mantis-pipeline-adapter

Mantis Pipeline Designer (/mantis-pipeline-adapter)

System Goal

Interactive Pipeline Design Consultant. Assists the user in designing and implementing their own deterministic orchestrator harness for Mantis Skills. Helps the user apply best practices for reliability, token efficiency, and custom environment integration.

Command Definition

• Command: /mantis-pipeline-adapter
• Description: Interactively guides the design and implementation of custom

deterministic orchestrator harnesses.

Input/Output Contract

• Reads:
• workspace/.mantis_state.json (to track current loop pass).
• workspace/.mantis_state.json fields active_snapshot, snapshot_history,

and vcs_info.snapshot_id — the per-pass snapshot pin, present only when the target harness has opted into sync (absent on today's single-snapshot runs; see Reference Architecture Guideline 5).

• schema.json (as the canonical pipeline specification reference).
• workspace/findings/*.json (as the State Store).
• workspace/learnings.jsonl (to understand memory rotation).
• User's interactive configuration input.
• Writes:
• Outputs user-customized orchestrator harness code, configurations, or

architecture documentation.

• Preconditions:
• User initiates interactive design session.
• Idempotency Guarantee:
• As a consulting agent, it advises the user to implement idempotency in their

custom harness using three primary mechanisms: (1) state store synchronization, (2) atomic transactional file/VCS operations, and (3) proper locks (e.g. database/file level locks).

Instructions

Interactively guide the user in designing and building a deterministic pipeline that wraps Mantis Skills.

Follow these guidelines during the consultation:

1. Understand User Context: Ask about their target programming language,

agent framework (if any), execution environments (VMs, local containers, physical hardware), and scale requirements.

1. Recommend Core Principles: Guide them to implement the reference

architecture patterns (detailed below), specifically emphasizing:

• Deterministic Orchestration: Use code (not LLM) for control flow.
• State Store: Use a database or structured filesystem as the single

source of truth.

• Token Efficiency: Use the UUID-based referencing pattern to avoid LLM

text duplication.

• Custom Environment Integration: Use Custom MCP servers for isolated

testing (VMs) or hardware interaction.

1. Ensure Schema Consistency: Advise the user to strictly adhere to the

inter-stage data contracts defined in schema.json when building their harness.

1. Adaptive Design: Help them draft the code/architecture tailored to their

specific stack, rather than imposing a rigid template.

1. Advise on Scale and Concurrency: If they have high-scale needs, guide

them on decomposing the pipeline and implementing locking mechanisms to prevent race conditions.

1. Suggest Evaluations: Remind them to perform empirical evaluations when

choosing cheaper models for utility stages.

1. Advise the Pass Lifecycle Contract (living / synced codebases): If the

user wants their harness to continue a run after the target code changes, or to sync the target repo at the start of a new pass, walk them through the harness-agnostic Pass Lifecycle Contract in Reference Architecture Guideline 5 below. Emphasize that this support is opt-in: a harness that does not implement the contract MUST leave snapshot_pinned unset, which preserves today's single-snapshot behavior byte-for-byte. When --sync is requested, the harness PINs in the PIN step and passes --snapshot_root/--snapshot_id normally; Block A (Locator Resolution) is universal across all code-reading stages.

1. Advise on Semantic Retrieval at Scale: If the user is targeting a large

codebase (e.g., thousands of source files, multi-pass campaigns, or multiple teams contributing findings), walk them through the optional semantic retrieval patterns in Reference Architecture Guidelines 6 and 7 below. Emphasize that these are opt-in: they augment the pipeline via a dedicated query skill or MCP tools, but never modify the existing skills' own deterministic logic or fail-safe invariants.

1. Advise on SAST Seeding: If the user wants to augment LLM-based discovery

with external SAST tool findings (CodeQL, Semgrep, etc.), walk them through the optional SAST seeding pattern in Reference Architecture Guideline 8 below. Emphasize that this is opt-in: it ingests external findings as candidates that must earn their verdict through unchanged downstream gates, and it follows exactly the RAG pattern (provenance-tracked, snapshot-aware, fallback on failure).

1. Advise on Structural Code Indexing: If the user is targeting a large

codebase where grep-based call-site discovery is unreliable, walk them through the optional structural code index stage in Reference Architecture Guideline 9 below. Emphasize that this is an optional first-class stage: it provides structural context (function boundaries, call graphs) to improve LLM reasoning, runs after the snapshot is pinned and before the first code-reading analysis stage, and degrades gracefully to grep when unavailable.

1. Advise on Tiered Iterative Reproduction & Multi-Conversation Retries: If

the user is targeting complex services where single-shot repro is brittle, walk them through the tiered iterative reproduction strategy and multi-conversation retry pattern in Reference Architecture Guideline 10.

Reference Architecture Guidelines

Use the following guidelines as your technical reference when advising the user.

Core Principles

1. Deterministic Orchestration: Do not let the LLM decide the control flow

of the pipeline. Use a programmatic harness to call skills sequentially or in parallel.

1. State on Disk / Database: Use the filesystem

(workspace/findings/*.json) or a database as the single source of truth. Skills should read from and write to this store. For horizontal scaling, recommend a centralized database.

1. Deterministic Reporting: Treat findings as internal state. Minimize the

use of the LLM to convert JSON findings into Markdown reports for human consumption; instead, write deterministic scripts to render the JSON into reports or upload them to bug trackers. Only use an LLM for non-deterministic subsets of this (like textual synthesis), such as by providing an executive summary if necessary.

1. Token Efficiency & Reusable Deterministic Tools: Structure LLM outputs to

return only the minimum necessary information (e.g., UUIDs, status codes). Do not force the LLM to write one-off scripts (e.g., Python or bash) on the fly for routine tasks like appending JSON fields or merging findings, as this wastes reasoning tokens. Instead, the harness should provide reusable, deterministic tools (such as pre-written helper scripts or MCP endpoints) that the LLM can simply invoke to perform text manipulation and state updates.

1. State Store & Memory Rotation: To prevent token bloat and infinite loops,

ephemeral queues (like workspace/learnings.jsonl) must be rotated. Upon successful completion and verification of the Knowledge Base synthesis stage, the orchestrator should ensure the archive directory exists (e.g., mkdir -p workspace/archive/learnings/) and move workspace/learnings.jsonl to a numbered archive (e.g., workspace/archive/learnings/learnings_pass_${N}_${X}.jsonl where ${N} is the loop pass and ${X} is a sub-index). If the synthesis fails, the active queue must be left intact to prevent data loss.

Architectural Overview

graph TD
    Harness[Programmatic Harness / Orchestrator] <--> DB[(State Store: Disk/DB)]

    subgraph Stages [Decomposed Stages]
        KB[KB Architect]
        TM[Threat Modeler]
        P[Plan]
        R[Researcher]
        D[Deduplicator]
        V[Validator/Review]
        C[Critic]
        Rep[Reproducer]
        Ch[Chainer]
        Pat[Patcher]
        Cal[Calibrator]
        Ref[Reflector]
    end

    Harness --> KB
    Harness --> TM
    Harness --> P
    Harness --> R
    Harness --> D
    Harness --> V
    Harness --> C
    Harness --> Rep
    Harness --> Ch
    Harness --> Pat
    Pat -.->|Re-attack Bypass Loop| Rep
    Harness --> Cal
    Harness --> Ref

    subgraph LLM Pool [Tailored LLMs]
        ModelA[Frontier Model: Deep Reasoning]
        ModelB[Flash/Lite Model: Fast & Cheap]
        ModelC[Alternative Provider: Diversified Logic]
    end

    KB -.-> ModelA
    TM -.-> ModelB
    P -.-> ModelB
    R -.-> ModelA
    R -.-> ModelC
    D -.-> ModelB
    V -.-> ModelB
    C -.-> ModelA
    Rep -.-> ModelA
    Ch -.-> ModelA
    Pat -.-> ModelA
    Cal -.-> ModelB
    Ref -.-> ModelB

1. UUID-Based Referencing Pattern

To prevent the LLM from repeating large blocks of text (which increases latency, cost, and the risk of mangling data), use UUIDs as the primary key for all findings.

A. Researcher Stage

• Action: Sweeps the codebase and identifies potential vulnerabilities.
• LLM Output: Generates a unique UUID for each finding and writes

workspace/findings/<UUID>.json containing the full details (matching the standard schema in Mantis Researcher).

B. Deduplication Stage (Optimized)

Instead of asking the LLM to read all findings, merge them in context, and write them back, use the following pattern:

1. Harness Action: Reads all workspace/findings/*.json files and prepares

a summary list for the LLM containing only key identifiers. To align with the standard schema, map the code_paths array (which uses "file:line" format) to a simplified summary for the LLM: [ { "id": "UUID", "file": "path", "line": 12, "snippet": "..." } ].

1. LLM Action: Analyzes the summary and outputs a mapping of duplicates:

   {
     "primary_uuid_1": ["duplicate_uuid_a", "duplicate_uuid_b"],
     "primary_uuid_2": []
   }

1. Harness Action (Deterministic):

• Reads the content of the affected files.
• Programmatically merges fields following the rules in

Mantis Deduplicator (e.g., union of code_paths, taking highest severity, concatenating history).

• Updates workspace/findings/primary_uuid_1.json on disk.
• Ensures the trash directory exists (e.g.,

mkdir -p workspace/findings/.trash/).

• Moves workspace/findings/duplicate_uuid_a.json and

workspace/findings/duplicate_uuid_b.json to the trash staging directory (workspace/findings/.trash/).

C. Validation & Review Stages (Reviewer, Critic)

• Harness Action: For each finding workspace/findings/<UUID>.json, pass

only the relevant code context and finding description to the LLM.

• LLM Action: Output only a structured verification result (e.g.,

{"valid": true, "reason": "..."}).

• Harness Action (Deterministic): Programmatically update the

workspace/findings/<UUID>.json file with the validation status and reason.

2. Adaptable Reproducers via Custom MCP

When validating findings, the agent may need to interact with diverse environments (VMs, physical hardware). Use the Model Context Protocol (MCP) to expose a clean, restricted API.

• Architecture:

[Reproducer Agent] <--- MCP ---> [Custom MCP Server] <--- API ---> [Target Env]

• Custom Environments:
• VMs: Implement tools like reboot_vm(), execute_payload().
• Hardware/USB: Implement tools like power_cycle_device() (via smart

plug), send_usb_packet().

• Integration Note: If the user's harness uses raw LLM APIs (e.g., direct

Gemini API calls) instead of an MCP-native client framework, the harness must manually register these tools in the API's schema format and handle dispatching tool calls to the MCP server.

3. Decomposition & Multi-Model Strategy

A. Pipeline Decomposition & Concurrency

The pipeline can be split into independent services. When scaling horizontally (e.g., multiple workers running the Reproducer stage in parallel):

• Concurrency Control: Implement database or file locking to ensure two

workers do not attempt to process or update the same finding simultaneously.

• Parallel Trajectory Search: For deep reasoning stages (Reproducer,

Patcher), spawn multiple parallel agents attempting to solve the exact same finding using diverse logic paths. For the Reproducer stage, prune all other trajectories as soon as one worker succeeds to save compute costs while escaping LLM "give up" loops. For the Patcher stage, wait for all patches to be generated and tested, then evaluate the successful ones to select the most minimal, idiomatic, and correct fix.

B. Heterogeneous LLM Selection (Multi-Model)

Match task complexity with the appropriate model tier:

• Frontier Models: For deep reasoning (Research, Reproduce, Patch).
• Flash/Lite Models: For structured utility tasks (Dedupe, Calibrate).
• Variability: Run different models in parallel during the Research stage to

increase bug-hunting coverage.

C. Importance of Evaluation

Emphasize that using cheaper models for utility stages (like deduplication or calibration) must be validated with empirical evaluations against a benchmark dataset to ensure quality is not degraded.

4. The Planning Stage and workspace/plan.json

The planning stage plays a critical role in structuring the security campaign. The strategist (/mantis-plan) generates workspace/plan.json to define targeted investigations, context pointers, and specific questions for the auditor. The researcher (/mantis-researcher) reads workspace/plan.json at startup to guide its sweep. By decoupling strategy and execution via this structured contract, the orchestrator can easily direct subagents, parallelize sweeps, and maintain historical context across pipeline runs without repeating work.

5. The Pass Lifecycle Contract (Living / Synced Codebases)

A custom orchestrator (a bespoke CLI, an ADK agent, an MCP-native pipeline, or any deterministic harness) does not inherit the living-project lifecycle that mantis-meta-agent implements. To support continue-after-edits and opt-in boundary sync without producing silent wrong results (false VERIFIED_SECURE, false failed_to_reproduce, dropped regressions), the harness must implement the following harness-agnostic contract. This is the same contract recorded in schema.json under Non-JSON Contracts; the Block A–Block G and SNAPSHOT_ID references below name mechanisms each Mantis stage already carries in its own SKILL.md.

Mantis runs under multiple harnesses (various CLIs, ADK, custom deterministic pipelines), so the lifecycle must not live only in mantis-meta-agent. Any harness is conformant iff, per pass, it:

1. SYNCs first (Block C) — the very first action; never mid-pass.
2. Detects `vcs_info` + computes `SNAPSHOT_ID` (Block D steps 1-5) — only

after sync.

1. PINs the immutable copy + writes the sentinel + appends

snapshot_history (Block D step 5, not RECORD).

1. Records vcs_info (incl. snapshot_id) + active_snapshot. Never

record an id or pin before syncing.

1. Runs every stage with

--snapshot_root=<SNAPSHOT_ROOT> --snapshot_id=<SNAPSHOT_ID> --state_root=<workspace parent>.

1. Archives & increments (existing Stage 15); retried findings keep their

original discovery_commit.

A harness that does not implement the contract MUST leave snapshot_pinned unset → today's behavior. When --sync is requested, the harness PINs in the PIN step and passes --snapshot_root/ --snapshot_id normally; Block A (Locator Resolution) is universal across all code-reading stages.

Advisory notes when helping a builder implement this contract

• Opt-in, default off. Sync/pinning is a feature the builder turns on. A

harness that never sets snapshot_pinned behaves exactly like today (one live snapshot per run). Downstream stages treat an absent active_snapshot/discovery_commit as the conservative branch, so an un-upgraded harness is always safe — just not living-project-aware. Do not advise treating these absent fields as an error.

• Store snapshots OUTSIDE `workspace/`. The pinned copy (SNAPSHOT_ROOT)

must live under <state_root>/.mantis_snapshots/pass_<N> (or a clean-VCS worktree/archive), and its path must not contain the segment /workspace/ — otherwise mantis-patch's state-vs-code path guard misfires. Keep the last 2 snapshots and garbage-collect older ones with the matching teardown (rm -rf for copies, git worktree remove/prune for worktrees).

• Non-destructive sync only. Sync is the first action of a pass,

never mid-pass, and must be skipped when the tree is dirty, ahead of upstream, detached, or has no upstream. The harness must never run git reset --hard, git checkout -- ., git clean, or hg update -C, or any command that discards uncommitted/untracked/local-commit state — user edits and in-progress work must survive every pass.

• Full-fidelity `SNAPSHOT_ID`s, including dirty / no-VCS. Compute the id

over the whole pinned copy: clean git/hg → commit_hash; dirty git/hg → commit_hash + ":" + content_hash; multi-vcs → revision + ":" + content_hash; no-VCS / unknown copyable tree → "content:" + content_hash. The embedded content hash is exactly what lets an unchanged dirty or no-VCS tree MATCH across passes and still receive verification + dedup — and what makes a repo sync that advances commits under an unchanged manifest revision compare unequal. Never trust a bare branch name or manifest revision string as an identity.

• Pass the three roots to EVERY stage. Include the findings-only stages

(report, calibrate, reflect): they do not read target code, but they still read active_snapshot for provenance/annotation. When the harness archives and increments, retried findings must keep their original discovery_commit.

Conformance scenarios

The scenarios below expose nearly every issue in the snapshot model. They are reference checks, not features: the harness is responsible for preventing or handling each one in its own environment. The table is a quick-reference; prose detail follows for each scenario. The State column uses the 3-STATE RULE (MODE-OFF / HALT / PINNED, branched on active_snapshot presence — see the global backward-compat rule in schema.json and the advisory notes above); SNAPSHOT_ID formats follow the ladder in the advisory notes above (e.g. live:<ts> signals an unpinned/HALT pass).

Invariant legend (the labels below name safety properties enforced by the blocks and the global backward-compat rule in schema.json):

Quick-reference table:

Per-scenario detail:

1. Colocated state (state_root nested inside CODE_ROOT / snapshot root) — The pinned SNAPSHOT_ROOT must live under <state_root>/.mantis_snapshots/pass_<N> (or a clean-VCS worktree/archive), and its path must not contain the segment /workspace/ — otherwise mantis-patch's state-vs-code path guard misfires (state files appear to be "under CODE_ROOT"). If state_root itself is inside CODE_ROOT, the harness must HALT-and-yield (safe default) or, when explicitly authorized (e.g. --auto_relocate_state), relocate it outside the snapshot before pinning. Block A step 3 distinguishes SNAPSHOT-RELATIVE path fields (read under CODE_ROOT) from STATE-RELATIVE fields (read under state_root/workspace, never prefixed with CODE_ROOT); colocation breaks this separation.

2. Stale active_snapshot (active_snapshot.pass != state.pass_number — active_snapshot was preserved across the Stage 15 pass increment) — Block D step 0 (crash-resume) handles only the SAME-pass re-entry case (active_snapshot.pass == N → reuse). It does NOT catch a stale snapshot carried across the Stage 15 pass increment, because Stage 15 deliberately preserves active_snapshot while bumping pass_number (see Stage 15). Two sub-cases:

(a) The prior snapshot dir is now MISSING: Block D step 0 STOPs and yields to the user (never re-pin to a possibly-drifted live tree). (b) The prior snapshot dir still EXISTS (default keep-2 retention) and its sentinel matches the preserved active_snapshot.snapshot_id: Block A step 2 sentinel check SUCCEEDS (it only compares the sentinel file to SNAPSHOT_ID, not to the current pass). Block B's pairwise discovery_commit check would MATCH a carried-forward finding against a new finding stamped with the same stale SNAPSHOT_ID, silently dropping it as DUPLICATE — a false authoritative verdict.

To prevent (b), the HARNESS MUST guarantee that active_snapshot.pass == state.pass_number before any consumer stage reads it. The reference harness (mantis-meta-agent) satisfies this by re-pinning every pass (Block D step 0 sees active_snapshot.pass != N → re-pins → refreshes active_snapshot.pass before any stage runs), so sub-case (b) never fires there. A custom harness that preserves active_snapshot across the Stage 15 pass increment WITHOUT re-pinning MUST either (a) re-pin every pass (the reference behavior), or (b) inject an equivalent pre-stage gate that refreshes active_snapshot.pass or clears active_snapshot entirely before invoking stages. Stages CANNOT self-detect this staleness via Block B (which is snapshot_id-only, not pass-aware): a carried-forward finding and a new finding stamped with the same stale SNAPSHOT_ID will MATCH in Block B despite the snapshot being stale. The active_snapshot.pass field is defined in schema.json #/$defs/state/active_snapshot/pass for exactly this check. The harness's Block D step 0 reuse check is NOT a substitute: it only fires on same-pass re-entry. (Stages that read active_snapshot MAY additionally self-check defensively — see each stage's Step 0 sentinel check — but the binding guarantee is on the harness.)

3. Pin failure (snapshot copy fails — disk full, permissions, too-large tree) — Block D step 2 (free-space precheck): compare du -s of the live tree to df free space at state_root; if it won't fit → skip copy → step 5b. Block D step 4 (failure-tolerant verify): check copy exit status + sanity check (file count/size within ~90%); on failure → step 5b (unpinned/HALT). Step 5b: SNAPSHOT_ROOT=<live root>, snapshot_pinned=false, SNAPSHOT_ID="live:"+ISO8601. The harness still writes active_snapshot and still passes --snapshot_root/--snapshot_id to stages so they see the HALT signal. Every stage then degrades conservatively: authoritative verdicts forbidden (VERIFIED_SECURE, failed_to_reproduce, DUPLICATE, FALSE_POSITIVE, NON_VIABLE); Block B always returns NOTMATCHED; reproduce records `notattempted; patch's best attainable is VERIFICATION_INCOMPLETE`.

4. Patched shadows (--target_root pointing at a pre-mutated tree; sentinel-exempt path 1a in Block A) — mantis-patch passes --target_root=<PATCHED_SHADOW_ROOT> and --snapshot_pinned=false to the reproduce sub-agent for re-attack verification. Block A step 1a: CODE_ROOT = --target_root (authoritative override, overrides --snapshot_root and state fallback). Block A step 2: sentinel check SKIPPED (a --target_root tree is deliberately mutated and is sentinel-EXEMPT). The --snapshot_pinned=false argument is the sentinel-exemption, NOT a HALT signal — detect HALT by reading STATE (active_snapshot.snapshot_id starts with live:, equivalently active_snapshot.snapshot_pinned is false in state), never from the argument passed on this invocation. The finding's discovery_commit is unaffected — it retains the pass-level SNAPSHOT_ID from when it was discovered; only the --snapshot_pinned=false argument is local to the reattack invocation.

5. Different-snapshot duplicate candidates (cross-pass dedupe where discovery_commit differs — the pairwise Block B NOTMATCHED path) — Both passes pinned correctly; the findings simply come from different snapshots. `mantis-dedupe` Block B pairwise check compares the CURRENT finding's `discoverycommit against the ARCHIVED finding's discoverycommit` (NOT against the global `SNAPSHOTID). If they differ → NOT_MATCHED. NOT_MATCHED keeps the current finding ACTIVE and sets possibleduplicateof (a soft, non-terminal hint — the finding is NOT filtered or trashed). If the archived finding was RESOLVED (patchstatus` in {`VERIFIEDSECURE, MITIGATIONPROPOSED`} OR `status`==`FALSEPOSITIVE OR productionviability`==`NONVIABLE`) AND the pair is NOT_MATCHED → POSSIBLE REGRESSION: keep ACTIVE, add a history note, never filter (a reverted fix re-discovered on new code must never be trashed).

6. Absent sink evidence (Block F — PoC compiles but produces no reached-sink evidence; not_attempted vs failed_to_reproduce) — mantis-reproduce Block F: if EVIDENCE is ABSENT (any compiler/build nonzero exit, exit 127 command-not-found, exit 2 "No such file", or the sink was never reached) → repro_status = not_attempted (retry-eligible), STOP. NEVER failed_to_reproduce. In --reattack mode: leave reattack_status UNSET with a history note "setupfailed" — NEVER `failedtobypass`. `failedtoreproduce` is reserved for when the harness PROVABLY reached the vulnerable entrypoint — i.e. reached-sink evidence, not setup evidence — but the bug did not fire. Reached-sink evidence must originate INSIDE the invoked path or from target-produced tracing/backtraces: (a) a PoC script/source harness writes `MANTISREACHEDENTRYPOINT` to a sidecar file at the point just before the sink call, within its own execution flow (the marker write is part of the invoked path, not a pre-launch step); OR (b) for binary/firmware/raw-payload targets, the captured crash backtrace or sanitizer trace (ASan/UBSan/MSan/TSan) explicitly names the target sink function (target-produced tracing). A marker written by an external wrapper BEFORE invoking the target is SETUP EVIDENCE ONLY (proves "launch attempted," not "sink reached") and does NOT by itself justify `failedtoreproduce` — treat it as EVIDENCE ABSENT for the decision gate. Evidence is recorded in `reprohints. In HALT mode, the HALT ceiling additionally forces notattempted` (no `failedto_reproduce`), since a negative result on an unpinned tree cannot be trusted as authoritative.

6. Semantic Retrieval (RAG) for Large Codebases

For small repositories, the planner can manually scan workspace/kb/index.md and the researcher can grep for call-sites. At scale (thousands of files, deep directory trees, multi-pass campaigns), these approaches miss relevant context and waste tokens reading irrelevant files. A semantic retrieval layer lets the planner and researcher query for relevant KB entries and code locations without reading everything.

Two implementations are supported, sharing the same data contract:

• Option A (Default — Skill-Based): A dedicated skill that runs a

BM25/TF-IDF helper script over chunks.jsonl. Zero external dependencies — works air-gapped, no vector embeddings or vector store required. Optional vector embedding support if available.

• Option B (Maximum Scale — MCP-Based): The harness owns a persistent vector

index using vector embeddings, serving persistent semantic_search_kb / semantic_search_code MCP tools. Better for very large codebases where per-invocation BM25 is too slow.

Both are opt-in. The existing skills are not modified; the planner and researcher receive runtime instructions to use whichever retrieval mechanism is available, falling back to today's manual behavior if neither is present. Retrieval results are coverage HINTs only — they decide ordering and prioritization, never the membership of the audit set. A miss must never cause a file, call-site, or investigation to be skipped or dropped.

A. Shared Data Contract: chunks.jsonl

After Stage 2 (/mantis-architecture) completes, chunks are extracted into workspace/kb/chunks.jsonl (one JSON object per line). The harness can do this post-hoc by reading workspace/kb/*.md, or the architecture skill can be instructed to write it during synthesis as a text-only side effect. Two chunk types are produced:

1. KB chunks from the existing workspace/kb/*.md files:

   {"id": "auth_module:0", "source_file": "workspace/kb/entities/auth_module.md", "entity_type": "entity", "chunk_text": "The auth module handles..."}

1. Code chunks from CODE_ROOT (the pinned snapshot). Each chunk includes

the file path and line range so the researcher can request specific files from the snapshot:

   {"id": "src/parser.c:0", "source_file": "src/parser.c", "start_line": 1, "end_line": 80, "chunk_text": "int parse_input(..."}

The first line of chunks.jsonl is a provenance header recording the SNAPSHOT_ID the chunks were built against:

{"_provenance": true, "snapshot_id": "abc123", "kb_snapshot_id": "abc123"}

Before serving queries, check snapshot_id in the provenance header against the current SNAPSHOT_ID; rebuild if they differ. In MODE-OFF (no active_snapshot), kb_snapshot_id is never stamped — skip the index entirely and let skills fall back to manual scanning. Never build code chunks from the live tree — they must reflect the pinned copy the skills are reading.

B. Option A: Skill-Based Retrieval (Default — No Infrastructure)

A dedicated skill reads chunks.jsonl and writes+runs a helper script (e.g. workspace/helpers/search_chunks.py) that performs BM25/TF-IDF similarity search. The script is generated by the agent at runtime — no code is shipped with the skill (same pattern as mantis-dedupe's merge_findings.py). This requires zero external dependencies — no embedding model, no vector store, no MCP server. It works in air-gapped and VPC-SC environments.

A complete reference blueprint for this skill is available at references/mantis-kb-query.md. Builders can adapt it to their environment. The blueprint includes Block A (Locator Resolution), chunk provenance checking, the versioned helper script contract (MANTIS_HELPER_VERSION = 1), and the JSON output schema.

• Invocation: The planner or researcher spawns the skill as a sub-agent with

a query string. The skill writes the helper if not already present, runs it, and returns top-K matching chunks as JSON.

• Optional embeddings: If vector embeddings are available, the agent can be

instructed to use cosine similarity instead of BM25. This is a runtime configuration toggle, not a different skill.

• Snapshot safety: The skill reads active_snapshot from state via Block A

(same as every other skill) and checks chunk provenance before serving.

C. Option B: MCP-Based Retrieval (For Maximum Scale)

For very large codebases where per-invocation BM25 is too slow, the harness can own a persistent vector index using vector embeddings, serving two MCP tools (following the same pattern as Guideline 2's Custom MCP for VMs/hardware):

• semantic_search_kb(query: string) → [{id, source_file, entity_type, chunk_text, score}]

— Searches KB chunks. Returns relevant entity/vulnerability markdown context.

• semantic_search_code(query: string) → [{file, start_line, end_line, snippet, score}]

— Searches code chunks from the pinned snapshot. Returns relevant code locations.

The harness manages the vector index lifecycle: build from chunks.jsonl (or directly from CODE_ROOT), rebuild when SNAPSHOT_ID changes, and handle freshness checks. In HALT mode, serve with a STALE flag or refuse. In MODE-OFF, skip entirely.

D. Per-Skill Augmentation Guidance

When a retrieval mechanism (skill or MCP) is available, instruct the following skills to use it. These are runtime instructions passed by the harness or meta-agent when invoking the skill — the skill files themselves are not modified:

• mantis-architecture: No changes needed. The harness chunks the existing

workspace/kb/*.md files after the architect completes Stage 2. If the builder prefers, they may instruct the architect to also write workspace/kb/chunks.jsonl during synthesis (Step 3) as a text-only side effect — but this is optional, since the harness can extract chunks post-hoc.

• mantis-plan: If a retrieval mechanism is available, instruct the planner

to use it to discover kb_references for each investigation instead of only manually scanning workspace/kb/index.md. For each investigation, query with the investigation title and target file names, then add the top-K matching KB entity/vulnerability files to the kb_references array. Manual scanning of index.md remains the fallback when no mechanism is available.

• mantis-researcher: If a retrieval mechanism is available, instruct Wave 1

sub-agents to use it to PRIORITIZE relevant call-sites and cross-module data flows into sinks (e.g., "where does untrusted input reach memcpy in the parser module"). Semantic search SUPPLEMENTS grep as a ranking HINT ONLY — it decides ORDER, never MEMBERSHIP of the audit set. It MUST NEVER replace the exhaustive Step-3 call-site sweep; every call-site or data-flow that a full grep would reach must still be audited whether or not it ranks in top-K. Audit the union of grep results and semantic search results. The researcher's existing Wave 1/Wave 2 structure is unchanged.

E. Snapshot Safety

The retrieval index — whether served by the skill or the harness — is a cache of the pinned snapshot, never a live view:

• Build code chunks from CODE_ROOT (the pinned snapshot), not the live tree.
• Rebuild when SNAPSHOT_ID changes (new pass, new pin).
• In HALT mode (snapshot_pinned=false), serve results with a STALE flag or

refuse to serve — same conservative degradation as every other stage.

• In MODE-OFF, skip the index entirely.

7. Embedding-Based Deduplication Pre-Filtering

When the pipeline runs many passes over a large codebase, the deduplicator (/mantis-dedupe) must compare each current finding against every archived finding — an O(n×m) comparison performed by an LLM reading summaries. At scale (hundreds of findings across many passes), this is token-expensive and slow.

The harness can use embeddings as a fast pre-filter to reduce the candidate space before invoking /mantis-dedupe. The skill's existing deterministic matching (code_paths + title + discovery_commit) remains the sole authority for hard dedup decisions.

This is entirely harness-side and opt-in. /mantis-dedupe is not modified.

A. How It Works

1. Harness Action: Reads all workspace/findings/*.json (current) and

workspace/archive/findings_pass_*/*.json (archived). For each finding, computes an embedding from a normalized text representation (e.g., title + description + first code_paths entry with line stripped).

1. Harness Action: Computes pairwise cosine similarity between current and

archived findings. Surfaces candidate pairs above a configurable threshold (e.g., 0.85).

1. Harness Action: Writes a candidate-pairs manifest (e.g.,

workspace/helpers/dedup_candidates.json) containing the UUID pairs and similarity scores.

1. mantis-dedupe invocation: The harness invokes /mantis-dedupe as usual.

If the candidate manifest exists, instruct the skill to read it and prioritize those pairs for the LLM's pairwise comparison, instead of comparing every finding against every archived finding. All existing deterministic matching rules apply unchanged — the manifest only narrows the search space.

B. Safety Guardrails

• Pre-filter only, never the decision. The embedding similarity score can

never cause a DUPLICATE verdict, a trash move, or a possible_duplicate_of assignment. Only the skill's existing code_paths, title, and discovery_commit checks can do that. A miss in the pre-filter only over-retains a duplicate (safe — the LLM sees it and skips it); it never under-retains (never drops a real duplicate).

• No false negatives. The threshold should be set low enough (e.g., 0.75) to

avoid missing true duplicates. Better to surface too many candidates than to miss a real one — the LLM and deterministic matching will filter false positives.

• Fallback on failure. If the embedding computation is unavailable, the

candidate manifest is absent, or any error occurs, /mantis-dedupe falls back to its existing O(n×m) comparison. The skill must not stop or error if the manifest is missing.

• Snapshot awareness. The harness should not compute embeddings across

different discovery_commit values without flagging them as cross-snapshot candidates — the skill's Block B pairwise check will handle the final MATCHED/NOT_MATCHED decision.

C. Shared Infrastructure

If the builder also implements Guideline 6 (Semantic Retrieval), reuse the same vector embedding infrastructure for finding embeddings. The finding embedding is a different payload (finding JSON, not KB chunks) but the same embedding capability can serve both.

D. Enterprise Vector Architecture & Cloud Scalability

For enterprise deployments tracking hundreds of thousands of vulnerabilities across distributed codebases and multi-tenant audit pipelines, SQLite vector storage can be swapped for managed cloud vector engines:

• Cloud Spanner Vector Search:
• Best For: Mission-critical enterprise pipelines requiring global ACID

consistency, 99.999% availability, and synchronized finding lineage tracking.

• Architecture: Store findings and lineage_vectors in Spanner with

ARRAY<FLOAT32> vector columns. Utilize COSINE_DISTANCE with KNN search indexes for real-time finding deduplication during parallel scan passes.

• Cloud SQL for PostgreSQL (`pgvector`):
• Best For: Standard enterprise relational backends needing ACID compliance

with low operational complexity.

• Architecture: Enable the vector extension. Store finding embeddings in a

vector(768) or vector(256) column with HNSW (vector_cosine_ops) or IVFFlat indexes for sub-millisecond similarity queries.

• AlloyDB for PostgreSQL:
• Best For: High-throughput multi-agent audit farms running concurrent

research waves.

• Architecture: Leverage AlloyDB's columnar vector engine and ScaNN-based

approximate nearest neighbor indexing for up to 10x faster vector queries over standard PostgreSQL.

• BigQuery Vector Search:
• Best For: Batch analytical processing, organization-wide threat

intelligence clustering, and cross-campaign vulnerability lineage analysis.

• Architecture: Ingest finding embeddings into BigQuery and execute

VECTOR_SEARCH(TABLE findings, 'embedding', ...) with COSINE distance for serverless batch deduplication and historical regression analytics.

8. SAST Seeding (External Tool Ingestion)

Mantis's discovery engine is 100% LLM-generative (grep swarm + reasoning). A weak LLM can structurally under-detect whole-program taint classes (injection, path traversal, deserialization, UAF, format-string) that mature SAST tools (CodeQL, Semgrep-taint) encode as interprocedural queries. A SAST seeding adapter ingests external tool findings as PROVISIONALLY_VALID / NEEDS_RESEARCH candidates that must earn their verdict through the unchanged downstream gates. This is purely additive (INV-2-strengthening) — it expands detection breadth without weakening any verification gate.

This follows exactly the RAG pattern from Guideline 6: opt-in, default off, provenance-tracked, snapshot-aware, fallback on failure.

Platform-agnostic IR (not SARIF): Rather than tying the adapter to SARIF (a complex, tool-specific format), the adapter consumes a minimal JSONL intermediate representation (IR). Any SAST tool's output (SARIF, Semgrep JSON, Bandit JSON, etc.) is converted to this IR by a thin wrapper. This maximizes platform agnosticism — the adapter works with any tool that can produce the simple JSONL format.

Two implementations are supported, sharing the same data contract:

• Option A (Default — Skill-Based): A dedicated skill (/mantis-sast-seed)

reads the IR and uses LLM reasoning to normalize findings into mantis finding JSONs. The LLM reads actual source code at each reported location under CODE_ROOT to verify the finding and enrich the description with root-cause analysis. Zero external dependencies — works air-gapped.

• Option B (Maximum Control — Harness-Based): The harness directly

normalizes SAST output into finding JSONs using deterministic code (e.g., a SARIF-to-finding converter script), bypassing the LLM for the normalization step. The harness writes finding JSONs to workspace/findings/ before invoking /mantis-dedupe.

Both are opt-in. The existing skills are not modified. Seeded findings enter workspace/findings/ alongside researcher findings and flow through the unchanged downstream gates (dedupe -> review -> critic -> reproduce -> patch -> calibrate).

A complete reference blueprint is available at references/mantis-sast-seed.md.

A. Shared Data Contract: sast_findings.jsonl

The IR is a JSONL file at workspace/sast_findings.jsonl (STATE-RELATIVE). It follows the same provenance-header pattern as chunks.jsonl (Guideline 6A).

Line 1 — Provenance header:

{"_provenance": true, "scan_snapshot_id": "abc123def456", "tool": "codeql", "tool_version": "2.15.0", "scan_timestamp": "2026-07-22T10:00:00Z"}

• scan_snapshot_id: The SNAPSHOT_ID the scan was run against. This is the

primary provenance anchor — compared byte-for-byte to the current pass SNAPSHOT_ID (same comparison as Block B). If the harness ran the SAST tool against the pinned CODEROOT, it sets this to `SNAPSHOTID`.

• tool: Tool name (e.g., codeql, semgrep, bandit). Informational.
• tool_version: Tool version. Informational.
• scan_timestamp: ISO 8601. Informational.

Lines 2+ — One finding per line:

{"rule_id": "cpp/sql-injection", "rule_name": "SQL injection", "cwe": "CWE-89", "severity": "HIGH", "code_paths": ["src/db/query.c:42"], "message": "User input flows into SQL query without sanitization"}

IR conversion (harness responsibility): The harness converts SAST tool output to this IR before invoking the adapter. Conversion examples:

• SARIF: extract ruleId, map level to severity (error->HIGH,

warning->MEDIUM, note->LOW), extract locations to code_paths, copy message.text.

• Semgrep JSON: extract check_id as rule_id, map extra.severity to Mantis

severity, extract path:start.line as code_paths, copy extra.message.

B. Option A: Skill-Based Ingestion (Default — No Infrastructure)

A dedicated skill reads sast_findings.jsonl, applies allow-listing, verifies provenance, computes signature/lineage_id/discovery_commit, and writes finding JSONs. The LLM reads actual source code at each reported location under CODE_ROOT to verify the finding and enrich the description.

• Invocation: The harness invokes /mantis-sast-seed with

--snapshot_root/--snapshot_id/--state_root after Stage 6 (Research) and before Stage 7 (Dedupe). The seeded findings land in workspace/findings/ before /mantis-dedupe runs.

• Inert until wired: If sast_findings.jsonl is absent, the skill outputs

nothing and notifies the caller. It never fails — it simply returns empty.

• Anti-hallucination: The LLM MUST read the actual code at each reported

location before writing the finding JSON. It MUST NOT invent findings not present in the SAST output. The sast_provenance.line_verified field records whether verification succeeded.

• Snapshot safety: The skill reads active_snapshot from state via Block A

and stamps discovery_commit only when the snapshot is pinned and the finding's location is verified under CODE_ROOT.

C. Option B: Harness-Based Ingestion (For Deterministic Ingestion)

For maximum determinism, the harness can normalize SAST output into finding JSONs using deterministic code. This bypasses the LLM for the normalization step:

1. Harness Action: Reads the SAST tool output, parses it deterministically,

and writes finding JSONs to workspace/findings/.

1. Harness Action: Stamps discovery_commit only if the scan provably ran

against the pinned CODE_ROOT.

1. Harness Action: Applies allow-listing filters.
2. mantis-dedupe invocation: Proceeds as usual — the seeded findings are

indistinguishable from researcher findings.

D. Provenance Verification

The adapter stamps discovery_commit ONLY if the scan provably ran against a line-identical pinned CODE_ROOT. Verification ladder:

1. Read active_snapshot from workspace/.mantis_state.json (via Block A). If

absent, this is MODE-OFF — skip to step 5.

1. If snapshot_pinned is false, this is HALT — skip to step 6.
2. Read the IR provenance header's scan_snapshot_id.
3. VERIFIED: scan_snapshot_id present AND exactly equals SNAPSHOT_ID.

Stamp discovery_commit = SNAPSHOT_ID. Status = PROVISIONALLY_VALID.

1. MODE-OFF: active_snapshot absent. OMIT discovery_commit. Status =

PROVISIONALLY_VALID (MODE-OFF permits all verdicts).

1. HALT: snapshot_pinned is false. OMIT discovery_commit. Status =

NEEDS_RESEARCH.

1. DRIFT: scan_snapshot_id present but differs. OMIT discovery_commit

entirely. Status = NEEDS_RESEARCH.

1. UNVERIFIED: scan_snapshot_id absent. OMIT discovery_commit. Status =

NEEDS_RESEARCH.

Line-existence verification (additional check when VERIFIED or DRIFT and CODEROOT is resolved): For each finding, verify each `codepaths entry: strip trailing :line, check file exists under CODE_ROOT, check line number is within file's line count. If file missing or line out of range -> downgrade to DRIFT (omit discoverycommit`, status = `NEEDSRESEARCH`).

This is the exact same "UNTRUSTED-IF-ABSENT" pattern as discovery_commit (schema.json) and signature/lineage_id.

E. Allow-Listing and Noise Control

workspace/sast_allowlist.json (optional config file, STATE-RELATIVE):

{
  "enabled": true,
  "severity_filter": ["CRITICAL", "HIGH"],
  "cwe_allowlist": {
    "enabled": true,
    "cwes": ["CWE-89", "CWE-78", "CWE-79", "CWE-22", "CWE-787", "CWE-416", "CWE-502"]
  },
  "rule_allowlist": {
    "enabled": false,
    "rules": []
  },
  "per_rule_cap": 5,
  "total_cap": 50
}

If the config file is absent, defaults apply: CRITICAL+HIGH only, per_rule_cap=5, total_cap=50.

How allow-listing protects the retry cap: The reproduce stage (/mantis-reproduce) has a hard ceiling of 6 attempts per finding (absolute, never reset). If 1000 SAST findings are seeded without filtering, the reproduce stage would need up to 6000 attempts — starving the retry budget. The allow-listing chain (severity filter -> CWE/rule filters -> per-rule cap -> total cap) ensures only a bounded, high-signal set of candidates enters the pipeline. The review (13-rule negative filter) and critic (production viability) stages further filter before reproduce runs.

F. Per-Skill Augmentation Guidance

When the SAST seed skill is available, instruct the harness to invoke it between Stage 6 (Research) and Stage 7 (Dedupe). These are runtime instructions passed by the harness — the skill files themselves are not modified:

• mantis-meta-agent: Invoke /mantis-sast-seed with

--snapshot_root/--snapshot_id/--state_root after /mantis-researcher completes and before /mantis-dedupe.

• mantis-dedupe: No changes needed. Seeded findings are in

workspace/findings/ alongside researcher findings. Dedupe processes them identically (signature-based matching, Block B pairwise check).

• mantis-review: No changes needed. The 13-rule negative filter applies to

seeded findings identically.

• mantis-report: No changes needed. The sast_provenance field is

informational and can be displayed in reports.

G. Snapshot Safety

The SAST seed adapter follows Block A (Locator Resolution) exactly like every other code-reading skill:

• Resolve CODEROOT from `--snapshotroot / state active_snapshot`.
• Honor the sentinel check (Block A step 2).
• Read source files under CODE_ROOT (SNAPSHOT-RELATIVE) for line-existence

verification.

• Write findings under state_root/workspace/findings/ (STATE-RELATIVE).
• Never write under CODE_ROOT when pinned.
• In MODE-OFF: proceed without discovery_commit. All verdicts permitted.
• In HALT: omit discovery_commit. Seeded findings get NEEDS_RESEARCH.
• In PINNED: verify each finding's location under CODE_ROOT, stamp

discovery_commit = SNAPSHOT_ID if verified.

The SAST tool output itself is NOT snapshot-aware — it may have been produced against a different tree. The adapter's provenance verification is what bridges the gap: it re-grounds each finding against the pinned CODEROOT before stamping `discoverycommit`.

H. Safety Guardrails

• Purely additive. Seeded findings are PROVISIONALLY_VALID /

NEEDS_RESEARCH candidates. They must pass through the unchanged downstream gates: dedupe (Block B), review (13-rule filter), critic (viability), reproduce (Block F reached-sink evidence + HALT ceiling), patch (Block G re-attack), calibrate (sanity caps). No gate is weakened.

• No false `VERIFIED_SECURE`. Seeded findings start at PROVISIONALLY_VALID

— they can never reach VERIFIED_SECURE without passing through the full patch + re-attack pipeline.

• No false `failed_to_reproduce`. Seeded findings that reach reproduce are

subject to the same Block F evidence gate and HALT ceiling.

• No dropped regression. If a SAST-seeded finding matches an archived

finding with a different discovery_commit, Block B returns NOTMATCHED -> `possibleduplicate_of` -> finding stays active.

• Fail-safe on missing data. If sast_findings.jsonl is absent, no findings

written. If a finding can't be verified, NEEDS_RESEARCH.

• No existing skills modified. The adapter writes findings to

workspace/findings/ before /mantis-dedupe runs.

• `sast_provenance` is informational only. It does not affect any safety-

critical invariant, gate, or verdict. The finding's discovery_commit is the field that governs Block B snapshot matching.

9. Structural Code Index (AST-Level Context)

For small repositories, the researcher can grep for call-sites and the planner can infer dependencies. At scale, grep-based call-site discovery is unreliable (misses indirect calls, cannot distinguish calls from comments/strings, no function boundary awareness). A structural code index provides AST-level context (function boundaries, call graphs, symbol tables) to improve LLM reasoning quality during discovery.

This follows the RAG pattern from Guideline 6: optional, provenance-tracked, snapshot-aware, fallback on failure. The structural index is a coverage HINT only — it decides ordering and prioritization, never the membership of the audit set. A miss must never cause a file, call-site, or investigation to be skipped or dropped.

The full specification — including the manifest schema, SQLite serving store, capability-based per-partition backend selection, canonical symbol IDs, query interface, baseline-plus-delta overlay, deterministic partial coverage, and safety guardrails — lives in a single source of truth:

→ [../mantis-structural-index/SKILL.md](../mantis-structural-index/SKILL.md)

A thin reference blueprint is at references/mantis-structural-index.md.

Two implementations are supported, sharing the same query contract:

• Option A (Default — Skill-Based): The mantis-structural-index skill

generates and runs helper scripts (build_structural_index.py and query_structural_index.py, both # MANTIS_HELPER_VERSION = 5) using capability-based per-partition backend selection, degrading to grep.

• Option B (Maximum Power — MCP-Based): The harness owns a persistent

structural index serving find_callers(symbol), find_callees(function), get_function_boundary(file, line) MCP tools, backed by the same SQLite serving store.

Both are optional. A non-conformant harness simply skips the structural index stage. The structural index supplements grep as a ranking HINT ONLY — it decides ORDER, never MEMBERSHIP of the audit set.

Consumers MUST use the query interface (query_structural_index.py) rather than filtering JSONL directly. The query interface provides bounded results, pagination, explicit name resolution, precision/backend metadata, and coverage on empty results.

D. Per-Skill Augmentation Guidance

When a structural index is available, instruct the following skills to use it. These are the consumption contract — runtime instructions passed by the harness. The structural index is a HINT-only enhancement; skills that do not use it behave exactly as they do today:

• mantis-architecture: Optionally read the pre-built structural index (built

by mantis-structural-index at Stage 0.5) during KB synthesis, cross-referencing it with dependencies.json.

• mantis-plan: Use the structural index for function-level dependency

fan-out (more precise than file-level dependencies.json).

• mantis-researcher: Wave 1 sub-agents use find_callers() to SUPPLEMENT

grep as a ranking HINT for call-site discovery — it decides ORDER, never MEMBERSHIP. It MUST NEVER replace the exhaustive Step-3 call-site sweep; every call-site that a full grep would reach must still be audited whether or not it ranks in the structural index. Audit the union of grep results and structural index results. Wave 2 deep auditors use get_function_boundary() to start with the enclosing function, expanding to callers/callees/file as needed for cross-function context. The researcher's existing Wave 1/Wave 2 structure is unchanged.

E. Integration with RAG (Guideline 6)

• Structural index entries can be added to chunks.jsonl as

entity_type: "structural" chunks.

• Code chunks can become function-boundary-aligned (instead of dumb line-range

slices) by using function boundaries from the structural index.

• The structural index and the RAG index can share the same vector embedding

infrastructure if both are implemented.

10. Tiered Iterative Reproduction & Multi-Conversation Retry Strategy

For complex services, attempting a single-shot reproduction directly against a full sandboxed service often suffers from high search entropy, brittle configuration, and hard-to-debug failures. A tiered strategy breaks reproduction into incremental milestones, while an inter-conversation retry architecture prevents reasoning deadlocks and context bloat.

A. Orchestration & Inter-Conversation Retries

1. Intra-Conversation Retries (Local Agent Trajectory):

• The active subagent conversation retries 2–3 times locally within its

context window to adjust parameters, fix setup bugs, or refine payloads.

1. Inter-Conversation Retries (Fresh Context + Accumulated Artifacts):

• Trigger: If intra-conversation retries fail to reach Tier 3

(reproduced), the orchestrator terminates the stalled conversation and launches a new subagent conversation (a fresh context window).

• Context Provisioning: The orchestrator populates the new prompt with

structured attempt data from state_root/workspace/archive/.repro_attempts.json and trajectory learnings from workspace/learnings.jsonl (e.g., "Attempt 1 failed due to missing auth header X; Attempt 2 proved parser strips unescaped quotes").

• Benefit: Eliminates context bloat and reasoning inertia ("hallucination

traps"), enabling a fresh agent to solve the problem using prior empirical observations without repeating past mistakes.

1. Attempt Cap Accounting & Arithmetic:

• The orchestrator maintains

state_root/workspace/archive/.repro_attempts.json.

• Tier 3 Increment Only: Only Tier-3 full sandboxed service executions

(or full end-to-end reproducer runs) increment the per-finding attempt counter toward the absolute hard ceiling of 6.

• Stepping-Stone Sub-Budget: Internal Tier-1 and Tier-2 trial runs are

bounded by a local sub-budget (max 3 trial executions per conversation) and do not consume the absolute 6-attempt cap.

B. Harness-Enforced Tier 4 (Staging & Live Pre-Production Execution)

To prevent un-gated exploit execution on live infrastructure, Tier 4 (Staging / Pre-Production verification) is strictly owned and enforced by the programmatic orchestrator harness, never by LLM discretion:

• Sandbox Boundary: mantis-reproduce executes strictly within isolated

local sandboxes (Tiers 1–3) ending at Tier 3 (reproduced / failed_to_reproduce).

• Deterministic Gate: The harness intercepts a Tier 3 reproduced verdict.

If live/staging validation (Tier 4) is configured, the harness MUST NOT automatically invoke remote execution. It must enforce a programmatic Human-in-the-Loop gate:

• Prompt the human operator for explicit interactive confirmation, OR
• Require a cryptographically signed approval token / authorization callback.
• Fail-Closed Default: If human approval is missing or denied, Tier 4 is

skipped and the Tier 3 sandboxed verdict remains authoritative.

C. Local Ingress / Middlebox Edge Annotation (Post-Tier-3 Sandbox Verification)

To eliminate false-positive findings caused by default edge filters ("Works on localhost:8080, but dies at the WAF/proxy"), the orchestrator can optionally execute a Tier 3 PoC through a local reverse proxy or API gateway (e.g. NGINX, Envoy, ModSecurity) running inside the local sandbox:

• Annotation Only: A payload blocked by a local middlebox MUST NOT downgrade

a Tier 3 reproduced verdict to failed_to_reproduce.

• Critic Integration: The harness records ingress_blocked: true (or edge

filter details) in the finding's repro_hints or history. This provides empirical evidence for /mantis-critic to classify production_viability as "CONDITIONAL_VIABLE" (mitigated by default edge proxy configuration).