Token Optimization

Reference for the settings that trim what Pando sends to the model. Step by step: Save tokens. Related: Tool Discovery.

All of these are fail-safe: if a compressed read would use more tokens than the raw one, Pando sends the raw one.

`[TokenOptimization]`

[TokenOptimization]
ReadModeDefault = "full"         # full | auto | signatures | map
ReadDedupDisabled = false        # true sends unchanged re-reads again in full
ReadModeLearning = false         # extra learning layer for auto mode
BuildCodeGraph = true            # record file relationships while indexing
RelatedFilesHint = false         # add a short list of related files to reads and searches
SavingsLedgerDisabled = false    # true stops recording savings

{
  "tokenOptimization": {
    "readModeDefault": "full",
    "readDedupDisabled": false,
    "readModeLearning": false,
    "buildCodeGraph": true,
    "relatedFilesHint": false,
    "savingsLedgerDisabled": false
  }
}

Setting (Web UI label)	Key	Default	Meaning
Default read mode	`ReadModeDefault`	`full`	`full`: the file as stored. `auto`: chosen by size and type. `signatures`: function and class names with their signatures. `map`: imports and top-level declarations
Deduplicate unchanged re-reads	`ReadDedupDisabled`	on	A re-read of an unchanged section returns a short reference
Adaptive auto-mode learning	`ReadModeLearning`	off	In `auto`, Pando always counts “bounces” (a compressed read followed at once by a full read of the same file) per file and per extension, and sends fuller reads where they repeat. This switch adds a statistical layer that predicts them before they happen
Build code property graph	`BuildCodeGraph`	on	Needed for impact analysis and related-file hints
Related-files hint	`RelatedFilesHint`	off	Costs a few tokens per read
Record token-savings ledger	`SavingsLedgerDisabled`	on	Feeds the savings widget and `pando gain`. Stored in `<data dir>/savings/ledger.jsonl`

The bounce tracker only applies to auto. With full, signatures or map chosen explicitly nothing is learned.

PANDO_READ_MODE_DEFAULT=auto overrides the read mode for one run.

Shell output (`[Bash]`)

[Bash]
OutputFilterDisabled = false   # true returns raw command output to the model
OutputFilterPaths = []         # extra TOML filter files; earlier paths win, all override the built-ins

Setting (Web UI label)	Default	Meaning
Enable output compression	on	Trims boilerplate, repeated headers and formatting from command output (git, builds, tests, linters). Exit codes and errors are kept
Extra filter files	empty	Your own filter definitions

Tool discovery (`[ToolDiscovery]`)

[ToolDiscovery]
Enabled = true
Mode = 'auto'            # auto | always | off
MaxDirectTools = 64      # threshold for auto
SearchLimit = 8          # results per tool_search
NonDeferredTools = []    # tools always visible
DeferredSources = []     # sources to keep back, e.g. "mcp", "lua"

Mode	Behaviour
`auto`	Starts when the number of tools goes above `MaxDirectTools`
`always`	Always keeps non-core tools back
`off`	All tools visible

Tool sources: core (built-in, always visible), internal, mcp, lua, mesnada, rag (knowledge base and code index), gateway (favourite MCP tools).

tool_search finds and runs tools. With a query it searches, ranking by how well the words match the tool’s name, aliases, server name, description and parameter names:

{ "query": "search code in repository" }

With a tool name it runs it, whether the tool is built in or lives on an MCP server:

{
  "tool_name": "github_create_issue",
  "parameters": { "title": "Fix login redirect" }
}

Other savers

[LLMCache]
Enabled = true     # provider prompt cache

Optimize images and Model Catalog (models.dev) are switches in Settings > General.

Savings report

pando gain                 # all-time summary (aliases: stats, savings)
pando gain --days 30       # a period
pando gain --price 3       # estimate $ saved at this price per 1M tokens
pando gain --json

The pando_stats tool gives the agent the same numbers during a conversation, with an optional days parameter.

Quick reference

Feature	Default	Recommended	Savings
Read mode	Full	Full or Auto	Medium to high
Deduplication	On	On	Medium
Adaptive learning	Off	Off	Low
Output compression	On	On	High
Code graph	On	On	Enables features
Related files	Off	Optional	Low
Savings ledger	On	On	None (tracking only)

[TokenOptimization]

Shell output ([Bash])

Tool discovery ([ToolDiscovery])