OverviewHistoryStatsSecurity
npx skills add ...
Documentation
SKILL.md
npx skills add everyinc/compound-engineering-plugin --skill ce-debug
Diagnosis loop for bugs and failing behavior. Use when asked to debug or fix failing or slow behavior.
npx skills add everyinc/compound-engineering-plugin --skill ce-debug
Find the root cause of a failure, then — when the user chooses to — fix it with test-first discipline.
Done when: the causal chain from trigger to symptom is stated with no gaps and file:line evidence, and either a verified fix has been handed off (PR, commit, or the user's chosen stop) or a diagnosis-only summary has been delivered. Escalate rather than persist: 2-3 hypotheses exhausted without confirmation, or 3 failed fix attempts, means diagnose why instead of trying again — that is the smart escalation references/investigate.md describes. One hypothesis, one change at a time; changing several to see what helps is shotgun debugging.
<bug_description> is whatever this skill was invoked with — a failure description, a mode: token, or an issue reference (#123, org/repo#123, an issue URL) — from the user or from a calling skill (ce-babysit-pr / lfg in mode:pipeline pass the failing jobs and log tails; lfg's defect route passes mode:return-to-caller and the user's own reference). Blank if nothing was provided.
Default is interactive: investigate, run the Phase 2 fix-choice gate, then the Phase 4 handoff.
mode:pipeline (set by an orchestrator such as ce-babysit-pr or lfg): run fully non-interactively and never call the blocking-question tool. Strip the token from <bug_description>, then read references/pipeline-mode.md and follow it — it overrides every "ask the user" point with a conservative default, replaces the Phase 2 fix-gate with "fix convergent bugs, defer divergent ones", and replaces the Phase 4 handoff with a structured return whose status is exactly one of fixed-and-pushed | fixed-not-pushed | diagnosed-no-fix | flaky-infra | needs-human. The caller branches on those exact spellings, so never rename, abbreviate, or add to them.
mode:return-to-caller (set by lfg on its defect route): run non-interactively, but the caller owns everything after the fix. Strip the token, then read references/return-to-caller.md and follow it — it keeps Phases 0-3 and the convergent-or-defer fix boundary, applies the Phase 3 branch rule so the fix lands on a feature branch, commits the fix-owned files without pushing, skips the post-fix polish and review steps, and returns a structured result whose status is exactly one of fixed | diagnosed-no-fix | needs-human | blocked. The same spelling rule applies.
Wherever this skill asks the user something, use the host's blocking question tool already in the current tool list (match by capability, not by a host-specific name). Presence in the current tool list is proof the tool exists; never call a user-facing question tool to discover whether it exists. If a matching tool is listed but unloaded, use the host's tool-discovery primitive to load that capability — do not search for another host's tool name. Fall back to numbered options on the host's chat surface only when no such tool is in the list or a real question call errors. Never silently skip the question, and never end a phase without a response.
Debugging surfaces raw output constantly — command results, captured payloads, log excerpts — and the harness may render a command's output the moment it runs, so check secrets when you construct the command, not afterward. Keep credentials in env vars rather than on the command line; when a command's output may carry a secret (verbose HTTP traces, dumped headers, config or environment prints), capture it to a file and surface only sanitized excerpts, writing <REDACTED> in place of each secret. No secret (credential, token, auth header, connection string) appears in anything shown, written, or committed. If sanitizing removes what the diagnosis needs, say so and ask the user rather than un-redacting.
Resolve <root> only when you first compose a <root>/ path — a run that composes none skips this entirely.
Resolve the CE artifact root <root> before composing any artifact path.
docs_root from <repo-root>/.compound-engineering/config.yaml only (<repo-root> = git rev-parse --show-toplevel). Do not read it from config.local.yaml. Unset -> <root> is docs, exactly as before..git/. Otherwise stop with an error naming docs_root and the value -- never fall back to docs.<root> as the sole artifact location: create it if absent, compose each path as <root>/<subdir> with this skill's own subdirectory, and never also read docs.Five phases in order: 0 Triage -> 1 Investigate -> 2 Root Cause -> 3 Fix -> 4 Handoff. Beyond Phase 0's trivial-bug fast-path there is no skipping and no complexity tiers. A hard bug spends longer in each phase; it does not enter fewer.
Read references/investigate.md now and follow it for Phases 0-2 — issue fetching, reproduction, environment sanity and the dirty-tree stash experiment, backward tracing, the tracker/PR-history search, hypothesis grounding, and the escalation table. Only the gates below are stated here.
The issue of record. If the user handed you a ticket or issue, that is where this bug already lives, whichever system it is in; a Sentry issue counts as much as a Linear ticket. Carry its identifier and URL through to Phase 4. If the input is only a stack trace, test path, or description, this run has no issue of record. That is an ordinary state, not a gap to fill: ship the fix without one, never open a ticket to manufacture a record, and never ask the user whether to. Phase 1's tracker search reads prior work and never establishes a new home for the bug. An existing ticket for this bug is one to link in Phase 4, never one to create.
The trivial-bug fast-path (cause readable from the input, one-line fix, no deep tracing) still runs Phase 2's fix-choice gate before editing: it saves investigation ceremony, not the user's choice over whether to apply a fix.
Choosing the regression test. The regression test for a confirmed defect belongs wherever existing coverage already owns that behavior: start from the tests that exist rather than from a new file. Read references/fix.md for the homes and the naming rule before writing Phase 2's recommendation, not only before Phase 3's edits. A test that fails because the change deliberately reverses the behavior it asserts does not have a wrong expectation — that is the divergent case below, deferred rather than updated.
Causal chain gate: do not proceed to Phase 3 until you can explain the full chain — trigger through every step to the observed symptom — with no gaps. "Somehow X leads to Y" is a gap. Only the user can authorize proceeding on a best-available hypothesis when investigation is stuck.
Once the root cause is confirmed, write the findings as a user-visible block: the causal chain with file:line references; the proposed fix and the files it changes; which tests to use, add, modify, or strengthen, and whether existing tests should have caught this; and any related ticket or PR and how it shapes the recommendation — if an open PR already fixes this, lead with that link instead of a fresh fix.
Same-turn presentation before the gate: do not open the fix-choice question until that findings block has been written in full — in this turn or the immediately preceding assistant message. The blocking question tool renders only its own stem on modal harnesses, so a question fired on "root cause confirmed" alone leaves the user choosing with none of the causal chain in front of them. Naming the options is not presenting the findings, and a promise to explain after the choice is too late.
When the request has not already authorized the next action, ask (per Blocking questions) which path to take, offering these three options. An explicit fix request is Phase 3. An explicit diagnosis-only request skips to Phase 4. mode:pipeline and mode:return-to-caller never ask. The test recommendations are part of the diagnosis either way.
ce-brainstorm) — only when the bug cannot be fixed within the current design: the root cause is a wrong responsibility or interface rather than wrong logic, the requirements themselves are wrong, or every candidate fix is a workaround around an assumption that no longer holds. Size alone is not a design problem.mode:pipeline and mode:return-to-caller: do not ask. Proceed to Phase 3 and apply a convergent fix; a divergent fix — one that would reverse a deliberate contract/behavior/product decision, including a "failing" test that asserts intended behavior — is deferred, not applied, per references/pipeline-mode.md or references/return-to-caller.md. Never route to ce-brainstorm here; a design problem becomes a needs-human residual.
If the user chose "Diagnosis only," skip to Phase 4's summary. If they chose "Rethink the design," control has transferred to ce-brainstorm and this skill ends.
Read references/fix.md before editing any file — the test-first sequence, the failed-fix rule, and the defense-in-depth and post-mortem triggers. Two rules decide whether the fix may start at all, so they stay here:
git status; if the user has unstaged work in files that need modification, confirm before editing. If the current branch is the default branch, create a feature branch without asking — derive a name from the bug, git checkout -b <name>, and say which branch you moved to. Detect the default by comparing against main, master, or git rev-parse --abbrev-ref origin/HEAD with its origin/ prefix stripped — the raw output is origin/<name>, so an unstripped comparison never matches.HEAD, whether git status --short is clean, and any pre-existing changed files. Then keep a list of fix-owned files (the tests and implementation changed for this bug) as you work. Phase 4 answers both of its questions from this record and cannot reconstruct it afterwards.mode:pipeline — skip this entire interactive handoff. No polish or review steps, no residual questions, no preview, no learning-capture offer. Commit and push the convergent fix per references/pipeline-mode.md, then emit that reference's structured return. Divergent / needs-human items are deferred there (open thread or the caller's run-report comment — never a PR-body section). mode:return-to-caller — skip it too: commit the fix-owned files on the feature branch, push nothing, and emit the structured return references/return-to-caller.md defines. In both modes the return is the last thing this skill writes. It ends this skill, not the turn. The caller runs in this same session, and its next step follows the return; do not announce a handoff to it. The rest of this section is the interactive path only.
Structured summary — always write this first:
If Phase 3 was skipped, stop after the summary — the user already said they were taking it from here. Do not prompt.
If Phase 3 ran, read references/post-fix-handoff.md now and follow it before routing below. It defines the quality steps after a fix: the contextual-override checks, the skip rule for mechanical fixes, the scoping that keeps ce-simplify-code and ce-code-review off unrelated branch work, what to do with leftover findings, the ## Post-Fix Quality block, and the criteria for offering to capture a learning. None of that appears in this body. The routing below names which action runs, never the scope rules that make it safe, so it cannot be improvised from. Skipping the read ships an unreviewed fix, lets review reach into unrelated branch work, and strands accepted findings in the session.
Land the fix without carrying along anything the user did not offer up — not into a commit, not into a push, not into a PR. Do not ask whether to open a PR; permission is not the gate. Two questions decide the handoff. Answer them from the pre-fix scope Phase 3 recorded, not from how the branch came to exist. Fire the action itself via the platform's skill-invocation primitive — never merely tell the user to type a command.
1. What may go into the commit — the fix-owned files and nothing else. This is a constraint on whichever skill commits in question 2, never an action of its own. It holds on every route, remote or not. Do not commit here.
ce-commit groups at file level and never splits a file. Ask (per Blocking questions) before anything commits whether to commit that file including their edits, leave the fix uncommitted, or stop. Only the first answer continues. The other two end the handoff, so question 2 never runs and nothing commits; say what was left and why. Every option loses something the agent cannot choose on the user's behalf, which is why this question survives. Phase 3's confirmation covered editing the file, never committing the user's edits with the fix.2. Who commits, and whether it ships. Exactly one of these runs.
Ships when all three hold: the pre-fix tree was clean, nothing on the branch is work the user has not already offered, and origin is PR-capable: somewhere gh can actually open a PR. Establish those however fits the repo in front of you. Two facts make it less obvious than it looks.
ce-commit-push-pr pushes the whole branch, and its PR spans every commit on it, not just your fix. So the question is about the branch, not your diff. It also pushes before creating the PR, so a remote gh cannot open a PR against leaves the branch published with no PR.If you cannot establish all three, take the local route instead; that is the safe direction, and the preview is not a substitute for it. Otherwise preview what will be committed, on what branch, and whether a PR opens or updates, then invoke the ce-commit-push-pr skill with branding:on. It commits under question 1's scope, so do not commit first. The preview is a statement, not a question. Surface the resulting PR URL.
Stays local when any of those three fails. Invoke the ce-commit skill under question 1's scope and push nothing. Say in one line what stayed local and why, and that you will push and open the PR on request. Do not ask first; a local commit is reversible.
Not a git repo: nothing commits. Stop after the summary and the quality block.
Contextual override ("don't open PRs from skills", "commit only", "stop after the fix") — follow what the user said, and Stop here without committing when that is what they asked for. A vague tonal cue is not an override.
After a PR is open — apply the reference's learning-capture criteria. Only when that gate qualifies the fix, offer capture; if the user accepts, invoke the ce-compound skill, then commit and push only the artifacts it actually wrote or updated so the open PR picks them up. If it writes nothing, end without a documentation commit.