mantis-patch

Patcher (/mantis-patch)

System Goal

Security Patching Expert. Generates minimal, correct code fixes, applies them to source code files, and verifies them inside isolated sandboxes before appending logs to long-term memory.

Command Definition

• Command: /mantis-patch [--target_root=<path>] [--state_root=<path>] [--snapshot_root=<path>] [--snapshot_id=<id>] [--snapshot_pinned=<true|false>]
• Description: Generates minimal security fixes using transactional isolation (shadow directories or file backups), applies patches, and verifies them.
• Parameters (all optional; absent ⇒ today's behavior):
  • --target_root: Root of the code snapshot to read (CODE_ROOT). When the orchestrator pins a snapshot it passes --snapshot_root; --target_root is reserved for a caller that hands you an already-prepared tree.
  • --snapshot_root: Root of the pinned immutable snapshot for this pass.
  • --snapshot_id: The SNAPSHOT_ID string of the pinned snapshot (used by the Snapshot Match Check).
  • --snapshot_pinned: true when a pinned immutable snapshot exists this pass; false/absent ⇒ degraded/legacy (see Backward-compat).
  • --state_root: Root of the Mantis state directory that contains workspace/ (defaults to the current directory). ALL workspace/... paths in this file are STATE-RELATIVE and resolve under <state_root>/workspace/....

Input/Output Contract

• Reads:
  • workspace/findings/ (reproduced finding JSON files where patch_status is not "VERIFIED_SECURE" or "MITIGATION_PROPOSED").
  • workspace/.mantis_state.json (to track current loop pass).
  • workspace/.mantis_state.json active_snapshot (root, snapshot_id, snapshot_pinned) — resolved via Block A; supplies CODE_ROOT and the SNAPSHOT_ID used by the Snapshot Match Check when no --snapshot_* flag is passed.
  • Target source code files.
  • Reproducer script path (repro_file_path) and command (run_command) from findings.
  • Pre-existing backup files matching finding ID (if Option B is used).
• Writes:
  • Source code modifications (applied transactionally and rolled back).
  • Updates finding JSON files in-place (sets "patch_status", "patch_diff", re-attack details including "reattack_status" and "reattack_variants", and history).
  • Appends to workspace/learnings.jsonl.
  • Reusable helper script workspace/helpers/append_patch.py.
• Preconditions:
  • Findings must exist in workspace/findings/.
• Idempotency Guarantee:
  • Snapshot-aware skip: Skips a finding whose patch_status is already "VERIFIED_SECURE" or "MITIGATION_PROPOSED" ONLY when its recorded patch_base_snapshot equals the current SNAPSHOT_ID (i.e. it was verified against THIS snapshot). If patch_base_snapshot is absent, empty, or different from the current SNAPSHOT_ID, the terminal status is stale — do NOT skip; re-open and re-verify the finding against the current snapshot. In legacy mode (no active_snapshot and no --snapshot_* flags) skip exactly as today (by patch_status alone).
  • Transactional isolation: modifies code inside uniquely generated temporary directories or creates temporary file backups (target.c.bak-[id]), restoring baseline state upon completion (using try...finally rollback mechanisms).
  • Reuses the existing append_patch.py script once created.

Instructions

Fix successfully reproduced security flaws without breaking standard code behavior.

Execute the patching and verification stage as follows:

0. Snapshot & Locator Resolution (run once, before touching any code):

   0a. Resolve CODE_ROOT, SNAPSHOT_ID, snapshot_pinned, and state_root via Block A (Locator Resolution):

   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).

   For this stage: mantis-patch READS target code (it is NOT a findings-only stage), so it runs Block A steps 1–6 in full. CODE_ROOT is the immutable pinned snapshot; you MUST NOT write, compile, or run reproducers under it (Block A step 4) — do all mutation in a private shadow (step 2 below).

   0b. Determine PATCH_MODE (mechanical, using the resolved values):

   • active_snapshot absent AND no --snapshot_root/--snapshot_id/ --target_root passed ⇒ LEGACY mode (today's behavior; the new ceilings in step 3 do NOT fire).
   • snapshot_pinned == true ⇒ PINNED mode (full gates below apply).
   • snapshot_pinned == false AND SNAPSHOT_ID begins with the literal live: ⇒ HALT/DEGRADED mode (a snapshot was attempted but could not be pinned; authoritative verdicts are forbidden this pass — see step 3).

   0c. For each finding you process, compute its MATCHED/NOT_MATCHED result via Block B (Snapshot Match Check):

   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.)
   

   The MATCHED/NOT_MATCHED result is consumed in step 2 (isolation-mode choice) and step 3 (verdict). It never on its own downgrades a verdict in PINNED mode: a NOT_MATCHED finding is RE-verified against the current snapshot via the step-3 gate and, if it passes, is stamped with the current patch_base_snapshot.

1. Load Findings to Patch: Read the JSON files in the workspace/findings/ directory. Filter for findings that are NOT already terminally verified AGAINST THE CURRENT SNAPSHOT — i.e. treat a finding as still-to-process unless its patch_status is "VERIFIED_SECURE" or "MITIGATION_PROPOSED" and its patch_base_snapshot equals the current SNAPSHOT_ID (Snapshot & Locator Resolution step 0). A terminal finding whose patch_base_snapshot is absent/empty/different is STALE — include it for re-verification. (In LEGACY mode, where there is no SNAPSHOT_ID, fall back to filtering on patch_status alone, exactly as before.) Among the included findings, keep those where (either repro_status is "reproduced" OR the finding is an exploit chain, e.g., the title starts with "Exploit Chain:", or history has an entry from the "chainer" stage, or the "constituent_findings" property is present and non-empty). If none exist, notify the user.

2. Generate and Apply Minimal Patches: For each reproduced security flaw:

   • Target Agnosticism (Binaries vs Source): If the target is source code, proceed with generating and applying a code patch as described below. If the target is a compiled binary or firmware blob without source code available, do not attempt to modify the binary or write binary patching scripts. Instead, skip the branch isolation/modification/diff steps and generate a general, high-level recommendation for how this issue could be mitigated in a production environment without requiring deep technical depth. Output this mitigation string in place of the patch_diff field, and set "patch_status" to "MITIGATION_PROPOSED".

   • Exploit Chains: If the finding is an exploit chain (identified by "Exploit Chain:" in the title, or history details, or if the "constituent_findings" property is present and non-empty), do not generate a code patch or diff. Instead, identify its sub-findings by reading the "constituent_findings" array of UUIDs. Monitor the patch status of these constituent findings (listed on disk as workspace/findings/<uuid>.json). Important: Defer evaluating exploit chains until all individual findings in the batch have been processed, so that the latest patch statuses of their constituents are available on disk.

     Evaluate the exploit chain status using these propagation rules (evaluated in order):

     1. Validity Check: Read the validity "status" of each constituent finding. If any constituent's "status" is "FALSE_POSITIVE", update the exploit chain finding's "status" to match it (e.g. "FALSE_POSITIVE") and immediately skip any further patching/verification for the chain. If any constituent's "status" is "DUPLICATE", resolve it to its canonical finding by recursively following its "duplicate_of" property.

        Duplicate Resolution Process:

        1. Locate the Finding File: The finding file for the duplicate (or any parent in the duplicate chain) may have been moved. Search for <uuid>.json in the following locations in order:
           • workspace/findings/<uuid>.json (active findings)
           • workspace/findings/.trash/<uuid>.json (de-duplicated trash)
           • workspace/archive/findings_pass_*/<uuid>.json or workspace/archive/loop*_findings/<uuid>.json (archives from previous passes) If the file cannot be found in any of these locations, treat it as a missing file error.
        2. Cycle Detection: Maintain a set of visited finding UUIDs during the resolution. If you encounter a UUID that has already been visited in the current resolution chain, raise a validation error (cycle detected).
        3. Maximum Depth: Limit the recursion depth to a maximum of 5 steps. If the chain is deeper, abort and report an error.
        4. Extract Status: Once you resolve to the canonical finding (one whose "status" is not "DUPLICATE" or does not have "duplicate_of"), use that canonical finding's "status" and "patch_status" for all downstream checks and propagation. Do not update the exploit chain finding itself to "DUPLICATE".

     2. Missing Files: If any constituent finding's JSON file is missing from the disk, set the chain's "patch_status" to "ERROR".

     3. Constituent Unset or Stale (Pending): A constituent counts as NOT-YET-VERIFIED if EITHER its "patch_status" is unset (null or missing, indicating it has not yet been reproduced/processed) OR its "patch_status" is terminal but its patch_base_snapshot is absent/empty/different from the current SNAPSHOT_ID (verified against a stale snapshot — Snapshot & Locator Resolution step 0). If ANY constituent is NOT-YET-VERIFIED, the chain's "patch_status" must remain unset (null or missing) and you must defer/suspend further evaluation of the chain until that constituent is re-verified on the current snapshot. (In LEGACY mode, where there is no SNAPSHOT_ID, apply only the "unset" half of this rule, exactly as before.)

     4. Constituent Errors: If any constituent's "patch_status" is "ERROR", set the chain's "patch_status" to "ERROR".

     5. **Consti

...