ctxmgr

func CapToolResult ¶

func CapToolResult(result string, maxBytes int, sessionID, toolCallID, baseDir string) (capped string, wasCapped bool, err error)

CapToolResult checks whether a tool result exceeds maxBytes. If it does, the full result is written to disk at:

{baseDir}/{sessionID}/{toolCallID}.txt

and SnipToolResult's head and tail of it are returned, around a marker that names the file.

If baseDir is empty, ToolResultsBaseDir() is used.

Returns:

• capped: the (possibly truncated) result string to send to the model
• wasCapped: true if the result was truncated
• err: any I/O error from writing the spill file

func EstimateTokens ¶

func EstimateTokens(text string) int

EstimateTokens approximates token count from text length. Uses the rough heuristic of 4 characters per token.

func FormatTokenCount ¶

func FormatTokenCount(tokens int) string

FormatTokenCount formats a token count with K/M suffix for display.

func GetModelLimit ¶

func GetModelLimit(model string) int

GetModelLimit returns the token limit for a model name. Falls back to the "default" entry if the model is not found.

func LoadSpilled ¶

func LoadSpilled(baseDir, id string) (string, error)

LoadSpilled returns the full tool result CapToolResult spilled under baseDir ("" is ToolResultsBaseDir) for a recall id <sessionID>/<toolCallID>, the id its notice names.

func LookupModelLimit ¶

func LookupModelLimit(model string) (int, bool)

LookupModelLimit returns the token limit and whether the model was actually found. Callers that validate a user's configured window need the second value: for an unrecognised model the returned limit is a conservative fallback, not knowledge, and treating it as an upper bound would reject a correct setting. Local models are never in the table — their window is whatever the server was started with.

func PruneToolResults ¶

func PruneToolResults(baseDir, keep string, now time.Time) error

PruneToolResults deletes spilled tool results under baseDir ("" is ToolResultsBaseDir): only spill session directories (spillSessionDir), never the compaction store beside them, whose bodies expire one by one after SpillKeepAge instead (expireCompactStore). Every session directory last changed more than SpillKeepAge ago, then the oldest others idle for spillActiveAge while all of them together are over maxTotalSpillBytes. keep (the session running now) always survives, and so does any session that spilled within spillActiveAge, as another celeste process may still be using it. A missing baseDir is nothing to prune (Aikido 806869375).

func SnipToolResult ¶

func SnipToolResult(result string, maxBytes int, note string) string

SnipToolResult cuts result to at most maxBytes in memory, keeping its head and tail around a "[...snipped N bytes...]" marker, without the spill file CapToolResult writes (2.0 F3). note, if not empty, goes into the marker: callers say there why the middle cannot be recalled and what to do instead. The loop uses it when the spill file cannot be written; sessions and checkpoints use it on tool results loaded from disk. Results at or under maxBytes come back unchanged. Cuts fall on UTF-8 character boundaries. The result is never over maxBytes: a cap too small for the marker gets a plain cut of the head.

func ToolResultsBaseDir ¶

func ToolResultsBaseDir() (string, error)

ToolResultsBaseDir returns the base directory for spilled tool results. Default: ~/.celeste/tool-results