thevibeworks/ccx/ccx-retro

Claude Code and agent skill by thevibeworks/ccx

Trust tier 3
  • community
Source check pending

Checking that the SKILL.md file is still in the repository.

Where did the agent go wrong, what fixed it, and what should change so it doesn't happen again? Mine a coding-agent session for mistakes, corrections, and wasted loops, then propose durable rule changes (CLAUDE.md / AGENTS.md / skills / memory) with evidence. Use after a frustrating session, when the human asks "what went wrong", "why did it do that", "retro", or "turn this into a rule". Routing: this skill answers "what should change"; for a plain "what happened" account, use ccx-recap.

Install

npx vskill@latest install thevibeworks/ccx/ccx-retro

SKILL.md

The same SKILL.md was published earlier as thevibeworks/ccx/ccx-retro.

ccx-retro

Recap tells you what happened. Retro makes the next run better. The real output is not a report — it is a proposed patch to the instructions the agent already reads (CLAUDE.md, AGENTS.md, a skill, memory), each line justified by evidence from the session.

This skill describes ccx >= 0.8. If a command or flag here is missing, upgrade ccx and run ccx skills install — the binary embeds skills matching its own CLI surface.

Process

  1. Outline. ccx trace --json. Failure signals are factual and

already in it:

  • steps with errors (tool failures, attributed to the step that

issued the call)

  • error streaks: consecutive steps failing at the same thing
  • turns where the user pushed back — read the actual user_text

and judge with your own comprehension; there is no correction flag, because keyword matching lies

  • interrupts and denials (turn/step/stats): the human pressed

stop or refused a tool call — the sharpest correction signal in the log; the denied call's evidence is marked denied

  • cost spikes without matching edits (spinning)
  • warnings (evidence gaps are findings too)
  1. Drill. ccx trace --turn N on each suspect; `ccx view

<session> --at <message_id> for the exact exchange. Reconstruct: what did the agent believe, what was actually true, what evidence was available at the time, what finally corrected it (user, test, self-check)? If the mistake was inherited from an earlier session (a stale handoff, a fork), ccx related <session>` finds it.

  1. Classify each finding.
  • mistake — agent had the evidence and got it wrong
  • friction — environment/tooling failed the agent
  • drift — agent misread intent; user had to steer
  • save — something caught the problem (name what: a test, a

verification habit, a user correction). Saves become rules too.

  1. Propose changes. For each finding worth keeping, draft the

exact edit: file, section, wording. A rule earns its place only if:

  • it would have prevented or shortened the observed failure, AND
  • the next agent cannot derive it from code, docs, or common sense

Fewer, sharper rules beat a rulebook. One-off accidents are reported, not legislated.

  1. Human gate. Present findings + proposed patches and STOP.

Never write to CLAUDE.md / AGENTS.md / skills / memory without explicit approval. Rejected proposals are dropped, not stashed.

  1. Land approved patches with provenance. When the human

approves, apply the patch and carry the citation with the rule — a trailing (ccx:<session-id> #turn.step) reference — so the next reader can re-open the receipt with ccx trace instead of trusting the rule on faith. The session id is the audit trail; the rule text travels inline. Findings the human read but did not turn into rules may still be worth a devlog/memory entry — offer once, cite the same way.

Output shape

Retro — <session> | <n> findings
# Kind Finding (cited) Proposed change
1 mistake Assumed X, evidence said Y [#12.3-12.7] CLAUDE.md: "check X before Y"
2 friction 3 loops on flaky sandbox error [#9.2] none — one-off, monitor
3 save Probe test killed a rewrite [#133.1-3] AGENTS.md: "profile before rewrite"
Patches for approval:
--- CLAUDE.md
+ <exact line(s)>

Every finding cites #turn.step. If nothing is worth changing, say so — "no durable lessons this session" is a valid retro.

Hard rules

  • Findings without receipts don't ship.
  • Judge corrections by reading the exchange, not by keywords.
  • Propose the smallest rule that prevents the failure class.
  • Do not create new knowledge stores; patch the ones the agent

already loads.

  • Preserve the user's exact wording when their correction becomes a

rule — their phrasing carries the intent.

Security

  • Security scanPassed, 95 of 100Pattern checks on the skill's files.
  • Model reviewPassed

Quality evaluation

  • Task evaluationNot runNo evaluation yet. Run one in Skill Studio.

Source

Popularity

Repository stars
11
Repository forks
3
7-day trend score
12

Stars and forks are for the whole GitHub repository, not this skill alone. Not refreshed yet.

Works with

Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.

  • Claude Code
  • Cursor
  • GitHub Copilot
  • Windsurf
  • Codex
  • Gemini CLI
  • Cline
  • Amp
  • Roo Code
  • Goose
  • Continue
  • Kilo Code
  • Junie
  • Trae
  • OpenHands
  • Qwen Code
  • AI Agent
  • Tabnine Chat
  • Sourcegraph Cody
  • Amazon Q
  • Firebase Genkit
  • Sweep
  • Devin
  • Phind
  • Mentat
  • Claude Desktop
  • Aider
  • GPT Pilot
  • Plandex
  • bolt.new
  • v0
  • Replit Agent
  • Anse
  • TypingMind
  • BetterChatGPT
  • LibreChat
  • Warp AI
  • Pieces
  • Fig AI
Embed badge
vskill VERIFIED badge
[![vskill VERIFIED](https://verified-skill.com/api/v1/skills/thevibeworks/ccx/ccx-retro/badge)](https://verified-skill.com/skills/thevibeworks/ccx/ccx-retro)

Something wrong with this skill or its page? Tell us.