Configuration and Tuning

Configuration and tuning

slipstream works with zero configuration. Everything below has a sensible default; you only touch it to change behaviour deliberately.

Dashboard settings

.claude/slipstream/dashboard.json in your project (src/dashboard/settings.ts):

{
  "enabled": true,   // auto-start the local server on SessionStart
  "autoOpen": true   // open the browser on first start
}

Environment overrides, for a single session without editing the file:

Variable	Effect
SLIPSTREAM_DASHBOARD=0	disable the dashboard (also stops the MCP server auto-starting it)
SLIPSTREAM_DASHBOARD=1	force enable
SLIPSTREAM_DASHBOARD_OPEN=0	keep the browser shut
SLIPSTREAM_DASHBOARD_OPEN=1	force open
SLIPSTREAM_MCP_EMIT=0	stop the MCP server emitting activity events (set by the plugin, since its PostToolUse hook already emits)

These last two gate the cross-IDE behaviour. Outside Claude Code both default on, so the MCP server feeds and auto-starts the dashboard with no setup; the plugin sets both to 0 because the hooks already do it.

Token-budget control

.claude/slipstream/budget.json is the shared, editable budget the dashboard gauge, sp_budget and the statusline all read (src/context/budget-config.ts):

{
  "targetTokens": 200000,  // the target the gauge fills toward
  "warnPct": 60,           // gauge turns amber here
  "compactPct": 85,        // gauge turns red here
  "actualTokens": 0        // optional: paste the editor's real count to calibrate
}

Edit it from the dashboard's "set budget" panel, with slipstream budget, or by hand. Thresholds are clamped and ordered on save so a bad edit cannot break the math.

Context budget

src/context/budget.ts holds the tunable constants:

Constant	Default	Meaning
BYTES_PER_TOKEN	3.6	bytes-per-token estimate; lower is more cautious
DEFAULT_WINDOW_TOKENS	200,000	model context window
COMFORT_FRACTION	0.6	the share treated as comfortably usable (ok below this)
LARGE_FILE_BYTES	16,000	the threshold above which a whole-file read is flagged

The window can be overridden per call: sp_budget(bytesRead, windowTokens) and slipstream budget --bytes N --window N. The levels are ok below COMFORT_FRACTION, warn up to 85%, compact above.

Recall budget

src/memory/recall.ts:

Constant	Default	Meaning
RECALL_TOKEN_BUDGET	1,200	the ceiling on the relevant subset reloaded at session start
WEIGHTS	branch 4, tag 3, file 3, description 2, haystack 0.5	the ranking weights

Raise the budget to reload more at session start, at the cost of more tokens; lower it to reload less. The weights are tuned against this project's own store (see Memory recall).

Digest bounds

src/memory/digest.ts bounds the digest so it stays a memory the next session can afford to reload: MAX_DECISIONS 8, MAX_FILES 20, MAX_LINE 200 characters.

Statusline

The statusline reads the same budget constants. To change what it shows, edit formatStatusline in src/statusline/index.ts (it is pure and unit-tested, so a change is caught by the test). To enable it manually, see Statusline.

Where state lives

slipstream writes only under .claude/slipstream/ in your project:

.claude/slipstream/
  map.md  map.json              # the project map (from /slipstream:map)
  memory/                       # one .md per fact + MEMORY.md index
  observations/<session>.jsonl  # auto-captured observation memory (+ .cursor, .counter)
  budget.json                   # editable token-budget target + thresholds
  savings.json                  # optimization tally: scoped reads vs whole-file baseline
  dashboard/<session>.jsonl     # append-only event log per session
  dashboard/server.json         # the running server's pid/port/url
  dashboard.json                # optional settings

Add .claude/slipstream/ to .gitignore to keep it local, or commit memory/ if you want the team to share durable facts.

See also

• Data formats for the exact shapes of these files.
• Memory recall for the recall tuning in context.
• Security model for what is written and what is redacted.

---

SarmaLinux . sarmalinux.com . Repository