steve
d0bd3ec3d9
All 3 cloud models converged on a real access-control bug; fixed it + the other genuine findings (the false-positives were dropped): Security (HIGH — all 3 models): - create_file_url skipped ValidateScope: a same-skill caller could mint a PUBLIC url for a file scoped to another user/run. Now runs ValidateScope (admin-aware), skipped only for the descendant-grant case — mirroring the read tools. Other real fixes: - ValidateScope hard-coded `false` at every call site (admin branch dead) -> pass inv.CallerIsAdmin (the executor sets it via the host AdminPolicy; still false/fail-closed when no admin). Stale "no admin flag" comment corrected. - create_file_url: ExpiresInSeconds clamped BEFORE the *time.Second multiply (huge values overflowed to a negative duration that slipped under the cap, minting already-expired tokens); swallowed json.Marshal error now returned. - RegisterMeta: build the default budget WITH the configured MaxPerRun (was NewInMemorySearchBudget(nil) -> hardcoded 10, ignoring MetaDeps.MaxPerRun). - classify: all-zero scores no longer return a false-positive top-1 winner; coerceClassifyScore uses strconv.ParseFloat (rejects trailing garbage like "50extra" that fmt.Sscanf silently accepted). - file_delete: honor the descendant grant (parent can clean up a worker's artifacts) — was the lone cross-skill-reject-outright file tool. - meta tools: input caps truncate at a UTF-8 rune boundary (truncateUTF8), not mid-rune. - think: removed the dead `var _ = fmt.Errorf` import-keeper; file_save default aligned to 16 MiB (matched RegisterStore). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
executus
⚠️ This project is vibe-coded. executus is written almost entirely by an AI coding agent (Claude), with a human steering at the design and review level rather than typing the code. That's a deliberate choice, stated up front — the same way gadfly is. Read the code before you depend on it, pin a version, and file issues if something looks off. It is offered as-is.
A batteries-included base for building LLM agent harnesses in Go. Import it, do a little wiring, and you have agentic capabilities: a bounded run loop, a tool registry with a suite of common tools, context compaction, config-driven model tiering and failover, structured output, and parallel fan-out — with sensible defaults so a brand-new project is agentic with almost no setup, and pluggable seams so a serious host can swap in its own storage, config, delivery, and tools.
executus sits strictly above majordomo — the lean LLM substrate (agent
loop, canonical llm types, providers, media normalization, model parsing /
failover / tiering). majordomo stays the substrate; executus is the opinionated,
batteries-included layer on top. executus requires no changes to majordomo.
Status
Early. Being extracted, phase by phase, from the agent layer of mort (a Discord
bot) — mort and gadfly are the first two consumers (heavy and light). See
CLAUDE.md for the architecture and the extraction roadmap (P0–P6).
Available today:
run/— executus is runnable.run.Executorties model resolution, the tool registry, majordomo's agent loop, context compaction, run-bounding, and step/audit instrumentation into oneRun(ctx, RunnableAgent, inv) Result, with every host concern behind a nil-saferun.Ports(Audit/Budget/Critic/ Checkpointer/PaletteSource/Delivery). Seeexamples/minimal.model/— config-driven tier resolution + failover over majordomo, with pluggableUsageSink/TraceSinkandGenerateWith[T]structured output.tool/— the tool registry + 3-stage permission model + SSRF guard.compact/— the per-run context compactor.lane/— bounded worker pool with fair-share queueing (run- and provider-concurrency).fanout/— programmatic N×M swarm with bounded global + per-key concurrency.config/,deliver/,identity/— host seams (config / output / identity), each with a shipped default.dispatchguard/,pendingattach/— run-safety primitives.
Design
Two tiers in one module (go.mod = majordomo + stdlib only):
- Core — everything a light host needs to be agentic: run loop, tool registry + common tools, model resolution, compaction, lanes, fan-out, structured output. No persistence, no scheduling.
- Batteries (opt-in sibling packages) — persona/agent nouns, saved skills, audit, run-critic, scheduling, budgets, checkpointing. Each is nil-safe and ships a default, so you add only what you use.
Persistence that needs a real database lives in a separate nested module
(contrib/store, pure-Go SQLite) so the core never drags in a DB driver — a
static-binary host (gadfly) stays static.
License
TBD.