Mantis Reproduce

Reproducer (/mantis-reproduce)

System Goal

Integration Test Engineer. Designs crash reproducers or inputs and executes them

inside isolated sandbox environments to empirically verify bugs.

Command Definition

• Command:

/mantis-reproduce [--reattack] [--finding_id=<uuid>] [--force] [--target_root=<path>] [--state_root=<path>] [--snapshot_root=<path>] [--snapshot_id=<SNAPSHOT_ID>] [--snapshot_pinned=<true|false>]

• Description: Generates and runs crash reproducers to verify security

flaws.

• Parameters:
• --reattack: When executing as part of patch verification to isolate

re-attack outcomes.

• --finding_id: The specific finding UUID to reproduce. Must be provided

and is required when --reattack is specified.

• --force: Override/bypass eligibility checks for targeted normal runs.
• --target_root: Path to the root of the target codebase under test

(defaults to .). AUTHORITATIVE when supplied — overrides --snapshot_root

(Block A step 1a); the sentinel check is skipped for this tree (e.g. a

patched shadow during re-attack verification).

• --state_root: Path to the root of the Mantis state directory containing

workspace/ (defaults to .).

• --snapshot_root: Root of the pinned immutable code snapshot for this pass.

Consumed by Block A (Step 0) when --target_root is not supplied.

• --snapshot_id: The SNAPSHOT_ID string of the pinned snapshot, consumed by

Block A (sentinel) and Block B (snapshot match check) in Step 0.

• --snapshot_pinned: When false (set by mantis-patch during re-attack on

a patched shadow), the reproduce sub-agent MUST skip the snapshot

sentinel/match check for this invocation — the --target_root tree is

authoritative and sentinel-exempt (Block A step 1a).

Input/Output Contract

• Reads:
• state_root/workspace/findings/ (viable/conditional findings).
• target_root/ (Repository source files to analyze trigger paths).
• state_root/workspace/archive/.repro_attempts.json.
• state_root/workspace/.mantis_state.json (to track current loop pass).
• Writes:
• PoC reproduction files (e.g. poc_[uuid].py or crash_[uuid].payload

inside state_root/workspace/reproducers/).

• If run normally: updates findings in-place under

state_root/workspace/findings/ (sets "repro_status",

"repro_file_path", "run_command", "repro_output", and appends

history). Updates status to "VALID" if provisionally valid.

• If run with --reattack: updates findings in-place under

state_root/workspace/findings/ (sets "reattack_status",

"reattack_file_path", "reattack_run_command", "reattack_output",

"reattack_variants", and appends history with stage "reattack"). Does

not modify "repro_*" fields or "status". Exception: may atomically

downgrade patch_status per INV-1 in Step 6 (never persist

VERIFIED_SECURE alongside a non-failed_to_bypass reattack_status).

• Updates state_root/workspace/archive/.repro_attempts.json atomically.
• Stamps "repro_snapshot_id" / "reattack_snapshot_id" on updated findings

and stores .repro_attempts.json values as {count,last_snapshot} objects

(bare integers still read correctly).

• Preconditions:
• Findings must exist in state_root/workspace/findings/.
• Sandbox/container runtime environment must be available.
• Idempotency Guarantee:
• Updates findings in place. Uses

state_root/workspace/archive/.repro_attempts.lock file locking and atomic

temporary file swaps (os.replace on

state_root/workspace/archive/.repro_attempts.json.tmp) to guarantee

concurrency safety and retry stability.

• Snapshot-aware: regenerates the PoC when the finding's snapshot no longer

matches; refuses to emit a negative verdict without reached-sink evidence.

Re-attack verdicts on a snapshot mismatch are governed by the C5

unpatched-baseline re-run (Step 6), which supersedes the legacy blanket

refusal — a failed_to_bypass verdict is only written after C5 has

confirmed the unpatched baseline still triggers on the current snapshot.

Instructions

Step 0: Locator Resolution + Snapshot Match (run first)

LOCATOR RESOLUTION (before reading ANY target code or artifact):
0. ROLE: If this skill NEVER reads target source (report, calibrate, reflect),
   you are a FINDINGS-ONLY stage: skip steps 2-6; still read active_snapshot from
   state for provenance/annotation; NEVER stop merely because a code root is unset.
1. Determine CODE_ROOT, in this priority order:
   a. If --target_root is passed on THIS invocation, CODE_ROOT = --target_root.
      It is AUTHORITATIVE and OVERRIDES SNAPSHOT_ROOT and the state fallback
      (used when a caller hands you a prepared tree, e.g. a patched shadow).
   b. Else if --snapshot_root (or SNAPSHOT_ROOT) is passed, use it.
   c. Else read state_root/workspace/.mantis_state.json (state_root from
      --state_root if passed, else ./workspace/... relative to the current dir)
      -> active_snapshot.root / .snapshot_id / .snapshot_pinned.
   d. Else (no arg AND no readable active_snapshot): CODE_ROOT = current directory,
      treat snapshot_pinned = false (MODE-OFF). Do NOT stop.
2. SENTINEL CHECK (only if snapshot_pinned is true AND you did NOT take path 1a):
   verify CODE_ROOT/.mantis_snapshot_id exists and equals SNAPSHOT_ID. If missing
   or different -> STOP "snapshot sentinel mismatch". (A --target_root tree (1a) is
   deliberately mutated and is sentinel-EXEMPT.)
3. PATH FIELDS:
   - SNAPSHOT-RELATIVE (read under CODE_ROOT): code_paths entries; plan target_files
     that are file paths. Strip ONLY a trailing ":<digits>". A code_paths entry
     containing "://" is a URL/endpoint, NOT a file read. A code_paths entry that is
     NOT of the form <existing-path>:<integer> is a non-source LOCATOR
     (symbol/offset/endpoint): only check that the artifact/symbol exists; skip ALL
     line-range and line-existence logic.
   - STATE-RELATIVE (read/write under state_root/workspace, NEVER prefix CODE_ROOT):
     kb_references, repro_file_path, reattack_file_path, helper scripts, report
     files, and all state/findings JSON.
4. Never WRITE under CODE_ROOT when snapshot_pinned is true. Any command that
   compiles, generates, or writes artifacts MUST run in a PRIVATE SHADOW copy
   (mktemp -d from CODE_ROOT), never with cwd=CODE_ROOT. Read-only inspection may
   cd into CODE_ROOT.
5. VCS-METADATA CARVE-OUT: history-log extraction and any VCS diff/blame command
   run in the LIVE repository root (which still has .git/.hg/.repo), NOT CODE_ROOT
   (the snapshot copy strips VCS metadata). Do NOT stop merely because CODE_ROOT
   lacks .git/.hg/.repo.
6. Every shell command uses ABSOLUTE paths and sets its own working directory on
   that call. Do NOT assume the working directory persists between calls.

> [!NOTE] **CURRENT-PASS CHECK (defensive; the binding guarantee is on the

> harness per mantis-pipeline-adapter Scenario 2):** if active_snapshot is

> present AND active_snapshot.pass != state.pass_number, treat the snapshot as

> STALE for this pass — STOP "stale active_snapshot: pass mismatch" or degrade

> as HALT (snapshot_pinned effectively false: no authoritative verdicts, Block

> B NOT_MATCHED, reproduce not_attempted). This catches a custom harness that

> preserved active_snapshot across the Stage 15 pass increment without

> re-pinning. The reference meta-agent re-pins every pass, so this check never

> fires there. Block B itself cannot detect this (it is snapshot_id-only, not

> pass-aware).

SNAPSHOT MATCH CHECK for finding F (decides MATCHED vs NOT_MATCHED):
1. If snapshot_pinned is false -> NOT_MATCHED. Stop.
2. Read F.discovery_commit:
   - missing OR empty OR the literal "MIXED" -> NOT_MATCHED.
   - not exactly equal to SNAPSHOT_ID          -> NOT_MATCHED.
   - exactly equal to SNAPSHOT_ID              -> MATCHED.
There is no other route to MATCHED; never fuzzy-compare. The global "default the
field and proceed" backward-compat rule does NOT apply to discovery_commit:
absent = NOT_MATCHED. (There is NO separate "dirty" gate: a dirty tree's
SNAPSHOT_ID already embeds the working-tree content hash, so within-pass findings
MATCH and cross-pass bare-commit findings do not.)

Notes: When invoked by the patcher with --target_root=<shadow> (a patched

copy), Block A step 1a makes that shadow the authoritative CODE_ROOT and SKIPS

the sentinel check (the shadow is deliberately mutated). Otherwise CODE_ROOT is

the pinned snapshot and the sentinel MUST match. Stamp repro_snapshot_id

(normal run) or reattack_snapshot_id (--reattack) = the current SNAPSHOT_ID

(from --snapshot_id or state active_snapshot.snapshot_id) on every finding

you update.

Write a Proof-of-Concept Reproduction Script (Repro) or raw input payload file

that reproduces a confirmed security flaw.

Execute the reproduction stage under these constraints:

• Load Viable Findings:

• If --finding_id is supplied:
• Load only that finding's file

(state_root/workspace/findings/<uuid>.json). Exit if it does not exist.

• If --reattack is specified: Enforce the **expected patch workflow

state** for the loaded finding:

• The finding's "status" must be "VALID" or "PROVISIONALLY_VALID".
• The finding's "repro_status" must be "reproduced".
• The finding's "patch_status" must NOT be "MITIGATION_PROPOSED".

(VERIFIED_SECURE IS allowed: C5 below atomically downgrades it when

the re-attack outcome is not failed_to_bypass.)

• Exit with an error if these conditions are not met, explaining the

invalid state.

• If --reattack is NOT specified (Targeted Normal Run):
• If --force is NOT specified, enforce standard eligibility filters:
• The finding's "status" must be "VALID" or

"PROVISIONALLY_VALID".

• The finding's "production_viability" must be "VIABLE",

"SAMPLE_OR_TEST", or "CONDITIONAL_VIABLE".

• Exit with an error if these conditions are not met, explaining the

invalid state.

• If --force is specified, bypass these eligibility checks.
• If --finding_id is not supplied:
• Constraint: Exit if --reattack is specified (it requires

--finding_id).

• Read the JSON files in the state_root/workspace/findings/ directory.
• Strict Eligibility Filter (Normal Runs): Include only findings where:
• "status" is "VALID" or "PROVISIONALLY_VALID".
• "production_viability" is "VIABLE", "SAMPLE_OR_TEST", or

"CONDITIONAL_VIABLE" (or skip this viability filter if not checking

viability, but always check status).

• If no applicable findings exist, notify the user and exit.

Tier 0 — Structural Reachability Pre-Check (Advisory Queue Sorting): If a

structural code index (mantis-structural-index) is available, you MAY query

query_structural_index.py (find_callers) before authoring code to check

whether an AST call path exists from a public entrypoint to the vulnerable

sink. Use this query to prioritize candidate execution order (process

findings with verified AST reachability first).

• CRITICAL HINT-ONLY GUARDRAIL: AST reachability is a ranking HINT ONLY.

Call graphs miss macros, function pointers, dynamic dispatch, and interface

tables. An absent call path MUST NEVER reject a finding, skip reproduction,

or set failed_to_reproduce.

Snapshot drift check: For each loaded finding, if it already has a

repro_snapshot_id and Block B (Step 0) returns NOT_MATCHED, treat any

stored PoC/offsets as STALE: regenerate the reproducer from scratch against

the current CODE_ROOT (do not reuse old line numbers/addresses). If Block B

is MATCHED you may reuse an existing PoC.

• Strict Host Isolation Constraint:

• Host command execution is strictly prohibited. Do not run commands directly

on your parent host terminal using terminal/shell execution tools.

• All reproducer executions must run isolated. Use the containerization or

sandbox execution tools provided by your environment. For memory-safety

PoCs, restrict network access and file system writes as much as possible.

For logic/auth functional tests, you may enable local network services as

needed, but never expose the environment to the external internet.

• Writing and Launching the Reproducer: Write a self-contained test script

(e.g., poc_[uuid].py or a C reproducer file in the same directory) or write

a raw crash input data payload (e.g., crash_[uuid].payload) that triggers

the target bug. **All generated PoC/re-attack scripts and payloads MUST be

written inside the state_root/workspace/reproducers/ directory (never in

the target_root directory).** You must ensure the parent directory

state_root/workspace/reproducers/ exists (e.g. using mkdir -p) before

writing any files. Analyze the code path and constraints carefully. If your

initial reproduction attempt fails, evaluate if the finding details (such as

input paths, parameters, or assumptions) are slightly incorrect based on your

observations, and adjust the finding details dynamically to attempt a fix. If

you cannot find a triggerable path after trying multiple approaches and

adjustments, abandon the attempt. **Do NOT directly mark it as

failed_to_reproduce** — route the abandon decision through the Step-5 Block

F (Reached-Sink Evidence) gate: if the harness provably reached the

vulnerable entrypoint but the bug did not fire, classify as

failed_to_reproduce; if evidence is absent (setup/build failure, exit 127,

"No such file", or the sink was never reached), classify as not_attempted

(retry-eligible). A raw negative from a setup/build failure burns the retry

cap and silently drops a real bug.

To run your script or payload, use the execution or containerization tools

available in your environment to execute the code safely. Select the most

appropriate runtime image and flags for the target. **All compilation and

test execution commands MUST be run in a PRIVATE BUILD SHADOW, never with

Cwd=CODE_ROOT (the snapshot is read-only — Block A step 4).** Before

compiling, create BUILD_ROOT=$(mktemp -d) and copy CODE_ROOT into it (e.g.

cp -a CODE_ROOT/. BUILD_ROOT/); run all compilation/test commands with

Cwd=BUILD_ROOT; delete BUILD_ROOT on teardown. Keep the generated PoC file

itself under state_root/workspace/reproducers/ (STATE-RELATIVE) and store

its ABSOLUTE path in "run_command"/"reattack_run_command".

Sanitizer compilation (C/C++ targets): When the bug class is

memory-safety or undefined-behavior, compile with

-fsanitize=address,undefined (ASan + UBSan) at a minimum. For data races,

add -fsanitize=thread (TSan) — TSan is also mutually exclusive with ASan,

so use a separate build. Use -fno-omit-frame-pointer for usable stack

traces. These flags surface bugs that would otherwise exit 0 silently (UBSan

defaults to recover-mode). MSan caveat: MemorySanitizer

(-fsanitize=memory) detects uninitialized-memory reads, but it requires the

ENTIRE dependency chain (including libc and the C++ runtime) to be

MSan-instrumented — on an arbitrary target that is usually not the case, so a

naive -fsanitize=memory build yields bogus WARNING: MemorySanitizer

traces. MSan is also mutually exclusive with ASan in one build (use a

separate build, not an additive flag). If a fully instrumented MSan

environment cannot be guaranteed, do NOT add -fsanitize=memory; fall back

to ASan+UBSan rather than trusting a possibly-bogus MSan trace. Recognizing

MSan output as evidence (Step 5) is still correct when a legitimate MSan

instrumented build is available. Sanitizer consistency guardrail: the

SAME sanitizer flags MUST be used for the baseline, attack, and re-attack

runs — a different sanitizer can mask the bug and produce a false negative

(INV-2) or false VERIFIED_SECURE (INV-1). See Block G.

{TARGET_ROOT} token substitution (numbered step):

• When writing run_command or reattack_run_command, use the literal

token {TARGET_ROOT} for any path that references the target tree.

• At run time, reproduce substitutes {TARGET_ROOT} with the actual root:
• On first execution: CODE_ROOT (the snapshot or --target_root).
• On re-execution (re-attack or retry): the current CODE_ROOT /

--target_root / active_snapshot.root.

• This ensures the stored command resolves correctly after snapshot GC —

never bake an absolute .mantis_snapshots/pass_<N> path into the stored

command.

Execute your reproduction using the appropriate environment: If the

target is firmware, you may write a script to boot it via qemu, unicorn,

or Firmadyne. If it's a binary, you may use dynamic instrumentation or

standard execution. Use your best judgment to construct a working harness for

the artifact.

• *Parallel Trajectory Search vs. Tiered Iterative Reproduction:*

• Parallel Trajectory Search (Breadth-First): When subagents are

available, deploy concurrent workers taking diverse logical approaches to

reproduce the bug. If any trajectory succeeds, immediately adopt its

payload and discard the others to escape potential "give up" loops and

prune compute costs.

• Tiered Iterative Reproduction (Depth-First Payload Refinement): Each

trajectory worker (or a single agent) uses a tiered escalation ladder

(Tier 1 -> Tier 2 -> Tier 3) to refine its trigger payload incrementally

rather than attempting a single-shot end-to-end launch.

• *Tiered Iterative Execution Ladder:*

• Tier 1 (Micro-Harness / Sink Logic Validation): Construct a

lightweight test calling the vulnerable function/module directly to

verify that the core bug hypothesis is sound in isolation.

• Tier 2 (Subsystem / Interface Validation): Pass the payload through

input serialization, parsers, routing, and auth wrappers to verify the

input survives intermediate processing without sanitization or

truncation.

• Tier 3 (Full Sandboxed Service / E2E Validation): Execute the

self-contained PoC against the target service via public APIs inside the

isolated sandbox (Docker, QEMU, VM). Yields the authoritative

reproduced verdict per Block F.

• CRITICAL STEP-4 TIER-1 HARD GATE (Fail-Closed):

• Tiers 1 and 2 are internal stepping stones only. You MUST NEVER

record repro_status = "reproduced" or "statically_confirmed" based on

a Tier-1 or Tier-2 execution.

• If a crash can ONLY be achieved by compiling a direct-call harness

that feeds a private/static function or bypasses the public API (Step 4),

and the payload cannot be escalated to trigger through Tier 3 (the public

API / sandboxed service), you MUST TERMINATE AND RECORD

repro_status = "failed_to_reproduce" with details citing

"Internal Invariant Protection".

• *Attempt Cap Accounting & Local Retries:*

• Sub-Tier-3 Stepping-Stone Sub-Budget: Internal Tier-1 and Tier-2

trial runs are bounded local execution steps (max 3 trial executions per

conversation) and DO NOT increment the absolute per-finding attempt

counter in state_root/workspace/archive/.repro_attempts.json.

• Absolute Attempt Cap Counting: Only Tier-3 full sandboxed service

executions (or full end-to-end reproducer runs) increment the absolute

attempt counter toward the hard ceiling of 6 (Section 6).

• Intra-Conversation Retries: When a tier fails, inspect logs, adjust

payload parameters, fix harness setup, and retry up to 2-3 times within

the active conversation before reporting back to the orchestrator.

Step 3a: Variant Hunting (re-attack only, MANDATORY)

When invoked with --reattack, you MUST author and execute **N ≥ 3

boundary-mutated variant inputs** in addition to re-running the original PoC.

The schema (schema.json) literally calls this the "variant-hunting re-attack"

— merely re-running the original PoC is insufficient. Over-narrow patches that

guard the exact PoC bytes are the dominant auto-repair failure mode; variant

hunting is the zero-infra-cost defense against them.

Legacy findings: Findings with absent reattack_variants (created before

this rule existed) are re-verified under the ≥3 variant requirement on their

first --reattack pass. This is intentional — legacy VERIFIED_SECURE verdicts

are upgraded to the stricter gate — but expect one-time churn across archived

findings on the first pass that runs variant hunting.

What to generate (bug-class-aware):

• **Memory-safety bugs (buffer overflow, OOB read/write, UAF, integer

overflow):** Author at least 3 of:

• Off-by-one: len = bound, len = bound + 1, len = bound - 1.
• Size mutations: len ± 1, len = 0, len = SIZE_MAX, sign flips on

signed lengths.

• Alternate paths to the same sink: If the vulnerable sink is reached via

multiple call sites, author a variant reaching it through a different path

(e.g., different API endpoint, file format variant, or protocol command).

• Type confusion / width mismatch: Exploit a different type path to the

same sink.

• Non-memory-safety bugs (logic, auth, injection, SSRF, path traversal):

Hunt for actual variants of the same class:

• Alternate endpoints/parameters: Try /api/v2/echo when PoC targets

/api/v1/echo, or alternate parameter names.

• Equivalent payloads: ..%2fetc%2fpasswd, ..\\..\\, URL-encoding,

double encoding, unicode normalization variants.

• Auth boundary variants: Different roles, empty/null tokens, alternate

privilege-escalation paths.

• Injection variants: '; EXEC--, UNION SELECT, blind variants,

alternate injection points.

• Parallel variant generation (if subagents available): You SHOULD spawn

subagents to author and test variants in parallel. Each subagent gets one

mutation strategy, writes its variant PoC, and reports whether it triggered.

Aggregate all results before setting reattack_status. If subagents are

unavailable, do them sequentially.

Execution: Write each variant as a separate script in

state_root/workspace/reproducers/ (e.g., reattack_variant_[uuid]_[N].py).

Execute each against the patched shadow (--target_root) using the same

isolation constraints as Step 3. Record each variant in the reattack_variants

array using EXACTLY the schema keys

{"description": "...", "triggered": true/false} (Step 6).

Verdict rule: reattack_status = "failed_to_bypass" requires a non-empty

reattack_variants array containing ≥ 3 valid variant inputs that ALL failed to

trigger the bug on the patched shadow. An empty or short set makes "all variants

failed" vacuously true — this is FORBIDDEN: if fewer than 3 meaningful variants

can be constructed after genuine effort, cap at VERIFICATION_INCOMPLETE

(history note insufficient_variants), NEVER failed_to_bypass. The [] case

may ONLY coexist with a non-failed_to_bypass status (e.g., C5 baseline failure

that halted before variant hunting).

Variant validity guardrail: A variant counts as a bypass (triggered = true

causing bypassed_patch) ONLY if it satisfies BOTH:

• Same vulnerability class: The variant reproduces the original bug class —

same sink function / sanitizer signature / crash type. A junk mutant (e.g.,

len=SIZE_MAX causing an unrelated OOM, an alternate endpoint 404-ing then

erroring, or a completely different crash) does NOT count as a bypass.

Discard it (set triggered = false with a description noting it was invalid)

and continue. For non-memory bugs (logic, auth, injection), compare the same

sink function or behavior (e.g., same unauthorized action succeeds, same

injection executes, reaches the same sink function) — NOT a ±line window

against a pre-patch line number (patches shift lines, and non-memory bugs

often have no precise sink line). If you cannot positively confirm a

triggering variant is a DIFFERENT bug, count it as a bypass (fail-closed:

prefer a false VERIFICATION_FAILED over a false VERIFIED_SECURE).

• Valid input per Step 4: The variant must be a valid exercise of the

public API or internal invariants — it must not rely on artificial harness

tricks (e.g., private-function direct calls with custom-allocated buffers)

that bypass the library's execution invariants.

Ideally, confirm each triggering variant still triggers on the unpatched

shadow (same baseline the original PoC ran against) to prove it exercises the

original bug rather than an artifact. If the unpatched shadow is unavailable

(e.g., snapshot mismatch), require the same-sink evidence (same sanitizer frame,

same crash address, same logic failure) as corroboration.

If ANY valid variant triggers the bug, set reattack_status = "bypassed_patch"

and apply INV-1 (downgrade VERIFIED_SECURE → VERIFICATION_FAILED).

• Reproduction Status Classification:

• reproduced: The PoC successfully triggered the vulnerability.
• failed_to_reproduce: The PoC was executed but did not trigger the

vulnerability.

• statically_confirmed: Reproduction was impossible due to environmental

constraints (e.g., missing hardware emulators, unavailable external

services) but the flaw is statically obvious (e.g., hardcoded credentials).

This is strongly discouraged and should only be used as a last resort.

• not_attempted: The reproduction stage was skipped entirely (e.g., due

to infrastructure setup failure, timeouts, or explicit skip configuration).

• Strict Public-API & Internal Invariant Constraints:

• Your crash reproducer should interact with the codebase through

public-facing APIs wherever possible, or strictly respect the library's

global execution invariants (such as allocator padding) to avoid generating

artificial, non-viable crashes.

• Do not declare a finding as "reproduced" if the crash can only be achieved

by compiling a direct-call harness that feeds a private/static function a

custom-allocated buffer (e.g., malloc(15)) that bypasses the library's

guaranteed allocator wrappers (e.g., png_malloc(rowbytes + 48)).

• If a crash cannot be triggered through the public API or with standard

allocation padding, classify the finding as "failed_to_reproduce" due to

"Internal Invariant Protection."

• Functional & Crash-Aware Validation: Analyze the output such as stdout,

stderr, and exit codes to classify reproduction success depending on the bug

class:

Before classifying ANY negative outcome (failed_to_reproduce, or in

--reattack mode failed_to_bypass), apply this gate:

   REACHED-SINK EVIDENCE GATE (mechanical):
   Each reproducer produces REACHED-SINK EVIDENCE via ONE channel, recorded in repro_hints:
      (a) script/source harness -> write the exact bytes MANTIS_REACHED_ENTRYPOINT to a
          sidecar file $SENTINEL_FILE and flush+fsync (or unbuffered write) BEFORE
          invoking the sink. (A file survives a crash that truncates buffered stdout.)
      (b) binary / firmware / raw-payload -> reached-sink evidence is a captured
          crash backtrace or sanitizer frame (ASan/UBSan/MSan/TSan) that
          explicitly names the target sink function (target-produced tracing). A
          marker written by a wrapper you author BEFORE invoking the target is
          SETUP EVIDENCE ONLY: it proves "launch attempted," not "sink reached,"
          and does NOT qualify as reached-sink evidence. If no in-path marker
          (channel a) and no target-produced backtrace/sanitizer trace
          (channel b) is achievable, the sink is unreached.
    EVIDENCE PRESENT (reached-sink) = (channel a) sidecar file contains
      MANTIS_REACHED_ENTRYPOINT written in-path, OR (channel b) target-produced
      backtrace/sanitizer output names the sink. A wrapper pre-launch marker alone is
      NOT evidence present.
   EVIDENCE ABSENT includes: any compiler/build nonzero exit; exit 127 (command not
     found); exit 2 with a "No such file" message.
   DECISION GATE (gate the DECISION, not specific verdict strings):
     - Record repro_status = reproduced OR statically_confirmed ONLY if EVIDENCE is
       PRESENT. If ABSENT -> repro_status = not_attempted (retry-eligible), STOP.
     - In patch verification, EVIDENCE is required on the UNPATCHED baseline (Block G),
       NOT on the post-patch attack run (a correct patch legitimately stops the input
       before the sink).
     - If NO evidence channel is achievable for this target, downgrade to
       not_attempted / VERIFICATION_INCOMPLETE. NEVER synthesize the marker.

Concretely: if the run produced NO reached-sink evidence (build/setup error,

exit 127, "No such file", or the sink was never reached), record

repro_status = not_attempted (retry-eligible) — NEVER

failed_to_reproduce; and in --reattack mode leave reattack_status UNSET

with a history note "setup_failed" — NEVER failed_to_bypass. Only classify

a negative when the harness provably reached the vulnerable entrypoint and

the bug did not fire.

HALT ceiling (3-state rule): If active_snapshot is present in state but

snapshot_pinned is false (HALT mode — the tree raced or could not be

pinned), you MUST NOT record failed_to_reproduce or failed_to_bypass at

all. In HALT, the code may have drifted and a negative reproduction result

cannot be trusted as authoritative. Instead, record

repro_status = not_attempted (retry-eligible) and, in --reattack mode,

leave reattack_status UNSET with a history note "HALT mode: snapshot

unpinned, negative result suppressed". This mirrors the authoritative-verdict

prohibition that applies to all stages in HALT. (In MODE-OFF — no

active_snapshot — classify negatives normally as today.)

• Logic & Authorization Bugs: A successful reproducer is a functional

unit test or script that explicitly demonstrates the logic failure (e.g.,

an unauthorized request returns 200 OK, or a test script successfully

bypasses validation and exits with 0).

• Memory Safety & Binary Crashes: If the sandbox execution produces a

crash signal or sanitizer trace in stdout/stderr, mark the reproduction as

"reproduced". **Scan stdout/stderr for sanitizer signatures regardless of

exit code** — UBSan defaults to recover-mode (exit 0), so an exit-0 run can

still contain a valid UBSan trace proving the bug fired. Check for:

• AddressSanitizer (ASan) error outputs (e.g. ERROR: AddressSanitizer).
• UndefinedBehaviorSanitizer (UBSan) runtime reports (e.g.

runtime error:, SUMMARY: UndefinedBehaviorSanitizer).

• MemorySanitizer (MSan) error outputs (e.g. WARNING: MemorySanitizer).

**MSan evidence is valid ONLY if repro_hints records that a

fully-instrumented MSan build was available** (Step 3 caveat). A naive

-fsanitize=memory build without full instrumentation produces bogus

traces — do NOT classify these as "reproduced".

• ThreadSanitizer (TSan) race reports (e.g. WARNING: ThreadSanitizer).
• Segmentation faults (SIGSEGV, exit code 139).
• Abort signals (SIGABRT, exit code 134).
• Crash or core dumps.

• Token-Optimized File Updates: To minimize LLM output tokens, **do not

re-emit or manually rewrite the entire JSON object in your output.** Instead,

use in-place editing tools (like a short script in your preferred language,

or jq) to programmatically append the new fields to the existing

state_root/workspace/findings/<id>.json file.

Additionally, you must Update the Reproduction Attempt Cache to help the

planner track attempts efficiently:

• Maintain a JSON cache file at

state_root/workspace/archive/.repro_attempts.json. Ensure the parent

directory state_root/workspace/archive/ exists (e.g.,

mkdir -p state_root/workspace/archive/) before creating, reading, or

locking the cache file.

• Key the cache by a stable identifier that persists across loop runs even if

UUIDs are regenerated. If the finding has a signature field, use it

directly as the cache key (it is already a deterministic content-identity

hash). If signature is absent, fall back to a computed stable key using

the finding's normalized title and its primary file path:

stable_key = normalized_title + "@" + primary_file_path.

• Compute normalized_title by converting the title to lowercase and

removing all non-alphanumeric characters.

• Compute primary_file_path by taking the first entry in code_paths and

stripping any line number suffixes (e.g., converting src/auth.c:120 to

src/auth.c).

• Snapshot-aware cap (value shape {count, last_snapshot}): store each

cache value as an object {count, last_snapshot}. When reading a value

V: if V is a bare integer, treat count=V and last_snapshot=UNKNOWN; if V

is an object, use V.count / V.last_snapshot (default UNKNOWN). Before

incrementing, run Block B comparing the finding's snapshot to the current

SNAPSHOT_ID: reset count=0 ONLY when Block B is NOT_MATCHED *because

the two snapshots are present and actually differ* (a genuine code change

earns a fresh budget). Do NOT reset on UNKNOWN (absent/pass_/unpinned)

— that would make no-VCS targets retry forever. Additionally keep an

absolute per-finding-id attempt counter that is NEVER reset, and stop

retrying once it reaches a hard ceiling (e.g. 6) regardless of snapshot

changes.

• To prevent race conditions during concurrent executions (including locking

bypasses caused by atomic file replacement) and protect lockless readers:

• Use a separate dedicated lock file

state_root/workspace/archive/.repro_attempts.lock which is never

deleted or replaced.

• Perform updates atomically using Python's fcntl.flock on this lock

file:

• Open the lock file state_root/workspace/archive/.repro_attempts.lock

(creating it if missing) and acquire an exclusive lock (fcntl.flock

with fcntl.LOCK_EX) inside a context manager (with statement).

• Read the current contents of the cache file

state_root/workspace/archive/.repro_attempts.json (treating it as {}

if missing or empty).

• Increment the count field of this finding's cache-key entry — keyed by

signature if present, else stable_key, the SAME key selection defined

above (the {count, last_snapshot} object) — by 1 ONLY for Tier-3 (full

end-to-end sandboxed service) executions. Internal Tier-1 and Tier-2

stepping-stone trials MUST NOT increment count (they are governed by

the sub-budget rule in Step 3).

• Write the updated JSON to a temporary file in the same directory (e.g.,

state_root/workspace/archive/.repro_attempts.json.tmp).

• Atomically replace the target cache file with the temporary file (e.g.,

os.replace in Python) to ensure readers never see a truncated or

incomplete file.

• Close the lock file descriptor to release the lock (automatically handled

by exiting the with context manager).

Depending on whether the --reattack flag is provided:

• If run normally (no --reattack flag): You must append or update the

following on the existing object:

• "repro_status" ("reproduced", "statically_confirmed",

"not_attempted", or "failed_to_reproduce").

• "repro_file_path"

• "run_command"

• "repro_output"

• "repro_snapshot_id": the current SNAPSHOT_ID this run executed against.

• "repro_hints": Record compilation and sandbox execution telemetry

(e.g., sanitizers_used: ASan+UBSan, assertions_disabled: true,

build_profile: release) to provide empirical execution evidence for

/mantis-critic.

• If reproduction succeeds (repro_status is evaluated as "reproduced"

or "statically_confirmed") and the finding's current "status" is

"PROVISIONALLY_VALID": BEFORE upgrading, scan the finding's

triage_checklist (if present). If ANY entry has outcome == "UNKNOWN"

(or passes == false), do NOT upgrade: leave status as

"PROVISIONALLY_VALID", still set repro_status to the success value

(reproduction DID succeed), and append a history note

upgrade-to-VALID-blocked: triage_checklist has UNKNOWN entries (re-review required).

This avoids violating the schema's VALID ⇒ no UNKNOWN allOf gate

(schema.json lines 471-507), which forbids UNKNOWN/passes:false on

any VALID non-chain finding's triage_checklist. Reproduce does NOT

touch triage_checklist entries (the checklist is review's artifact;

only review may resolve UNKNOWN entries). If triage_checklist is

absent (no reviewer history entry, e.g. a legacy finding), or NO entry

is UNKNOWN/passes:false, you must update "status" to "VALID".

• An entry to the "history" array:

     {
       "stage": "reproduce",
       "action": "reproduced",
       "details": "Reproduction status evaluated as [reproduced/failed_to_reproduce] using command: [run_command]",
       "pass_number": <current_pass_number>,
       "timestamp": "<current_iso8601_timestamp>"
     }

• If run with --reattack: You must append or update the following on

the existing object (do not touch repro_* or status):

• "reattack_status" ("bypassed_patch", "failed_to_bypass",

"inconclusive_baseline_changed").

• "bypassed_patch": The PoC bypassed the patch and triggered the bug. If

patch_status is VERIFIED_SECURE, atomically set

patch_status = "VERIFICATION_FAILED" in the same write (a bypass

defeats the patch). This is an explicit exception to "do not touch

status" and applies to both the C5 step-3 path and same-snapshot runs.

• "failed_to_bypass": The PoC was run but failed to bypass the patch. (No

downgrade needed — failed_to_bypass is the value the allOf gate

requires for VERIFIED_SECURE.)

• "inconclusive_baseline_changed": The unpatched baseline was re-run (see

C5 below) and the bug NO LONGER TRIGGERS on the current unpatched code.

Do NOT claim failed_to_bypass. If patch_status is VERIFIED_SECURE,

atomically set patch_status = "VERIFICATION_INCOMPLETE" in the same

write.

• INV-1 (single source — the schema's VERIFIED_SECURE allOf gate):

VERIFIED_SECURE => reattack_status must be failed_to_bypass, and

failed_to_bypass requires a non-empty reattack_variants array with ≥

3 valid variant inputs (Step 3a) that ALL failed to trigger the bug on

the patched shadow. An empty or short variant set makes "all failed"

vacuously true — this is FORBIDDEN: cap at VERIFICATION_INCOMPLETE

instead. Any other outcome MUST atomically downgrade patch_status in

the same write — VERIFICATION_INCOMPLETE for

inconclusive_baseline_changed or an insufficient variant set,

VERIFICATION_FAILED for bypassed_patch. Never persist

VERIFIED_SECURE alongside a non-failed_to_bypass reattack_status.

• Snapshot-mismatch -> governed by C5 below: When

reattack_snapshot_id != repro_snapshot_id, do NOT independently

record a verdict — follow C5, which re-establishes the unpatched baseline

on the CURRENT snapshot. If C5 cannot run, fall back to leaving

reattack_status UNSET with a SNAPSHOT_MISMATCH history note.

• C5 — Unpatched-baseline re-run (Phase 2): Before running the attack

on the patched build, if reattack_snapshot_id != repro_snapshot_id (a

genuine snapshot change, NOT a HALT-mode live: tree), FIRST

re-establish the unpatched baseline:

• Run the reproducer against a FRESH UNPATCHED copy of the current

snapshot — a fresh mktemp -d copy of active_snapshot.root, NOT the

patched shadow. (This is the same unpatched baseline that

mantis-patch Block G step 1 runs; reproduce has no Block G.)

• If the unpatched baseline does NOT trigger (evidence absent per Block

F): set reattack_status = "inconclusive_baseline_changed", do NOT

proceed with the attack, do NOT claim failed_to_bypass. You MUST

still populate reattack_file_path, reattack_run_command, and

reattack_output with the baseline re-run's details. Apply INV-1

(downgrade VERIFIED_SECURE -> VERIFICATION_INCOMPLETE).

• If the unpatched baseline DOES trigger: proceed with the attack

(bypassed_patch or failed_to_bypass). Apply INV-1 (downgrade

VERIFIED_SECURE -> VERIFICATION_FAILED if bypassed_patch).

• Sanitizer consistency: the baseline re-run, attack, and re-attack

MUST use the same sanitizer flags that made the original baseline

trigger. If the baseline triggered via UBSan (exit 0), recompiling the

attack without UBSan would mask the bug and produce a false negative

(INV-2) or false failed_to_bypass (INV-1).

• HALT-mode guardrail: Detect HALT by the PASS's snapshot id in state

(active_snapshot.snapshot_id starts with live:), NOT the

--snapshot_pinned=false argument (that's only the sentinel-exemption

for the patched shadow). In HALT, skip C5 and follow the existing HALT

ceiling (leave reattack_status UNSET with the HALT note).

• This does NOT change the VERIFIED_SECURE gate — Block G still requires

the unpatched baseline to trigger.

• "reattack_file_path"

• "reattack_run_command"

• "reattack_output"

• "reattack_variants": An array of objects, one per variant input

attempted during Step 3a. Each object MUST contain EXACTLY the two

required keys "description" (string) and "triggered" (boolean). Do

NOT invent free-form or custom keys (such as input, result,

bypassed, label, name):

       "reattack_variants": [
         {
           "description": "off-by-one: len=bound+1",
           "triggered": false
         },
         {
           "description": "alternate path via /api/v2/echo",
           "triggered": false
         },
         {
           "description": "boundary mutation: max INT_MAX",
           "triggered": false
         }
       ]

• "description": What the variant does (e.g.,

"off-by-one: len=bound+1", "alternate path via /api/v2/echo").

• "triggered": boolean — whether this variant triggered the original

vulnerability class on the patched shadow (per the variant validity

guardrail in Step 3a). Invalid/junk mutants that produced an unrelated

crash or error are recorded as triggered = false with a description

noting invalidity. reattack_variants MUST be non-empty (≥ 3 entries)

for reattack_status = "failed_to_bypass". An empty array [] may

ONLY coexist with a non-failed_to_bypass status (e.g., C5 baseline

failure that halted before variant hunting, or

inconclusive_baseline_changed).

• "reattack_snapshot_id": the current SNAPSHOT_ID this run executed

against.

• An entry to the "history" array:

     {
       "stage": "reattack",
       "action": "reproduced",
       "details": "Re-attack status evaluated as [bypassed_patch/failed_to_bypass] using command: [reattack_run_command]",
       "pass_number": <current_pass_number>,
       "timestamp": "<current_iso8601_timestamp>"
     }

• Criticism of Reproduction Validity: To ensure the reproduction is a valid

example of reproducing the reported vulnerability, have a subagent with a

fresh context window review and criticize the generated PoC. Seek genuine

criticism to ensure false reports are never surfaced later.

Variant criticism (re-attack only): When --reattack is specified, the

critic subagent MUST also verify that the variant inputs (Step 3a) are

genuinely diverse — not trivially identical mutations (e.g., changing a

comment while keeping the same payload). If the critic finds the variants are

not meaningfully diverse, record a history note

variant_diversity_insufficient and re-author variants before finalizing

reattack_status.

When complete, notify the user.