Configuration

Three layers, in precedence order: environment variables beat settings, and settings beat defaults. Project trust gates everything project-local.

Local inference (the bundled fleet extension)

Variable	Default	Purpose
HUMMIN_COLIBRI_INSTANCES	none	Comma-separated server base URLs. Order is preference; first server serving a model wins, later duplicates become fallbacks
COLI_API_KEY	none	Bearer token sent to every local server. A placeholder is sent when unset, so keyless servers work
HUMMIN_COLIBRI_CTX	16384	Fallback context window, used only when a server does not report one via /props

Cloud (Z.ai)

Variable	Purpose
ZAI_API_KEY	Z.ai API key (or hummin auth login zai)

Memory

Setting / env	Purpose
memoryEnabled / HUMMIN_MEMORY	Master switch; HUMMIN_MEMORY=0 forces off (set automatically in spawned subagents)
memoryMode	lessons or vault
memoryVaultDir	Vault directory (git-backed, Obsidian-compatible)
memoryProvider, memoryModelId	Model used for distillation

Fleet

Setting	Purpose
fleet.servers[]	Ordered server registry: enables /fleet Start/Stop/Restart, health probes, start-from-picker
fleet.autoStart	Skip the confirm when starting a server from the model picker

Replay trimming

Completed turns are rewritten per request before the provider sees them, so GLM-family endpoints stop resending reasoning and stale tool output (z.ai recommends not resending reasoning). Session history is never modified; the active tool loop, redacted thinking, and thinking blocks with structured replay signatures always pass through.

Setting	Default	Purpose
replayTrim.thinking	true	Drop replayed reasoning from completed turns
replayTrim.toolResults	true	Prune old tool results to head + marker + tail (false = lossless replay)
replayTrim.toolResultCap / toolResultTail	2000 / 200	Head/tail characters kept when pruning

Applies to anthropic-messages and openai-completions models; other APIs are left untouched.

System-1 decision layer

One decision engine backs four automatic hooks (bash gate, destructive-intent steer, test-failure triage, memory intake). A single engine, no failover; every read fails open, so an unreachable engine never blocks a session.

Setting / env	Purpose
decision.url	Engine endpoint (default http://127.0.0.1:9987/v1/systemone)
decision.apiKey	Engine auth (COLI_API_KEY works as fallback)
decision.gateThreshold / steerThreshold / triageThreshold / intakeThreshold	Calibrate the four hooks
sys1Gate.extraSafe / extraDestructive	Regex lists layered onto the bash gate
HUMMIN_SYS1_RIGHTSIZE / HUMMIN_SYS1_RIGHTSIZE_SWING	Model right-sizing for child dispatches

Telegram bridge (cc-tg-hub)

Requires a running cc-tg-hub broker (one per machine; install with bunx cc-tg-hub setup). See docs/features/integrations.md.

Setting / env	Purpose
tgHub.enabled / TG_HUB=1	Opt in. The env switch is per-session - prefer it (the setting is global and would register child/cron sessions too)
tgHub.socketPath / TG_HUB_SOCKET	Broker socket (default ~/.claude/cc-tg-hub/broker.sock)
tgHub.askTools / TG_HUB_ASK_TOOLS	Tools that raise Allow/Deny buttons in Telegram (Deny blocks; Allow never bypasses a terminal dialog)

Behavior

Setting	Purpose
editorMode	Vim editing mode
compactPrompt	Condense tool descriptions and system-prompt guidance (useful for slow-prefill local models)
cacheWarming	off / streaming / idle / always provider-side prompt-cache warming
statusline.left, statusline.right	Statusline segment tokens (dir, repo, branch, model, ctx, tokens, cost, queue, background, sandbox, mcp, git, diff, ...)
providers.showAll	Show the full upstream provider catalog in /model, not just GLM-first curation
httpIdleTimeoutMs	Set to 0 for local models: slow prefills otherwise look like dead connections
defaultProjectTrust	Trust fallback for un-trusted projects

Sandboxing, guardrails and other env switches

HUMMIN_* variables cover the fork surface: memory (8 variants), instances/ctx, fleet, sandbox, bashguard, guardrails and budgets, agents, cron (HUMMIN_CRON=0 disables), mcp, lsp, and offline mode. Run /doctor to see what hummin actually picked up - settings, fleet, credentials, extensions and vault are all checked there.

Files

Path	Purpose
~/.hummin/agent/settings.json	Main settings
~/.hummin/agent/extensions/	TypeScript extensions (autoloaded, hot reload)
project AGENTS.md / SYSTEM.md	Context files fed to the model
project hooks.json	Shell hooks on four lifecycle events