everyinc--compound-engineering-plugin/AGENTS.md

Agent Instructions

This repository is the root of the compound-engineering coding-agent plugin and the marketplace/catalog metadata used to distribute it.

It also contains:

• the Bun/TypeScript CLI that converts Claude Code plugins into other agent platform formats
• shared release and metadata infrastructure for the CLI, marketplace, and plugin

AGENTS.md is the canonical repo instruction file. Root CLAUDE.md is a symlink to AGENTS.md so Claude Code and other tools that look for CLAUDE.md still find it at the expected path. Keep that symlink (do not replace it with a regular file): a real root CLAUDE.md makes claude plugin validate --strict fail because this checkout is also the plugin root.

Quick Start

bun install
bun run test              # full test suite (also runs in CI; `--parallel` across worker processes)
bun run release:validate  # plugin/marketplace consistency (also runs in CI)
bun run plugin:validate   # Claude marketplace + plugin schema (also runs in CI; needs `claude` on PATH)

Codex Local Plugin Development

When testing current skill files in Codex, run the repository workflow from the checkout or worktree you intend to test:

bun run codex:dev -- local    # link this worktree's skills and remove CE plugin installs
bun run codex:dev -- status   # show local/remote state and checkout provenance
bun run codex:dev -- remote   # restore the official marketplace-backed plugin
bun run codex:dev -- remove   # remove both supported CE installation surfaces

refresh is an idempotent alias for local. Local mode manages only the exact $CODEX_HOME/skills/compound-engineering-local symlink and Compound Engineering plugin IDs; it must not alter unrelated user skills. The symlink includes modified and untracked files from the selected worktree. Start a new Codex session after switching installation modes. Current Codex versions detect direct skill edits automatically; restart only if an edit does not appear. For live local testing, use this workflow instead of adding the repository as a marketplace: a marketplace install caches a snapshot, while local mode links the current skill files.

Working Agreement

• Branching: Create a feature branch for any non-trivial change. If already on the correct branch for the task, keep using it; do not create additional branches or worktrees unless explicitly requested.
• Merge policy: All changes to main go through pull requests. Direct pushes and direct merges are not allowed; branch protection on main enforces this by requiring the test status check to pass. The direct path bypasses release:validate, the test suite, and PR title validation — past direct merges have caused version drift requiring multi-PR recovery (see docs/solutions/workflow/release-please-version-drift-recovery.md).
• Contribution gate (non-maintainers): If you are not a repository maintainer or admin, do not open a PR without a linked issue — file the issue first and reference it from the PR. Adding a new skill has a stricter gate: non-maintainers and non-admins must raise a discussion in an issue and get explicit maintainer approval before starting the work; do not open a new-skill PR that has not been approved this way. Maintainers and admins are exempt from both gates but still follow the merge policy above.
• PR disclosure: .github/pull_request_template.md ends with ## Security Disclosure and ## Agent Disclosure sections. Fill both when opening a PR — including PRs authored via gh pr create --body/--body-file, which bypass the template so nothing pre-fills them. State any security-relevant changes (or "No security-relevant changes"), and the model that did the bulk of the work — your harness plus the most specific model identity your own context gives you, e.g. Claude Code · claude-opus-4-8 or Codex CLI · GPT-5. Copy an exact model ID verbatim when your harness states one; when it exposes only a generic family, report the family and stop. Measured 2026-07-24: Codex and Cursor agents cannot see their running model at all (Codex's "based on GPT-5" is fixed boilerplate), so do not upgrade a family to a version, and do not read config files for one — the configured default is often not the model actually running. Never invent a version or variant. The body above those sections stays freeform — add whatever sections best explain the change.
• Safety: Do not delete or overwrite user data. Avoid destructive commands.
• Testing: Run bun run test after changes that affect parsing, conversion, output, skill conventions, or other mechanical guards. Local bun run test is the same suite CI runs — there is no separate local-only unit-test lane. Prefer it over bare bun test: the package script carries --parallel, which is where the suite's speed comes from. Bare bun test <file> is still the right tool for iterating on one file.
• Release versioning: Releases are prepared by release automation, not normal feature PRs. The repo has one root plugin/package release component (compound-engineering) plus marketplace components (marketplace, cursor-marketplace). GitHub release PRs and GitHub Releases are the canonical release-notes surface for new releases; root CHANGELOG.md is only a pointer to that history. Use conventional titles such as feat: and fix: so release automation can classify change intent, but do not hand-bump release-owned versions or hand-author release notes in routine PRs.
• Output Paths: Keep OpenCode output at opencode.json and .opencode/{agents,skills,plugins}. For OpenCode, commands go to ~/.config/opencode/commands/<name>.md; opencode.json is deep-merged (never overwritten wholesale).
• Scratch Space: Default to OS temp. Use .context/ only when explicitly justified by the rules below.
  • Default: OS temp — covers most scratch, including per-run throwaway AND cross-invocation reusable, regardless of whether a repo is present or whether other skills may read the files. A stable OS-temp prefix handles cross-skill and cross-invocation coordination equally well as an in-repo path; repo-adjacency is rarely the relevant property.
    • Per-run throwaway: mktemp -d "${TMPDIR:-/tmp}/<prefix>-XXXXXX" (OS handles cleanup). Use for files consumed once and discarded — captured screenshots, stitched GIFs, intermediate build outputs, recordings, delegation prompts/results, single-run checkpoints. Always pass an explicit template under ${TMPDIR:-/tmp}. Do not use bare mktemp, bare mktemp -d, mktemp -t, or mktemp -d -t: those forms ignore $TMPDIR on macOS and can resolve outside a sandbox's writable temp directory.
    • Cross-invocation reusable: use a stable, effective-user-owned prefix under /tmp/compound-engineering-<effective-uid>/<skill-name>/ — not mktemp -d — so later invocations by the same OS user can find prior outputs without sharing a writable root with other users. Derive the effective UID with id -u, reject a symlink or path not owned by the current user, and create or repair the top-level root to mode 0700 before use. Probe before committing to /tmp: when that root cannot be created, is not yours, or is not writable, use ${TMPDIR:-/tmp}/compound-engineering-<effective-uid> instead — the same rule, in the same order, in every shell preamble and Python default, so a later invocation resolves the same root. Claude Code's macOS sandbox allowlists writes under $TMPDIR (/tmp/claude-<uid>) but not /tmp itself, so without the fallback every skill's scratch setup aborts there with Operation not permitted; and an existing root from an unsandboxed session passes mkdir -p as a no-op yet refuses the first write, which is why the probe is a writability check ([ -w ]), not creation alone. Copy the block from any shipped skill (for example skills/ce-code-review/SKILL.md) rather than re-deriving it; tests/scratch-root-preamble-executes.test.ts runs every copy, including the fallback. The default layout is one <scratch-root>/<skill-name>/<run-id>/ directory per run; use it for caches keyed by session, checkpoints meant to survive context compaction, intermediate state, and outputs whose lifecycle or mutation belongs to one run.
      • Discoverable collection exception: omit the per-run directory only when later invocations intentionally enumerate multiple sibling final artifacts as core product behavior and run isolation would materially worsen discovery or the user-facing path. Use a stable collection namespace (for example, repository identity plus a general fallback), descriptive immutable filenames, metadata that supports ranking, and no-overwrite collision handling that atomically reserves the final filename and retries with the next suffix on collision; never check availability and then write. Do not use this exception for caches, checkpoints, intermediate files, or merely to shorten a path.
      • Prefer /tmp over $TMPDIR so paths stay accessible: $TMPDIR on unsandboxed macOS resolves to /var/folders/64/.../T/, which is hostile for users who want to inspect checkpoints, grep them, or copy them out — which is why $TMPDIR is the fallback, taken only when /tmp cannot host the root, and never the first choice. The explicit effective-UID segment supplies the required cross-user boundary while preserving a readable path. Agents running as the same OS user intentionally remain in one discretionary-access-control principal.
  • Exception: .context/ — use only when the artifact is genuinely bound to the CWD repo AND meets at least one of:
    • (a) User-curated: the user is expected to inspect, manipulate, or manually curate the artifact outside the skill (e.g., a per-repo TODO database, a per-spec optimization log that survives across sessions on the same checkout).
    • (b) Repo+branch-inseparable: the artifact's meaning is inseparable from this specific repo or branch (e.g., branch-specific resume state that a user expects to pick up again in the same checkout).
    • (c) Path is core UX: surfacing the artifact path back to the user is a core part of the skill's output and that path is easier to communicate as a repo-relative location than an OS-temp one. Namespace under .context/compound-engineering/<workflow-or-skill-name>/, add a per-run subdirectory when concurrent runs are plausible, and decide cleanup behavior per the artifact's lifecycle (per-run scratch clears on success; user-curated state persists). "Shared between skills" is not by itself sufficient — OS temp handles that equally well.
  • Durable outputs (plans, specs, learnings, docs, final deliverables) belong in docs/ or another repo-tracked location, not in either scratch tier.
  • Cross-platform note: /tmp is writable on macOS (symlink to /private/tmp), Linux, and WSL. For per-run throwaway files, use an explicit ${TMPDIR:-/tmp} template so macOS and sandboxed hosts honor the selected temp parent. Skills authored here assume Unix-like shells (bash on macOS/Linux, or Git Bash on Windows). Native Windows is a supported target for Python interpreter resolution and peer-job detach — never hardcode python3; probe execution per docs/solutions/conventions/resolve-python-interpreter-not-python3.md.
• Character encoding:
  • Identifiers (file names, agent names, command names): ASCII only -- converters and regex patterns depend on it.
  • Markdown tables: Use pipe-delimited (| col | col |), never box-drawing characters.
  • Prose and skill content: Unicode is fine (emoji, punctuation, etc.). Prefer ASCII arrows (->, <-) over Unicode arrows in code blocks and terminal examples.

Directory Layout

src/              CLI entry point, parsers, converters, target writers
skills/           Compound Engineering plugin skills
.claude-plugin/   Claude plugin manifest and marketplace catalog metadata
.codex-plugin/    Codex plugin manifest
.cursor-plugin/   Cursor plugin manifest and marketplace catalog metadata
.opencode/        OpenCode package entrypoint and install docs
.pi/              Pi extension entrypoint
tests/            Converter, writer, and CLI tests + fixtures
docs/             Requirements, plans, solutions, and target specs
CONCEPTS.md       Shared domain vocabulary (glossary of project-specific terms)

Repo Surfaces

Changes in this repo may affect one or more of these surfaces:

• root plugin content under skills/, AGENTS.md, README.md, and platform manifests
• marketplace catalogs under .claude-plugin/, .cursor-plugin/, and .agents/plugins/
• the converter/install CLI in src/ and package.json

Do not assume a repo change is "just CLI" or "just plugin" without checking which surface owns the affected files.

Plugin Maintenance

When changing plugin content:

• Update substantive docs like README.md when the plugin behavior, inventory, or usage changes.
• When adding a user-facing skill, document it: create a docs/skills/<skill-name>.md page (purpose, novel mechanics, when to use, chain position — follow the shape of the existing pages) and add a catalog row under the right category in docs/skills/README.md, alongside the root README.md inventory row and the skill-count bump in tests/release-metadata.test.ts. Keep these in sync when a skill's purpose or inventory changes. This is convention, not yet validated by a test, so it is easy to miss — most skills have a page; the few that don't (e.g. lfg, ce-dogfood-beta) are the exception, not the rule.
• When adding, removing, renaming, or changing the meaning/default/consumer of a .compound-engineering/config.yaml option, update skills/ce-setup/references/config-template.yaml, its byte-identical .compound-engineering/config.example.yaml copy, the centralized docs/skills/configuration.md reference, and the affected consumer skill docs in the same change. Ordinary keys may also live in optional checkout-local config.local.yaml (overrides the repo file). docs_root belongs only in config.yaml. Durable team instructions still belong in the project's normal agent-instructions mechanism.
• Do not hand-bump release-owned versions in plugin or marketplace manifests.
• Do not hand-add release entries to CHANGELOG.md or treat it as the canonical source for new releases.
• Run bun run release:validate if agents, commands, skills, MCP servers, or release-owned descriptions/counts may have changed.
• When removing a skill, agent, or command, add its name to both cleanup registries so stale flat-install artifacts are swept on upgrade:
  • STALE_SKILL_DIRS / STALE_AGENT_NAMES / STALE_PROMPT_FILES in src/utils/legacy-cleanup.ts
  • EXTRA_LEGACY_ARTIFACTS_BY_PLUGIN["compound-engineering"] in src/data/plugin-legacy-artifacts.ts

Useful validation commands:

bun run release:validate
cat .claude-plugin/marketplace.json | jq .
cat .claude-plugin/plugin.json | jq .

Runtime vs Authoring Context

AGENTS.md, CLAUDE.md (symlink to AGENTS.md), and GEMINI.md are authoring context for this source repository. Skills are installed into end-user environments, where they run against the user's local instruction files, not this repo's. Behavioral rules that must affect a skill at runtime belong in that skill's SKILL.md or files under its own references/ directory.

Working on Skills

This repository authors each skill once and distributes it across multiple agent models and harnesses. A skill is a set of goals, not a state machine: it hands the agent the goal, the done condition, the safe failure direction, and the facts it cannot derive from the repo in front of it, then gets out of the way. docs/solutions/skill-design/portable-agent-skill-authoring.md is the standard; the rules in this file supplement it and take precedence where more specific.

Before creating, editing, reviewing, or acting on review feedback for anything under skills/**, invoke the repo-local ce-skill-work skill (.agents/skills/ce-skill-work/, which Codex and Cursor discover directly; .claude/skills is a symlink to .agents/skills for Claude Code). It carries the procedures for each of those four activities, the audit questions, the provenance rule for removals, and the validation contract. This file states only what must be always-loaded; when the two disagree, fix the disagreement rather than following the shorter one.

Three rules that hold regardless of whether the skill was invoked:

• State conditions, not procedures or cases. When a block keeps absorbing "add the case we just found" — in authoring, in a review round, or in your own fix to a finding — the representation is wrong. Delete the additions and restate the goal, then re-verify against every path the additions served; a restatement that no longer names a path is a new defect, not a simplification.
• Prescribe a mechanism only where it is owned. A delegating skill states the condition, the safe failure direction, and the non-derivable callee facts, never a re-derivation of the callee's commands (docs/solutions/skill-design/skill-gates-state-conditions-not-prescribed-git-commands.md).
• Bring the block you touch up to the standard; leave untouched blocks alone and name them as follow-up. Skills predate the standard and evolve toward it.

User-Facing Skill Invocations

Keep agent-to-agent or skill-to-skill routing semantic: format formal skill names as inline code (for example, ce-plan) and invoke the named skill through the active harness's callable skill mechanism. When a skill prints or copies a user-runnable invocation, default to /skill-name; use $skill-name only when the active harness is Codex or explicitly documents dollar-prefixed skill invocation. On oh-my-pi (omp), keep the default form for model-visible targets; use native /skill:<name> only when the target is not model-visible because it declares disable-model-invocation or hide (for example, /skill:ce-polish). In prose, render only the invocation as inline code; use a fenced block only when the command stands alone. Output exactly one form. Do not apply this rendering rule to built-in commands such as /goal.

At runtime, put the smallest self-contained rendering rule immediately before the smallest section that contains all affected user-copy seams. Do not repeat it in every step; repeat it only in a separately loaded reference that independently owns output.

Reviewing a skill change (bots and humans)

Review bots read this file when reviewing a PR here. On skills/**:

• A finding is a gap in the goal, the done condition, or the safe failure direction; over-prescription that degrades degrees of freedom; or a mechanism at the wrong owning layer — commands prescribed in a skill that delegates that work, repeated command blocks where one parameterized recipe would decide the same behavior, a model-invoked description that opens with identity boilerplate or catalogs one branch, per-step done checks not protecting a fragile gate, repeated ask-first gates not marking a different external/destructive/scope/user-only boundary, a rule placed where it will not fire, a Claude-only construct in a cross-host skill, a rendering that breaks on another harness.
• A case a stated condition already covers is not a finding. Before filing "what if X" against a rule, check whether the rule's condition decides X. If it does, do not file; if the condition is wrong or missing, file that.
• State the requested fix as a condition or an owning-layer move, never as a case to add. "Command X fails in state Y" against a delegating skill is a finding about the representation; the fix is to drop the command and state the condition, not to correct the command.
• A block restated to the standard is the expected shape of an edit, not scope creep, when the restatement covers every path the old text served.
• Ordinary code under src/, tests/, and scripts/ gets ordinary code review; these rules are about instruction prose.

Acting on review feedback

This governs any agent or tool that acts on review feedback in this repository — ce-resolve-pr-feedback, another vendor's resolver, or a person — and on skills/** it takes precedence over the tool's own "default to fixing". Skill prose is not code: a natural-language instruction can always be made more specific, so a reviewer can produce a valid-looking edge case against any condition indefinitely, and patching each one dilutes the instruction (#1397: 24 findings over nine rounds on a two-condition step). A case the stated condition already decides is answered with the condition, not patched; only a wrong or missing condition, or a mechanism at the wrong owning layer, is a fix. On the second round against the same block, restate it rather than qualify it. The full procedure — Evidence, Owning layer, Mechanism, Reconcile, Stop the accretion loop — is ce-skill-work's respond mode.

Referencing Project Conventions in Skills

When a skill needs to discover a project convention at runtime — the issue tracker, coding standards, commit format, lint command, scope constraints, etc. — describe what to look for in the agent's existing context, not which file to open.

On the read path, do not name instruction files (AGENTS.md / CLAUDE.md / GEMINI.md / .cursor/rules). Phrase it as "the project's active instructions and conventions already in your context." Three reasons:

• Redundant. Every major harness auto-injects the project's root instruction file into context at session start (Claude Code loads CLAUDE.md, Codex AGENTS.md, Gemini GEMINI.md). Telling the agent to "read AGENTS.md" asks it to re-open content it already has.
• Brittle / not portable. The filename differs per harness, and this plugin is authored once and converted to all of them. A hardcoded "read AGENTS.md (or CLAUDE.md)" silently finds nothing on a harness that uses a different name.
• Security smell. Instructing an agent to go read named instruction dotfiles is the exact shape that prompt-injection defenses in some agent frameworks (e.g., Hermes) flag. Referencing context rather than filenames avoids tripping those guards.

Name a concrete file only where the skill must do something a context reference can't express:

• Writing a convention back (e.g., persisting project_tracker: linear) needs a target — name it minimally and as an example ("the project's root agent-instructions file, e.g., AGENTS.md; if it @-includes another, write to the substantive one").
• Reading content that is genuinely not auto-loaded — a subdirectory-scoped instruction file governing the area being changed, an optional project doc like STRATEGY.md / CONCEPTS.md / README.md, or any file a fresh subagent (which does not inherit the parent's loaded instructions) must open to do its job. Auditing tools that must enumerate every standards file (e.g., ce-code-review's project-standards reviewer globbing all CLAUDE.md/AGENTS.md) are a legitimate exception — they review the files, they don't re-read them for context.

Describe the capability, not the tool. Pair this with naming the category of thing rather than a closed set: "the project's issue tracker (e.g., GitHub Issues, Linear, Jira)" and "whatever interface that tracker exposes (connector/MCP, documented API, or a documented CLI)" — never assume a specific CLI exists, and never treat a missing binary / env var / MCP server as proof the capability is unavailable.

Validating Agent and Skill Changes

Behavioral changes to a plugin skill or skill-local persona (anything under skills/) need a different validation path than mechanical code changes, because of how Claude Code loads plugins.

• Test prose changes by injecting the current on-disk skill into a fresh agent. Plugin skills cache at session start, so invoking the edited skill in the authoring session tests stale content. This repository does not ship an eval skill. The portable path is the host CLI you already have, against a skill dir extracted from a ref: bun run test:skill-eval-cell. Named scenarios (pre-change contract vs current tree) run with bun run test:skill-eval-pack -- --skill <name> --arm ab. That uses claude / codex / grok on PATH and bills those products, not a separate API key. Anthropic's skill-creator is an optional Claude-hosted helper, not required, and does not replace the cell.

• Plugin agent and skill definitions both cache at session start. Once a Claude Code session is open, dispatching a typed plugin agent runs the in-memory copy that was loaded when the session began. The same applies to skills: invoking a skill goes through the cached skill loader, so edits to skill scripts are also not tested via that path. File edits to either layer after session start do not propagate within the same session. Any iteration loop built around typed-agent dispatch or Skill-tool invocation in the same session is testing pre-edit content, not your changes.

• Do NOT edit ~/.claude/plugins/cache/ or ~/.claude/plugins/marketplaces/ to try to force a reload. Those paths are user machine state, not repo-managed. Modifying them does not reliably bypass the in-session cache (it didn't, in observed behavior), risks being silently overwritten by plugin updates, and is the wrong layer to test from. Inject current disk content into a fresh agent instead; if you genuinely need fresh-loaded behavior of the typed-agent dispatch path, restart the session.

• A version-matched cache is not automatically stale — confirm by content, not by version. When this working tree is the local marketplace source, a session (re)start re-copies it into ~/.claude/plugins/cache/.../compound-engineering/<version>/ (a plain copy, no .git; <version> is the working tree's .claude-plugin/plugin.json version), so the loaded plugin can be identical to — and as current as — your edits. Do not assume the running copy is stale just because it lives under the cache path; equally, do not assume a matching <version> means it includes your latest change. Version match is necessary but not sufficient: edits within a release do not bump the version, so a matching segment proves only that the cache was built from this release, not that it captured your most recent edit. To know which copy is actually loaded, diff the specific cache file against the working-tree file — identical means the running plugin is your current edit and you can trust it; differing means the session predates the edit, so restart or inject the working-tree files into a fresh agent. Never infer "stale" or "current" from the version segment alone.

• Mechanical changes do not have this restriction. Skill scripts (e.g., extract-metadata.py), parser logic, conversion code, and anything bun test exercises always run the current source. The caching issue only affects LLM-driven skill prose behavior dispatched through the plugin loader.

CI and Quality Gates

PR CI (.github/workflows/ci.yml) is the merge gate. It runs, in order: PR-title lint (PRs only), bun run release:validate, bun run plugin:validate, and bun run test. Do not invent a parallel local-only mechanical suite — if a check is deterministic and should block merges, put it in one of those steps (usually bun run test).

The test script runs bun test --parallel, which distributes test files across worker processes (one file still runs its own tests serially, and --parallel implies --isolate). This is the single biggest lever on CI wall time, because most of the suite is spent blocked on subprocesses — python3, bash, git, and bun run src/index.ts — not on CPU. Keeping it in the package script rather than the workflow means CI and a contributor's local run cannot drift apart.

That makes cross-file isolation load-bearing rather than incidental: a test file may not depend on another file's leftovers, and any test that writes outside its own mktemp directory is a latent flake. There are no exceptions — a test that needs a dirty tree builds a throwaway git repo for it.

A test that runs a bundled script which inspects the repository must point that script at a throwaway repo, never this checkout. Otherwise the developer's uncommitted work becomes test input. tests/skills/ce-code-review-cross-model-routes.test.ts ran the real review script against git diff HEAD in the checkout, so any uncommitted change over roughly 160KB crossed the script's large-diff threshold and failed 31 of its tests for reasons unrelated to the change under test. CI never saw it, because CI runs on a committed tree. The fixture pattern is dirtyFixtureRepo() in that file: git init a temp dir, two commits so HEAD~1 resolves, then one staged edit.

Do not pin a worker count. --parallel with no value tracks the runner's core count, which is what you want. Raising it looks free — the suite is idle-bound, so more workers should pack better — but it was measured on CI and it is not: at --parallel=8 on a 4-core runner, wall time improved ~9% (102s -> 93s) while total test-CPU inflated from 223s to 343s, and five tests crossed the 5000ms default per-test timeout. That converts runner busyness into red builds. A file that legitimately runs for seconds should call setDefaultTimeout instead, as the subprocess-heavy suites do.

A file never splits across workers, so an oversized file sets a floor. tests/skills/ce-work-unit-workspace.test.ts was 4,564 lines and 86 tests under one describe; it is now five ce-work-unit-workspace-*.test.ts files sharing tests/skills/helpers/ce-work-workspace-harness.ts. Measured with three workflow_dispatch runs per ref in the same window, the Run tests step went from a median of 88s (87/112/88) to 81s (83/80/81).

Splitting bought ~8% of CI wall time and most of the run-to-run variance — baseline spread 25s, split spread 3s. That second effect is the durable one: a 60s serial file makes wall time depend on which worker takes it, so a busy runner produced the 112s outlier. Locally the same split is much larger (160s -> 75s), because a developer machine has enough cores for the long file to be the whole critical path.

Do not use bun test --parallel=4 on a many-core laptop as a CI proxy. It predicted a 26% CI win where the real number was 8%: capping bun to four workers still leaves the other cores absorbing the git / python3 / bash subprocess load, so it does not behave like a 4-core runner. Dispatch the real workflow on both refs instead.

At 81s the suite is nearer CPU-bound on a 4-core runner than bounded by its slowest file, so further file splitting has small returns; tests/ce-babysit-pr-snapshot.test.ts (~36s) is the largest remaining file and was deliberately left intact.

Size a test file by its measured time, not its line count. A file is a wall-time problem when it approaches the suite's slowest-file ceiling. Roughly a thousand lines under one describe is a smell worth measuring, never a threshold to split at on sight — splitting a file that already runs well under the ceiling buys nothing. Get per-file times before deciding:

bun test --parallel --reporter=junit --reporter-outfile=/tmp/t.xml

The ce-work-unit-workspace-* shards run 10-23s each against that ~36s ceiling, so two of them sitting just over a thousand lines is fine and they are deliberately left whole. Put shared fixtures in tests/skills/helpers/.

What belongs where

Kind of check	Where it lives	Notes
Deterministic invariants (frontmatter, parity, path safety, script behavior, converter/writer output, greppable skill contracts)	bun test / release:validate / plugin:validate	Must pass in CI
Skill prose behavior (routing judgment, restraint, cross-model peer outcomes)	Fresh-agent eval (on-disk skill injected), local / PR evidence	Not a CI job; non-deterministic and needs a model

That split is intentional. See docs/solutions/skill-design/portable-agent-skill-authoring.md ("Evaluate proportionally"). Mechanical checks belong in CI; behavioral agent evals are best-effort evidence, not an exhaustive CI matrix.

Right-size new mechanical guards

When a review bot or human finds a greppable invariant that bun test missed:

1. Prefer tightening an existing guard over adding a new suite (e.g. widen a regex that already documents the rule).
2. Pin the smallest falsifiable unit — a token, enum, path, heading, or one fixture that would have failed on the regressing diff. Do not snapshot whole skill bodies or pin incidental wording.
3. If the failure needs an LLM to judge, keep it as a behavioral eval; do not fake it as a brittle string test.

Maintaining plugin:validate

• package.json plugin:validate must validate both the marketplace catalog and the plugin manifest, with --strict on each. Paths: .claude-plugin/marketplace.json and .claude-plugin/plugin.json. Do not use claude plugin validate . — that resolves this repo as a marketplace only (because .claude-plugin/marketplace.json exists with source: "./") and skips plugin-root checks.
• CI pins @anthropic-ai/claude-code for reproducible schema rules. Bump the pin deliberately when adopting new upstream rules; do not float @latest.
• Root CLAUDE.md must remain a symlink to AGENTS.md (path stays at the repo root where contributors expect it). Upstream warns on a regular-file plugin-root CLAUDE.md because it is not loaded as end-user project context; the symlink avoids that warning so --strict can stay on. Do not replace the symlink with a regular @AGENTS.md shim or relocate the file just for validators.
• If --strict starts failing again on CLAUDE.md after an upstream bump, check whether the symlink was materialized into a regular file (Windows/core.symlinks=false checkouts) or whether the validator started following symlinks — fix the layout or pin, do not silently drop --strict.

When CI comments look "stale"

If CI claims a deferred warning or a missing gate, reproduce with the pinned claude version against .claude-plugin/plugin.json before treating the comment as current. Marketplace-only validation can hide plugin warnings.

Coding Conventions

• Prefer explicit mappings over implicit magic when converting between platforms.
• Keep target-specific behavior in dedicated converters/writers instead of scattering conditionals across unrelated files.
• Preserve stable output paths and merge semantics for installed targets; do not casually change generated file locations.
• When adding or changing a target, update fixtures/tests alongside implementation rather than treating docs or examples as sufficient proof.

Commit Conventions

• Prefix is based on intent, not file type. Use conventional prefixes (feat:, fix:, docs:, refactor:, etc.) but classify by what the change does, not the file extension. Files under skills/ and plugin manifests are product code even though they are Markdown or JSON. Reserve docs: for files whose sole purpose is documentation (README.md, docs/, CHANGELOG.md).
• Type selection — classify by intent, not diff shape. Where fix: and feat: could both seem to fit, default to fix:: a change that remedies broken or missing behavior is fix: even when implemented by adding code, and net additions do not turn a fix into a feat:. Reserve feat: for capabilities the user could not previously accomplish where nothing was broken. Other conventional types (chore:, refactor:, docs:, perf:, test:, ci:, build:, style:) remain primary when they describe the change more precisely than either. Heuristic: if a regression test you could write today would have failed before the change, it's fix:. The user may override this default for a specific change.
• Include a component scope. The scope appears verbatim in the changelog. Pick the narrowest useful label: skill/agent name (document-review, learnings-researcher), CLI or marketplace area (cli, marketplace), or shared area when cross-cutting (review, research, converters). Never use compound-engineering — it's the entire plugin and tells the reader nothing. Omit scope only when no single label adds clarity.
• Never use ! or a BREAKING CHANGE: footer without explicit user confirmation. These markers trigger release-please's automatic major version bump — a decision the user may not want even when a change is technically breaking. If a change appears breaking, surface that to the user and let them decide whether to apply the marker.

Adding a New Target Provider

Only add a provider when the target format is stable, documented, and has a clear mapping for tools/permissions/hooks. Use this checklist:

1. Define the target entry

   • Add a new handler in src/targets/index.ts with implemented: false until complete.
   • Use a dedicated writer module (e.g., src/targets/codex.ts).

2. Define types and mapping

   • Add provider-specific types under src/types/.
   • Implement conversion logic in src/converters/ (from Claude → provider).
   • Keep mappings explicit: tools, permissions, hooks/events, model naming.

3. Wire the CLI

   • Ensure convert and install support --to <provider> and --also.
   • Keep behavior consistent with OpenCode (write to a clean provider root).

4. Tests (required)

   • Extend fixtures in tests/fixtures/sample-plugin.
   • Add spec coverage for mappings in tests/converter.test.ts.
   • Add a writer test for the new provider output tree.
   • Add a CLI test for the provider (similar to tests/cli.test.ts).

5. Docs

   • Update README with the new --to option and output locations.

Specialist Prompt Assets in Skills

The compound-engineering plugin no longer ships standalone agent definitions under agents/. When a skill needs a specialist persona, store it inside that skill directory, usually under references/agents/ or references/personas/, and have the calling skill dispatch a generic subagent with that file's contents in the prompt.

Internal prompt asset file names should be descriptive and unprefixed because they are not externally exposed agent names.

Example:

• references/agents/learnings-researcher.md (correct)
• references/agents/ce-learnings-researcher.md (wrong for an internal prompt asset)

These prompt assets must not include YAML frontmatter. Model selection, tool constraints, and dispatch policy belong in the calling skill's SKILL.md, not in the prompt asset.

File References in Skills

Each skill directory is a self-contained unit. A SKILL.md file must only reference files within its own directory tree (e.g., references/, assets/, scripts/) using relative paths from the skill root. Never reference files outside the skill directory — whether by relative traversal or absolute path.

Broken patterns:

• ../other-skill/references/schema.yaml — relative traversal into a sibling skill
• /home/user/compound-engineering-plugin/skills/other-skill/file.md — absolute path to another skill
• ~/.claude/plugins/cache/marketplace/compound-engineering/1.0.0/skills/other-skill/file.md — absolute path to an installed plugin location

Why this matters:

• Runtime resolution: Skills execute from the user's working directory, not the skill directory. Cross-directory paths and absolute paths will not resolve as expected.
• Unpredictable install paths: Plugins installed from the marketplace are cached at versioned paths. Absolute paths that worked in the source repo will not match the installed layout, and the version segment changes on every release.
• Converter portability: The CLI copies each skill directory as an isolated unit when converting to other agent platforms. Cross-directory references break because sibling directories are not included in the copy.

If two skills need the same supporting file, duplicate it into each skill's directory. Prefer small, self-contained reference files over shared dependencies.

Note (March 2026): This constraint reflects current Claude Code skill resolution behavior and known path-resolution bugs (#11011, #17741, #12541). If Anthropic introduces a shared-files mechanism or cross-skill imports in the future, this guidance should be revisited with supporting documentation.

Lean Repo Grounding

Use the project's active instructions already in the main agent's context, then go directly to task-specific current evidence. Pass fresh subagents the relevant project and task context, or have them read the applicable current instruction source when operational rules affect their work. If a task cannot be scoped from that context, use one targeted probe. Do not create a reusable generic repo profile or run a default root, stack, or layout scan.

Platform-Specific Variables in Skills

This plugin is authored once and converted for multiple agent platforms (Claude Code, Codex, Gemini CLI, etc.). Do not use platform-specific environment variables or string substitutions (e.g., ${CLAUDE_PLUGIN_ROOT}, ${CLAUDE_SKILL_DIR}, ${CLAUDE_SESSION_ID}, CODEX_SANDBOX, CODEX_SESSION_ID) in skill content without a graceful fallback that works when the variable is unavailable or unresolved.

How a bundled-file reference resolves depends on who resolves it and whether a shell is involved, so references fall into three tiers. Do not assume a bare scripts/… path behaves the same in all three.

Tier 1 — Read-time file references (relative, no anchor): When skill content points the agent at a co-located file to read (e.g., "read references/schema.yaml"), use a relative path from the skill root. The skill loader resolves these against the skill's own directory on all major platforms — no variable prefix needed. This is the rule in File References in Skills above.

Tier 2 — Prose pointers to a bundled file the agent acts on (relative + a "from this skill's directory" cue): When skill prose names a bundled file the agent will use but does not put it in an executed shell command (e.g., "drive the loop with scripts/hitl-loop.template.sh" or "generate the package with scripts/review-package BASE HEAD"), use a relative path plus an explicit "from this skill's directory" phrase. The cue tells the agent what to resolve against without the verbosity of an anchor.

Tier 3 — Executed shell commands (the SKILL_DIR anchor): When skill content puts a bundled script in a command the agent runs through the Bash tool — a fenced ```bash block or an inline bash … / python … — anchor it to the skill dir. The Bash tool's working directory is the user's project, not the skill directory, on Claude Code, Codex, and Cursor alike, so a bare bash scripts/my-script.sh resolves to <project>/scripts/…. Relative paths here often still work — a capable agent resolves them against the skill dir it loaded (which is how the agentskills.io spec and other ecosystems ship them) — but that relies on the agent translating the path, and the failure mode is a fenced block copied verbatim into a Bash call, which runs literally and misses (exit 127; recovery is a wasted round-trip that weaker models / mid-tier subagents botch). Anchoring bakes the resolution into the command, so it is deterministic. Use the anchor for executed shell as the house default — a conservative choice, not a claim that bare relative cannot work (recurring bug class: #764 ce-worktree, #811 ce-code-review, #898 ce-compound):

# set inline in the SAME command (shell state does not persist between Bash calls):
SKILL_DIR="<absolute path of the directory containing the SKILL.md you just read>";
bash "$SKILL_DIR/scripts/my-script.sh" ARG

Keep the trailing ; on the assignment line. Some hosts (observed on Codex) flatten a fenced multi-line block into a single line by replacing the newline with a space before executing it. Without the ;, SKILL_DIR="…" + newline + bash "$SKILL_DIR/…" collapses to the env-var-prefix form SKILL_DIR="…" bash "$SKILL_DIR/…", where the shell expands $SKILL_DIR before the prefix assignment takes effect — so it expands to empty and the script path becomes /scripts/my-script.sh (No such file or directory). The trailing ; makes the assignment a complete statement that survives flattening; it is load-bearing, not a style choice, so do not remove it.

An existence guard (if [ -f "$SKILL_DIR/scripts/my-script.sh" ]; then … else echo "not found — re-check the SKILL.md path"; fi) is optional — useful when there's a real fallback, but see the permission caveat below before guarding a pinned call.

SKILL_DIR is a model-filled value, not a harness variable: every harness loads SKILL.md from a real absolute path the agent knows, so the skill instructs the agent to set SKILL_DIR to that directory. This works identically on Claude Code, Codex, and Cursor precisely because it depends on no host-specific variable — SKILL_DIR, CLAUDE_SKILL_DIR, CODEX_SKILL_DIR, AGENT_SKILL_DIR are not env vars on any of them, yet the script runs because the agent supplies the path. This is the production pattern used by widely-installed cross-host skills (e.g. last30days). Two constraints: (1) shell state does not persist between separate Bash-tool calls, so SKILL_DIR cannot be set once and reused — each invocation must carry the absolute path (set it inline in the same command). (2) A script that needs its own directory (to read a sibling file) derives it from BASH_SOURCE, not SKILL_DIR, since SKILL_DIR is the orchestrator's shell var and is not exported to the child process — see skills/ce-code-review/scripts/cross-model-adversarial-review.sh for the reference implementation. last30days adopted this anchor for its critical multi-host engine after a path-resolution regression; it is the right tool when a script must run reliably, which is why it is the tier-3 default — but tiers 1 and 2 deliberately stay lighter.

Avoid ${CLAUDE_SKILL_DIR} here — in this cross-agent plugin it is a footgun, not a neutral alternative. Every skill in this repo is authored once and installed across Claude Code, Codex, Cursor, and Gemini, and ${CLAUDE_SKILL_DIR} is a Claude-Code-only SKILL.md content substitution (not an env var) that is empty on every other host. So a ${CLAUDE_SKILL_DIR}-guarded call's then branch quietly never fires off-Claude — the genuine silent skip — and a Claude-only mechanism breaks on Codex/Cursor because the converter doesn't rewrite these paths and the native Codex install loads raw SKILL.md (no ce_platforms filtering). The model-filled SKILL_DIR anchor works on every host, so it is the right replacement wherever a ${CLAUDE_SKILL_DIR}-guarded executed-shell call exists today (tier 3). Do not reach for ${CLAUDE_SKILL_DIR} as a "portable" option — it isn't. Reach for it only for behavior that is genuinely Claude-Code-only and will never run on another harness — which, given this plugin's cross-host install model, is essentially never; treat any new use as a smell to justify or remove.

So: a skill's core behavior can live in a bundled script across hosts — invoke it via the SKILL_DIR-from-read-path anchor. You no longer need to avoid bundled scripts for portability; anchor them instead. Read-time references (references/*.md) still resolve against the skill dir on all targets and need no anchor.

Permission caveat (Claude Code). Claude Code's permission checker evaluates every subcommand of a compound command, and a bare [ -f … ] test is not pre-approved — so wrapping a pinned bash "…sh" call in an if … then … fi guard defeats a narrow Bash(bash *…sh) allow-rule and prompts on every run. If a bundled-script call must stay auto-approved via such a pin, keep it a single pinned command rather than guarding it inline. Note the model-filled SKILL_DIR anchor produces a dynamic absolute path that won't match a static Bash(bash /…/scripts/x.sh) pin regardless of guarding — so for the anchor, expect a one-time approval prompt per distinct command (or use a broader allow-rule); the static-pin trick mainly applies to the fixed ${CLAUDE_SKILL_DIR} form.

Do not use ! load-time pre-resolution in skills. The !cmd`` SKILL.md syntax runs cmd at skill load and inlines its stdout, but it is banned here (enforced by tests/skill-shell-safety.test.ts) for two unfixable reasons: it runs only on Claude Code — on Codex, Cursor, Gemini, and Grok the line is inert literal text — and on Claude Code a command that exits non-zero aborts skill load with a user-facing error. Every real use was git context (git rev-parse …, gh pr view …) whose non-zero exit is a normal state (no PR yet, no origin/HEAD, detached HEAD, not a repo), so the ordinary case broke the skill. The POSIX guards that force exit 0 (2>/dev/null || echo SENTINEL) then fail to parse under Windows PowerShell 5.1, which broke skill load there instead (issue #1066). No single command string both exits 0 on the expected-failure states and parses under both POSIX sh and PowerShell, so the construct cannot be made safe.

Gather context at runtime instead. Have the agent run one argv-style command per shell tool call (git …, gh …) — no ;, &&, ||, pipes, $(…), or redirects — and interpret each exit status as control flow. This parses identically under POSIX sh and PowerShell because it is a single external-program invocation, and a non-zero exit becomes data the agent reads rather than a load-time abort. See ce-commit / ce-commit-push-pr for the pattern.

When a platform variable is unavoidable: resolve it at runtime with a single shell tool call and include explicit fallback instructions, so the agent knows what to do if the value is empty, a literal command string, or an error — e.g. run jq -r .version "${CLAUDE_PLUGIN_ROOT}/.claude-plugin/plugin.json"; if it resolved to a semantic version use it, otherwise fall back to the versionless behavior. This applies equally to any platform's variables — a skill converted from Codex, Gemini, or any other platform will have the same problem if it assumes platform-only variables exist without a fallback.

Repository Docs Convention

• Plans live in docs/plans/ — unified plan artifacts. New ce-brainstorm outputs are requirements-only unified plans (artifact_readiness: requirements-only); ce-plan enriches them to implementation-ready plans (artifact_readiness: implementation-ready). Historical docs/brainstorms/*-requirements.* files remain readable legacy inputs and should not be migrated just because a new plan is created.
• Brainstorm evidence / legacy requirements may live in docs/brainstorms/ — historical requirements docs and specialized analysis artifacts such as docs/brainstorms/riffrec-feedback/. Do not treat this as the canonical output path for new ce-brainstorm artifacts.
• Solutions live in docs/solutions/ — documented solutions to past problems (bugs, best practices, workflow patterns), organized by category with YAML frontmatter (module, tags, problem_type). Relevant when implementing or debugging in documented areas.
• Specs live in docs/specs/ — target platform format specifications.

Solution categories (docs/solutions/)

This repo builds a plugin for developers. Categorize solutions from the perspective of the end user (a developer using the plugin), not a contributor to this repo.

• developer-experience/ — Issues with contributing to this repo: local dev setup, shell aliases, test ergonomics, CI friction. If the fix only matters to someone with a checkout of this repo, it belongs here.
• integrations/ — Issues where plugin output doesn't work correctly on a target platform or OS. Cross-platform bugs, target writer output problems, and converter compatibility issues go here.
• workflow/, skill-design/ — Plugin skill and agent design patterns, workflow improvements.

When in doubt: if the bug affects someone running bun install compound-engineering or bun convert, it's an integration or product issue, not developer-experience.