install-improve-audit

Get an unfamiliar repo installing and building locally, fix only what blocks it, and open a PR carrying the fixes plus an append-only change log. The deliverable is a working build and a PR — never a

SKILL.md

GitHub

Install Improve Audit

Get the target repo building locally, fix what blocks it, and open a PR. The name is the priority order, and the order is strict:

  • Install — the only thing that counts as success. Non-negotiable.
  • Improve — a bonus tier, opened only after the build passes, hard-capped, always droppable.
  • Audit — the append-only log of what you changed and why. A change log, not a code audit.

Three outputs and no others: a working build, a committed fix set, an open PR.

Input

The target is the repo path or URL the user named, or the current working directory if they named none. If they stated a <slug>, use it for the branch and doc names; otherwise derive it from the repo directory name. Honor a docs-path override and a GitHub issue number if given. Ask for nothing else — if the target resolves, start.

Acceptance gate — read first

Your final message is invalid unless it contains all four:

  1. a verdict: INSTALLED, INSTALLED-DEGRADED (<what is off>), or BLOCKED (first blocker: <reason>)
  2. the commands that now work, or the exact command that failed
  3. the last 15 lines of install-check output — from the passing run if DONE was reached, from the failing run if it was not. If those lines are a dependency-resolver package wall, keep the lines that prove pass/fail and elide the package list mid-line, saying so — the requirement is evidence, not bulk
  4. the PR URL — or, if you lack push access, the branch name and the exact push command in its place. That is the only permitted substitute.

When more than one degradation applies, semicolon-join every one of them in this fixed order — do not pick a winner, and never drop one to keep the line short:

  1. excluded components, with the count (2 of 10 workspace members excluded)
  2. scripts disabled — untested at runtime
  3. relaxed version, with what it was relaxed from
  4. no build command
  5. container build only; native install documented but not verified
  6. tests failing, build clean
  7. start-check failure, with the reason

So: INSTALLED-DEGRADED (2 of 10 workspace members excluded; scripts disabled — untested at runtime). A most-severe-wins rule would hide a partial install behind a lesser reason, which is the exact thing the verdict exists to surface.

An issues list is not a deliverable. A code-quality assessment is not a deliverable. A repo-is-bad summary is not a deliverable. If the repo cannot be installed, ship BLOCKED with the first blocking error and the smallest next action — not a broad assessment.

If the caller asks for a repo analysis, say this skill installs first, and that a code-review or repo-health tool is the right one for analysis. The "Audit" in the name is the change log, not a code audit.

DONE — binary, nothing else counts

From a clean tree, all three of these exit 0:

  1. the repo's dependency install command
  2. the repo's build command
  3. a minimal load check of what was installed — python -c "import <package>", node -e "require('<entry>')", or the ecosystem's equivalent, appended to install-check.sh. Loading only: no app start, no config, no network. This is the cheapest proof that the resolver's exit 0 produced something that executes — a pin that installs clean but cannot import fails here instead of in the first user's session.

Write all three into install-check.sh at the repo root at step 4 of the order of operations — right after branching, so the file is born on the install branch. If that file exists or the root is not writable, use .install-audit/install-check.sh and name that path in your final output.

The first two lines are #!/usr/bin/env bash and set -euo pipefail. Without set -e the script's exit status is the last command's, so a failed build followed by a passing load check exits 0 and the gate certifies a broken install. This is not optional and not a style preference.

You are not finished until the check exits 0 and you have pasted its last 15 lines verbatim. "It builds now" is not acceptance.

Edge cases, decided in advance — do not deliberate:

  • No build command exists or can be inferred: dependency install plus the load check is DONE. Verdict INSTALLED-DEGRADED (no build command).

  • The repo's install path is containerized: see the ## Containers section below. A container-native repo reaches plain INSTALLED.

  • The documented build command also runs tests, lint, or e2e: narrow it to the compile/build step only. Tests are not part of DONE. If it cannot be narrowed and only the test phase fails, verdict INSTALLED-DEGRADED (tests failing, build clean).

  • The target moves mid-run (upstream commits land, or the operator calls for a fresh tree): commit whatever has landed, reset to a clean tree at the new state, then cherry-pick back only the commits whose errors still recur. Keep the audit-log rows already earned and restart the loop at step 7. The budget does not reset. Because every fix is committed as it lands (step 7), a reset costs you nothing that was already earned.

  • Optional, after DONE: start the app for about 10 seconds, then kill it. Start it in the background and kill it by PID — do not reach for timeout, which is not present on macOS/BSD by default. Do not source an example env file blindly either; unquoted values with spaces word-split and the failure looks like a repo bug when it is your harness. If the app dies, verdict is INSTALLED-DEGRADED with the reason. Do not fix runtime config unless it is a one-liner.

    Worth doing rather than skipping: the one real run found a repo that documents no build command but whose run command silently compiled a production build. The start-check is where an implicit build shows itself.

Containers — a first-class install path

Some repos are built on containers. For those, the container path is the install, and this skill's job is to get it working — not to refuse it and report that the repo has no build command. A container-native repo that builds reaches plain INSTALLED.

Build is not run. docker build and docker compose build start nothing: no services, no ports, no long-running process. They exit 0 or non-zero, and that is exactly what the gate needs. docker compose up starts services and is not a gate command — if the documented command is docker compose up --build, narrow it to docker compose build.

When to use the container path. Use it when the repo documents no native install, when the native path is explicitly unsupported, or when the container build is the README's primary route and the native path fails. Prefer native when both work — it is faster to re-run and proves more about the dependency situation.

The gate, substituted. The three-part native gate collapses to two commands, both in install-check.sh:

  1. docker compose build (or docker build -t <tag> .) — dependency install and build are fused inside the image; there is no separate install step to assert.
  2. The load check, run inside the image: docker compose run --rm <service> python -c "import <package>", or docker run --rm <tag> node -e "require('<entry>')". Same rule as native: loading only, no app start, no network.

Verdicts. No new verdict vocabulary — the existing grammar already covers it:

  • Container build and load check pass, and the repo is container-native → INSTALLED.
  • Repo documents both paths and only the container one was verified → INSTALLED-DEGRADED (container build only; native install documented but not verified).
  • Docker is not installed, or the daemon is not running → BLOCKED (first blocker: docker unavailable). That is a genuine blocker, not a degradation — say so plainly rather than falling back to a native path the repo does not support.

Cost. A cold image build is slow, but the budget excludes unattended waiting and layer caching makes re-runs fast — which is what you want from a repeatable gate. Image build time is never a reason to refuse the container path.

Budget — counted in attempts, not minutes

8 failed fix attempts, total, for the whole install. One attempt is one pass of the step-7 loop: one first-error, one smallest change, one re-run. A pass that clears its error does not count against the 8 — only failures do.

Attempts, not minutes, because you have no clock. A wall-clock budget requires stamping a start time and subtracting subprocess duration from it, and in the one real run of this skill none of that happened: the session ran roughly three times its stated budget and no status line was ever emitted. An attempt counter is something you can actually hold. Note you cannot buy headroom by batching several changes into one attempt — the loop takes the first error and makes the smallest change that clears it, so a batched attempt breaks a rule that is already there.

  • After attempt 4, print one line: STATUS 4/8 | last error | next move.

  • After attempt 8, if the check has never passed, all fixing stops. From that point the only permitted actions are:

    1. commit anything not yet committed
    2. finish the audit log
    3. write ## Blocked and ## Recommendations
    4. open the draft PR
    5. produce the final output

    No further source reading. No alternate-architecture research. No issue list. No improvements — a repo that never built has nothing to improve.

Unattended waiting — dependency downloads, cold image builds — is free and always was. The Improve box and the PR steps sit outside this count and have their own limits.

Order of operations — do not reorder

  1. Environment, one line each, no prose: OS/arch; the runtime version the repo declares (.nvmrc, .python-version, engines, rust-toolchain.toml, go.mod); the version actually installed; which version manager is available; package manager and version; whether Docker is installed and its daemon is running (docker info). Docker is a gate input, not trivia — a container-native repo cannot be installed without it.

  2. Recon — one pass, delegated. Launch the Explore subagent, read-only, with this exact return contract: "Return the documented dependency-install command and build command from README, INSTALL, Makefile, and CI workflow files. Return the declared runtime version. Return nothing else. Do not read source files. Do not assess the code." If Explore is unavailable, do the same recon yourself, limited to those same files. The file list is the bound, not a clock.

  3. Branch install/<YYYY-MM-DD>-<slug>.

  4. Write install-check.sh with all three parts DONE requires (or the two-command container substitute), set -euo pipefail at the top, then prove the gate can fail before trusting it:

    Replace the first command with an unmistakable marker — a command that cannot exist, e.g. zzz_gate_marker_zzz — run the script, and confirm two things together: it exits non-zero, and the error names your marker. Then restore the command. Do not run the unmodified gate here; step 6 does that.

    Both halves matter. A non-zero exit alone proves nothing at this point, because the install has not run yet and is presumed failing — that is the premise of this whole skill. Only the marker appearing in the output proves the script actually reaches and reports the command you think it does. Breaking the first command keeps this to seconds instead of a cold dependency resolve.

    The gate is self-authored under time pressure — exactly the condition this skill says produces polarity inversions in fix_probes — and an inverted gate reads every install as done, including through the pasted output lines.

  5. Create the audit doc: frontmatter and an empty log table only. Nothing else goes in it yet.

  6. Run the gate's install path — the same command install-check.sh holds, not whatever the README says verbatim. Native repo: the documented dependency-install command. Container repo: the narrowed build from ## Containers (docker compose build / docker build), never docker compose up, even when up --build is exactly what the README documents. Do not pre-diagnose. Let it fail.

  7. Loop until DONE or budget: take the first error → smallest change that clears it → re-run → append one log row → commit that fix on the spot, one commit per logical fix, message naming the error it cleared. Repeat.

    Committing inside the loop is what makes "one commit per logical fix" real instead of reconstructed at the end, and it is what lets a mid-run reset cost nothing. Every repo file you change gets a log row and a commit — including a change you made and then decided was unnecessary, which gets a row saying so.

  8. On DONE: open the Improve box, if anything on its list qualifies.

  9. Write the audit doc's status table, ## Blocked, ## Improvements, and ## Recommendations, then open the PR.

What you may read during the fix loop

Only files directly implicated by the error in front of you: build scripts, config files, env templates, manifests, lockfiles, version pins, and the file the failing command names. Do not browse the codebase looking for problems.

Improve — the bonus tier

Opens only after install-check.sh exits 0. Everything in it is droppable, and dropping all of it costs the run nothing.

Box: 3 items maximum, and at most 2 failed attempts across the whole box. Separate from the install attempt budget, and it runs before the PR steps. When either cap is hit, stop mid-item, revert anything unfinished, and move what is left to ## Recommendations.

Eligible — only these, and only when the install itself put it in front of you:

  1. Record a runtime version the repo needs but does not declare (.nvmrc, .python-version, engines) — the version you had to discover the hard way.
  2. Correct an install or build command in the repo's own docs that you proved wrong.
  3. Add the step you had to work out yourself to the existing install docs.
  4. Add a key you had to guess to an existing .env.example.
  5. Wire install-check.sh into a CI workflow that already runs on push — one line, and only if such a workflow already exists. This is the one item the local re-run rule cannot verify; the reviewer sees the result on the PR, which is what CI is for. On a run with no push access it goes to ## Recommendations instead.

Not eligible, however tempting: dependency upgrades the build did not need, refactors, formatting or lint sweeps, new tests, new CI files, new doc files, renaming or restructuring anything, or an improvement to code you never had to open. If it is not on the list above, it is a recommendation, not an improvement.

Rules that keep this skippable:

  • Each improvement is its own commit, prefixed improve:. A reviewer must be able to drop every one of them and still merge a working build.
  • Re-run install-check.sh after each one. If an improvement breaks it, git revert that commit and move on — never debug an improvement.
  • Improvements never change the verdict, never appear in ## Blocked, and never delay the PR.
  • BLOCKED runs skip this tier entirely.
  • Zero improvements is a normal, complete run. Do not manufacture one to fill the section; an empty ## Improvements is omitted, not apologized for.

PDDA compliance — deliberately partial

Full PDDA compliance is not a goal of this run and is not achievable inside the budget. Produce exactly what is listed below and nothing more. Do not open PDDA.md or ROUTER.md to check whether more is required — the answer is no.

Location: PROJECT/2-WORKING/INSTALL-<slug>.md. Override with the docs path if one was given. If the repo has no PROJECT/ tree, fall back to docs/install/INSTALL-<slug>.md, or repo root. Do not create a PDDA lifecycle structure in a repo that lacks one.

Frontmatter — exactly these keys:

title, status, created, updated, owner, doc_type: install-audit
ratings_provisional: true
gh_issue          # only if an issue number was given

Body — exactly these sections:

Status table (what was completed | what's next)
the append-only fix log table
## Blocked
## Improvements     # omit if none landed
## Recommendations

Deliberately omitted — do not add: Swarm Preflight Contract, fix_probes, phases, complexity / risk / effort, non_goals, roadmap_exempt, MARATHON.yaml stanzas, lifecycle folder transitions.

fix_probes specifically: install-check.sh is the gate for this work. A probe set authored under time pressure is where polarity inversions come from, and an inverted probe reads as already-done. A later pass adds probes if this graduates.

ratings_provisional: true is load-bearing — a single-pass run's ratings are guesses, and the flag keeps this doc out of automated selection until a human reviews it.

Never move the doc to 3-COMPLETED. That transition belongs to the closeout chain after the PR merges, not to this run.

Two rules hold regardless of location:

  • Append-only during the run. Each row is written after the fix is applied. No narrative sections until step 9.
  • The log is a table: | time | step | error (first line only) | change made | file | result |

Blocker ladder — in this order

Two fix attempts per distinct error, then move down. Every failed attempt counts against the 8:

  1. Honor the lockfile exactly (npm ci, pip install -r with hashes, poetry install, cargo build --locked) before touching any version.
  2. Switch to the runtime version the repo declares, via the version manager. Do not edit the declaration.
  3. If a postinstall script is the blocker, retry with scripts disabled (--ignore-scripts or equivalent). The log row names the skipped scripts and what they were supposed to produce, and the verdict reason says scripts disabled — untested at runtime. If a skipped script builds the package's own native component, disabling it is a broken install wearing a green exit code — that is not a degradation; keep moving down the ladder.
  4. Relax the declared version. Log it as a degradation with the reason.
  5. Exclude the failing optional component, workspace member, or extra; build the rest. Log what is excluded, and carry it into the verdict parenthetical with the count — INSTALLED-DEGRADED (2 of 10 workspace members excluded) — and into the PR title. The log table alone is not enough visibility for a partial install; a reader of the verdict must see the hole without opening the log.
  6. BLOCKED. Append to ## Blocked: the error, what was tried, and the single smallest thing that would unblock it. The verdict names it as the first blocker — serial fixing means later blockers are unknown, and the reader must not assume the named one is the only one.

Before rung 6, if the repo has a container path you have not tried, try it — see ## Containers. A repo that builds in its own container is installed, not blocked.

Scope fence

During the fix loop this fence is absolute. The Improve tier widens it by exactly five listed items, and only after DONE.

May edit freely: dependency manifests and lockfiles, config and env files, build and run scripts, install-check.sh, install docs, version-manager pins.

May patch minimally — only when the error is clearly build-shaped: a bad import path, a missing shim, a syntax break from a version bump, a hardcoded absolute path. Smallest possible diff. Log every one.

May NOT: refactor logic, bump majors to chase a nicer API, change public APIs, reformat, fix lint or type errors that do not block the build, edit tests except to make them runnable, or repair broken business logic. If the application genuinely does not compile against its own declared runtime, that is BLOCKED with a recommendation for a separate code-fixing run — not this skill's job.

Anything outside the fence is logged as a recommendation. You do not do it.

If the repo is not owned by your organization: use no credentials it did not ship with, no internal registries, and run nothing beyond its documented install and build commands.

The PR — the run is not finished at DONE

  1. Commit what is left — the audit doc, and anything not already committed. The fixes themselves were committed inside the loop at step 7, one per logical fix, so the history already shows which change cleared which error; do not squash it now. Improvement commits stay separate and keep their improve: prefix, so they can be dropped independently.
  2. CHANGELOG.md — if the repo already has one, add one entry summarizing the landed fixes. If the repo has no CHANGELOG, do not create one.
  3. Push the branch. If you lack push access to origin, stop here and hand back the branch name and the exact push command. That pair substitutes for the PR URL in the acceptance gate and in Final output — it is the one case where a run is complete without a URL. Everything else in the final output is still required.
  4. Open the PR. Title: install: <verdict> — <slug>. Body: your final output, verbatim. It is already exactly what a reviewer needs; do not author a second description.
  5. Do not merge. No auto-merge, no reviewer assignment, no labels.
  6. BLOCKED still ships a PR — as a draft. install-check.sh, the fixes that did land, and the blocked doc are what let the next attempt resume rather than restart. A blocked run with no PR wastes the whole session.

Final output — exactly this, nothing more

  • Verdict: INSTALLED | INSTALLED-DEGRADED (<what is off>) | BLOCKED (first blocker: <reason>)
  • PR: URL, and draft if drafted. With no push access: the branch name and the exact push command instead — the only permitted substitute for a URL
  • Last 15 lines of install-check output, verbatim — the passing run on DONE, the failing run on BLOCKED. Never omitted: a run with no output is not a report. Package-wall lines may be elided mid-line with the elision marked; the status lines that prove the exit are the part that must survive.
  • The copy-pasteable commands that now work, or the command that failed
  • git diff --stat for the branch
  • ## Improvements — one line each, with the commit that carries it (omit if none)
  • ## Blocked and ## Recommendations — ranked. Recommendations: maximum 3, plus a +N more not listed line if more qualify — never drop one silently. Each must be something you actually saw during recon or the run, not something you inferred about a repo you did not read. The fence limits where you look, never what you may say about what you saw. The cap is what keeps this from becoming the findings report this skill exists to prevent; a provenance test does not, and would delete the recon-sourced observations that are often the most useful lines in the report
  • ### Security — incidental findings only (below; omit entirely if empty)

Security — incidental findings only

Written after DONE, at the bottom. Not an audit. Never changes the verdict.

Only what the install put in front of you — no scanning, no grepping for secrets, no extra tool calls, no time spent looking. Typical qualifying finds: a credential or private key in a tracked file you opened; a committed .env with real values; an install script fetching and executing remote code; a dependency pulled over plain HTTP or from an unpinned source; no lockfile at all.

Maximum 3 items, one line each: what it is, the path, why it matters. No severity ratings, no remediation plans, no CVE lookups. If you find yourself investigating, stop — write the line and move on. If more than 3 qualify, write the 3 that matter most and state the count of the rest in one line — never drop a qualifying find silently.

An empty security section is the expected outcome, not a gap.

Not wanted

Architecture review, code quality assessment, deliberate security scanning, coverage analysis, dependency hygiene review, or any finding you did not encounter while making the build pass.

Do not deliver a static-analysis report. A final message containing findings but no verdict and no PR is incomplete work.