opencode-conversations

name

opencode-conversations

description

View opencode CLI conversations/sessions from Claude Code or opencode — pull the FULL transcript (every turn, never a truncated window) or a COMPACTED summary produced by opencode's own configured compaction agent (agent.compaction in opencode.json) using opencode's built-in compaction prompt, run non-destructively on a throwaway fork. Use when the user wants to see / read / inspect / summarize an opencode session, check what happened in an `opencode -s <id>` session, or get a compact recap of a long opencode conversation. Triggers on: opencode session, opencode conversation, opencode -s, ses_, view opencode, read opencode chat, summarize opencode session, compact opencode, what happened in opencode.

opencode-conversations — view opencode sessions (full or compacted)

Engine: oc-convo.py (next to this file). It shells out to the opencode CLI and, for compaction, to opencode's local HTTP server. Works from Claude Code and the opencode CLI — the same skill is installed in both:

Agent	Skill path
Claude Code	~/.claude/skills/opencode-conversations/
opencode CLI	~/.config/opencode/skills/opencode-conversations/

Resolve the script path (copy-paste this first)

OC_CONVO="$HOME/.claude/skills/opencode-conversations/oc-convo.py"
[ -f "$OC_CONVO" ] || OC_CONVO="$HOME/.config/opencode/skills/opencode-conversations/oc-convo.py"

(If the installer created the convenience launcher, oc-convo on PATH works too.)

Commands

python3 "$OC_CONVO" list [--limit N]            # recent sessions: id / title / updated
python3 "$OC_CONVO" full   <sessionID>          # the ENTIRE conversation, readable
python3 "$OC_CONVO" compact <sessionID>         # summary via the configured compaction agent
python3 "$OC_CONVO" raw    <sessionID>          # clean raw JSON (for further processing)

full — the whole conversation

Prints every message in order: user/assistant text in full, assistant reasoning, tool calls with their arguments, and tool results. Tool results are truncated to keep it readable (they are often huge logs); text is never truncated.

• --tool-chars N — cap tool-result length (default 800; -1 = show everything).
• --no-reasoning — omit assistant reasoning blocks.

This always pulls the full session, never a segment.

compact — opencode's own compaction agent

Produces a summary using opencode's configured compaction agent — it reads agent.compaction.model from ~/.config/opencode/opencode.json (e.g. deepseek/deepseek-v4-flash) and calls opencode's POST /session/{id}/summarize, which applies opencode's built-in compaction prompt.

• Non-destructive by default: the session is forked, the fork is summarized, the summary is printed, and the fork is deleted — the original session is never modified.
• --in-place — instead add a real compaction checkpoint to the original session.
• --model provider/model — override the compaction model.
• --timeout S — summarize timeout (default 600s).

How to use (agent guidance)

1. If you don't know the session id, run list (or the user gives an ses_… id, e.g. from opencode -s ses_…).
2. Default to full when the user wants to read/inspect what happened.
3. Use compact when the user wants a recap/summary of a long session, or asks for the "compacted" version. Mention it ran via opencode's compaction agent.
4. For very long sessions, prefer piping full to a file and reading that, or use compact to get the gist first.

Requirements & notes

• opencode CLI on PATH (or at ~/.opencode/bin/opencode) and python3 (stdlib only).
• compact needs the compaction provider authenticated in opencode (opencode auth list) — e.g. DeepSeek for deepseek-v4-flash.
• Robustness baked into the engine: exports are written to a temp file (opencode truncates piped stdout at 64 KiB), and run with --pure to disable plugins so the JSON is clean.