grahama1970/battle

STOP. READ THIS ENTIRE SKILL.MD BEFORE CALLING ANY ENDPOINT.

Battle Skill

Red vs Blue Team Security Competition Orchestrator

Pits a Red Team (attack) against a Blue Team (defense) in a long-running competitive loop. Each team leverages all .pi/skills to attack or defend a target codebase.

Invariant battles (the common case)

Battle's default scoring targets exploitation (system-down, command injection), but the most frequent real use of an adversarial Red/Blue loop is verifying a PROJECT-SPECIFIC INVARIANT: "no PII value leaks", "the ledger balances", "the parser drops no record", "the authz check cannot be bypassed".

Supply a pluggable invariant Judge -- a small independent module judge(target_dir, params) -> {passed, violations, evidence} (schema battle.invariant_result.v1). Red's objective becomes "produce an input that makes the Judge fail"; Blue's is "make it pass"; the scorekeeper reads the Judge result, never an agent's self-report. Judges are fail-closed: a judge that errors is a FAILED invariant, never a silent pass.

python3 -m battle_skill.invariant_judge \
  --judge fixtures/reference-judges/no_data_leak_judge.py \
  --target <released-output-dir> \
  --params '{"policy": "policy.json", "output_subdir": "corpus"}'

fixtures/reference-judges/no_data_leak_judge.py is the anonymizer confidentiality invariant as a Judge: it independently scans released output (JSON scalars incl decoded escapes, numeric expansion, SQLite cells + schema DDL + header integers, text/CSV, report.json, and captured stdout/stderr) for any policy value in any representation.

Contract variation-family research (generic Battle composition)

Battle is not project-specific. For any project or skill with an acceptance_contract.bundle.v1, Battle should expand each contract item through Dogpile before claiming comprehensive adversarial coverage:

./run.sh contract-variation-plan \
  --acceptance-bundle /path/to/acceptance_bundle.json \
  --dogpile-source brave-search \
  --dogpile-source arxiv \
  --out /tmp/battle-variation-plan.json

The output is battle.contract_variation_plan.v1. It is a three-phase Battle contract, not a loose note:

1. Phase 1: pass the frozen acceptance_contract.bundle.v1 floor.
2. Phase 2: after that pass, gather $project-state, current Battle receipts, source-filtered $dogpile research, and $ask one-shot reviewer proposals.
3. Phase 3: freeze useful Phase 2 findings into deterministic cases, run them, and retain adaptive-lineage proof for Red wins, Blue fixes, and replay.

The plan maps every acceptance case to source-bearing Dogpile research lanes and reusable variation families. Use repeatable --dogpile-source filters when a Battle phase needs only selected Dogpile providers such as brave-search, arxiv, github-search, youtube, brave-questions, feeds, wayback, or context7 instead of the full source fanout. The reusable variation families include representation equivalence, encoding/normalization, parser differentials, release-surface boundaries, lossy conversion, split/composed facts, scale, retry/concurrency, authorization, and failure-leak boundaries. Dogpile and Ask are research input only: they discover meaningful variation families and source evidence. Battle then freezes selected families into deterministic generators, runs the real target in Docker/QEMU/digital-twin evidence gates, emits battle.case_receipt.v1 per case, aggregates with battle.campaign_aggregate.v1, and retains the workflow classes as $agentic-evals cases. A serious release gate should expect hundreds of deterministic cases and roughly 20-30 retained eval classes when the contract surface is broad; smaller smoke gates must label themselves as smoke.

Invariant campaigns (test the contract floor, then attack beyond it)

An invariant battle judges one output. An invariant campaign has Red generate the whole MATRIX of input "versions" the target's spec names -- every format x every representation x the documented edge cases -- PLUS random fuzz, runs the real target on each, and the Judge scores every output. Every case emits a battle.case_receipt.v1 with fixture precheck, execution, rejection, security Judge, optional functional Judge, and verdict fields. A campaign PASSES only if all case receipts pass, required MUST_ACCEPT cases are actually accepted and functionally judged, required MUST_REJECT cases are safely rejected, and the computed battle.campaign_aggregate.v1 has no failed or incomplete cases. One failing version is a concrete, reproducible Red win.

For anonymization/privacy targets, the acceptance contract is only the floor. When an acceptance_contract.bundle.v1 exists, the arena build may pass it as acceptance_floor to the production adapter. Battle then fails closed unless every acceptance_cases[] id maps to at least one campaign generator case in the profile's required_case_ids; those are the bare-minimum adversarial versions that MUST pass before any extra fuzz or beyond-contract cases matter. Battle must also run a beyond-contract campaign that attacks surfaces a client brief often omits: JSON keys and duplicate keys, CSV headers/dialects/multiline cells, SQLite identifiers/defaults/generated values/partial indexes/triggers, filenames, report.json, stdout/stderr, alternate encodings, and same-identity representation traps. A clean brief-matrix replay alone is a smoke proof, not a comprehensive Battle proof.

python3 -m battle_skill.invariant_campaign \
  --generator fixtures/reference-generators/anon_brief_matrix.py \
  --target-run-cmd 'docker run --rm -v {input}/corpus:/trial/input/corpus:ro -v {input}/policy.json:/trial/input/policy.json:ro -v {output}:/trial/output anonymization-trial run' \
  --judge fixtures/reference-judges/no_data_leak_judge.py \
  --gen-params '{"fuzz": 20}' --judge-params '{"output_subdir": "corpus"}'

fixtures/reference-generators/anon_brief_matrix.py yields the anonymization brief's versions: the four formats, JSON string/int/float/scientific, SQLite TEXT/INTEGER/REAL, Unicode NFC/NFD, BOM, JSON \u-escape, SQLite CHECK-literal, plus fuzz. It also carries the oai-trial roundtable edge cases: formatted policy phone values stored as digit-only JSON/SQLite numerics, the same identity seeded across every in-scope format, and lossy leading-zero / large-float traps. fixtures/reference-generators/anon_beyond_brief_matrix.py is the required next rung for privacy/anonymization proof: it tries non-obvious schema/path/encoding/ log/release-boundary surfaces that go beyond the literal acceptance contract. A generator + target-run-cmd + judge is a pluggable trio: point it at any project's spec matrix and invariant.

The no-data-leak Judge also supports an explicit opt-in interpretation profile for transformation semantics. These guarantees are OFF unless declared in --judge-params, so Battle does not silently expand the contract after seeing a failure:

{
  "interpretation_profile": {
    "decoders": ["base64", "base64url", "hex"],
    "record_local_reconstruction": true,
    "max_decoded_bytes": 4096
  }
}

With that profile, whole scalar/token base64/base64url/hex values are decoded once and record-local adjacent JSON/CSV/SQLite scalar fields may reconstruct a complete policy value. Arbitrary recursive decoding, global field joins, and visual-confusable character folding remain out of the default blocking gate.

After Red finds failing cases and Blue patches the target, emit the replayable lineage receipt instead of summarizing in prose:

./run.sh invariant-lineage-receipt \
  --red-campaign /tmp/red-result.json \
  --replay-campaign /tmp/replay-result.json \
  --target oai-trial \
  --out /tmp/battle-lineage.json

A battle.invariant_adaptive_lineage.v1 PASS proves Red found contract edge cases, Blue removed those Red wins, and the independent Judge replay passed.

Then Battle must produce a $create-report-validated report with $project-state context and an explicit exploits table. The report is the human-readable decision artifact; receipts remain the authority. Generate fresh project state first, then render the report:

PROJECT_STATE_ROOT=/path/to/target ../project-state/run.sh report --json --output /tmp/project-state.json
./run.sh invariant-report \
  --campaign /tmp/battle-brief-fuzz.json \
  --campaign /tmp/battle-beyond-brief.json \
  --adaptive-lineage /tmp/battle-lineage.json \
  --project-state /tmp/project-state.json \
  --target oai-trial \
  --out-json /tmp/battle-report.json \
  --out-md /tmp/battle-report.md

For project-agent terminal review, use cards when cases have long evidence:

./run.sh invariant-report \
  --campaign /tmp/battle-brief-fuzz.json \
  --campaign /tmp/battle-beyond-brief.json \
  --project-state /tmp/project-state.json \
  --target oai-trial \
  --out-json /tmp/battle-report.json \
  --out-md /tmp/battle-report.md \
  --terminal-cards

invariant-report is a Typer CLI command. It writes machine JSON to stdout and keeps the human Battle report on stderr. Use --terminal-cards for normal project-agent review of Battle evidence: it groups cards under Acceptance contract floor, Beyond-contract exploits, Adaptive lineage, and fallback Other campaign cases headings, then prints one ============== block per case with Scope, Case, Acceptance parent, Expect, Result, Example, Why Battle checks this, Related research, Adaptive lineage, and Judge evidence. If a production-adapter receipt carries acceptance_floor.case_map, contract-floor cards show the parent AC-* id(s); otherwise Acceptance parent is not recorded in case receipt. If a receipt has example, why_chosen, research_refs, or source_refs, cards show those exact fields; otherwise examples/rationales are deterministic from the case id and research is reported as not recorded in case receipt.

--terminal-table remains the compact overview alias for --terminal-summary: both print contract floor, Red pressure, Scorekeeper call, and one case-table row per attack. In an interactive terminal Battle uses Rich's terminal-aware colored table (RED_WIN red, clean passes green, fail-closed stops yellow) and honors NO_COLOR; under capture/CI it falls back to deterministic plain text so logs and tests stay stable.

invariant-report writes create_report.report.v1, validates it through skills/create-report/run.sh validate, renders Markdown through skills/create-report/run.sh render, and appends ## Exploits Table. The table must be plain-spoken and scannable: one row per attack case, with columns for Scope, Contractual?, Adaptive lineage?, Case, Exploit / attack, Why chosen, Expectation, Result, and Judge evidence. Scope separates contractual acceptance-floor cases from beyond-contract probes; Why chosen explains non-contractual probes; Adaptive lineage? marks Red wins that were fixed and replayed; RED_WIN blocks release. The table is derived from battle.case_receipt.v1 when present, falling back to legacy case_log only for older campaign receipts. A Battle closure without that report is missing the decision surface even if campaign receipts pass.

Purpose Boundary

Battle's purpose is the Red/Blue security competition backend: authorized target setup, isolated execution, Red attack generation, Blue defense generation, independent Judge replay, scorekeeper receipts, adaptive lineage, and durable learning. Adaptive lineage is a backend learning loop that spawns, evaluates, selects, and promotes or rejects child Red/Blue evidence from Judge-backed receipts.

PixiJS is only a spectator/replay surface for Battle receipts. A PixiJS pass proves that recorded receipts can be inspected in a fun replay; it does not prove the Battle orchestrator, provider-driven subagents, Docker/QEMU isolation, overnight scheduler, scorekeeper, or memory learning works. Do not close core Battle readiness from PixiJS evidence alone, and do not block backend adaptive lineage on replay polish beyond truthful receipt inspectability.

Architecture

Production Battle is an orchestration skill, not a large bespoke security engine. The host-side process should schedule rounds, choose personas, dispatch subagents, provision Docker runtimes, collect receipts, score hard runtime signals, write reports, and persist learning. Target code and team-generated code must execute only inside Docker.

Required production invariants:

• Red and Blue are subagent teams. Each dispatched subagent must include an explicit persona selected by the orchestrator for that turn. Multiple personas per team may run concurrently when the turn benefits from breadth.
• Battle execution requires a security.target_authorization.v1 manifest before Docker, QEMU, target runtime setup, Hack delegation, proof replay, or patch replay starts. The manifest binds project/operator scope and target identity; it is not a legal opinion and does not prove exploit success or patch effectiveness.
• Red-team $hack execution is a subagent responsibility, not a Battle Python import. Battle performs or schedules scan/research/memory recall, builds the candidate exploit list, chooses the Red persona, dispatches an agent-skills/agents/Tau subagent with that contract, and records the returned exploit receipt.
• Subagent handoffs and receipts should follow the compact Tau-style JSON contract shape used by tau.agent_handoff.v1 and tau.subagent_receipt.v1, with Battle-specific fields layered on top rather than a separate ad hoc protocol.
• Battle calls modular Tau subagent contracts. Tau and the loop/agentic harness own subagent execution and use scillm as the LLM/model caller. Battle owns team selection, persona assignment, Docker runtimes, scorekeeping, artifacts, and memory promotion.
• SciLLM is a Tau-owned provider sidecar. Battle operators and project agents must not call $scillm, /scillm, http://localhost:4001, /v1/chat/completions, or /v1/scillm/* directly for Battle proof work. Express provider work as Tau DAGs, Tau command-loop nodes, or Tau skill nodes, then consume Tau receipts and node outputs.
• Model choice is strategic but routed through Tau/loop: SOTA models for planning, small fast models for high-throughput mutation generation and triage, specialist models for language/security niches, and batch calls for broad candidate generation.
• Red and Blue have free research access through approved agent-side research skills, including dogpile, brave-search, memory, GitHub/code search, docs, papers, CVEs, and public writeups.
• All target apps, exploit probes, fuzzers, payloads, repro scripts, patch builds, tests, migrations, dependency installs, and replay checks run in Docker. The host is control plane only.
• Docker target runtimes may be rebuilt and relaunched between rounds. Persist only controlled volumes and artifacts that must survive a rebuild; store durable strategic context and learnings in $memory.
• Docker runtimes must support dynamic language/toolchain selection. Any code language required by the target may be added to the runtime image or selected adapter; Battle should not hard-code one language.
• Battle should be high-throughput when the target runtime is warm: thousands of exploit/defense mutations may be attempted with tight 10-15 second Docker execution windows on capable workstation hardware.
• Battle should use combinatorial mutation. Red tries every plausible exploit family and combination within safety/time budgets; Blue tries every plausible patch, hardening, configuration, test, detection, and mitigation combination. Successful combinations receive stronger promotion than isolated tactics.
• Research may burst concurrently from the agent side. Red and Blue may fan out multiple brave-search and dogpile calls, including 10x concurrent Brave search batches when needed, then store useful results and negative evidence in $memory.
• Dogpile research receipts are design input for Battle, not proof. Use them to seed Red exploit-family menus, Blue hardening/detection menus, GitHub security-tool candidates, DARPA/AIxCC context, and follow-up research questions. Exploit success, patch effectiveness, tool safety, and repository adoption still require Battle-owned Docker/QEMU execution, hard runtime signals, and Judge replay.
• Security repositories found through Dogpile must flow through $github-search evaluation criteria first. Any adopted repo code, PoC, scanner, or payload still runs only inside Battle's isolated target/runtime gates; do not execute untrusted repo-provided install scripts or payloads on the host.
• Target containers default to no network. External research happens from the agent side through controlled skills unless a scenario explicitly grants target-container network access.
• The scorekeeper records objective outcomes: system down, system still up after the allotted time, exploit success, crash artifacts, patch timing, regression behavior, resource limits, and replay results. It is not an LLM judge.

Based on research into RvB framework, DARPA AIxCC, and Microsoft PyRIT:

┌─────────────────────────────────────────────────────────┐
│                 Battle Orchestrator                      │
│  - Game loop (RvB pattern)                              │
│  - Concurrent Red/Blue execution                        │
│  - Entropy-driven termination                           │
│  - Checkpointing for overnight runs                     │
└─────────────────────────────────────────────────────────┘
         │                              │
    ┌────┴────┐                    ┌────┴────┐
    │ Red Team │                   │ Blue Team│
    │ (Thread) │                   │ (Thread) │
    ├──────────┤                   ├──────────┤
    │ Skills:  │                   │ Skills:  │
    │ - hack   │                   │ - anvil  │
    │ - memory │                   │ - memory │
    └──────────┘                   └──────────┘
         │                              │
         └──────────┬───────────────────┘
                    │
    ┌───────────────┴────────────────────┐
    │           Digital Twin              │
    │  ┌─────────────────────────────┐   │
    │  │ Mode: git_worktree          │   │
    │  │   - Red attacks arena       │   │
    │  │   - Blue patches workspace  │   │
    │  │   - Cherry-pick to test     │   │
    │  ├─────────────────────────────┤   │
    │  │ Mode: docker                │   │
    │  │   - Isolated containers     │   │
    │  │   - Battle network          │   │
    │  ├─────────────────────────────┤   │
    │  │ Mode: qemu                  │   │
    │  │   - Emulated firmware       │   │
    │  │   - GDB attach points       │   │
    │  └─────────────────────────────┘   │
    └────────────────────────────────────┘

Digital Twin Modes

The battle skill supports multiple target types through its Digital Twin system:

1. Source Code (git_worktree)

For battling over git repositories. Creates isolated git worktrees for each team.

./run.sh battle /path/to/repo --rounds 100

2. Docker Container (docker)

For battling over containerized applications. Spins up separate containers for each team.

# Using a Docker image
./run.sh battle --docker-image nginx:latest --rounds 100

# Using a Dockerfile in the target directory
./run.sh battle /path/with/Dockerfile --mode docker

3. Firmware/Microprocessor (qemu)

For battling over firmware and embedded systems. Boots firmware in QEMU emulator.

# Auto-detect architecture from ELF header
./run.sh battle firmware.elf --rounds 100

# Specify machine type explicitly
./run.sh battle firmware.bin --qemu-machine arm
./run.sh battle firmware.bin --qemu-machine riscv64
./run.sh battle bios.rom --qemu-machine x86_64

Supported QEMU machines:

• arm - ARM Cortex-M (STM32, etc.)
• aarch64 - ARM64
• riscv32/riscv64 - RISC-V
• x86_64/i386 - x86
• mips - MIPS (routers, embedded)

3b. Full-system VM (qemu-vm)

For battling over host-level policy (systemd units, AppArmor profiles, cgroup-BPF IP filters) that Docker twins cannot host because containers share the host kernel. Boots a base cloud image (e.g. Ubuntu 24.04 qcow2) with a fresh overlay per round, KVM acceleration, ephemeral SSH key via NoCloud cloud-init, executes one payload inside the VM, and writes a battle.vm_round.v1 receipt. Requires a security.target_authorization.v1 manifest binding the image path AND sha256 before QEMU starts.

./run.sh vm-round --image /mnt/storage12tb/skills/battle/images/noble-server-cloudimg-amd64.img \
  --command 'uname -r; echo VM_CANARY_OK' \
  --authorization-manifest fixtures/vm-round/authorization.json --out-dir /tmp/battle-vm

Snapshot discipline: each round boots base.qcow2 + throwaway overlay (qemu-img backing file), so rounds never accumulate state; persistent learning belongs in $memory, not the disk image.

4. Copy Mode (fallback)

For non-git directories. Creates simple file copies for each team.

Commands

# Start a battle (10 rounds for testing)
./run.sh battle /path/to/codebase --rounds 10

# Start overnight battle (1000 rounds)
./run.sh battle /path/to/codebase --overnight

# Battle a Docker container
./run.sh battle --docker-image myapp:latest --rounds 100

# Battle firmware with QEMU
./run.sh battle firmware.bin --qemu-machine arm --rounds 100

# Check battle status
./run.sh status

# Resume interrupted battle
./run.sh resume <battle-id>

# Generate report from completed battle
./run.sh report <battle-id>

# Run the deterministic Battle v0 fixture proof
./run.sh battle-fixture battle-001 --out /tmp/battle-001

# Run the reactive Blue + independent Judge Docker proof
./run.sh prove-reactive-judge-round \
  --authorization-manifest skills/battle/fixtures/reactive-judge/authorization.json \
  --out /tmp/battle-reactive-judge-round

# Run canonical BATTLE-004 with parent-spawn lineage requested
./run.sh arena-parent-spawn-proof battle-004 --out /tmp/battle-004-parent-spawn --red-workers 2 --blue-workers 2

Battle v0 Fixture Proof

Battle v0 is a narrow, deterministic proof rung for the Battle artifact contract. It runs one local Red -> Blue -> Judge fixture and emits replayable receipts:

• red-receipt.json
• blue-receipt.json
• judge/judge-receipt.json
• scoreboard.json
• monitor-index.json
• run-receipt.json

The Battle v0 scoreboard is derived from the independent Judge receipt, not from Blue-side self-certification fields. This addresses the current battle loop gap where a Blue patch can carry verified and functionality_preserved claims without a separate Judge phase.

The fixture proof is intentionally limited:

mocked: no
live: local_deterministic_fixture
agentic: false
models_used: []

It proves the receipt boundary and monitor artifact rendering for the local fixture only. It does not prove real Red or Blue agent behavior, scillm, OpenCode, anvil, code-runner, memory learning, Docker, QEMU, or multi-round campaign readiness. See docs/BATTLE_V0.md for the validation commands and artifact-backed monitor proof path.

Reactive Judge Round Proof

prove-reactive-judge-round is the deterministic local Docker proof rung for the default Battle round authority boundary. It runs a small authorized fixture with one command-injection behavior and one candidate patch:

mocked: no
live: local_docker_fixture
agentic: false
models_used: []

The proof emits:

• authorization-validation.json
• immutable-baseline-manifest.json
• event-ledger.json
• red-hack-observation.json
• judge-1/judge-1-receipt.json
• blue/proactive-blue-input.json
• blue/reactive-blue-input.json
• blue/candidate-patch-receipt.json
• judge-2/judge-2-receipt.json
• scorekeeper-receipt.json
• round-receipt.json
• artifact-hash-manifest.json

The required phase order is authorization, immutable baseline, concurrent Red and proactive Blue, Red observation, Judge #1 confirmation, reactive Blue, candidate patch, Judge #2 replay, scorekeeper, and round receipt. Proactive Blue must receive no private Red finding. Reactive Blue may receive only the Judge-confirmed finding and replay contract. The scorekeeper derives Red/Blue scores only from Judge receipts; Blue verified, success, and functionality_preserved fields are advisory and not score authority.

The ordinary in-process battle round now fails closed at this same authority boundary: it preserves proactive overlap, dispatches reactive Blue only for Judge-confirmed findings, and does not award Blue score without a Judge #2 success verdict. The local Docker proof is the executable receipt path for the complete reactive/Judge behavior. It does not prove provider-driven Red/Blue quality, arbitrary target exploitability, production deployment readiness, or overnight scheduler readiness.

Scoring System (AIxCC-style)

Metric	Weight	Description
Vulnerability Discovery	1x	Red team finds vulnerability
Exploit Proof	+0.5x	Red team proves exploitability
Successful Patch	3x	Blue team patches vulnerability
Time Decay	Variable	Faster responses score higher
Functionality Preserved	Required	Patches must not break code

Scores

• TDSR (True Defense Success Rate): Vulnerabilities fixed AND code works
• FDSR (Fake Defense Success Rate): Attack blocked but code broken
• ASC (Attack Success Count): Total unique exploits discovered

Game Loop (Learning-Based)

Each round follows a learn → act → reflect pattern:

Round k:

┌─────────────────────────────────────────────────────────────┐
│                    1. RESEARCH PHASE                         │
├─────────────────────────────────────────────────────────────┤
│ Red Team:                      Blue Team:                    │
│ - Recall past attack attempts  - Recall past defenses        │
│ - Query /dogpile for new       - Query /dogpile for          │
│   exploitation techniques        hardening strategies        │
│ - Review opponent's patterns   - Analyze attack evolution    │
│ (Budget: 3 research calls max)                               │
└─────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────┐
│                    2. ACTION PHASE                           │
├─────────────────────────────────────────────────────────────┤
│ Red Team Attack:               Blue Team Defense:            │
│ - Execute learned strategy     - Apply patches via anvil     │
│ - AFL++ fuzzing with coverage  - Verify via QCOW2 overlay    │
│ - Collect crashes/findings     - Run regression tests        │
│ - Tag findings with /taxonomy  - Tag patches with /taxonomy  │
└─────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────┐
│                   3. REFLECTION PHASE                        │
├─────────────────────────────────────────────────────────────┤
│ Both Teams:                                                  │
│ - Archive round episode (actions, outcomes, learnings)       │
│ - Store successful strategies in /memory                     │
│ - Update belief about opponent's capabilities                │
│ - Evolve strategy for next round                            │
└─────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────┐
│                   4. SCORING & CHECKPOINT                    │
├─────────────────────────────────────────────────────────────┤
│ - Calculate AIxCC-style scores                               │
│ - Check termination conditions                               │
│ - Save checkpoint (QEMU state + team memories)              │
└─────────────────────────────────────────────────────────────┘

Memory Architecture

Each team maintains isolated knowledge:

battle_red_<battle_id>/           battle_blue_<battle_id>/
├── strategies/                   ├── strategies/
│   ├── successful_attacks        │   ├── successful_patches
│   └── failed_attempts           │   └── broken_defenses
├── research/                     ├── research/
│   └── dogpile_results           │   └── dogpile_results
├── episodes/                     ├── episodes/
│   ├── round_001.json            │   ├── round_001.json
│   └── round_002.json            │   └── round_002.json
└── taxonomy/                     └── taxonomy/
    ├── cwe_classifications       ├── mitigation_types
    └── severity_scores           └── effectiveness_scores

Teams cannot access opponent's memory - this creates true adversarial learning.

Termination Conditions

Battle ends when ANY condition is met:

1. Null Production: Both teams fail to generate new findings for 3 rounds
2. Maximum Rounds: Configured limit reached
3. Metric Convergence: Scores stable for 5 consecutive rounds
4. Kill Switch: Manual termination via ./run.sh stop

Task Monitor Integration

Battles register with task-monitor for overnight progress tracking:

# View battle progress in TUI
.pi/skills/task-monitor/run.sh tui --filter battle

Report Output

After battle completion, generates:

• Executive Summary: Winner, key metrics, risk score
• Vulnerability Report: By severity, category, remediation status
• Attack Evolution: How Red team adapted over rounds
• Defense Timeline: Blue team improvements over time
• Recommendations: Prioritized security improvements

Memory + Taxonomy Integration

The skill integrates with the shared memory and taxonomy systems via memory_integration.py for cross-battle learning:

• Pre-hook (recall_prior_battles): Before starting a battle, recalls prior battle findings for the same target or technique. Enables teams to build on accumulated security knowledge across battles.
• Post-hook (learn_battle): After battle completes, stores the full outcome (target, red findings, blue defenses, winner, scores, TDSR) to memory with taxonomy bridge tags.
• Bridge keywords: Precision, Resilience, Fragility, Corruption, Loyalty, Stealth (tuned to security/exploitation domain).
• Tags: ["battle", "security"] + bridges

Gracefully degrades if common.memory_client or taxonomy/taxonomy.py are unavailable.

Nondeterministic Exploit Specimen Synthesis

Battle Red agents may behave as nondeterministic exploit-code authors. A Red exploit subagent may combine high-level web, protocol, MITM, packet, timing, source-level, binary, assembly, fuzzing, and obscure research-derived techniques into generated exploit specimens.

Most generated specimens may be bad ideas: they may fail to compile, fail at runtime, combine irrelevant methods, or produce no useful target signal. Battle treats those outputs as genetic material, not proof.

Battle owns deterministic selection and evidence:

• generated code must be materialized as an artifact;
• generated code must run only in Docker;
• stdout, stderr, HTTP observations, timing, exit code, and optional packet summaries must be captured;
• runnable code is not exploit success;
• target contact is not exploit success;
• Judge replay is required before any exploit-success claim;
• memory promotion requires replayable evidence.

The first backend proof rung is exploit-combiner-proof. It is fixture-backed, agentic:false, and proves the specimen lifecycle only: bad generated code, Docker execution, captured failure observations, target contact, runnable unproven code, and fail-closed non-claims. Live Tau generation, child materialization, packet capture, Blue adaptation, memory promotion, and Judge exploit-success replay are later rungs.

The second backend proof rung is spawn-architect-proof. It is fixture-backed, agentic:false, and proves the DAG birth contract only: Battle loads a spawn-policy decision, constructs a child knowledge packet from parent specimen evidence, authors a tau.dag_contract.v1 child exploit-synthesis DAG, validates private-artifact exclusions, and records that Tau execution is deferred to PR3. It does not run Tau, materialize a child exploit, generate live exploit code, compile child code, contact the target, or claim exploit success.

The current live Tau child DAG canary is live-tau-child-dag-canary. It is non-mocked and invokes the existing local Tau DAG runtime without fixture fallback. The PR3b/PR3c boundary is:

lineage-summarizer PASS
research-scout PASS with Tau-validated source-bearing design-input receipts
method-combiner PASS with a deterministic exploit genome candidate
exploit-code-author PASS only when Tau/SciLLM returns provider_live:true
provider-authorship evidence; otherwise BLOCKED at the precise attestation gap

The PR3c boundary may materialize provider-authored child exploit code, but it does not compile child code, run a child specimen in Docker, or claim exploit success. Compile repair, Docker execution, and Judge replay are later gates.

File Structure

battle/
  SKILL.md                   # This file
  run.sh                     # Shell entry point; launches package through uv
  sanity.sh                  # Deterministic fixture and structure sanity gate
  pyproject.toml             # Dependencies
  .ask/browser-oracles.yaml  # WebGPT project mapping for browser-oracle walk-up
  src/battle_skill/
    cli.py                   # Typer CLI entry point
    config.py                # Constants and paths
    state.py                 # Data classes and BattleState
    memory.py                # Team-isolated memory system
    scoring.py               # AIxCC-style scoring
    digital_twin.py          # Git worktree, Docker, QEMU isolation
    red_team.py              # Red Team attack agent
    blue_team.py             # Blue Team defense agent
    orchestrator.py          # Game loop orchestrator
    battle_fixture.py        # Deterministic fixture proof runner
    judge.py                 # Deterministic scorekeeper verifier
    receipts.py              # Receipt dataclasses and JSON writer
    report.py                # Report generation
    qemu_support.py          # QEMU emulator support
    qemu_peripherals.py      # QEMU peripheral emulation
  fixtures/battle-001/       # Deterministic local fixture
  spectator/               # Self-contained BATTLE-004 spectator UI + Pixi engine
  monitor/battle/            # Artifact-backed React monitor

Leveraged Skills

Skill	Team	Purpose
hack	Red	Scanning, auditing, exploitation
anvil	Blue	Multi-agent patching (Thunderdome)
memory	Both	Recall prior strategies
treesitter	Blue	Code structure analysis
taxonomy	Both	Classify findings
task-monitor	Orchestrator	Progress tracking
ops-docker	Both	Container management

Example Battle

# Start 100-round battle on current project
./run.sh battle --target . --rounds 100

# Output:
# Battle ID: battle_20250128_221500
# Target: /home/user/project
# Rounds: 100
#
# Registering with task-monitor...
# Starting Round 1/100...
# [Red] Scanning target with hack...
# [Red] Found 3 potential vulnerabilities
# [Blue] Analyzing attack logs...
# [Blue] Generating patch for SQL injection...
# [Blue] Patch applied, running verification...
# Round 1 complete. Red: 3 pts, Blue: 9 pts
# ...
#
# Battle Complete!
# Winner: Blue Team (847 pts vs 423 pts)
# Report: ./reports/battle_20250128_221500.md