws). Only a lock whose owner pid is dead is reclaimed (plus an ownerless directory older than 2s); there is NO age-based reclaim, a live pid is never broken. Reclaimers serialize behind a short-lived pi-dev-tools-worktree.lock.reclaim guard directory and re-verify the same token and a dead pid inside it. The guard records its owner with an immutable <pid>.<token> filename. Only a confirmed dead guard owner is reclaimed, never a live owner based on age. Release/reclaim unlink only the token-owned filename, then remove the empty guard directory, never a replacement owner's contents. Empty or malformed guards (including a crash before owner publication) require manual inspection; bounded lock-timeout errors name the guard path. Residual risk: PID reuse. A recycled pid makes a stale lock look live; waiters then fail with lock-timeout naming the pid, and a human can remove the lock directory. If owner.json cannot be written the directory is removed and the call fails with lock-failed. Waiting is bounded at 15s, then the call fails with lock-timeout naming the holder pid. list takes no lock. list and prune text is capped like every other tool (truncation notice, full text spilled to a temp file); structured content stays complete.

• git_worktree_list: parses git worktree list --porcelain -z into path, head, branch (short name), detached, bare, main (first entry), locked/lockReason, prunable/prunableReason. Text is one line per worktree
• git_worktree_add {branch, base?, path?}: default path is <main worktree>-<slug> next to the main worktree, where slug is the branch lowercased with anything outside [a-z0-9._-] -> -. A missing branch is created with -b from base (default HEAD); an existing branch is checked out, base is refused for it, and a branch checked out in another worktree is refused with that path (branch checked out elsewhere). Also refused: invalid branch name, unknown base, existing non-empty path. branch and base may not start with - and go after --end-of-options
• git_worktree_remove {path, force?}: refuses the main worktree, the session cwd (or cwd) and any parent of it (cwd is inside), a locked worktree even with force, and a dirty one (modified or untracked) unless force. A worktree whose directory is gone points at prune. The branch is always kept
• git_worktree_prune {dry_run?}: git worktree prune --verbose; reports {name, reason} per record, dry_run changes nothing

Preflight checks (branch exists, base resolves, path state) separate "no" from "could not tell": a timeout, abort, spawn failure or unexpected exit is reported as such, never as a missing branch or bad revision. An unreadable target path is path unreadable (EACCES); a path removed mid-check counts as free. If the final HEAD lookup fails after the worktree was created, the error says head-unreadable and that the worktree exists.

States: branch checked out elsewhere, branch exists, path exists, path unreadable, dirty worktree, locked worktree, main worktree, lock-timeout, lock-failed, head-unreadable, plus the shared not a git repository, unknown revision. The Bash guard does not steer raw git worktree ...; it only blocks hook-skipping and force pushes, as for every other git command.

ssh

ssh_exec {host, command, cwd?, timeout?=30} runs the user's own ssh with -T -a -x -o BatchMode=yes -o ConnectTimeout=8 -o ClearAllForwardings=yes -o RequestTTY=no -- <host> 'exec sh -s', so the per-host ForwardAgent/ForwardX11 in ~/.ssh/config is overridden and nothing can prompt.

• host must match ^[A-Za-z0-9][A-Za-z0-9._-]*$ and be a concrete Host alias in ~/.ssh/config (Includes followed, wildcards and ! entries skipped). Option injection is impossible by construction and the -- keeps it that way
• The command travels on stdin: written to a remote mktemp file through a quoted heredoc with a per-call random delimiter, run as sh "$tmp" </dev/null inside ( ... ) (so a cat or read in the command cannot eat the wrapper), then a printf '\n__PI_RC=%d' sentinel. A trap removes the temp file even when the local ssh dies first. Inside the command $0 is the temp path and there are no positional arguments. No remote argv quoting, no command in ps, multi-line scripts and heredocs work, an exit N ends the command rather than the shell, and unbalanced quotes or parentheses stay inside the file (sh exits 2 with a syntax error)
• States: ok, command-failed (remote exit, including 255; NOT a tool error, grep exiting 1 is normal), cwd-missing (cd failed: marker plus exit 97, so neither a command exiting 97 nor one printing the marker can fake it), ssh-failed (no sentinel: connection failure or a remote without mktemp), timeout, aborted. A timeout or abort kills the local ssh only; the remote command may keep running
• Output and the spill file are redacted for Bearer ..., token|secret|password|api_key style assignments in any case or quoting, and JWT signatures. Count-like keys (TOKEN_COUNT=) are over-redacted on purpose
• No ControlMaster: measured steady-state gain was ~70ms per call, not worth stale sockets
• No stdin parameter (the plan had one): the script already owns stdin, a heredoc in command covers the need, and Plan mode's gate refuses the key anyway
• Plan mode: the pi-plan-mode fork admits ssh_exec as annotated read-only and re-validates {host, command} against safeSubcommands.ssh and the remote reviewed policy, so rm, cd, redirects and pipes to interpreters are blocked there. Outside Plan mode the tool runs whatever it is given, like bash

ssh_pull {host, remote_path, local_path, overwrite?} and ssh_push {host, local_path, remote_path, overwrite?} use scp -O -q with the same options and refuse to overwrite unless asked (ssh_push probes the remote path with test -e first; a small race remains between probe and copy). Legacy scp hands remote_path to the remote shell, so it is restricted to [A-Za-z0-9._/+@%:=,-]: no spaces, quotes, ~ or metacharacters. ssh_pull is read-only, ssh_push is not. Failures are named: missing, permission denied, connection failed.

~/.ssh/config parsing follows OpenSSH: relative Include paths resolve against ~/.ssh, not the including file; Host=x, quoted names, trailing comments and CRLF are handled.

py

All five run uv run --frozen <tool> from the nearest pyproject.toml above cwd (the walk stops at a .git boundary, so a stray ~/pyproject.toml cannot capture a repo), so the project's pinned ruff/ty/pytest apply and uv.lock is never rewritten as a side effect (plain uv run re-locks when pyproject.toml drifted). uv's own failures (no lockfile, tool not in the environment) are reported as such, not as findings. Calls on one project are serialized through the shared lock; a waiting call says so. NO_COLOR and UV_NO_PROGRESS are set. Every positional the model supplies (paths, targets, select, packages) must be a plain value and goes after --; -k/-m travel as -k=EXPR.

• py_lint {paths?, fix?, select?, limit?=50}: ruff check --output-format concise; fix adds --fix --ignore-noqa (the Makefile gate's form) and reports fixed. States: clean, findings (exit 1), tool error (exit 2). One path:line:col CODE [fixable] message per finding; lines the parser does not recognise are kept under unparsed instead of dropped
• py_format {paths?, mode: check|diff|apply}: ruff format --diff for check (file list) and diff (plus the capped diff), write mode for apply. States clean, would-reformat, reformatted
• py_typecheck {paths?, limit?}: ty check --output-format concise, parsed as file:line:col severity[rule] message
• py_test {targets?, k?, markers?, maxfail?=5, last_failed?, coverage?}: pytest --verbosity=-1 --tb=short -p no:cacheprovider --maxfail=N (--verbosity=-1 is absolute, so a project addopts = "-q" cannot hide the summary; --lf re-enables the cache plugin; coverage adds --cov=src --cov-report=term-missing). Exit codes map to passed, failed, interrupted (2, collection error), usage-error (4), no-tests (5, not an error). Only failures are listed, from the short summary, with the failure block below them; stoppedEarly says the run hit --maxfail. There is no extra_args: verbatim flags such as --basetemp or -p would let a read-only tool write, and Plan mode admits this tool
• py_uv {action: sync|lock|add|remove, packages?, group?, dev?, all_extras?=true, all_groups?=true, locked?}: sync defaults to --all-extras --all-groups (the Makefile uses --all-extras; groups are added so dev tools are present too); locked fails loudly on a stale lock

Annotations: py_typecheck and py_test are read-only and admitted in Plan mode. py_lint and py_format carry a rewriting mode, so they are not (known cost: no lint or format check in Plan mode; splitting the fix modes into their own tool would lift it); py_uv reaches the network. All five are deferred by default (found via search_tools), per the always-active budget.

cargo

All nine run from the outermost Cargo.toml with [workspace] above cwd (else the nearest manifest), default to --all-features, accept package (-p) and toolchain (rustup run), set CARGO_TERM_COLOR=never, and serialize per workspace through the shared lock (keyed on the workspace root, not the target dir); cargo's Blocking waiting for file lock line and Compiling progress are relayed as updates. Every bare value the model supplies (package, features, bin, ...) must be a plain identifier-ish token; crates for cargo_add may also carry version requirements (serde@^1); filters only need to not start with -. A relative CARGO_TARGET_DIR resolves against the workspace root. Criterion lookups honor that environment variable, but do not consult build.target-dir from .cargo/config.toml.

Compile diagnostics (cargo_check, cargo_clippy, cargo_build, cargo_doc) come from --message-format=json-diagnostic-short: keyed on reason, errors and warnings only, deduped on (level, code, primary span, message) because cargo emits one copy per target, with the ~900-char code.explanation dropped and the first suggested_replacement shown. One file:line:col: level[code]: message per finding, errors first, capped by limit. States clean, warnings (exit 0 but counted), errors. A non-zero exit with no compiler errors (manifest error, missing toolchain) is a tool failure with cargo's stderr.

• cargo_check {all_targets?=true} mirrors make check
• cargo_clippy {pedantic?=true, fix?, deny_warnings?} mirrors make lint (--all-targets --all-features -- -W clippy::pedantic); fix adds --fix --allow-dirty --allow-staged
• cargo_fmt {check?=true}: cargo fmt --all -- --check --color never (rustfmt colours even when piped); Diff in <file>:<line>: blocks give the file list
• cargo_test {filter?, no_fail_fast?=true, doc?=true, retries?, nocapture?}: cargo nextest run with a pi-dev-tools profile layered UNDER the repo's own nextest config via --tool-config-file (repo retries, slow timeouts and groups still apply), junit written to an absolute path in a per-call temporary directory, removed after the run. Report lookup is independent of CARGO_TARGET_DIR, Cargo's build.target-dir and nextest's [store].dir (nextest's store does not follow Cargo's target directory). A repo override of the pi-dev-tools profile's junit path is reported as missing output. <failure> and <error> both count, flaky retries do not; then cargo test --doc --no-fail-fast, skipped (and said so) when a filter is set. A filterset expression (any of ( ) | & ! or a leading not) goes to -E, a bare string is a substring match. nextest's exits map to passed, failed (100), no-tests (4), compile-error (101). Passing tests are never listed; a doctest failure flips passed to failed
• cargo_build {bin?, example?, release?} returns the executables from compiler-artifact
• cargo_run {bin?, example?, args?, stdin?, release?}: stdin closed unless given, output capped; a non-zero exit is nonzero-exit, not a tool error, and only cargo's own could not compile counts as compile-error (a program may print error: and exit 101)
• cargo_doc: --no-deps with RUSTDOCFLAGS=-D warnings, like the docs gate
• cargo_bench {bench?, filter?, save_baseline?, baseline?}: reads criterion's new/estimates.json (ns) and change/estimates.json (fraction) written since the run started; baseline flags imply --benches because libtest harnesses reject them
• cargo_add {crates, dev?, build?, features?, optional?, dry_run?} reports name@version

Annotations: cargo_check, cargo_test, cargo_doc are read-only; cargo_add reaches the network; the rest write the tree or target/. All nine are deferred (found via search_tools).

Development

Wrapper coverage measurement, executable-boundary scope, fixture isolation, and Node/Bun commands: coverage evidence. External tool implementations are not part of this package's coverage denominator.

# From the monorepo root:
make install
npm run check --workspace @nredd/pi-dev-tools  # biome, tsc --noEmit, vitest, bun test test/scenarios

Harness gate before any tag: HARNESS_LOCAL=pi-dev-tools make verify in ~/code/pi-harness-e2e. Pin lives in chezmoi third_party/pi-agent-settings.base.json; see ~/.pi/agent/README.md.

Layout

• src/core/ shared runner, output bounding, lock, result shaping
• src/<group>/ one module per tool group, registered from src/index.ts
• test/<group>/ vitest, test/fixtures/<group>/ recorded real tool output
• test/scenarios/ the extension inside a real AgentSession (@earendil-works/pi-coding-agent/testing): faux model tool calls, pi's validation, tool_call hooks, the real bash tool and result serialization. Run under Vitest and bun test, so they use no runner-specific API