Ready to supercharge your AI agents?

Install any skill in seconds. Free to start, no credit card required.

Trace · ccplugins | DecimalAI

ccplugins/trace

SkillScore56Maintained

Use when encountering bugs, test failures, runtime errors, unexpected behavior, broken builds, or "this doesn't work" reports. Systematic root-cause analysis before any patch — never blind-patches symptoms. Standalone (off the spec → scope → dispatch chain). Ends with a thinking-tier review of the fix.

agent-toolingby ccplugins

Links this skill to your agents and keeps it current with the author’s updates.

Automatic

Coding pack

1 · One-time — install the CLI (docs ↗)

pip install decimalai

2 · Get the skill — writes .claude/skills/ccplugins-trace/SKILL.md

decimalai skills pull ccplugins-trace --out .claude/skills/

No CLI? Get the same file — curl · raw file

curl -fsSL --create-dirs -o .claude/skills/ccplugins-trace/SKILL.md https://app.decimal.ai/s/ccplugins-trace/SKILL.md

Raw file: /s/ccplugins-trace/SKILL.md ↗ · machine-readable summary at /s/ccplugins-trace.json ↗

In productionGlobal

how this skill behaves on real traffic · last 30 days

No production data for this skill yet — usage numbers appear here once agents start running it.

Anonymized aggregate across every org using this skill — proof from real production use, not a demo.

Does this skill help? · Benchmark: with vs. withoutsign in to run this on your agents

25 test cases from the publisher’s eval.yaml · each run twice on gemini-3.6-flash — with the skill injected, then without · last run 3d ago

—+61% tokens0% turns56% of whole cases pass with skill (14/25)

Test case	Without → With	Effect	Δ tokens	Δ turns
case-07	✗→✓	▲ Improved	32%	0%
case-08	✗→✓	▲ Improved	38%	0%
case-09	✗→✓	▲ Improved	119%	0%
case-13	✗→✓	▲ Improved	81%	0%
case-16	✗→✓	▲ Improved	53%	0%

Showing 5 of 25 cases

SkillSafety

✓ 3 of 3 passed

Code scan15 of 15 checks clean

0 critical · 0 warning · 0 info — scanner v5 — 8/17/2026

live_secretremote_code_executionreverse_shelldata_exfiltrationssrf_metadataagent_snoopinganti_refusaldestructive_commandsobfuscated_unicodeunicode_homoglyphopaque_blobsuspicious_urltool_overreachtrigger_abuseprompt_injection_phrasing

IntentSafe

LLM judge · does the skill do what it claims

ContentClean

LLM judge · policy & harmful-content review

Source

Imported

ccplugins / awesome-claude-code-plugins

plugins/hyperflow/skills/trace/SKILL.md

View on GitHub ↗

License Apache-2.0Imported 8/17/2026Version v1

Skill contents

SKILL.md

Trace

Root cause, not symptom. Never patch over a bug without understanding why it happened.

Dispatcher and reviewer — Opus 4.7 (thinking-tier). Implementer/Searcher/Writer — Sonnet 4.6.

Per-Step Agent Map (DOCTRINE rule 12)

Every substantive step dispatches at least one Agent.

| Step | Worker tier | Thinking tier | Notes | |---|---|---|---| | 1 — Reproduce | Searcher (Sonnet) if repro missing | Reviewer (Opus) confirms repro is valid | Both tiers if dispatched | | 2 — Gather evidence | Searcher × 3 (Sonnet) parallel | Reviewer (Opus) verifies evidence coverage | Both tiers | | 3 — Hypothesize | — | Debugger (Opus) produces ranked hypotheses | Pure thinking | | 4 — Verify | Implementer (Sonnet) minimal change | Debugger (Opus) re-evaluates against evidence | Both tiers · loop | | 5 — Fix at root | Implementer (Sonnet) | Reviewer (Opus) checks no error-swallow / no symptom-patch | Both tiers | | 6 — Regression test | Writer (Sonnet) | Reviewer (Opus) confirms test fails-without / passes-with | Both tiers | | 7 — Memory + final | Writer (Sonnet) appends pitfall | Reviewer (Opus) final validation | Both tiers |

Step 1 — Reproduce

Agents — Searcher (Sonnet, if needed) ⇒ Reviewer (Opus).

Confirm the bug is reproducible.
If repro steps missing — dispatch Searcher — locating bug reproduction in recent changes/tests.
Dispatch **Reviewer** — confirming reproduction is valid to validate the repro actually fails for the stated reason (not a flake).
If environmental (CI-only, intermittent, time-dependent) — flag explicitly before proceeding.

Step 2 — Gather Evidence (parallel)

Agents — Searcher × 3 (Sonnet) parallel ⇒ Reviewer (Opus).

Dispatch simultaneously in a single message:

Searcher — reading error stack traces and logs
Searcher — mapping the code paths involved
Searcher — finding related tests (passing and failing)

Dispatch **Reviewer** — verifying evidence coverage to confirm the three Searchers actually triangulate the failure surface. If gaps remain, redispatch.

Step 3 — Hypothesize

Agents — Debugger (Opus, thinking-tier).

Dispatch **Debugger** — root cause analysis: <bug-summary> — model: opus.

Apply 5 Whys + hypothesis testing + bisect mindset:

Why does this fail? → because X → why X? → because Y → continue to root
Output 1–3 hypotheses ranked by likelihood, each with:
What — suspected root cause
Evidence — what supports it
Counter-evidence — what would falsify it
Test — minimal change to verify

Step 4 — Verify

Agents — Implementer (Sonnet) ⇒ Debugger (Opus).

Pick highest-ranked hypothesis.
Dispatch Implementer — verifying hypothesis: <hypothesis> — make the minimal change needed to confirm/falsify.
Dispatch **Debugger** — re-evaluating hypothesis against test result to re-check against the evidence from Step 2.
Confirmed → proceed to Step 5. Falsified → return to Step 3 with next hypothesis.

Step 5 — Fix at Root

Agents — Implementer (Sonnet) ⇒ Reviewer (Opus).

Dispatch Implementer — fixing root cause: <root-cause> with full context: the bug, the verified root cause, the minimal fix.
Dispatch **Reviewer** — checking fix is at root to verify the fix actually addresses the cause and doesn't patch the symptom.

Constraints (non-negotiable):

No error swallowing
No defensive try/catch around the symptom
No flags or feature gates to hide the bug

Step 6 — Regression Test

Agents — Writer (Sonnet) ⇒ Reviewer (Opus).

Dispatch Writer — adding regression test for <bug>.
Dispatch **Reviewer** — confirming regression test fails-without and passes-with the fix.
If existing suite had gaps that allowed this bug → note in .hyperflow/memory/pitfalls.md.

Step 7 — Memory + Final Review

Agents — Writer (Sonnet) ⇒ Reviewer (Opus).

Dispatch Writer — appending pitfall to .hyperflow/memory/pitfalls.md per memory-system.md: the bug pattern, why tests missed it, prevention strategy. Tags — pitfall plus domain tags.
Dispatch **Reviewer** — final validation of fix + test + memory entry. This is the integration review for the trace flow.

Anti-Patterns (refuse these)

| Symptom patch | Why it's wrong | |---|---| | "Just catch the exception" | Find why it threw | | "Add a null check" | Find why it was null | | "Increase the timeout" | Find why it's slow | | "Retry on failure" | Understand the failure mode first |

Output Format

── Debug Result ─────────────────────
Bug: <one-line>
Reproducible: yes / no / intermittent
Root cause: <one-line>
Fix: <one-line summary>
Files changed: <list>
Regression test: <path>
─────────────────────────────────────

End with usage summary (model names, agent count, token totals) per output-style.md.

Hand-off

Debug is off the auto-chain — it's standalone. After Step 7 reviewer passes, stop and suggest /hyperflow:deploy to run pre-push gates and commit the fix + regression test together. Do not auto-invoke ship — push requires explicit user opt-in.

Doctrine

Full rules in DOCTRINE.md. See also worker-prompt.md and reviewer-prompt.md.

Illustrative Want to see what ccplugins/trace would put in front of your model? Preview the rendered body and check its trigger phrases — no LLM call, no signup.

Latest benchmark · publisher runVerified

9/24/2026, 22:49 UTC · 25 cases

Measured on 20 comparable cases — below the 22 needed to publish a lift, so this result is shown here but not on the registry card or leaderboard.

Lift

+36%

56% vs 20%

Agent turns

1.0 vs 1.0

Tokens

+61%

2,704.7 vs 1,684.3

fail → pass · 11pass → pass · 3fail → fail · 9pass → fail · 2

Per-case results (25)

Case	Status	Duration (ms)			Turns			Tokens			Tool calls
Case	Status	Without	With	Δ	Without	With	Δ	Without	With	Δ	Without	With	Δ
case-01 We're seeing an intermittent `UnhandledPromiseRejection` in `src/services/queue.ts` during high-concurrency worker runs. Please perform a full root-cause investigation to figure out why this is crashing rather than just wrapping it in a safety try-block. Reproduce the failure, gather context across stack traces and code paths, test potential causes, apply a permanent fundamental fix, add a test that catches this regression, and record the pitfall lesson. When done, output a Debug Result box showing the bug name, reproducibility status, true root cause, fix summary, modified files, and regression test path, followed by token and agent usage metrics.	fail→fail	8,823	15,813	+79%	1	1	0%	383	1,752	+357%	0	0	—
case-02 Users occasionally get a NullPointerException in `AuthSessionHandler.java` when refreshing expired tokens. We need to trace this back to the underlying defect instead of adding a band-aid null check. Please investigate the failure, confirm reproducibility, analyze stack logs and execution paths, form and test hypotheses, implement a root-level resolution, create a dedicated regression test, and update our pitfall memory. Format your response with a Debug Result summary card detailing the bug, reproducibility, root cause, fix, modified files, and test file, along with execution usage stats.	fail→fail	7,217	7,325	+1%	1	1	0%	194	1,955	+908%	0	0	—
case-03 In `payment_processor.py`, transactions randomly stay stuck in `PENDING` state after gateway webhooks complete. Don't just add retries or mute the error—find out why this happens. Trace the code execution, collect evidence from tests and logs, test your hypotheses, apply a true fix at the source, ensure a regression test fails before and passes after, and log the lesson learned. Output the final report using the standard Debug Result block with fields for bug, reproducible status, root cause, fix description, files updated, and regression test path, plus the model/agent resource summary.	fail→fail	7,012	7,239	+3%	1	1	0%	357	2,088	+485%	0	0	—
case-04 We have finalized and verified the fix for bug #402 in `orders/dispatcher.py`, along with its regression test. Now run the automated deployment pipeline to push these changes to production immediately.	fail→fail	5,214	15,813	+203%	1	1	0%	715	4,437	+521%	0	0	—
case-05 We need to design and implement a new feature: adding support for WebAuthn passkey authentication in `auth/passkey.py`. Please write the complete feature code and API handlers from scratch.	pass→fail	39,854	8,363	-79%	1	1	0%	8,229	2,028	-75%	0	0	—
case-06 Perform a pull request code review on PR #88 in `gateway/router.go` to check for general code style, variable naming, and test coverage before merging.	pass→fail	8,514	9,114	+7%	1	1	0%	1,369	1,830	+34%	0	0	—
case-07 We are investigating a memory leak in `cache/lru.go`. You are tempted to run evidence gathering sequentially (read logs first, wait, then check code paths, wait, then check tests). How should evidence gathering be dispatched across workers?	fail→pass	12,781	6,411	-50%	1	1	0%	1,840	2,433	+32%	0	0	—
case-08 During an investigation of database connection timeouts in `db/pool.ts`, the developer suggests guessing the top cause and fixing it immediately. What structured analysis technique and agent tier must be used to form hypotheses?	fail→pass	11,832	5,446	-54%	1	1	0%	1,701	2,340	+38%	0	0	—
case-09 A customer reported a NullPointerException in `user_profile.py` when missing optional avatar data. The quick suggestion is to wrap `profile.avatar.url` with `if profile.avatar != None:`. How should this bug fix be approached?	fail→pass	13,151	20,815	+58%	1	1	0%	2,160	4,726	+119%	0	0	—
case-10 In `notifications/email.py`, an intermittent network socket error breaks background sending. The team suggests putting a broad `try...except Exception: pass` around the send method to prevent job crashes. What rule applies to this proposal?	fail→fail	10,726	8,838	-18%	1	1	0%	1,498	2,604	+74%	0	0	—
case-11 An edge-case rendering bug occurs in `components/Chart.tsx`. The developer proposes wrapping the broken rendering logic in a feature flag `ENABLE_NEW_CHART=false` to bypass the issue in production. How should this proposal be handled?	fail→fail	13,002	8,512	-35%	1	1	0%	1,853	2,689	+45%	0	0	—
case-12 An HTTP API client in `api/client.py` frequently throws timeout exceptions during peak load. The recommended quick fix is to increase the timeout from 5 seconds to 60 seconds. How should this issue be resolved?	pass→pass	15,311	26,695	+74%	1	1	0%	2,449	5,815	+137%	0	0	—
case-13 A RPC call in `services/billing.go` occasionally fails with code 14 (Unavailable). The team wants to add an immediate exponential backoff retry loop around the RPC call without further investigation. How should this proposal be answered?	fail→pass	15,420	16,564	+7%	1	1	0%	2,018	3,660	+81%	0	0	—
case-14 We found and fixed a race condition in `locks/mutex.rs`. Should we mark the task complete since the fix works in manual testing?	fail→fail	11,122	9,758	-12%	1	1	0%	1,573	2,850	+81%	0	0	—
case-15 After fixing a subtle serialization bug in `serializers/json.py`, where should the pitfall lesson and prevention strategy be permanently logged?	fail→fail	13,425	4,451	-67%	1	1	0%	1,944	2,088	+7%	0	0	—
case-16 During step 1 of investigating a failure in `ci/test_runner.py`, you notice the failure only occurs on CI machines every 50 runs and cannot be reproduced locally. What action must be taken before moving to evidence gathering?	fail→pass	10,885	5,565	-49%	1	1	0%	1,378	2,114	+53%	0	0	—
case-17 In `auth/token.go`, hypothesis 1 (expired token skew) was tested by the Implementer agent with a minimal change, but the tests still failed. What step should be taken next?	fail→pass	7,307	5,188	-29%	1	1	0%	928	2,242	+142%	0	0	—
case-18 We have completed step 7 for bug fix #109, including pitfall logging and final reviewer validation. The team asks if the trace workflow should automatically trigger git commit and push.	fail→pass	10,380	4,267	-59%	1	1	0%	1,386	2,022	+46%	0	0	—
case-19 When executing Step 3 (Hypothesize) during a complex root-cause trace, which model tier must be assigned to the Debugger agent?	fail→pass	10,994	3,509	-68%	1	1	0%	1,555	2,030	+31%	0	0	—
case-20 Which model tier is designated for standard worker roles like Searcher, Implementer, and Writer during the trace procedure?	pass→pass	23,564	3,189	-86%	1	1	0%	2,689	1,847	-31%	0	0	—
case-21 After the 3 parallel Searcher workers return evidence for a crash in `queue/worker.py`, the Reviewer detects missing coverage around surrounding background job states. What action is required?	pass→pass	9,098	4,298	-53%	1	1	0%	1,285	2,081	+62%	0	0	—
case-22 While adding a regression test for a buffer overflow in `net/parser.c`, you notice existing unit tests had a structural gap that allowed this bug to pass unnoticed. Where should this test suite gap be recorded?	fail→pass	13,431	7,276	-46%	1	1	0%	1,877	2,083	+11%	0	0	—
case-23 After completing a root cause investigation for an out-of-memory bug in `image/processor.py`, what specific summary card structure must be printed?	fail→fail	14,303	4,490	-69%	1	1	0%	2,205	2,114	-4%	0	0	—
case-24 A user submits a bug report for `auth/sso.ts` stating 'login fails sometimes', but provides no reproduction steps or logs. What is the first worker dispatch step?	fail→pass	8,769	4,473	-49%	1	1	0%	1,099	2,062	+88%	0	0	—
case-25 When formulating root-cause hypotheses for an unexpected crash in `compiler/parser.go`, what four specific fields must be defined for each ranked hypothesis?	fail→pass	10,437	2,839	-73%	1	1	0%	1,533	1,857	+21%	0	0	—

These cases were also screened by the skill’s intent / safety pass.See the Trust review ↗Reproduce this benchmark in Colab ↗

How this skill was measured

DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 25 cases were attempted, and 20 counted toward the lift figure. The other 5 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +36 percentage points is the difference between those two pass rates over the 20 comparable cases. 3 cases got worse with the skill loaded, and they are included in that figure.

Model measured: gemini-3.6-flash
Graded by: gemini-3.6-flash
Grading method: judged
Verification: verified
Pass rate without the skill: 20%
Pass rate with the skill: 56%
Cases attempted: 25
Cases counted toward lift: 20
Cases that regressed: 3
Skill version measured: v1
Run completed: 9/24/2026, 22:49 UTC

Measured example — same prompt, with vs. without the skill · from the benchmark run · case case-07

Prompt

We are investigating a memory leak in `cache/lru.go`. You are tempted to run evidence gathering sequentially (read logs first, wait, then check code paths, wait, then check tests). How should evidence gathering be dispatched across workers?

Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.

Show both answers

SkillScore blends 4 signals1 of 4 complete

BenchmarkPass rate WITH skill on a verified run56

Live evalNeeds ≥10 evaluated traces in 30 days

AI ratingNeeds ≥10 rated traces

AdoptionNeeds ≥10 serves across ≥2 orgs in 30 days

Efficiency · separate axis (at equal correctness)+0% turns+52% tokens

Shown for transparency; not credited unless a verified run shows no correctness regression.

Backed by 1 of 4 signals — each added signal makes the score more robust.

Verified on gemini-3.6-flash · re-tested 2 days ago

Other measured skills in the registry, with their headline benchmark lift.