opencode-smartsnip

Cuts shell-output tokens in opencode.

Wraps only what snip can filter. Everything else runs byte-identical.

Install · Numbers · How routing works · Configuration

 

- git status && cat data.json | jq '.users | length'
+ snip git status && cat data.json | jq '.users | length'

Only git status gets the snip prefix. The jq pipe is left exactly as the agent wrote it. smartsnip works out which parts of a command are safe to wrap, wraps just those, and does nothing when it isn't sure.

Install

brew install edouard-claude/tap/snip

// ~/.config/opencode/opencode.json
{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["opencode-smartsnip"]
}

That's it. No new tools, no prompt overhead. If snip isn't on PATH the plugin disables itself with a warning, so it's safe in a shared repo config.

Install from source instead

git clone https://github.com/carson2222/opencode-smartsnip.git
mkdir -p ~/.config/opencode/plugins
printf 'export { default } from "%s/src/index"\n' "$PWD/opencode-smartsnip" \
  > ~/.config/opencode/plugins/smartsnip.ts

Numbers

One audited day of real work on a real project: ~600 bash calls, every credited output replayed through the actual snip filters — no command re-runs.

git output	−70%
pnpm output	−69%
all bash output	−11% ¹
context traffic avoided	~1M tokens

¹ Conservative floor: only single-segment commands are credited; the 158 chained commands that day counted as zero savings. Real reduction is higher.

Tool output gets re-sent on every turn that follows it, so a token saved at the source stays saved for the rest of the session — and through compaction. That multiplier is why an 11% cut on raw output erased ~1M tokens of cumulative context traffic in a single day. It was a browser-automation-heavy day; run the same conservative replay on your own history:

bun scripts/measure-savings.ts --days 7        # measured, replayed through snip
bunx opencode-smartsnip discover --days 30     # estimated, plus what to filter next

How routing works

flowchart LR
    A([bash command]) --> B{heredoc, loop,<br>unbalanced quotes?}
    B -- yes --> P([run as-is])
    B -- no --> C[split into<br>top-level segments]
    C --> D{builtin, piped,<br>redirected, subshell,<br>denied?}
    D -- yes --> P
    D -- no --> E{in snip's<br>filter table?}
    E -- no --> P
    E -- yes --> W([prefix with snip])

Loading

The allowlist comes from snip's own filters (131 built-in, plus anything you drop in ~/.config/snip/filters/). exclude_flags and require_flags are honored, so git log --format=... stays raw. #nosnip anywhere in a command skips the whole thing.

A wrong passthrough costs a few tokens. A wrong wrap breaks a command. The bias follows.

Configuration

Optional. ~/.config/opencode/smartsnip.json, overridable per project in .opencode/smartsnip.json:

{
  "deny": ["pnpm", "git diff"],
  "allow": ["curl", "mytool"],
  "toast": true,
  "stripMimicry": true
}

• deny — never wrap these ("cmd" or "cmd subcommand")
• allow — force wrap-eligibility, wins over deny
• toast — once per session, a small TUI toast with tokens saved
• stripMimicry — strip stray snip prefixes the agent picked up from history before re-deciding (default on). Turn off only if you wrap commands via snip filter dirs that smartsnip doesn't scan

ssh, curl, wget, psql, jq are denied by default. snip has filters for them, but they're blunt truncations and agents usually need that output verbatim. A truncated API response forces a re-run, which costs more than it saves. "allow": ["curl"] brings any of them back.

Getting raw output back

snip can tee the original to a local file and append [full output: /path.log] to the filtered result. The agent already has a Read tool, so nothing is ever lost. In ~/.config/snip/config.toml:

[tee]
mode = "always"

smartsnip doctor checks this, along with the rest of your setup.

One line in AGENTS.md makes agents use it well:

Shell output is auto-compressed. Append `#nosnip` to a command if you need raw output.
If you see `[full output: <path>]`, Read that file instead of re-running.

Finding what to filter next

bunx opencode-smartsnip discover --days 30

Replays your real opencode bash history (read-only, local) through the router:

2106 commands, ~893.5k tokens of raw output

FILTERED by snip:
  git        418 calls   182.4k est. tokens
  pnpm       184 calls   164.7k est. tokens

NO FILTER (biggest missed savings first):
  python3     49 calls    73.2k est. tokens
  agent-browser 83 calls  33.2k est. tokens

A snip filter is ~10 lines of YAML. bunx opencode-smartsnip install-command adds a /snip-filter command that writes and tests one for you — say /snip-filter python3 and the new filter is picked up automatically.

Why not wrap everything?

That's what opencode-snip does, and it's where this project started. The failure modes are all known issues there:

	wrap everything	smartsnip
snip: no filter for "X" noise	#16	never from routing ¹
snip snip stacking	#15	normalized away per segment
jq '.a | .b' quoted pipes	#8	quote-aware parser
VAR=$(cmd) corruption	#22	detected, passthrough
heredocs	#6	detected, passthrough
permission rules see rewritten commands	#7	only filterable commands change

The router is validated against 23k+ real bash commands from actual opencode sessions — 65% of calls still get filtered, with none of the breakage.

¹ One feedback loop is unavoidable at this layer: opencode stores the rewritten command, so agents start typing snip themselves — sometimes on things snip can't filter (snip sed, … | snip python3). smartsnip strips those stray prefixes back off before running (stripMimicry, on by default), which also collapses any stacking. For the rare command too complex to parse, set quiet_no_filter = true under [display] in ~/.config/snip/config.toml as a backstop — smartsnip doctor checks for it.

Development

bun install
bun test                  # includes a replay of 656 sanitized real-world commands
bun run typecheck
bun run generate:filters  # re-sync allowlist from upstream snip filters
bun run measure --days 7  # replay your real bash history through snip (the Numbers)

License

MIT