What the binary includes

Subagent orchestration, unified structural search, autonomous deep review, kernel resource limits, and deterministic line-anchored editing. Developer detail: handbook under docs/.

Subagent Cockpit (/agents)

Full visibility into every subagent. The /agents command opens an interactive inspector across the entire session tree. Stream live child transcripts, inspect intermediate tool calls and outputs, check real-time token spend, and jump directly into any subagent history with history://<id>.

Live Stream Inspector: monitor concurrent worker thoughts, running tool executions, and stream outputs without blocking parent flow.
Session Tree Traversal: drill down into nested subagents at any depth, switch thread focus, and audit child tool histories.
Transcript Resolution: internal history://<id> and agent://<id> URLs let parent sessions and tools extract structured data from past workers.
Concurrency Governance: bounded worker execution via dynamic session spawn semaphores and provider rate limits.

# interactive subagent cockpit
/agents

Active Spawns (3 running, 1 completed):
  [1] DynamicsAgent   running  tool: edit (src/physics.ts)
  [2] RenderAgent     running  tool: search (type: structure)
  [3] FlightAgent     running  tool: bash (bun test)
  [4] TypesWorker     complete exit: 0 (14.2k tokens)

# inspect worker history directly
read history://subagent-4-types

Unified Search (search)

Replaces fragmented grep, find, and ast-grep tools with a single unified search tool. Combines full-text PCRE2 regex, gitignore-aware glob layout, and tree-sitter AST syntax queries under one interface. Cuts context tokens drastically and eliminates model tool-selection ambiguity.

Three Modes, One Tool:
• type: "text" — PCRE2 / regex full-text search with ripgrep speed and path bounds.
• type: "files" — Fast parallel file and layout discovery respecting .gitignore.
• type: "structure" — Tree-sitter / ast-grep syntax and semantic AST structure search across types, methods, and callsites.
Context-Preserving Output: returns matched lines, structural hierarchy, and file references without dumping redundant raw bytes into the prompt.
Zero Tool Confusion: the model never hesitates between distinct grep or find tools; one tool handles every discovery task.

# 1. Structural AST search for method signatures
search { type: "structure", input: "class $NAME { #flushUsageLocked($$$) { $$$ } }" }

# 2. Fast regex text search with path filter
search { type: "text", input: "goalTokenDelta", path: "src/goals" }

# 3. Workspace file layout discovery
search { type: "files", input: "packages/coding-agent/src" }

Autonomous Auto-Review (afterEditCheck)

Automatic deep code inspection after complex multi-file changes. When enabled in settings (afterEditCheck: review) or triggered via /review, Veyyon autonomously reads back modified files and audits changes before concluding the turn.

Cross-File Contracts

Catches exported function signature mismatches, interface drift, missing type definitions, and broken import/export barrels across workspace packages.

Regressions & Invariants

Audits boundary conditions, unhandled exceptions, lifecycle state transitions, concurrency races, and guarantees termination on queues and retry loops.

Security & Secret Leaks

Verifies that credentials, private tokens, absolute host paths, and unsanitized command inputs never leak into public files, logs, or commit histories.

Kernel CPU & Resource Limits (/cpu-limit)

Machine-wide resource protection. Enforces hard kernel-level CPU quotas across the entire session process tree (main agent, subagents, bash commands, workers, and MCP servers) so heavy parallel builds never freeze your workstation.

Kernel-Level Quotas: Linux cgroup v2 quotas and Windows Job Objects throttle the entire process tree as a single unified budget group.
Tree-Wide Scope: limits apply to the session and every subagent spawned beneath it; delegating work to 10 workers cannot multiply host CPU consumption.
Runaway Protection: session.cpuLimitKill: true automatically sends SIGTERM to processes that peg cores past saturation, reporting a clean budget action instead of a crash.
Interactive Tuning: adjust on the fly with /cpu-limit 4 or remove with /cpu-limit remove without editing persistent config.

# limit entire session tree to 4 CPU cores
/cpu-limit 4

# kernel enforcement active
cgroup v2: /sys/fs/cgroup/veyyon-session-38a1/
cpu.max: 400000 100000 (4 cores hard ceiling)
processes: 14 (main CLI, 3 workers, 2 compilers)

Zero-Leak Secrets (/secret)

Vault-stored credentials referenced as #NAME# tokens. Real values are injected right before execution and never leak into prompts, logs, or transcripts.

Token References: write #NAME# where the credential belongs — a header, an env var, a CLI flag. The real value is substituted in just before the tool runs, and you never see it.
Vault Storage: secrets live in an encrypted vault on disk, separate from project config. A secret is created, updated, and deleted through /secret without touching tracked files.
No Prompt Leakage: the model's history contains the token, not the value. Logs, transcripts, exports, and subagent returns all carry the opaque token — the real value never crosses the model boundary.

# reference a secret in a tool call
bash: curl -H "Authorization: Bearer #STAGING_TOKEN#" https://api.example.com

# the model sees #STAGING_TOKEN#, never the value
# the value is injected at execution time

Autoswarm (/autoswarm)

Autoresearch with breadth. Each iteration builds several candidate arms instead of one change, rejects the ones that cannot be trusted, has the survivors cross-review each other, and keeps at most one.

Multi-Arm Search: breadth (1–8, default 3) controls how many candidate arms each iteration produces. Arms share one worktree — they are built, measured, and reverted one at a time, so breadth costs iteration time, not disk.
Mechanical Rejection: before a reviewer sees anything, four rejections fire automatically — empty (changed nothing), scope (edited an off-limits path), opaque (unreadable diff: binary patch or 512+ base64 chars), and duplicate (same diff as another arm).
Cross-Review Ring: with 3+ survivors, each arm is reviewed by another in a ring where no pair reviews each other. A reviewer flags an arm when the metric moved for the wrong reason — a hardcoded answer, a cache keyed on benchmark inputs, a narrowed input space, or work relocated out of the timed region. A flagged arm cannot win, however good its number.
Relocated-Cost Detection: a change that moves work out of the timed region lowers the metric without making anything faster. If the harness reports a cold-start metric (cold_ms), growth above 25ms against baseline is stated to the reviewer as a measured fact.
Setup Console: /autoswarm opens an interactive console for the goal, breadth, attempts, and certification toggle. Text after the command prefills the goal: /autoswarm make the tokenizer faster.

# autoswarm setup console
/autoswarm

Autoswarm setup
Autoresearch with breadth. The model derives the metric from your harness.

› Goal          make the tokenizer faster
  Breadth       3     candidate arms per iteration
  Attempts      1     retries before an arm is abandoned
  Certification on    arms cross-review before one is kept

3 arms in a review ring: each arm is reviewed by another,
and no pair reviews each other.

↑↓ field   ←→ adjust   space toggle   enter start   esc cancel

Approvals, Session Trees, and MCP

Configurable permission tiers, non-destructive session branching, and broad extensibility via MCP servers, custom skills, and native tools.

Approval Tiers

Permission modes: plan, ask, auto-edit, and yolo. Project rules refine auto-run commands; untrusted directories require explicit trust.

Session Branching

Session files are non-destructive trees. /tree navigates history; /fork and /branch create new session timelines without modifying previous state.

MCP, Skills, Browser

Integrate external MCP servers, declare local SKILL.md playbooks, and drive headless web automation with the built-in browser tool.

Developer detail lives in the handbook

Settings, tools, modes, and architecture are documented under docs/. Source: github.com/santhreal/veyyon.