agentsclimarketplace

Circuit surface test

Skill petekp/circuit/.claude/skills/circuit-surface-test

Comprehensively manually test the Circuit plugin's user-facing surface in the current host, Claude Code or Codex. Use this skill whenever the user asks to "manually test Circuit", "QA the Circuit plugin", "exercise the Circuit surface", "run the Circuit checklist", "smoke test Circuit", "find regressions in Circuit", "test the Circuit plugin", "test the Claude Circuit plugin", "test the Codex Circuit plugin", or when preparing a Circuit release for marketplace publication. If no host is named, automatically test the host you are currently running in. Before testing, refresh generated plugin output and sync local host plugin caches from the current Circuit checkout so the run never exercises a stale cached plugin. Produces a Markdown report with per-command pass/fail, exploratory findings ranked by severity, run-folder evidence links, and a concise terminal summary. Use even if the user does not say the word "test" - phrases like "go through every Circuit command" or "make sure Circuit still works end-to-end" should also trigger.From its SKILL.md

Install
npx -y skills add petekp/circuit --skill circuit-surface-test

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 3 stars3 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

23.9 KB, as published. Nobody here has run it

Circuit Surface Test

You are about to act as a QA tester for the Circuit plugin. The user has asked you to walk the user-facing surface of the plugin in the current host (Claude Code or Codex), looking for regressions, broken progress rendering, missing summaries, and other defects a real user would hit.

This skill is a protocol, not a script. It tells you what to test, in what order, what counts as a finding, and what to put in the report. The source inventory and two host-specific checklists in references/ carry the per-command details.

Host selection

Default to the host you are currently running in:

  • claude — the Claude Code plugin at plugins/claude/. The published slash commands are /circuit:run, /circuit:handoff, and /circuit:pursue. Built-in flows (build, explore, fix, prototype, review) are routed through Run; they ship as compiled flow mirrors under plugins/claude/skills/<flow>/, not as separate slash commands. Pursue is both Run-routable and a direct slash command, and it also ships a compiled mirror like the other flows.
  • codex — the Codex plugin at plugins/codex/. The published Codex commands/skills are run, handoff, and pursue; the other flows are routed through Run and ship as compiled mirrors under plugins/codex/flows/<flow>/. Pursue ships a compiled mirror too.

If the user does not name a host, do not ask. Infer it from the execution context: Codex means codex; Claude Code means claude. If the user explicitly names the other host, honor that as a cross-host test and record the limit in the report. If the execution context is genuinely unknowable, stop with a blocker instead of guessing.

Native vs. cross-host execution

Circuit's surface is meant to be triggered from the host the plugin targets. The normal path is native: current host equals tested host. Cross-host testing only happens when the user explicitly asks for the other host.

You are running inHost to test by defaultExecution mode
Claude CodeclaudeNative. Use real /circuit:* slash commands.
CodexcodexNative. Use real Codex Circuit skills.
Explicit overrideExecution mode
Claude Code testing codexCross-host. Slash commands won't work for the Codex package. Drop to CLI-only and inspect the Codex skill manifests for surface coverage. Note this clearly in the report.
Codex testing claudeCross-host. Same situation, reversed.

When you can't natively invoke a host's surface, you are testing a strict subset (CLI behavior + manifest correctness). State that limitation in the report's Environment block so the operator isn't misled.

Why this skill exists

Circuit's automated tests cover runtime contracts, generated surfaces, and focused connector paths. They do not prove every installed host package path, native rendering path, wrapper injection, and real worker subprocess shape in one user-facing pass. The whole point of this protocol is exercising the installed or packaged host surface against strict schemas.

Freshness is part of the surface. Before any checklist row runs, rebuild the current plugin package from the Circuit checkout, replace the local host plugin caches from that package, and prove the installed caches match. Do not continue with a stale cache and downgrade it to a finding. If cache sync, cache check, or installed-plugin doctor fails after sync, stop before the checklist and report a blocked test setup. If the current host caches command or skill metadata at session start, reload the host surface after cache sync before native rows. If you cannot reload it in the current environment, use the refreshed wrapper for CLI-backed rows and mark native UI rows partial-skip instead of treating a pre-sync loaded command as authoritative.

Concrete consequence: before declaring a clean run, you must have executed at least one default-axis invocation per public flow against the real connector, reached either through Run (host model recommendation — routing is model-only, there is no deterministic router) or the explicit CLI flow start; pursue is the only flow that also ships its own host command. If a default-axis flow aborts on its first relay step, that is the headline finding and the rest of the surface coverage is moot until it's fixed. Section A0 of the checklist exists for exactly this reason — run it first, stop on structural failure.

Why you should be checklist-first, exploratory-second

Circuit is a tool with many entry points, modes, and outcomes. Skipping straight to exploratory testing tends to find one or two showy bugs and miss the boring regressions that matter for a release. Inverting the order — run a fixed checklist first, then circle back to anything that looked off and probe deeper — gives the operator both a baseline pass-rate signal and a ranked issues list.

Run the checklist for the chosen host. Mark each item pass / pass-with-finding / fail / skipped / partial-skip with a one-line note and a pointer to its run folder. Then, for any failures or anything that smelled wrong (slow, ugly progress output, vague error, missing field in the summary), do a focused exploratory pass.

Test environment

Code-changing Circuit flows mutate files. Running them against the user's working repo would damage the actual project. Before starting, set up an isolated scratch repo unless the user has told you to test against the live repo.

# macOS mktemp -t treats the suffix as a literal — drop the trailing X's
SCRATCH=$(mktemp -d "${TMPDIR:-/tmp}/circuit-surface-test.XXXXXX")
cd "$SCRATCH"
git init -q
echo "# scratch fixture for circuit-surface-test" > README.md
echo "function add(a, b) { return a + b; }" > add.js
echo "function buggyAdd(a, b) { return a - b; }" > bug.js

# Stub package.json so flows that emit a verify step have something to run.
# Without this, Build's verify step has nothing to invoke and may abort the
# run even when the implementation step landed correct edits.
cat > package.json <<'JSON'
{
  "name": "scratch",
  "private": true,
  "scripts": {
    "verify": "echo ok",
    "test": "echo ok",
    "check": "echo ok",
    "lint": "echo ok"
  }
}
JSON

git add -A && git -c [email protected] -c user.name=qa commit -q -m "fixture"
echo "Scratch: $SCRATCH"

The bug.js file is a planted bug for exercising Fix. The add.js file gives Build something innocuous to act on. The stub package.json ensures verify-style steps have an executable target. Add more fixture files only if a checklist item asks for one.

Skill-Hooks setup (Section F only). Project skill-hook policy goes in $SCRATCH/.circuit/config.yaml (user-global alternative: ~/.config/circuit/config.yaml). Run Section F rows through the host wrapper with cwd at $SCRATCH — bare bin/circuit from a scratch cwd fails flow resolution before config even loads. For auto-injection rows, pre-install a resolvable local skill at ~/.agents/skills/<id>/SKILL.md (or ~/.claude/skills/<id>/SKILL.md); Circuit discovers only flat <root>/<id>/SKILL.md folders, so namespaced plugin/marketplace skills are not discoverable. With no skill_hooks config the Skill-Hooks surface is silent by design (the opt-in gate), so do not flag the absent section as a regression on a default run.

For Explore and Review you can stay inside the scratch repo (or, if the user explicitly asks, the live Circuit repo) — those flows do not mutate.

Development override when debugging a source checkout

Surface tests should use the refreshed installed plugin root produced by the cache-sync gate. The wrapper uses its bundled runtime by default. If you explicitly need to debug a source checkout while operating on the scratch repo, pass the source CLI as a development override and keep the shell cwd at $SCRATCH. Mark that run as source-debug coverage, not installed-surface coverage:

cd "$SCRATCH"
CIRCUIT_CLI="$WORKTREE/bin/circuit" \
  node "$WORKTREE/plugins/claude/scripts/circuit.ts" run <flow> ...

Do not set project-root environment variables just to make the wrapper find a runtime. The project root the flow operates on should stay the scratch repo unless the operator explicitly asked for live-repo coverage.

Report path

The report and per-run-folder evidence must outlive the scratch repo, since operators sometimes want to re-read findings days later and mktemp -d output gets cleaned. Default location:

REPORT_ROOT="$HOME/circuit-surface-test/$(date -u +%Y-%m-%dT%H-%M-%SZ)"
mkdir -p "$REPORT_ROOT"

If the user passed an explicit --keep-report-at <path> (or set CIRCUIT_SURFACE_TEST_REPORT_ROOT), use that instead. Pass --run-folder "$REPORT_ROOT/<flow-id>" to every Circuit invocation so all evidence lands inside the report tree, not inside the scratch repo. Write the final report to $REPORT_ROOT/report.md.

The protocol

Follow these steps in order. Do not improvise the order of phases — the phases exist to keep the report honest.

1. Select the host and refresh the installed package

Set HOST before reading the checklist:

# Default this from your execution context; explicit user host names override it.
# Codex session: HOST=codex
# Claude Code session: HOST=claude
HOST="<claude|codex>"

Read references/current-surface-inventory.md in full, then run its evidence commands from the current Circuit checkout. The evidence commands now include the latest-state gate:

  1. Run the single local refresh command (npm run plugins:refresh-local).
  2. Confirm its output shows fresh Claude and Codex cache targets and installed doctor success.

If any of those commands fails, stop before the checklist. Do not continue against the repo wrapper or an older host cache and call it a surface test. Record the failed command, stdout/stderr path, and the stale or missing cache details in the report.

Record the observed commit, dirty state, host packages, cache sync targets, CLI help, wrapper versions, wrapper doctor summaries, and installed-package doctor result in the report. Local refresh is a test precondition, not a public marketplace publish.

The generated source map, generated host packages, local CLI/help output, runtime bundles, contracts, and specs are authority. If the checklist and the inventory disagree, trust the observable source, mark the checklist row stale, and record a finding against this skill rather than inventing expected behavior.

2. Read the host checklist

Read either references/checklist-claude.md or references/checklist-codex.md in full. Each checklist enumerates every command/skill, the axis flags to test, the expected outcomes, and what counts as a finding. Do not paraphrase items from memory — read them.

3. Set up the environment

Create the scratch repo as described above. Resolve the plugin root from the refreshed installed cache for the chosen host. Do not point $PLUGIN_ROOT at $CIRCUIT_REPO/plugins/<host> for canonical surface rows; that bypasses the cache freshness proof.

CIRCUIT_REPO="${CIRCUIT_REPO:-/Users/petepetrash/Code/circuit}"
PLUGIN_VERSION="$(node -e 'const fs=require("node:fs"); const path=require("node:path"); console.log(JSON.parse(fs.readFileSync(path.join(process.argv[1],"plugins/version.json"),"utf8")).version)' "$CIRCUIT_REPO")"
case "$HOST" in
  claude)
    PLUGIN_ROOT="${HOME}/.claude/plugins/cache/circuit/circuit/${PLUGIN_VERSION}"
    ;;
  codex)
    CODEX_HOME="${CODEX_HOME:-$HOME/.codex}"
    PLUGIN_ROOT="${CODEX_HOME}/plugins/cache/circuit-local/circuit/${PLUGIN_VERSION}"
    ;;
  *)
    echo "unknown HOST: $HOST" >&2
    exit 2
    ;;
esac

Sanity-check the plugin actually exists and the CLI runs:

node "$PLUGIN_ROOT/scripts/circuit.ts" --help
node "$PLUGIN_ROOT/scripts/circuit.ts" version --json
node "$PLUGIN_ROOT/scripts/circuit.ts" doctor --json

If any of those fail, stop. Record the failure in the report's Environment block and surface it to the user — the rest of the checklist is moot.

Then snapshot the version under test so the report names exactly what got exercised:

git -C "$CIRCUIT_REPO" log -1 --format='%H %s'
git -C "$CIRCUIT_REPO" status --short
node "$PLUGIN_ROOT/scripts/circuit.ts" version --json

Record the commit, plugin version, and any uncommitted-files list in the report's Environment block. If git status shows uncommitted changes that affect plugin behavior (anything in plugins/, src/, commands/, scripts/), call them out explicitly — the test only exercises committed state if you ran npm run build after switching worktrees, so divergence between working-tree intent and what's actually being tested is itself a finding worth surfacing.

3b. Wrapper vs. bare CLI

The plugin ships a wrapper at $PLUGIN_ROOT/scripts/circuit.ts which injects packaged-flow paths and other host-specific arguments before forwarding to the underlying CLI. The wrapper has its own injection logic and is a separate failure surface from the bare bin.

Whenever a checklist row is exercised through the host, you are testing the wrapper. CLI fallback rows in the checklists already use the wrapper path, so most coverage is automatic. But if you spot a wrapper-specific bug (e.g. wrapper injects a flag the underlying subcommand rejects), confirm it also fails through the host and reproduces against the wrapper from a plain shell — not just against bin/circuit directly. Wrapper-only bugs are easy to miss if the bare bin is your default debugging path.

4. Phase 1 — run the checklist

Walk the checklist top to bottom. For each item:

  1. Construct the invocation per the checklist. Native invocations go through the host (slash command or Codex skill). Cross-host or "host can't expose this" items go through the CLI directly with the same arguments the host would have constructed.
  2. Run it.
  3. Capture stdout/stderr (or the host-rendered output) and the run folder path.
  4. Apply the checklist's pass criteria. Use this five-state rubric:
    • pass — every criterion holds.
    • pass-with-finding — the user-observable goal of the row was met (e.g., the file edit landed correctly, the summary rendered) but a non-fatal criterion broke (e.g., the run reported outcome: aborted even though the implementation step succeeded). Record the row as pass-with-finding and file a separate exploratory finding for the broken criterion. Without this tier, every "edits landed but verify cycle aborted the run" turns into a Fail and obscures the real regression count.
    • fail — the row's goal was not met.
    • skipped — the item cannot run in this environment. Note why.
    • partial-skip — a file-backed or CLI-backed degraded check ran, but native host rendering or host UI behavior still needs an interactive session. Use this for operator-summary file checks when you cannot observe the real host render.
  5. Move to the next item. Do NOT pause to fix or investigate failures during this phase — finish the checklist first, then come back. This keeps the pass-rate signal clean and prevents the test session from rabbit-holing on the first regression. The one exception is Section A0's structural smoke matrix: a schema-validation abort there compounds across many downstream rows, so A0 explicitly authorizes stopping and consulting the operator before continuing. Anywhere else in the checklist, take the note and move on.

5. Phase 2 — exploratory pass

After the checklist is complete, look at every fail and every "smelled wrong" pass (slow, weird progress text, surprising summary, ambiguous error, sluggish render between events). Pick the most suspicious 3-5 and probe deeper. The probes that tend to find real issues:

  • Adversarial input: pass tasks containing single-quotes, double-quotes, backticks, $(cmd)-shaped substrings, embedded newlines. The shell-escape rule in every command requires single-quote wrapping with '\'' for literal apostrophes — verify it actually holds.
  • Axis combinations the checklist did not cover. Try valid combinations such as Explore --process high --tournament 2, power dials (--power low, --power auto), and supported autonomous runs. Also test fail-closed combinations such as Build --tournament, Review --autonomous, and --tournament 5 outside the v1 range.
  • Checkpoint resume: trigger a checkpoint-waiting outcome, kill the session, and resume with --checkpoint-choice from a fresh shell.
  • Operator-summary completeness: the host commands render the readable digest (operator_summary_markdown_path) verbatim when present, falling back to run_surface_markdown_path only when the digest is absent. Verify the rendered summary matches the digest file, not the run surface. When operator_summary_status_text or operator_summary_html_path is present, verify host behavior against docs/contracts/host-rendering.md.
  • Outcome legibility: drive a non-complete outcome (the runtime-proof internal flow aborts deterministically offline; goal closes stopped) and confirm the digest reads as an honest failure with a reason, never as a neutral or complete result. escalated in particular must not render as a false success.
  • Skill hooks: if the build under test has Skill Hooks, author a small .circuit/config.yaml policy (see Section F) and confirm a fired hook is disclosed in the digest's ## Skill hooks section — and that a run with no config shows no skill-hook surface at all.
  • Untracked content: in Review, confirm --include-untracked-content changes behavior (default omits content; flag includes it).
  • Cross-flow handoff: save a handoff in the middle of a Build, resume it, confirm the saved state is faithful.

For each probe, write down what you tried, what happened, and what the expected behavior was if the result looks wrong.

6. Run one local protocol smoke

Before writing the final report, run at least one local smoke that proves the protocol can be followed without depending on a live worker connector. Good minimum smokes:

node "$PLUGIN_ROOT/scripts/circuit.ts" version --json
node "$PLUGIN_ROOT/scripts/circuit.ts" doctor --json
node "$PLUGIN_ROOT/scripts/circuit.ts" handoff save \
  --goal 'surface-test protocol smoke' \
  --next 'confirm wrapper accepts utility subcommand args' \
  --state-markdown 'scratch session' \
  --debt-markdown 'none' \
  --progress jsonl

Record the smoke transcript and result in the report. This smoke is not a replacement for the checklist; it only proves the revised protocol's local setup and wrapper path can be followed.

7. Write the report

Use assets/report-template.md as the structure. Fill every section. The report must be self-contained — someone reading it without your terminal should be able to assess plugin health. In particular:

  • Ranked findings. Stack-rank exploratory findings by severity (Critical, High, Medium, Low). Do not bury big problems behind small ones.
  • Repro for every finding. Every finding gets a one-paragraph repro: command run, expected, actual, run folder path or transcript snippet.
  • Source-backed inventory. Include the current commands, public/internal flows, generated host packages, axis allow-list, runtime bundle result, operator-summary fields observed, and any host-only limits.
  • Pass-rate. Surface the checklist pass rate (e.g. 27/31 pass, 3 fail, 1 pass-with-finding, 1 skipped, 2 partial-skip) at the top.

Save the report to the path described in "Report path" above.

8. Print a concise terminal summary

After saving the report, print to the terminal — verbatim, no preamble:

Circuit surface test — host: <claude|codex>
Pass rate: <P>/<T> (<F> failed, <PWF> pass-with-finding, <S> skipped, <PS> partial-skip)
Critical: <count> | High: <count> | Medium: <count> | Low: <count>
Top findings:
  1. <highest-severity finding> — <one-line>
  2. <next highest> — <one-line>
  3. <next highest> — <one-line>
Report: <absolute report path>
Scratch repo: <absolute scratch path>

Top findings are the three highest-severity findings overall — Critical first, then High, then Medium — not one-per-bucket. If there are fewer than three findings, list however many there are. The operator wants to glance and know the health number without opening the file.

What counts as a finding

A finding is anything a real user would notice and complain about. Concretely:

  • A command crashes or returns a non-zero exit code where success was expected.
  • The host fails to render visible presentation.status_text, or when presentation is absent, a visible display.text event.
  • The final operator summary is missing, truncated, or differs from the file it claims to render.
  • A documented flag is rejected, silently ignored, or behaves opposite to its description.
  • An unsupported flag such as --mode or --rigor appears in a generated host surface or checklist row.
  • A generated public flow is missing from the surface inventory, or a host direct command/skill is claimed for a flow that only has packaged CLI JSON.
  • Shell-escape handling is wrong — special characters in task text either blow up the command or appear mangled in the output.
  • A checkpoint-waiting outcome cannot be resumed cleanly.
  • An aborted outcome surfaces no abort reason or surfaces the wrong one.
  • A failure outcome (aborted, escalated, failed, blocked) reads as a neutral or complete result instead of an honest failure — escalated rendering as a false success is the specific regression the failure-legibility work fixed.
  • The host renders the compact run surface (CIRCUIT / ⎿ <status>) as the final answer while the readable digest exists on disk (digest-first is the contract).
  • A configured-and-firing skill hook produces no skill_hook_activations entry and no ## Skill hooks digest section (silent injection — a legibility regression), or the section appears on a run with no skill_hooks config.
  • Time-to-first-useful-output is dramatically worse than expected (subjective but worth noting — log it as Medium and let the operator decide).

A finding is NOT:

  • The flow doing something you personally would not have done. You are testing the surface, not the flow's judgment.
  • The flow taking a long time on a real task — only worth flagging if the CLI itself is unresponsive (no progress events for >30s when work is ongoing).

Voice and posture

You are testing on the operator's behalf. Be terse and concrete in the report. Do not soften findings to be polite. If something is broken, say it is broken and show the repro. If everything passed, say so plainly — a clean report is a real result.

Authority

  • references/checklist-claude.md — Claude Code surface checklist
  • references/checklist-codex.md — Codex surface checklist
  • references/current-surface-inventory.md — current source-backed surface map
  • assets/report-template.md — report structure
  • The Circuit plugin source under plugins/claude/ and plugins/codex/
  • The CLI source at src/cli/circuit.ts

Keep looking

Skills are one crate of 609,856. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.