mantis-dedupe

Deduplicator (/mantis-dedupe)

System Goal

Duplicate Finding Merger. Evaluates lists of raw findings to cluster and
consolidate identical or highly overlapping issues into singular, descriptive
records.

Command Definition

• Command: /mantis-dedupe
• Description: Consolidates raw security findings to eliminate redundant
  reports.
• Arguments (optional; supplied by the orchestrator, consumed by Block A):
  --snapshot_root/--snapshot_id/--state_root. All absent ->
  MODE-OFF/legacy mode (behaves as today; snapshot gating disabled).

Input/Output Contract

• Reads:
  • workspace/findings/ (raw finding JSON files, ignoring .trash/).
  • workspace/archive/findings_pass_*/*.json and
    workspace/archive/loop*_findings/*.json (to skip findings already
    evaluated and triaged in previous passes).
  • workspace/.mantis_state.json (to track current loop pass).
• Writes:
  • Moves duplicate findings to workspace/findings/.trash/ after setting
    "status": "DUPLICATE" and "duplicate_of".
  • Sets "possible_duplicate_of" (soft, non-terminal) on NOT_MATCHED matches
    and stamps "discovery_commit" on current findings that lack it. Reads
    active_snapshot/snapshot_pinned from .mantis_state.json.
  • Appends transaction logs to workspace/.tx_log.jsonl.
  • Generates/executes merging script workspace/helpers/merge_findings.py.
  • Updates primary finding workspace/findings/<primary_id>.json (merges
    fields and history).
• Preconditions:
  • workspace/findings/ must exist and contain finding files.
• Idempotency Guarantee:
  • Cross-references against archived findings in workspace/archive/ to filter
    out any findings already processed in previous passes of this run.
    Snapshot-gated: a resolved archived finding on a differing snapshot is
    flagged as POSSIBLE REGRESSION (never silently filtered). Logs transactions
    to workspace/.tx_log.jsonl to support tracking and potential rollbacks.
    Deterministic merging rules implemented in merge_findings.py.

Instructions

Step 0: Locator Resolution (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).

Notes: workspace/findings/, workspace/archive/, workspace/.tx_log.jsonl,
workspace/helpers/ and .mantis_state.json are STATE-RELATIVE (under
--state_root). Any code snippet you inspect for a finding is SNAPSHOT-RELATIVE
(under CODE_ROOT). Never write under CODE_ROOT.

Review a list of security findings and merge duplicate findings that refer to
the exact same security flaw or adjacent code paths.

Execute your task as follows:

1. Load Raw Findings & Archived Findings Queue:

   • List the contents of the directory and read the files in
     workspace/findings/. If the directory is empty or does not exist, notify
     the user and exit.
   • Important: Ignore hidden files and directories (such as the .trash/
     subdirectory) when listing or processing findings.
   • Locate and load all archived finding JSON files from previous loop passes,
     if they exist, under workspace/archive/findings_pass_*/*.json and
     workspace/archive/loop*_findings/*.json. These files represent
     vulnerabilities that have already been fully evaluated, triaged, and
     potentially patched in previous passes.
   • Important: Do NOT read or deduplicate against
     workspace/historical_learnings.jsonl (VCS history), as we want to catch
     regressions if old bugs were reintroduced.

2. Filter Loop Duplicates (snapshot-gated). First, stamp discovery_commit
   on any CURRENT finding that lacks it, using active_snapshot.snapshot_id
   from .mantis_state.json (skip when unpinned). Then, for each current
   finding that matches an archived finding (by code_paths+title
   similarity), run:

   Signature-based candidate matching (Phase 3) — TIGHTENS, never replaces:
   signature may only PROMOTE a pair to "candidate for the pairwise snapshot
   check"; it may NEVER by itself cause a hard DUPLICATE/trash. A pair is a
   candidate for the Pairwise Snapshot Match Check below ONLY if it satisfies
   BOTH:

   • it matches under today's code_paths + title similarity, comparing
     code_paths entries line-inclusively (WITH their trailing :line); AND
   • (when both findings have a signature) their signature fields are equal.
     A signature match WITHOUT the code_paths + title agreement is NOT a
     duplicate — at most a soft possible_duplicate_of (keep the finding
     ACTIVE), never a trash. Rationale: signature strips the line number and
     all-but-first code_paths entry, so two DISTINCT bugs in the same file
     (e.g. parser.c:100 vs parser.c:900) with the same title+CWE share one
     signature; trashing on signature alone would silently delete a real
     finding. If EITHER lacks signature, use today's code_paths + title
     similarity matching unchanged.

   PAIRWISE SNAPSHOT MATCH CHECK (decides MATCHED vs NOT_MATCHED) — compares the
   CURRENT finding's discovery_commit against the ARCHIVED finding's
   discovery_commit for this pair (NOT against SNAPSHOT_ID):

   1. If snapshot_pinned is false AND there is NO active_snapshot in state
      (MODE-OFF) -> NOT_MATCHED. Stop. (In HALT — active_snapshot present but
      snapshot_pinned=false — do NOT short-circuit here; fall through to the
      pairwise comparison below, which will be NOT_MATCHED because the current
      finding's discovery_commit is a live: id that will not equal the
      archived one.)
   2. Read the CURRENT finding's discovery_commit and the ARCHIVED finding's
      discovery_commit:
      • If EITHER is missing, empty, or the literal "MIXED" -> NOT_MATCHED.
      • If they are NOT byte-for-byte equal to each other -> NOT_MATCHED.
      • If they ARE byte-for-byte equal to each other (both present, non-MIXED)
        -> 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.) Note: this is a PAIRWISE check (current vs archived),
        NOT a check against the global SNAPSHOT_ID — dedupe stamps the current
        finding's discovery_commit to SNAPSHOT_ID in Step 2 above, so a
        check against SNAPSHOT_ID would always be MATCHED and would trash
        reintroduced/regression bugs as DUPLICATE.

   Then, using the idempotency rule (Input/Output Contract → Idempotency
   Guarantee) to avoid double-writes, decide mechanically:

   • MATCHED (both present and equal): soft-delete the current finding as a
     loop-duplicate exactly as before — set "status": "DUPLICATE" and
     "duplicate_of": "<archived_uuid>", clear possible_duplicate_of if
     present, ensure mkdir -p workspace/findings/.trash/, move it there, and
     log a loop_filter transaction in workspace/.tx_log.jsonl. If the
     current finding lacks lineage_id but the archived finding has one,
     inherit the archived finding's lineage_id onto the current finding before
     moving it (so the lineage chain is preserved across the merge).
   • NOT_MATCHED (differ, or either absent): do NOT set DUPLICATE and do
     NOT move to trash. Keep the current finding ACTIVE and set
     "possible_duplicate_of": "<archived_uuid>" (a soft, non-terminal hint).
     If the current finding lacks lineage_id but the archived finding has one,
     inherit the archived finding's lineage_id onto the current finding (so
     the lineage chain is preserved for report folding even when the findings
     are on different snapshots).

   [!IMPORTANT] STATUS & DUPLICATE INVARIANTS:

   • A finding MUST NOT carry both duplicate_of and possible_duplicate_of
     pointing to the same target UUID.
   • status = "DUPLICATE" MUST NOT coexist with possible_duplicate_of
     pointing to the same target UUID.
   • Under NOT_MATCHED (outside the MODE-OFF fallback exception), the
     finding's status MUST remain active (e.g. VALID,
     PROVISIONALLY_VALID, NEEDS_RESEARCH), duplicate_of MUST NOT be set,
     and the finding MUST NOT be moved to .trash/. Setting
     possible_duplicate_of is a non-terminal hint only.

   • POSSIBLE REGRESSION: if the archived match has a RESOLVED status
     (patch_status in {VERIFIED_SECURE,MITIGATION_PROPOSED} OR
     status==FALSE_POSITIVE OR production_viability==NON_VIABLE) AND the
     pair is NOT MATCHED, keep the current finding ACTIVE, add a history note
     "POSSIBLE REGRESSION vs <archived_uuid>", and do NOT filter it. (A
     reverted fix re-discovered on new code must never be trashed.)
   • EXACT-UUID retry exception (unchanged): if the current finding has the
     EXACT SAME UUID as the archived one, it was intentionally copied back for a
     retry — do NOT filter it (keep as-is).
   • Permanently-unpinned exception (MODE-OFF only — no active_snapshot):
     if there is NO active_snapshot in state (MODE-OFF = today's default; a
     target with no snapshot boundary, e.g. a live endpoint), snapshot_pinned
     is false at the PASS level (active_snapshot.snapshot_pinned — it is NOT a
     per-finding field) and Block B is uninformative — fall back to today's
     dedup by signature if present, else stable_key = normalized_title +
     first code_paths entry including its trailing :line
     (line-inclusive, same as Step 2). Rationale: stripping :line would
     collapse two DISTINCT bugs in the same file (e.g. parser.c:100 vs
     parser.c:900) with the same title+CWE into one — silently deleting a real
     finding. Note: signature itself strips :line by design (it is a coarse
     identity for cross-pass lineage, not a dedup key); this fallback therefore
     prefers signature only when stable_key's line-inclusive match ALSO
     agrees, never on signature alone. This preserves dedup for targets that
     can never MATCH. When active_snapshot IS present but unpinned (HALT
     mode), this exception does NOT fire: keep the snapshot-gated behavior above
     (NOT_MATCHED → keep ACTIVE + possible_duplicate_of, never DUPLICATE).
     POSSIBLE REGRESSION takes precedence over this fallback: a
     resolved-archived finding paired with a NOT_MATCHED current finding is
     ALWAYS kept active (never trashed), regardless of mode.

3. 3-Tier Deduplication Ladder (Deterministic & Semantic Matching): When
   resolving finding lineages and evaluating deduplication candidates, the
   deduplicator uses a hierarchical 3-tier ladder designed for sub-millisecond
   fast-path execution with semantic RAG fallback:

   • Tier 1 (Fast-Path Exact Heuristic Anchors — < 1ms, 0 tokens):

     1. Exact content identity signature match:
        hashlib.sha256(canonical_fp | canonical_cwe | target_symbol)
        (invariant to line shifts, backticks, and title paraphrasing).
     2. Exact canonical_filepath + normalized CWE + target_symbol match.
     3. Strict line proximity window ($\le 3$ lines) on exact canonical
        filepath when target symbol is empty.
     • Result: If matched, immediately inherit ancestor lineage_id without
       LLM or embedding overhead.

   • Tier 2 (RCA Normalization): If Tier 1 heuristic matching does not
     produce an exact anchor, synthesize a standardized 5-line Root Cause
     Analysis (RCA) summary:

     • Component: Normalized canonical filepath and symbol.
     • Vulnerability Class: Canonical CWE taxonomy identifier and name.
     • Root Cause Mechanism: Underlying programming or logic defect.
     • Failure Condition: Specific input, state, or boundary condition.
     • Taint Dataflow: Source-to-sink dataflow trajectory.

   • Tier 3 (Vector Embedding & Cosine Similarity Scan):

     • Project the standardized RCA summary into dense vector embeddings using
       the configured multi-provider embedding engine (default:
       vertex_ai/gemini-embedding-001).
     • Perform nearest-neighbor scan over historical lineage_vectors in the
       database using bounded cosine similarity.
     • CWE Family & Class Structural Guard: When comparing query finding
       vectors against candidate lineage records, if both findings have explicit
       CWE classifications (e.g. CWE-89 vs CWE-78), normalize them. If they
       belong to distinct, incompatible CWE IDs, skip candidate vector
       comparison entirely to structurally prevent false-merging distinct
       vulnerability classes regardless of cosine score.
     • Positive threshold ($\ge 0.90$): Semantically equivalent findings
       with differing phrasing, scanner labels, or line shifts merge into the
       same ancestor lineage_id (default: 0.90, configurable via
       EMBEDDING_SIMILARITY_THRESHOLD).
     • Negative threshold ($< 0.70$): Distinct vulnerability classes (e.g.,
       SQLi vs Command Injection, Stored vs Reflected XSS) maintain low
       similarity and fail closed, minting a fresh unique lineage_id.

4. Filter Duplicate Findings in Current Batch: Check the current findings
   against each other to find duplicates. Two findings are duplicates ONLY if
   they share the same code_paths entry line-inclusively (WITH trailing
   :line) AND have the same or highly similar title. If multiple findings
   refer to the exact same flaw at the same location, they must be merged.
   Findings at different lines in the same file are DISTINCT — never merge them.

5. Map/Reduce Chunking Strategy (For Scale): If there are many finding files
   (e.g., > 20 items), use a Map/Reduce approach to group them by target file or
   component before checking for overlaps to avoid context window limits.

6. Token-Optimized Consolidation and Merging: To minimize LLM output tokens
   and prevent data loss, do not manually rewrite or output the merged JSON
   files in your response. Instead, follow this pattern:

   1. Identify Duplicates: Internally map which findings are duplicates of a
      primary finding.
   2. Reusable Deterministic Scripting (versioned): Write a reusable helper
      script (e.g. workspace/helpers/merge_findings.py) whose FIRST LINE is
      exactly # MANTIS_HELPER_VERSION = 2. Before reusing an existing helper,
      grep its first lines for MANTIS_HELPER_VERSION = 2; if that marker is
      absent or a different integer (a helper left by an older pipeline
      version), REGENERATE the helper. Only reuse it when the marker matches.
      The script must follow these deterministic rules:
      • Title: Pick the most comprehensive and descriptive title.
      • ID: Preserve the unique "id" of the primary finding being kept.
      • Severity: Pick the highest severity level specified among the merged
        items.
      • Privileges Required: Inherit the most severe privilege requirement
        (priority: NONE > LOW > HIGH).
      • Attacker Position: Inherit the most critical position requirement
        (priority: EXTERNAL > INTERNAL_NETWORK > IN_CLUSTER > LOCAL >
        HOST_SYSTEM > SUPPLY_CHAIN > PHYSICAL_TEMPORARY >
        PHYSICAL_LONG_TERM).
      • User Interaction: Inherit the most severe user interaction
        requirement (priority: NONE > REQUIRED).
      • Code Paths: Collect and deduplicate all file paths and line numbers
        into a single unique array.
      • Description, Mitigation, & Impact: Concatenate cleanly.
      • History: Concatenate and preserve all "history" entries from the
        merged findings. Append a new entry to the "history" array for this
        merge action conforming to the schema (containing "stage": "dedupe",
        "action": "merge",
        "details": "Merged duplicate findings: [comma-separated-ids]",
        "pass_number": <current_pass_number>, and
        "timestamp": "<current_iso8601_timestamp>").
      • Unknown keys & provenance (MANDATORY): The script MUST copy through
        EVERY key it does not explicitly handle (including discovery_commit,
        repro_snapshot_id, patch_base_snapshot, possible_duplicate_of,
        signature, lineage_id, cwe) from the primary finding onto the
        merged object — never drop unknown fields. It MUST REFUSE to merge two
        findings whose discovery_commit values differ (they describe different
        code versions); leave them separate and log the refusal. When merging
        findings that all share one discovery_commit, preserve it unchanged.
   3. Execute the Script: Run your script to update the primary finding's
      file (workspace/findings/<primary_id>.json) on disk.

7. Transactional Staged Clean Up: Do not permanently delete redundant files.
   Ensure the trash directory exists (e.g.,
   mkdir -p workspace/findings/.trash/). Before moving, the script must update
   the duplicate finding files, setting "status": "DUPLICATE" and
   "duplicate_of": "<primary_uuid>". Move the merged duplicate .json files
   to the trash staging directory (workspace/findings/.trash/). For every file
   moved, append a transaction record to workspace/.tx_log.jsonl.

   Transaction Log Schema Format (workspace/.tx_log.jsonl)

   Each line must be a self-contained JSON object documenting the transaction:

   {"timestamp": "2026-07-14T15:13:00Z", "action": "loop_filter | dedupe_merge", "primary_uuid": "[UUID] (or null for loop_filter)", "moved_uuid": "[UUID]"}
   

   This cleans up the directory for downstream stages while preserving rollback
   capability.

When complete, notify the user.