| name | beastmode |
| description | Multi-agent orchestration framework for high-intensity feature implementation. Routes work across model tiers: frontier models (Claude Fable, Kimi 3, Opus/Codex) own design, architecture, and review sign-off, while economy models (MiniMax M3, Qwen/Gwen) handle implementation and mechanical validation in isolated worktrees, with a self-improving learning loop that promotes lessons back into skills. Harness-agnostic: works with Ultraswarm, GSD, delegate_task, or manual orchestration.
|
| version | 2.1.0 |
| author | Luis Calderon |
| tags | ["beastmode","orchestration","multi-agent","cost-optimization","model-routing","self-improving","worktrees"] |
| related_skills | ["ultraswarm","gsd","subagent-driven-development","self-improvement"] |
| agents | ["hermes","codex","openclaw","claude-code","fable","kimi","minimax"] |
Beastmode: Multi-Agent Orchestration Framework
Beastmode is a structured approach to multi-agent software development that separates high-judgment work (planning, architecture, review) from routine execution (implementation, tests, docs) across different model tiers, with strict cost discipline and a self-improving learning loop. It is designed for MofA (Mixture of Agents) orchestration and pairs well with MemroOS/memroos-style durable memory for long-running agent goals.
Harness-agnostic: Works with Ultraswarm, GSD, delegate_task, Claude Code subagents, or manual orchestration. No specific tool required.
Core Principle
Frontier models design. Cheap models build and validate. The loop learns.
- Director/Lead (Design tier): Frontier model (Claude Fable, Kimi 3, Opus, or Codex) owns intent, architecture, creative judgment, and final sign-off.
- Watcher/Reviewer: Adversarial reviewer (frontier or mid-tier: Codex/GSD/Kimi 3) challenges plans, gates merges, catches scope creep.
- Executor (Execution tier): Economy model (MiniMax M3, Qwen 3.7 Plus / Gwen) handles routine implementation and mechanical validation (running tests, lint, typecheck, diff summaries) in isolated worktrees.
- Harness: Any orchestration tool (Ultraswarm, GSD,
delegate_task, Claude Code subagents, or manual git workflow).
- Memory: Self-improvement loop records lessons and promotes repeated patterns into skills/config.
Model Tiers & Routing
Beastmode routes every unit of work to a tier, not a specific model. Pick the best available model in each tier for your environment.
| Tier | Example models | Owns |
|---|
| Design (frontier) | Claude Fable (claude-fable-5), Kimi 3, Claude Opus, Codex/GPT frontier | Intent interpretation, architecture, API/data-model design, tradeoff decisions, acceptance contracts, final review sign-off, escalations |
| Execution (economy) | MiniMax M3, Qwen 3.7 Plus / Gwen, Haiku-class | Implementation, tests, docs, refactors, scripts, mechanical validation (run test suites, lint, typecheck, build, produce structured pass/fail reports) |
The routing principle: verification cost, not task type.
A task is safe for a cheap model exactly when its output can be cheaply and objectively verified — tests pass, schema validates, the diff matches a concrete spec. A task needs a frontier model when verification is expensive or subjective — "is this the right architecture?", "does this match user intent?". Phase labels (design/implement/validate) are just the common case of this rule, not the rule itself.
This reframes the frontier model's job: its primary output is not code or even plans — it is verifiability. The design phase converts an unverifiable goal ("build the feature well") into verifiable tasks (concrete interfaces + acceptance contract + verification commands). Once that conversion happens, the cheap tier can do everything downstream, because failures are caught by the verifier, not by expensive review.
Routing decision, per task:
- Is there a cheap objective verifier for this task's output? → Economy tier, cheap-first cascade (retry once on failure, then escalate).
- No verifier exists? → Don't route it to frontier by default. First ask: can the frontier tier create a verifier (tests, contract, checklist) and then delegate? Only work that resists verifier-creation — genuine judgment calls — is done directly by the frontier model.
Validation is split in two:
- Mechanical validation (cheap): The executor tier runs verification commands, collects results, and produces a structured report (what passed, what failed, diff stats). This is deterministic work — never spend frontier tokens on it.
- Judgment review (frontier): The design tier reads the report + diff, not the raw logs, and makes the merge decision. One frontier pass per phase, at the gate.
See references/model-routing.md for the verification-cost routing rule in full, the per-phase routing table it implies, provider configuration examples, and the escalation ladder.
When to Use Beastmode
Use beastmode for complex tasks that need:
- Multi-phase workflows with planning and review gates
- Multiple workstreams or files
- High-stakes architecture/product/creative judgment
- Cost-efficient execution under an expensive lead model
- Strong QA and merge gates
- A reusable learning loop
Do not use beastmode for:
- Trivial one-file edits (use the cheap executor directly)
- Simple questions or information retrieval
- Tasks that don't benefit from role separation
Two Beastmode Variants
Variant A: Frontier-Led Beastmode (Fable / Kimi 3 / Opus)
Use when: You have a frontier model available as the lead (Claude Fable, Kimi 3, Claude Code / Opus) and need maximum judgment for product/creative/architecture decisions.
Role split:
- Director (Fable / Kimi 3 / Opus): Intent, architecture, design docs, creative judgment, final sign-off
- Watcher (Codex/GSD or a second frontier model): Adversarial planning, scope/cost review, merge gating — pairing two different frontier models (e.g. Fable designs, Kimi 3 challenges) catches blind spots a single model family misses
- Executor (MiniMax M3 / Qwen / Gwen): Implementation, tests, docs, scripts, mechanical refactors, and mechanical validation (running verification commands, producing pass/fail reports)
Key rule: The frontier lead must aggressively avoid spending tokens on routine implementation or on watching test output scroll by. Delegate file edits, test writing, docs, refactors, command execution, and validation runs to the executor tier; the lead only reads the structured validation report and the diff.
Variant B: Codex-Led Beastmode
Use when: You don't have a frontier lead, or the task doesn't require frontier-level judgment. Codex/GSD leads, with an economy model executing routine work.
Role split:
- Director/Reviewer (Codex/GSD or current session): Planning, review, merge decisions
- Executor (MiniMax M3 / Qwen / Gwen): Implementation, tests, docs, scripts, mechanical validation
- Escalation: Codex or a frontier model (Fable / Kimi 3) handles security, auth, payments, data-loss, production incidents, or failed executor attempts
Key rule: Delegate routine work to the executor tier, but don't merge until the lead verifies acceptance.
Hard Rules
- Main tree stays clean. Executors work in isolated worktrees/branches. Never let cheap executors directly mutate the main working tree unless the task is tiny and explicitly approved by the lead.
- Lead reviews, executor implements. The lead can plan, inspect, test, and merge. Routine work goes to the executor.
- Every phase has an acceptance contract. Define goal, non-goals, verification commands, and escalation triggers before delegation.
- Every phase improves the loop. Record learnings, errors, routing mistakes, and token/cost surprises. Promote repeated lessons into skills/config.
- Escalation doesn't skip self-improvement. Record why the cheap route failed and whether routing rules should change.
- Usage is reported per phase, not just at the end. Every phase closes with a usage report: requested vs actual model per task, tokens used vs phase budget, actual vs estimated time (see
references/autonomy-levels.md for the format). If the harness doesn't expose a value, say "unavailable" — never omit the report.
- Model drift always surfaces. If a task was served by a model other than the requested
provider/model (router fallback, harness default, silent substitution), flag it as MODEL DRIFT in the phase report immediately, at every autonomy level. Drifted work is not validated until re-validated under the correct tier.
- Gates are blocking below high autonomy. At
low and medium autonomy (medium is the default), the run stops at each phase gate — report, then wait for approval before the next phase or any merge. Only --autonomy high proceeds through gates automatically, and even it halts on its always-surface events.
Choosing Your Harness
Beastmode works with any orchestration harness. Choose based on your environment:
Runner CLI (bm)
For one-shot goals without writing a full plan: bm "<goal>" from any repo.
Parses --gsd, --frontier <alias>, --economy <alias>, --on local|<host>,
--autonomy low|medium|high (default medium). Tier aliases resolve via
references/tier-aliases.json — kimi3 → kimi-coding/k3, fable →
anthropic/claude-fable-5, minimax → minimax/MiniMax-M3, etc. Override
per-repo with <repo>/.beastmode/tier-aliases.json. See scripts/bm and
references/autonomy-levels.md.
Before invoking pi, bm runs a model-availability preflight that checks
each resolved provider/model against pi --list-models on the local host.
If any are missing, bm exits with code 2 and prints the available
alternatives, so a goal never starts against an unresolvable model. Skip
with BM_SKIP_MODEL_CHECK=1 (CI / scripted runs). The check is also
skipped when --on is not local (the remote host owns availability).
Harness 1: Ultraswarm (Preferred for Git Repos)
Use when: You have Ultraswarm installed and want worktree isolation, adaptive QA, merge gates, and cost reporting.
Commands:
ultraswarm run "<task + acceptance contract>" --repo . --provider auto --mode direct
ultraswarm qa <task-id>
ultraswarm merge <task-id> --repo . --approved
ultraswarm report
For multi-phase work:
ultraswarm plan "<goal>" --repo . --mode gsd
ultraswarm run "<goal or phase>" --repo . --provider auto --mode gsd
Harness 2: GSD (Get Shit Done)
Use when: The repo already uses GSD for planning/phase management.
Commands:
gsd-plan-phase "<phase goal>"
gsd-execute-phase "<phase>"
gsd-verify-work
gsd-ship
Let GSD handle planning/phase gates, and delegate routine implementation units to the executor tier (MiniMax M3 / Qwen) via Ultraswarm or delegate_task.
Harness 3: delegate_task (Hermes/OpenClaw)
Use when: You're in Hermes or OpenClaw and need subagent orchestration without worktrees.
Example:
delegate_task(
goal="<tight task with acceptance contract>",
context="Repo, acceptance contract, files, verification commands, commit requirement",
toolsets=['terminal', 'file']
)
Note: delegate_task doesn't provide worktree isolation. Use for small parallel tasks or when worktrees aren't needed.
Cache note: delegate_task gives each subagent a fresh context, so the shared
prefix (repo path, acceptance contract, verification commands) is re-sent every call.
Keep that context= block byte-identical across delegations and fire them
consecutively — the first pays a 1.25x write, the rest read at 0.10x while warm.
Re-wording the context per task, or interleaving delegations with other model calls,
forfeits that.
Harness 4: Claude Code Subagents
Use when: You're in Claude Code and want to spawn subagents for routine work.
Example:
Task("<tight task with acceptance contract>")
Note: Claude Code subagents don't provide worktree isolation by default. Use git branches manually if needed.
Harness 5: Manual Git Workflow
Use when: No orchestration tool is available, but you still want isolation.
Workflow:
git checkout -b beastmode/<task-id>
git diff main...beastmode/<task-id>
git checkout main
git merge beastmode/<task-id>
Note: Manual workflow requires discipline. Don't skip the review step.
Harness Selection Guide
| Harness | Worktree Isolation | QA Gates | Cost Reporting | Best For |
|---|
| Ultraswarm | ✅ Yes | ✅ Adaptive | ✅ Yes | Git repos, multi-phase work |
| GSD | ❌ No (uses branches) | ✅ Phase gates | ❌ No | Repos already using GSD |
| delegate_task | ❌ No | ❌ No | ❌ No | Small parallel tasks, no repo |
| Claude Code subagents | ❌ No | ❌ No | ❌ No | Claude Code environments |
| Manual git | ✅ Yes (branches) | ❌ Manual | ❌ No | No orchestration tool available |
Default recommendation: Use Ultraswarm if available. Fall back to GSD if the repo uses it. Use delegate_task or Claude Code subagents for small tasks. Use manual git workflow as last resort.
The Beastmode Loop
Step 0: Preflight
cd "$REPO"
git status --short
ultraswarm doctor
ultraswarm report || true
bm --help
pi --list-models | head
If your harness is unavailable, fall back to a simpler harness (e.g., delegate_task or manual git), then record the failure in the self-improvement log.
Model availability preflight (bm). Before invoking pi, bm validates every --frontier and --economy alias against pi --list-models on the local host. If a resolved provider/model is not present, bm exits with code 2 and prints the available alternatives, so a goal never starts against an unresolvable model. The check is skipped when BM_SKIP_MODEL_CHECK=1 (CI / scripted runs) or when --on is not local (the remote host owns availability).
Step 1: Define Acceptance Contract
Before any delegation, write:
Goal: <user-visible outcome>
Non-goals: <scope boundaries>
User-visible acceptance: <what the user will see/test>
Files/areas likely touched: <paths>
Verification commands: <unit/integration/e2e commands>
Manual QA: <visual/security checks>
Escalation triggers: <auth/security/payments/data-loss/architecture-uncertainty>
Self-improvement log path: <.learnings/BEASTMODE.md or project-local path>
Step 2: Design (Frontier Tier — With Challenge)
Design is the highest-leverage phase — this is where frontier tokens are worth spending. Do not skimp here to save cost; a bad design multiplies executor waste downstream.
For frontier-led (Fable / Kimi 3 / Opus):
- The lead drafts intent, constraints, architecture, and interface contracts
- A second model (Codex/GSD, or the other frontier model) turns it into phases and tries to find gaps
- The lead resolves tradeoffs and approves the phase map
- The output is a design package the executor can implement without judgment calls: file-level plan, interfaces/signatures, acceptance contract, verification commands
For Codex-led:
- Lead writes the plan directly, or uses harness planning commands
Planning commands by harness:
- Ultraswarm:
ultraswarm plan "<goal>" --repo . --mode gsd
- GSD:
gsd-plan-phase "<phase goal>"
- Manual: Write plan in markdown, commit to
.planning/ or similar
Step 3: Delegate Routine Work
Use tight task specs. One task should be reviewable in a single diff.
Batch delegations to the same lane consecutively. Every worker starts with a cold
prefix and pays a cache write (1.25x); subsequent workers on the same lane, model, and
system prompt read that prefix at 0.10x while it is warm (5-minute TTL, refreshed on
each hit). Ten workers fired back-to-back on one lane pay roughly one write; the same
ten interleaved across three lanes pay a write each time you switch back. Group by
lane, don't round-robin.
Keep the worker contract byte-identical across workers. The universal worker
contract (allowed files, forbidden commands, required output) is the shared prefix.
Inject the per-task objective after it, never inside it — a task ID or timestamp
spliced into the contract header gives every worker a unique prefix and forfeits the
discount entirely.
Delegation by harness:
- Ultraswarm:
ultraswarm run "<task>" --repo . --provider auto --mode auto
- delegate_task:
delegate_task(goal="<task>", context="...", toolsets=['terminal', 'file'])
- Claude Code:
Task("<task>")
- Manual: Executor works in branch, commits changes
Step 4: Validate (Cheap), Then Review (Frontier)
Stage 1 — Mechanical validation (executor tier): the economy model (MiniMax M3 / Qwen) runs the contract's verification commands and produces a structured report:
## Validation Report <task-id>
- Commands run: <each command + exit code>
- Tests: <passed>/<total> (list failures with one-line reasons)
- Lint/typecheck: pass | fail (<count> issues)
- Diff stats: <files changed, +/- lines>; unrelated files touched: yes/no
- Contract checklist: <each acceptance item: met / not met / can't verify>
Stage 2 — Judgment review (frontier tier): the lead or Codex reviews the report + diff against the contract. Frontier tokens go to reading the diff and making the call, not to re-running or watching tests.
Review commands:
ultraswarm qa <task-id>
git diff --stat main...<branch>
git diff main...<branch>
Reject if:
- Tests fail
- Diff includes unrelated files
- Scope expanded beyond the acceptance contract
- Code uses nondeterminism or network calls where not allowed
- Secrets/credentials were exposed
- The executor made decisions reserved for the lead
Step 5: Merge Gate
Merge commands by harness:
- Ultraswarm:
ultraswarm merge <task-id> --repo . --approved
- GSD:
gsd-ship (after verification)
- Manual:
git checkout main && git merge <branch>
Never merge on executor self-report alone. The lead or Codex watcher must verify.
Step 6: Self-Improving Checkpoint
After every phase, append a learning entry before continuing.
Preferred locations (in order):
- Project-local
.learnings/BEASTMODE.md
.planning/LEARNINGS.md
- Relevant skill patch if the lesson is immediately reusable
Template:
## BM-YYYYMMDD-HHMM <phase/task-id>
- Director/Lead: <model/agent>
- Watcher/Reviewer: <model/agent>
- Executor: <model/agent>
- Harness: <ultraswarm/gsd/delegate_task/claude-code/manual>
- Acceptance checks: <commands run>
- Result: pass | fail | partial
- Token/cost note: <estimate or harness report>
- What worked: <specific observations>
- What failed / drifted: <specific observations>
- Routing rule to change: <if applicable>
- Skill/config update needed: yes | no
- Promoted to: <skill/config/file or none>
Promotion rules:
The self-improvement loop writes notes only during a beastmode run. Any lasting change to agent behavior belongs in a separate user-approved maintenance task after the run is complete.
- Same routing mistake twice → record a proposed routing-rule change
- Same QA gap twice → record a proposed addition to the acceptance contract checklist
- Same tool failure twice → record a proposed troubleshooting entry
- Reusable workflow discovered → draft reusable procedure notes for later review
- User correction → record immediately and flag whether a lead-approved future update is needed
Cost Discipline
Frontier-Led Cost Rules (Fable / Kimi 3 / Opus)
Keep the frontier tier for:
- Interpreting user intent
- Product/creative judgment
- Architecture tradeoffs and design packages
- Judgment review of validation reports + diffs
- Final sign-off
- Escalation decisions
Move to MiniMax M3 / Qwen / Gwen:
- Code generation
- Tests
- Docs
- Data transformations
- Scripts
- Asset assembly
- Repetitive refactors
- Command execution
- Mechanical validation (running test suites, lint, typecheck, build; summarizing results into a validation report)
Use Codex (or the second frontier model) for:
- GSD planning
- Adversarial scope/cost review
- Merge gating
- High-risk analysis
- Debugging failed executor attempts
Anti-patterns that burn frontier tokens:
- Frontier lead re-running tests the executor already ran
- Frontier lead reading raw test/lint output instead of the validation report
- Frontier lead writing boilerplate ("just this once") because delegation feels slow
- Sending the executor an underspecified design so the frontier lead has to answer clarifying questions mid-implementation
- Restarting the director session between phases — the frontier prefix is the most
expensive thing in the run and the most valuable to keep cached; a restart re-pays it
in full at 1.25x
- Varying the frontier system prompt per phase — a phase label injected into the
system prompt makes every phase a cache miss on the entire prefix
Tier routing and prompt caching are independent levers — apply both. Routing moves
work to a 10–50x cheaper model; caching takes up to 90% off the repeated prefix on
whatever tier you land on. Caching matters most exactly where routing helps least:
the frontier director, whose long stable prefix is re-sent on every turn.
Codex-Led Cost Rules
Keep Codex for:
- Planning and architecture
- Adversarial review
- Security/auth/payments/data-loss risk
- Production incidents
- Failed executor attempts
Move to MiniMax M3 / Qwen / Gwen:
- Everything else (implementation, tests, docs, refactors, scripts, commands, mechanical validation)
Required Final Report
End every beastmode run with:
✅ Beastmode complete: <goal>
Variant: frontier-led | codex-led
Harness: <ultraswarm/gsd/delegate_task/claude-code/manual>
Phases completed: <n>
Director / watcher / executor split: <summary>
Models: Frontier (Fable/Kimi 3/Opus) <x%>, Codex/GPT <y%>, Executor (MiniMax M3/Qwen) <z%>
Token/cost report: <harness report or estimate>
Verification: <commands and results, per validation report>
Self-improvement: <learning entry path + promoted updates, if any>
Merge status: <merged/branch ready/blocked>
Choosing Your Variant
Use frontier-led when:
- You have a frontier model available (Claude Fable, Kimi 3, Claude Code / Opus)
- The task requires maximum product/creative/architecture judgment
- You're willing to pay for frontier-level design decisions but want implementation and validation on the cheap tier
Use Codex-led when:
- You don't have a frontier lead, or the task doesn't require frontier-level judgment
- Codex/GSD is sufficient for planning and review
- You want the cheapest possible lead with strong gates
Both variants share:
- The same worktree isolation, QA, merge, cost-report, and self-improvement gates
- The same acceptance contract requirements
- The same escalation rules
- The same final report format
Escalation Rules
Escalate from the executor tier (MiniMax M3 / Qwen / Gwen) to the frontier tier (Fable / Kimi 3 / Opus / Codex) when:
- Security, auth, payments, data-loss, legal/financial data, or production incident risk appears
- The work requires non-obvious architecture tradeoffs
- The executor fails the same acceptance check twice
- The diff is too broad to review cheaply
- The user explicitly asks for frontier reasoning
Escalate a task, not the whole phase — the rest of the phase keeps running on the cheap tier. See references/model-routing.md for the full escalation ladder.
Escalation does not skip self-improvement. Record why the cheap route failed and whether the routing rule should change.
Implementation Notes
Worktree isolation is non-negotiable. Executors must work in branches/worktrees, never directly in the main tree (unless the task is tiny and explicitly approved by the lead).
If your preferred harness is unavailable:
- Fall back to a simpler harness (e.g.,
delegate_task or manual git)
- Record the harness failure in the self-improvement log
- Don't skip isolation — use git branches manually if needed
The goal is portability. Beastmode should work in any agent environment (Claude Code, Hermes, OpenClaw, Codex) with any available harness. The principles (role separation, acceptance contracts, self-improvement) are constant; the harness is flexible.
Self-Improvement Philosophy
Beastmode is not just an execution framework—it's a learning system. Every run should make the next run better.
During a run:
- Record observations, errors, and routing mistakes
- Note token/cost surprises
- Flag acceptance gaps
After a run (in a separate maintenance task):
- Review learning entries
- Promote repeated patterns into skills/config
- Update routing rules
- Improve acceptance contract checklists
- Draft reusable procedures
The goal: Beastmode should get cheaper, faster, and more reliable over time as the learning loop promotes lessons into permanent improvements.
Context Management (Critical)
Beastmode runs accumulate context fast — subagent outputs, tool results, file diffs, planning docs. Without active management, you'll hit 300-500KB in 10-15 minutes, causing compression timeouts and /compact failures.
Hard rules:
- Compact every 5-10 minutes — don't wait for context to break. Run
/compact after each major phase (planning, execution, QA, merge).
- Limit sessions to 30 minutes — save state (commit work, write learnings), start fresh, resume from saved state. Note the tension: a restart also discards a warm prompt cache, so restart on context pressure, not on the clock. If context is still healthy at 30 minutes, keep going.
- Subagent output summarization — instruct subagents to return only final results (files changed, tests passed/failed, issues), not intermediate tool outputs. One subagent task should add <10KB to context, not 100KB.
- Break large tasks into small units — one subagent = one small, bounded task. "Implement auth system" = 200KB output. "Create User model" + "Implement /login" + "Add password hashing" = 3x 20KB outputs.
- Never compress prompts at the API layer — prompt caching is prefix-keyed and bills cache reads at 0.10x. Any middlebox that rewrites the payload (LLMLingua/headroom-style) turns a 0.10x read into a 1.00x uncached read. At a measured 5% compression rate, break-even is only a 5.6% cache hit rate — agent workloads run far above that, so compression is a net loss.
- Compress tool output, not prompts — squeez on CLI/tool results shrinks text before it enters context and leaves the prefix byte-stable. That is cache-safe and worth doing.
- Compact after 3+ subagent delegations — rule of thumb, but note each
/compact rewrites history and resets the cached prefix. Prefer bounded subagents that keep the orchestrator prefix small over frequent compaction.
Keep the prefix byte-identical. Cache is keyed on an exact byte prefix, so one changed byte early invalidates everything after it:
- Order context stable → volatile: system prompt, tool definitions, reference docs, conversation, new turn last.
- No timestamps, session IDs, or rotating text in the system prompt.
- Don't edit or re-order earlier messages mid-conversation.
- Cache TTL is 5 minutes, refreshed on each hit — one continuous session stays warm; ten scattered one-shots pay the write cost ten times.
Alert thresholds:
- Context size > 200KB → compact now
- Cache hit rate < 30% on a long session → something is invalidating the prefix
- Session duration > 30 minutes → save state and restart
What NOT to compress:
- Error messages and stack traces (need full context for debugging)
- Small files (< 100 lines) — compression overhead > savings
- Structured data the agent needs to parse exactly (JSON APIs, CSV)
Verify caching survives your proxy chain with scripts/cache-hitrate. See
references/context-rot-mitigation.md for the break-even math, architectural fixes,
and monitoring.
References
- Model routing: See
references/model-routing.md for the per-phase tier routing table, provider configuration examples (Fable, Kimi 3, MiniMax M3), the mechanical-vs-judgment validation split, and the escalation ladder.
- Tier aliases: See
references/tier-aliases.md (and tier-aliases.json) for the friendly-name → provider/model map consumed by scripts/bm. Verified against pi --list-models on oracle-1 / maeve-u1.
- Autonomy levels: See
references/autonomy-levels.md for low / medium (default) / high autonomy — what surfaces, what runs silent, blocking-gate semantics below high, the per-phase usage report format, model-drift detection, and how to map levels to pi --approve / --no-builtin-tools.
- Context rot mitigation: See
references/context-rot-mitigation.md for detailed analysis of context accumulation, architectural fixes, and monitoring strategies.
- Orchestration comparison: See
references/orchestration-comparison.md for the evolution from early prototypes to v2.0.
- Public sharing checklist: See
references/public-sharing-checklist.md for sanitization guidelines when publishing beastmode skills publicly.