ccbox-insights（ccbox 见解）

ccbox-insights

Use ccbox session logs to produce an evidence-based "lessons learned" memo for the agent (and humans), plus copy-ready instruction snippets that improve future sessions.

Tool-call failures are an important signal, but the goal is broader than errors: capture what worked, what did not, and what should become standing behavior in the future.

Requirements

• ccbox on your $PATH.
• Access to the local sessions directory scanned by ccbox (see ccbox --help if discovery looks empty).

Quick start (single session)

1. Find the latest session for the current folder:

ccbox sessions --limit 5 --offset 0 --size

1. Inspect the latest session timeline (increase --limit if needed):

ccbox history --full --limit 200 --offset 0

Workflow (recommended)

Follow a staged pipeline: collect -> filter -> summarize -> label -> aggregate -> synthesize -> propose instructions.

Stage 0: Choose scope

• Session: one .jsonl log for deep root-cause analysis.
• Project: last N sessions for one project to find recurring failure patterns.
• Global: a sample across projects to find cross-project patterns.

Stage 1: Collect evidence with ccbox

• Project discovery: ccbox projects
• Session listing: ccbox sessions [project-path] --limit N --offset 0 --size
• Timeline capture: ccbox history [log-or-project] --full --limit N --offset 0
• Skill spans (optional): ccbox skills [log-or-project] --json

When triaging quickly, scan the timeline for failure signals (examples): error, failed, non-zero, rejected, permission, timeout.

Stage 1.5: Identify clarifications, corrections, and interruptions

Treat user clarifications/corrections as high-signal evidence of workflow breakdowns. Your goal is to pinpoint where the agent went off-track and what rule would prevent it next time.

Look for user messages that:

• Clarify intent ("I meant X", "not that", "use Y instead").
• Correct mistakes ("this is wrong", "stop", "revert", "you didn't follow the instructions").
• Restate constraints after the fact ("do not run X", "no emojis", "do not use cargo", "do not change version", "do not release automatically").
• Interrupt the session due to friction (hangs, repeated retries, "cancel", abandoning the thread).

For each such moment, capture a small "course-correction record":

• Trigger: what the agent did immediately before (tool call, plan, edit, or assumption).
• Correction: the exact user sentence(s) that clarified/corrected.
• Fix: what changed after the correction (new approach, different tool, narrower scope).
• Lesson: one rule that would have prevented the detour (copy-ready, scoped project/global).

If the session ends without a clear resolution (or the user abandons it), mark the outcome accordingly and explain the likely interruption reason using evidence (for example: hang/timeout, repeated invalid tool use, conflicting constraints).

Stage 2: Summarize long timelines (only if needed)

If the timeline is too large to analyze in one pass, summarize in chunks:

• Focus on: user request, tool calls, tool outputs/errors, and outcome.
• Preserve: tool names, command lines, error messages, and user feedback.
• Keep each chunk summary to 3-5 sentences.

Stage 3: Extract per-session tool-call facets

For each session in scope, produce one JSON object matching references/facets.md.

If skill spans are available (via ccbox skills --json), annotate each failure with the active skill context when possible (e.g., "this failure happened inside the commit skill span").

Hard rules:

• Use only evidence from the session log; do not guess missing details.
• Separate "tool failed" from "wrong approach" (a tool can succeed but still be the wrong move).
• Count explicit user rejections as their own category (the tool did not fail; the action was declined).

Stage 4: Aggregate and analyze

Aggregate the facet set to produce:

• Top failing tools and failure categories.
• Three root-cause themes with concrete evidence snippets.
• Repeated user constraints that should become standing instructions.
• Engine-neutral recommendations that reduce tool-call failures and improve UX.

Use references/report.md as the output template.

Stage 5: Propose instruction updates (project + global)

Produce two additive sets of copy-ready snippets:

• Project-level: bullets to add to AGENTS.md.
• Global: bullets for the user's global agent instructions.

Guidelines:

• Do not include local paths, repository names, or one-off incident details.
• Prefer rules that prevent repeated errors (2+ occurrences) over one-time fixes.
• Each instruction should include: what to do, what to avoid, and why (1 sentence).

Stage 6: Deliverables

Deliver:

• A concise lessons learned memo (Markdown), following references/report.md.
• Proposed additive AGENTS.md snippet (project-level).
• Proposed additive global instruction snippet.

Optional (for future memory systems):

• "AutoMemorial candidates": a short, structured list of rules that should become standing agent memory, each backed by evidence and scoped (project vs global).

References

• Facet schema + taxonomy: references/facets.md
• Report template + instruction templates: references/report.md