debug

Debug mode

How to run the reverse proxy with pruning debug output and inspect which tools were kept vs dropped.

cyt-indexer CLI

Check how the app mutates the requests by decomposing tools and then re-composed based on survivors to see what the tool will change using cyt-indexer-cli.sh script.

Install and run

From the repo (or any machine with the project deps):

uv tool install 'clear-your-tools[all]'

uv run cyt proxy --debug

--debug logs each transformed request to .debug/anthropic.log (see network.proxy.reverse.debug_log_dir in config) and still forwards to upstream. It also enables full tool JSON in stats (store_full_tools is turned on while debug is active).

Dry-run (no upstream call):

uv run cyt proxy --debug-dry-run

Point Claude Code at the proxy:

export ANTHROPIC_BASE_URL="http://localhost:8834/anthropic"
export OPENROUTER_API_KEY="..."
export ANTHROPIC_AUTH_TOKEN="${OPENROUTER_API_KEY}"

Pruning needs API keys for the configured pipeline (e.g. OPENROUTER_API_KEY, DEEPINFRA_API_KEY). See scripts/proxy.sh for a local example.

SQLite: full tool lists before and after pruning

Default DB path: ~/.config/cyt/stats.db (stats.database.path in config).

Each proxied request is recorded in proxy_request. Related tables: model_request (per-stage upstream/pruning calls) and tokens (token counts per model request).

proxy_request columns (current schema)

Column	Type	Meaning
id	TEXT	UUID7 primary key (not an integer; do not use MAX(id) for “latest”)
endpoint	TEXT	Route name (e.g. anthropic)
tools_in	INTEGER	Tool token count before pruning
tool_count_in	INTEGER	Number of tools before pruning
tool_properties_count_in	INTEGER	Optional property count before pruning
tools_out	INTEGER	Tool token count after pruning
tool_count_out	INTEGER	Number of tools after pruning
tool_properties_count_out	INTEGER	Optional property count after pruning
tools_pruned	INTEGER	Tool tokens saved
tool_count_pruned	INTEGER	Tools removed
tool_properties_count_pruned	INTEGER	Optional properties removed
ts_ms	INTEGER	Request time (Unix ms); use for ordering
prune_status	TEXT	e.g. applied, skipped, error
pipeline	TEXT	JSON array of pruning stages
query	TEXT	Text used for retrieval/rerank
error	TEXT	Pruning error message, if any
tools_accepted_json	TEXT	Full tools JSON before pruning (store_full_tools or --debug)
tools_final_json	TEXT	Full tools JSON after pruning (store_full_tools or --debug)

Only some rows store full tool JSON (proxy run with --debug or stats.store_full_tools: true). Use this filter on every investigation query:

WHERE tools_accepted_json IS NOT NULL
  AND length(tools_accepted_json) > 0

Example queries (all scoped to rows with stored tool JSON):

# Last few investigable rows (pick an id for the queries below)
sqlite3 ~/.config/cyt/stats.db "
SELECT id, datetime(ts_ms / 1000, 'unixepoch') AS ts
FROM proxy_request
WHERE tools_accepted_json IS NOT NULL
  AND length(tools_accepted_json) > 0
ORDER BY ts_ms DESC
LIMIT 10;
"

# One row by id (paste id from the list query above) and save into a .json file
ID=019e6c0a-035c-7bd6-ab8b-75492296b553
sqlite3 ~/.config/cyt/stats.db "
SELECT tools_accepted_json
FROM proxy_request
WHERE id = '${ID}'
  AND tools_accepted_json IS NOT NULL
  AND length(tools_accepted_json) > 0;
" > tools_accepted.json

sqlite3 ~/.config/cyt/stats.db "
SELECT tools_final_json
FROM proxy_request
WHERE id = '${ID}'
  AND tools_final_json IS NOT NULL
  AND length(tools_final_json) > 0;
" > tools_final.json

CLI summaries (token totals, not per-tool JSON):

cyt stats totals
cyt stats events --limit 20

To persist full tool JSON without --debug, set stats.store_full_tools: true in ~/.config/cyt/config.yaml.

.debug/anthropic.log: snapshot per request

Each block is a JSON snapshot (timestamp header, then payload). Useful fields:

• body.tools — tools after pruning (what would be sent upstream); these are the tools that “survived.”
• pruning — metadata for the same request:
  • input.tools — full copy of request tools before pruning (unchanged from the client payload)
  • tools_in / tools_out — tool counts before and after pruning
  • status — e.g. applied, skipped, error
  • query — text used for retrieval/rerank
  • token and decomposed breakdown fields when pruning ran

For side-by-side JSON outside the log file, tools_accepted_json / tools_final_json in SQLite still apply when store_full_tools is enabled (on by default with --debug).

Console lines while --debug is on also print a one-line summary, e.g. pruning status: applied (tools 30 -> 28).

Related

• README: Debug without calling upstream (--debug-dry-run)
• Default listen port: 8834
• Log rotation: large anthropic.log files may appear as anthropic.log.1 under .debug/

Request Example

To investigate the request prior and after pruning:

3 main parts to look at:

• User Query saved in content.messages: (role: user)
• Final tools after pruning is in tools:
• Original request tools are in pruning.input.tools
• Decomposed optional properties per stage are in pruning.decomposed_catalog.<stage>.json; enums are in md

See as short example in example.json