Agent settings

Profile → Agents is combo agent’s deepest tuning surface. Plan Mode chooses how to plan; here you choose what to remember, forget, compress, and retrieve.

Concepts: Observability, Multi-tenancy, Architecture.

1. Page overview

Entry: Avatar → My settings → Agent (brain icon).

Six collapse panels

Toolbar:

Button	Role
Unsaved / Saved	Status lamp — orange when dirty
View JSON	Debug / backup
Reset defaults	Factory reset (confirm)
Save	Persist for future sessions

Rendering diagram…

These are defaults for all your Agents. Single-Agent overrides exist; Plan Mode per session overrides further. Priority: session Plan Mode > Agent override > global default.

2. Panel ① Compaction

Near context limits the system summarizes older turns to free tokens — key to long chats.

Compaction panel

2.1 `compactionMode`

Mode	Description	Use
Standard `default`	Summarize older messages; keep last N + summary	Default daily driver
Safeguard `safeguard`	Conservative; triggers closer to limit	Compliance / audits
Off `off`	No compaction — hard truncation	Debug only

2.2 Key fields

Field	Range	Meaning
`reserveTokensFloor`	1K–100K	Minimum free tokens after compaction
`keepLastMessages`	1–100	Recent turns untouched
`summaryMaxTokens`	100–10K	Summary cap

2.3 Memory flush

Side effect: compacted turns + summaries can flush into LTM for cross-session recall.

Field	Meaning
Enable flush	On compaction → optional LTM write
`softThresholdTokens`	Begin flush earlier
System prompt	How to extract long-term facts
User prompt	Extra extraction instructions

3. Panel ② Context pruning

Compaction handles history; pruning trims individual tool payloads (e.g., a giant SQL dump).

Pruning panel

3.1 `pruningMode`

Mode	Description
Off `off`	No trimming
Adaptive `adaptive` (default)	Soft trim + optional hard clear
Aggressive `aggressive`	Earlier trims for tight windows

3.2 Key fields

Field	Range	Meaning
`keepLastAssistants`	0–20	Recent assistant turns protected
`softTrimRatio`	0–1	When soft-trim begins
`hardClearRatio`	0–1	Threshold to replace tool output with placeholder
`minPrunableToolChars`	1K–500K	Skip tiny returns

3.3 Soft-trim tuning

Keeps head/tail excerpts.

3.4 Hard clear placeholder

When enableHardClear fires:

[Cleared: tool X return (N chars). See Ledger for full output.]

3.5 Tool allow/deny lists

Field	Role
`toolsAllow`	Trim only listed tools (empty = all)
`toolsDeny`	Never trim (e.g. `graft_get`, `ledger_read`)

Typical toolsDeny: correctness-critical integrations (doors_query, alm_create_requirement). Leave toolsAllow empty to allow adaptive trimming broadly.

4. Panel ③ Memory search

Configure LTM retrieval side (compression writes; this reads).

4.1 Switches

Field	Meaning
`memorySearchEnabled`	Master switch
`memorySources`	`memory` / `sessions` / `final_results`

4.2 Chunking LTM excerpts

Field	Range
Chunk tokens	50–2000 (default 512)
Overlap	0–500 (default 64)

4.3 Sync cadence

Field	Meaning
Sync on session start	Incremental pull each open
Sync before search	Freshest / costliest
Interval minutes	1–1440
Byte / message deltas	Burst triggers

4.4 Query tuning

Field	Range
`queryMaxResults`	1–50
`queryMinScore`	0–1

Mode	Meaning	Typical
per-sender	User isolation	Personal assistant
per-agent	Per-Agent thread	Multi-agent desk
shared	Shared conversation	Team handoff

5.2 Program memory

Toggle	Meaning
Count runs	Invocation stats
Task history	start/end/params
Index finals	Push answers to LTM
Max history	1–500 FIFO

5.3 Reset policy

Field	Meaning
`resetMode`	`idle` timeout vs `daily` cron
Idle minutes	Inactivity clears
Daily hour	Scheduled reset

6. Panel ⑤ Storage

Where LTM physically lives.

Storage panel

6.1 `storageDriver`

Driver	Notes
Elasticsearch	Production default
Infinity	Lightweight single-node
SQLite	Dev / POC

6.2 Index naming

Suggested: combo_ltm_<env>_<tenant>.

6.3 Embedding (LTM)

Separate from KB embeddings.

Field	Meaning
Model name	e.g. `bge-m3`
Dimensions	Match model

7. Panel ⑥ Agent basics

Basics panel

Field	Range	Meaning
`workspace`	path	Sandbox root
`primaryModel`	provider/model	e.g. `openai/gpt-5`
`fallbackModels`	list	Failover chain
`contextTokens`	1K–1M	Window budget
`thinkingDefault`	low/medium/high	Reasoning budget
`timeoutSeconds`	10–3600	Hard cap per turn

Thinking depth strongly affects latency/cost — Fast + low vs Full + high.

8. Presets

8.1 Daily writing / Q&A

compactionMode     : default
pruningMode        : adaptive
memorySearchEnabled: true
sessionScope       : per-sender
primaryModel       : openai/gpt-5
thinkingDefault    : medium
contextTokens      : 128000

8.2 Patent / compliance (strict detail)

compactionMode     : safeguard
pruningMode        : off
memorySearchEnabled: true
memorySources      : [memory, sessions, final_results]
primaryModel       : anthropic/claude-4.7-sonnet
thinkingDefault    : high
contextTokens      : 200000
timeoutSeconds     : 600

8.3 Large-scale code review / ASPICE

compactionMode     : default
pruningMode        : aggressive
hardClearRatio     : 0.7
toolsDeny          : [graft_get, alm_create_requirement]
sessionScope       : shared

9. Save / effect

Click Save — orange disappears
Applies next turn
Running jobs unaffected until stopped

10. Next steps

Cron & events → Other settings
Team sharing → Collaboration
Internals → Architecture