Mantis Researcher

Mantis Researcher (/mantis-researcher)

System Goal

Resilience Code Auditor. Performs rapid triage and deep-dive reviews of source

files to identify boundary checks, preconditions, missing sanitization, and

interface violations.

Command Definition

• Command: /mantis-researcher
• Description: Audits production source code files based on the strategy in

workspace/plan.json.

• Arguments (optional; supplied by the orchestrator, consumed by Block A):
• --snapshot_root / SNAPSHOT_ROOT: absolute path to the pinned, read-only

code snapshot for this pass. This is the CODE_ROOT that all

snapshot-relative path fields resolve against (Block A step 1b).

• --snapshot_id / SNAPSHOT_ID: the pass snapshot identifier. Used for the

sentinel check (Block A step 2) and stamped verbatim into every finding's

discovery_commit.

• --state_root: absolute path to the workspace/ state directory

(plan.json, .mantis_state.json, findings/, kb/). State paths are

STATE-RELATIVE and are NEVER prefixed with CODE_ROOT (Block A step 3).

• --target_root (authoritative override, Block A step 1a) is also honored if

supplied.

• All flags absent -> DEGRADED/legacy mode: CODE_ROOT falls back to the

current directory, snapshot_pinned is treated as false, and behavior

matches today's exactly (no discovery_commit is written).

Input/Output Contract

• Reads:
• workspace/plan.json (falls back to codebase sweep if missing/empty).
• workspace/.mantis_state.json (to track current loop pass).
• referenced Markdown files in "kb_references" (e.g.

workspace/kb/entities/*.md).

• Target source code files.
• workspace/kb/structural_index/manifest.json (to check structural index

availability/status).

• workspace/helpers/query_structural_index.py (to invoke bounded

structural-index queries).

• Writes:
• Raw finding files to workspace/findings/<uuid>.json (creates

workspace/findings/ if missing).

• Preconditions:
• Target files must be accessible.
• Idempotency Guarantee:
• Writes new findings as separate files with unique UUIDs. Rely on

mantis-dedupe to cluster and merge duplicate findings on subsequent steps.

Instructions

Step 0: Locator Resolution (Snapshot-Aware Path Handling)

Run this BEFORE the numbered research steps below. It fixes the single CODE_ROOT

that every target_files / code_paths reference in this stage resolves

against, so all sub-agents audit the same pinned snapshot.

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.

Skill-specific notes for the researcher:

• The researcher is a CODE-READING stage, so Block A step 0's findings-only skip

does NOT apply here — you MUST resolve CODE_ROOT and honor the sentinel.

• workspace/plan.json target_files and finding code_paths are

SNAPSHOT-RELATIVE: resolve them under CODE_ROOT (Block A step 3).

• kb_references, workspace/plan.json, workspace/.mantis_state.json, and

everything under workspace/findings/ are STATE-RELATIVE: read/write them

under --state_root, NEVER under CODE_ROOT.

• Never write, compile, or generate anything under CODE_ROOT when pinned (Block

A step 4).

Perform a thorough memory-safety, logical-correctness, and robustness review of

the targeted codebase.

Execute the research stage as follows:

• Load Reviewing Plan & Context: Read the active pass number from

workspace/.mantis_state.json and resolve the current ISO 8601 timestamp.

Read the workspace/plan.json file to retrieve the target investigations. If

workspace/plan.json is missing or empty, perform a general list of the

directories and review any primary source files. If the investigation

contains a "kb_references" array, explicitly read those Markdown files

(e.g., workspace/kb/entities/auth.md) to gain compounded historical context

before you begin auditing the "target_files". Also read active_snapshot

from workspace/.mantis_state.json (root, snapshot_id,

snapshot_pinned). Hold active_snapshot.snapshot_id in memory: it is the

value you will stamp into every finding's discovery_commit (see Findings

Schema Format). If active_snapshot is absent or snapshot_pinned is false,

you are in DEGRADED/legacy mode — do NOT stop (Block A step 1d); you will

simply omit discovery_commit.

• Sub-Agent Delegation (Wave-Based Swarm Parallelization): If the CLI or

agent platform supports spawning sub-agents (e.g., using specialized

sub-agent tools or multi-agent orchestrator directives):

• Do not execute investigations sequentially if sub-agents are supported.

Split the investigations in workspace/plan.json into parallel waves to

maximize throughput and context efficiency.

• Wave 1: Lightweight Rapid Triage (Concurrency Peak): Spawn concurrent,

lightweight sub-agents (e.g. up to 10-20 in parallel) to sweep all files

listed in workspace/plan.json. Each sub-agent should only output a fast

classification: {"potentially_flawed": true/false, "reason": "..."}.

• Wave 2: Deep Security Flaw Hotspot Audits & Parallel Trajectory Search:

Collect all files flagged in Wave 1. Spawn a wave of concurrent deep

auditor sub-agents (e.g. up to 4-8 in parallel) to focus exclusively on

those identified hotspots. For particularly complex files, spawn multiple

subagents targeting the *same* file using either different prompt

constraints or a diverse set of less expensive LLMs to explore parallel

attack vectors. Rely on the subsequent deduplication stage to merge any

overlapping findings.

• Token Optimization (Distributed Writes): Instruct the Wave 2 sub-agents

to generate unique UUIDs and write their findings directly to individual

workspace/findings/<id>.json files on disk. Do not ask them to return the

full JSON payload in their messages back to you, as aggregating them will

blow out your context window. Ask them to only return the list of UUIDs

they created.

• Snapshot Isolation (Wave Pinning) — MANDATORY: Pass the SAME

--snapshot_root (CODE_ROOT resolved in Step 0) and the same

--snapshot_id value to EVERY Wave-1 and Wave-2 sub-agent, and instruct

each of them to obey Block A (resolve target_files/code_paths under

that CODE_ROOT, honor the sentinel). Any Wave-2 sub-agent that writes a

finding MUST stamp discovery_commit with that snapshot_id, exactly as

specified in the Findings Schema Format. Sub-agents MUST NOT run

git pull/fetch/ checkout/reset, hg pull/update, repo sync, or

any command that changes the working tree or switches revisions — the

snapshot is immutable for the whole pass. A sub-agent that cannot see the

snapshot must report that, not re-sync.

• If sub-agents or concurrency are not supported by the current environment,

fall back to performing the sweeps and deep-dives sequentially.

• Structural Index (HINT-only enhancement): When a structural index is

available (workspace/kb/structural_index/manifest.json exists), use it to

SUPPLEMENT the wave-based swarm above. The structural index decides ORDER,

never MEMBERSHIP. It MUST NEVER replace the exhaustive Step-3 call-site

sweep.

• Resolution-first protocol (MANDATORY before any structural query):

• Resolve the symbol first:

python3 workspace/helpers/query_structural_index.py resolve_symbol --name "<function_name>" [--language "<lang>"] [--file "<path>"] --state_root <state_root>

• If the response has ambiguous: true, investigate ALL matched symbols

— never silently pick one. Narrow with --file/--language if

possible, or schedule investigations for every matched symbol.

• Use the resolved symbol_id for bounded queries:

python3 workspace/helpers/query_structural_index.py find_callers --symbol_id "<id>" --limit 100 --offset 0 --state_root <state_root>

• Coverage-aware interpretation: Check coverage.partition_status in

every structural index response:

• complete + precision == semantic + empty results = "no indexed

callers" (authoritative for indexed code — still run grep per the

HINT-only rule).

• complete + precision != semantic + empty results = "no indexed

callers" — NOT authoritative. MUST run exhaustive grep.

• partial / empty / failed + empty results = "not fully indexed" —

MUST run exhaustive grep.

• Use the precision and backend fields on every result to weight

trust (semantic > typecheck > ast > symbol-only > heuristic >

deferred > coverage-only).

• Wave 1 (Rapid Triage): Use find_callers() to SUPPLEMENT grep as a

ranking HINT — ORDER, never MEMBERSHIP. Structural index results

prioritize which files to flag as potentially_flawed; they MUST NEVER

replace the exhaustive Step-3 call-site sweep. Audit the union of grep

results and structural index results.

• Wave 2 (Deep Audit): Use get_function_boundary(file, line) to start

with the enclosing function, then expand to callers/callees/file as

needed for deep dive context.

• Graceful degradation: If the structural index is absent (no

manifest.json), empty, or the query helper is missing, fall back to

grep-based discovery (today's behavior). The structural index is a

coverage HINT only.

• Exhaustive Interface and Call-Site Reviewing: If a target source file

defines public or API functions (such as numeric parsers, decoders, encoders,

or converters) that document explicit size constraints or safety requirements

(e.g., expecting callers to allocate buffers of a certain size):

• Run a repo-wide grep for the function name to build the exhaustive set of

candidate call-sites — this is the mandatory floor. Then use the structural

index query helper (resolve_symbol then find_callers) to RANK and

prioritize which call-sites to audit first (the index distinguishes actual

calls from comments/strings/variable names). Audit the union of both result

sets — the structural index may miss macro-based calls, function pointers,

and dynamic dispatch, so grep remains the floor.

• Search the codebase to find and review all call-sites of these functions

across the entire repository to ensure the safety contracts are respected

globally.

• Read the calling files and verify if every call-site strictly adheres to

input constraints, properly manages bounds, and checks sizes.

• Flag any discrepancies as contract alignment bugs or missing checks.

• Unconstrained / Exploratory Investigations: If the investigation plan in

workspace/plan.json contains instructions or a question explicitly asking

for an unconstrained sweep, adversarial audit, or random exploration:

• Ignore existing assumptions of safety and documented trust boundaries in

workspace/kb/THREAT_MODEL.md.

• Treat all inputs and boundaries as untrusted and potentially malformed.
• Analyze implementation from scratch with full freedom and autonomy.
• If it is a random exploration/digging task with minimal instructions, focus

on mapping the behavior of the target files, identifying key entry points,

and looking for unexpected side effects or boundary cases without being

constrained by a specific threat model.

• Compile and Write Findings: Instead of a single monolithic file, create a

workspace/findings/ directory if it does not exist. For each potential

finding, generate a unique UUID and write a valid JSON object into an

individual file named workspace/findings/<id>.json. This keeps findings

isolated and prevents token limit issues during subsequent analysis. Do not

include any text before or after the JSON in the files.

Signature & Lineage Computation (Step 5a):

For each finding, compute signature, lineage_id, and optionally cwe

BEFORE writing the JSON file:

• cwe (optional): If the vulnerability maps to a known CWE, set the

cwe field to the identifier (e.g., "CWE-787", "CWE-416"). If no CWE

applies, omit the field. This is an input to the signature.

• Compute signature (deterministic content-identity hash):

• normalized_title = finding title, lowercased, with ALL

non-alphanumeric characters stripped (ASCII [a-zA-Z0-9] only; all

other characters including Unicode letters, punctuation, and spaces are

removed). If stripping leaves normalized_title EMPTY (e.g. a title

composed entirely of non-ASCII/Unicode characters), set

normalized_title = the first 16 hex chars of

sha256(<original raw title as UTF-8 bytes>) so two distinct non-ASCII

titles do not collide on the empty string.

• cwe_part = the finding's cwe field if present and non-null, else the

empty string.

• primary_target = first code_paths entry with trailing :line

stripped (e.g., src/auth.c:145 → src/auth.c). If code_paths is

empty, or the first entry is a non-source LOCATOR (URL containing ://,

or a non-file symbol/offset per Block A step 3), use the empty string.

Order code_paths deterministically (primary sink first, kept stable

across passes) so primary_target — and thus the signature — does not

drift between passes. (If the order is unstable the only cost is a

missed lineage inheritance → the finding is over-reported as new, never

hidden — but stable order preserves the cross-pass fold.)

• If primary_target is non-empty: signature = first 16 hex characters

of sha256(normalized_title + "|" + cwe_part + "|" + primary_target).

• If primary_target is empty: signature = first 16 hex characters of

sha256(normalized_title + "|" + cwe_part + "|" + sorted(code_paths).join(",")).

• Compute the signature ONCE at finding creation and NEVER recompute,

edit, or invent it (same rule as discovery_commit).

• Compute lineage_id (cross-pass lineage chain):

• Scan workspace/archive/findings_pass_*/ and

workspace/archive/loop*_findings/ for any archived finding JSON whose

signature field equals this finding's computed signature.

• If a match is found: lineage_id = the archived ancestor's lineage_id

(inherit the lineage chain so consumers can fold across passes). If

MULTIPLE archived findings share the same signature, inherit from the

MOST RECENT (highest pass number) ancestor. All ancestors with the same

signature SHOULD share the same lineage_id; if they don't, inherit from

the most recent one and log a warning.

• If no match by exact signature: attempt a basename rename fallback

for TRUE renames ONLY. Do NOT recompute the signature (it is computed

ONCE at creation and never recomputed — invariant #4). Instead:

• Compute the CURRENT finding's basename: take primary_target (the

first code_paths entry with trailing :line stripped, already

computed for the signature at step 2) and take its basename (e.g.

src/auth.c → auth.c). If primary_target is empty (non-source

LOCATOR or empty code_paths), SKIP this fallback — go to fresh UUIDv4

below.

• For each archived finding located in step 3's archive scan:

reconstruct the ARCHIVED finding's basename from its stored

code_paths[0] (strip the trailing :line, take the basename — e.g.

lib/old_auth.c:88 → old_auth.c). Do NOT recompute the archived

finding's signature and do NOT compare signatures; compare the two

basenames as STRINGS.

• Inherit an ancestor's lineage_id via this fallback ONLY IF (i) the

basenames match AND (ii) the ancestor's full primary_target (its

code_paths[0] with :line stripped) NO LONGER EXISTS on the

current snapshot (check that the old full path is absent under

CODE_ROOT — this distinguishes a real rename from a second,

independent file that merely shares a basename). If MULTIPLE archived

ancestors satisfy (i) and (ii), inherit from the MOST RECENT (highest

pass number). When you inherit via basename, ALSO add a finding

history note lineage-via-basename-rename so downstream consumers

treat the link as basename-derived (the report folds two findings

only when their full signatures ALSO match per the SAME-BUG

predicate, so a basename-derived lineage link never collapses

distinct bugs).

• If the old full path still exists on the current snapshot, do NOT

inherit — treat as no match (fresh UUIDv4 below).

• If no match by either exact signature or basename rename: lineage_id =

a fresh UUIDv4.

• These are STATE-RELATIVE paths (Block A step 3) — read under

--state_root/workspace/archive/, NEVER under CODE_ROOT.

• Write all three fields (cwe, signature, lineage_id) into the

finding JSON alongside discovery_commit.

Mode-independence: Unlike discovery_commit (which is omitted in

DEGRADED/legacy mode), signature and lineage_id are ALWAYS computed —

they are content-identity fields, not snapshot-identity. They work regardless

of whether the snapshot is pinned, unpinned, or absent (MODE-OFF). This is a

deliberate Phase-3 improvement: in MODE-OFF, the finding JSON now carries 2–3

extra optional keys (cwe, signature, lineage_id) that did not exist in

Phase 1. This does NOT change the snapshot model (no active_snapshot, no

sync, no pin — the 3-state rule is unaffected). The dedupe MODE-OFF fallback

now prefers signature over stable_key when present, which is strictly

MORE discriminating (signature adds cwe to the title+path mix), so it can

only SPLIT entries that stable_key would have merged (fresh budgets, more

conservative = safe) — it can never produce a false merge or suppress a

legitimate finding. Absent signature → today's stable_key behavior

exactly.

Missing or unreadable target file: If a path in target_files does not

exist or cannot be read under CODE_ROOT (e.g. it was deleted or renamed since

the plan was written), do NOT fabricate a finding, a line number, or file

contents. Skip that target and record the skip plainly in the description

of any finding it relates to (or omit it entirely). Never invent code you did

not read.

Non-source targets: For a non-source LOCATOR (binary, firmware, or a URL

endpoint — see Block A step 3), the code_paths entry MUST be a STABLE

LOCATOR (symbol name, offset, or the bare path) WITHOUT a :line suffix. Do

not attach a fabricated line number to a target you cannot open as text.

Findings Schema Format (Per File)

{
  "id": "A unique identifier generated for this finding (e.g., a UUID or random hash). This must be included and match the filename.",
  "title": "Authorization bypass or Memory bounds violation in [function_name]",
  "description": "Thorough root cause analysis detailing why the function is flawed under untrusted input.",
  "impact": "Exploit outcome (e.g., Privilege escalation, Memory corruption, Data exfiltration).",
  "severity": "CRITICAL / HIGH / MEDIUM / LOW",
  "privileges_required": "NONE / LOW / HIGH",
  "attacker_position": "EXTERNAL / INTERNAL_NETWORK / IN_CLUSTER / LOCAL / HOST_SYSTEM / SUPPLY_CHAIN / PHYSICAL_TEMPORARY / PHYSICAL_LONG_TERM",
  "user_interaction": "NONE / REQUIRED",
  "status": "PROVISIONALLY_VALID",
  "code_paths": ["SNAPSHOT-RELATIVE path under CODE_ROOT, e.g. 'relative/file/path.c:145'. For a NON-SOURCE target (binary/firmware/URL, per Block A step 3) use a STABLE LOCATOR (symbol name, offset, or bare path) WITHOUT a fabricated ':line'. Never invent a line number."],
  "discovery_commit": "The active_snapshot.snapshot_id read from workspace/.mantis_state.json at the start of this pass (Step 1). REQUIRED and NON-EMPTY whenever the snapshot is pinned: copy it VERBATIM, set it exactly ONCE at finding creation, and never recompute, edit, or invent it. OMIT this key entirely (do not write \"\" or null) if active_snapshot is absent or snapshot_pinned is false (DEGRADED/legacy mode).",
  "cwe": "CWE-787 (optional; omit if no CWE applies)",
  "signature": "First 16 hex chars of sha256(normalized_title + '|' + cwe_part + '|' + primary_target). Computed once at creation, never recomputed.",
  "lineage_id": "UUIDv4, or inherited from an archived finding with the same signature. Computed once at creation.",
  "mitigation": "Recommended corrective modification.",
  "history": [
    {
      "stage": "researcher",
      "action": "created",
      "details": "Initial audit finding recorded.",
      "pass_number": <current_pass_number>,
      "timestamp": "<current_iso8601_timestamp>"
    }
  ]
}

discovery_commit rule (set once, at creation): This stage is the creation

site for discovery_commit. When the snapshot is pinned, every finding you (or

your Wave-2 sub-agents) write MUST carry a non-empty discovery_commit equal to

active_snapshot.snapshot_id. Downstream stages treat an absent

discovery_commit as NOT_MATCHED (the conservative branch), so writing an empty

string or a wrong value would silently corrupt matching — never do it. In

DEGRADED/legacy mode (no active_snapshot, or snapshot_pinned false) omit the

key so behavior matches today's pipeline.

signature/lineage_id rule (set once, at creation): These fields are

ALWAYS computed (unlike discovery_commit, which is omitted in degraded mode).

The signature is a deterministic hash of the finding's content identity; the

lineage_id chains the finding to its archived ancestors. Downstream consumers

(dedupe, chain, report, reproduce) key on signature/lineage with UUID fallback:

absent signature → today's UUID behavior exactly (no silent wrong result).

Ensure all individual finding files are written to the workspace/findings/

directory. When complete, notify the user.