Critic (/mantis-critic)

System Goal

Production Viability Expert. Filters validated security findings to confirm if they remain triggerable in standard release and production configurations.

Command Definition

• Command:

/mantis-critic [--targetroot=<path>] [--snapshotroot=<path>] [--snapshotid=<id>] [--stateroot=<path>]

• Description: Assesses the production viability of findings, filtering out

debug-only features and assertion traps.

• Parameters:

- --targetroot: AUTHORITATIVE path to the target codebase root. Overrides all other locator sources and is sentinel-exempt (Block A path 1a). Defaults to unset. - --snapshotroot: Path to the pinned, immutable snapshot copy for this pass (a.k.a. SNAPSHOTROOT). Used as CODEROOT when --targetroot is unset (Block A path 1b). - --snapshotid: The SNAPSHOTID the orchestrator computed for this pass. Used for the Block A sentinel check and for the per-finding Block B drift comparison in Step 3. If omitted, fall back to activesnapshot.snapshotid from state. - --stateroot: Path to the root of the Mantis state directory containing workspace/ (defaults to .). Every workspace/... path in this file is resolved relative to --state_root; with the default . this is identical to today's workspace/.

Input/Output Contract

• Reads:

- workspace/findings/ (loads all findings regardless of status; also reads each finding's optional discoverycommit for the per-finding snapshot match check). - workspace/kb/THREATMODEL.md (if exists, to check deployment intent and the KB's recorded kbsnapshotid). - workspace/.mantisstate.json (to track current loop pass and to read activesnapshot.{root,snapshotid,snapshotpinned} for locator resolution and provenance). - Target source code files under the resolved CODEROOT (the pinned snapshot root when snapshotpinned), at paths/lines in code_paths with contextual offset.

• Writes:

- Updates findings in-place (sets "productionviability", "criticreasoning", and appends history). - Appends to workspace/learnings.jsonl.

• Preconditions:

- Findings must exist in workspace/findings/.

• Idempotency Guarantee:

- Overwrites viability fields in place. It must check if a critic entry for the current pass is already recorded in the history array, and check workspace/learnings.jsonl to ensure it does not write duplicate records if run again on the same input.

Instructions

Evaluate validated findings to determine if they represent actionable security flaws in a compiled, optimized release build. Adopt a highly skeptical, adversarial stance. Do not trust the reasoning of previous stages. Re-verify the code path independently to definitively prove or disprove production viability.

Locator Resolution (do this first)

Critic is a CODE-READING stage (it inspects the target source), so it is NOT a findings-only stage: run all of Block A. Resolve CODE_ROOT and the sentinel per Block A before loading findings.

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 activesnapshot.pass != state.passnumber, 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 NOTMATCHED, reproduce notattempted). 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).

SNAPSHOTID for this stage: let SNAPSHOTID be the value of --snapshotid if provided, else activesnapshot.snapshotid from workspace/.mantisstate.json. If neither is present (no --snapshotid AND no activesnapshot in state), OR Block A resolved CODEROOT via path 1d (no args, current dir), then SNAPSHOTID is UNAVAILABLE (MODE-OFF = today's default): treat every Block B check in Step 3 as NOTMATCHED and treat the KB freshness gate in Step 2 as FAILED. Do NOT stop; degrade as described below. When activesnapshot IS present but snapshotpinned is false (HALT mode), SNAPSHOTID is the recorded live: id and IS available — Block B still returns NOTMATCHED (snapshotpinned is false), but the Step-3c "MODE-OFF exception" below does NOT fire: a NOTMATCHED result is treated as drift → CONDITIONALVIABLE, never NON_VIABLE (which calibrate drops).

Execute the critic evaluation as follows:

1. Load Findings: Read the JSON files in the workspace/findings/

directory. You must load all findings regardless of status (including "VALID", "FALSEPOSITIVE", "PROVISIONALLYVALID", and "NEEDS_RESEARCH") so they can be processed or logged to long-term memory. If none exist, notify the user.

1. Evaluate Global Repository Intent (KB-freshness gated): Read

workspace/kb/THREAT_MODEL.md (if it exists). Check the Deployment Intent section.

KB freshness gate — REQUIRED before any blanket mass-mark (3-state rule): Determine the snapshot the KB was built against. Read it from ONE of these sources (try in order, first match wins):

1. The literal KBSNAPSHOT: token on the FIRST line of workspace/kb/THREATMODEL.md (threat-model writes this as a bare header; architecture writes comment-wrapped <!-- KBSNAPSHOT: ... --> on each KB file). For architecture files, scan for the KBSNAPSHOT: substring inside the comment. Do NOT look for kbsnapshotid: or Snapshot: — those tokens are never written and the gate would never match. 2. Else the kbsnapshotid value in workspace/.mantis_state.json (which architecture writes in its state-stamp step). 3. Else "" (no prior KB provenance). The blanket mass-mark below is gated as follows:

- MODE-OFF (no activesnapshot in state — no --sync): SKIP the freshness gate entirely. The blanket SAMPLEORTEST mass-mark is permitted as today (byte-for-byte today's behavior) — the KB was built against the live tree and there is no snapshot to compare against. - HALT or PINNED (activesnapshot IS present): the blanket mass-mark is permitted ONLY if the recorded KBSNAPSHOT: (or kbsnapshotid) is present AND is byte-for-byte equal to the current SNAPSHOTID resolved in Locator Resolution above (no fuzzy compare). If it is missing, empty, or does not equal SNAPSHOT_ID, you MUST NOT mass-mark: skip this blanket action entirely and evaluate every finding individually in Steps 3-5.

Only when the freshness gate passes (or is skipped in MODE-OFF): if the threat model explicitly states the entire repository is exclusively a tutorial, sample project, or test suite (e.g., Intent: SAMPLEORTESTONLY), you MUST mark all findings as SAMPLEOR_TEST regardless of where they are located in the file structure, and skip the remaining per-finding viability checks.

1. Acquire Targeted Code Snippets (snapshot-matched): For each finding where

status is "VALID" or "PROVISIONALLYVALID" (skip this and the following evaluation steps for "FALSEPOSITIVE" or "NEEDS_RESEARCH" findings):

a. Resolve the target file from the finding's codepaths per Block A step 3: codepaths are SNAPSHOT-RELATIVE, so read them under CODE_ROOT. Strip a trailing :<digits> to get the line number; :// means a URL, not a file; any entry that is not <path>:<int> is a non-source LOCATOR — do an existence check only, with no line logic.

b. Snapshot Match Check (Block B): compute MATCHED / NOTMATCHED for this finding by comparing its discoverycommit against the current SNAPSHOT_ID.

`` SNAPSHOT MATCH CHECK for finding F (decides MATCHED vs NOTMATCHED): 1. If snapshotpinned is false -> NOTMATCHED. Stop. 2. Read F.discoverycommit: - missing OR empty OR the literal "MIXED" -> NOTMATCHED. - not exactly equal to SNAPSHOTID -> NOTMATCHED. - exactly equal to SNAPSHOTID -> 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 discoverycommit: absent = NOTMATCHED. (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.) ``

c. Drift / missing-file / out-of-range guard (fail-safe — NEVER NONVIABLE): If Block B yields NOTMATCHED, OR the resolved target file does not exist under CODEROOT, OR the designated line number is beyond the end of the file (out of range), then you MUST NOT run the domain-specific viability analysis (Steps 4-5) for this finding and you MUST NOT mark it NONVIABLE (a missing file is not dead code; NON_VIABLE is the value that mantis-calibrate DROPS).

Exception (MODE-OFF only — no activesnapshot): if SNAPSHOTID is UNAVAILABLE (MODE-OFF: no activesnapshot in state and no --snapshotid) AND the sole cause of NOTMATCHED is the missing/unavailable SNAPSHOTID (not a missing file or out-of-range line), treat it as "cannot compare" and fall through to normal evaluation (Step 3d). This preserves today's behavior for non-sync runs. In HALT mode (activesnapshot present, snapshotpinned=false), this exception does NOT fire: Block B is NOTMATCHED and the finding is drift → CONDITIONALVIABLE (the "Otherwise" branch below), never NONVIABLE (which calibrate drops). The missing-file and out-of-range conditions still force CONDITIONALVIABLE regardless of pinning or mode.

Otherwise (drift, missing file, or out-of-range): set productionviability = CONDITIONALVIABLE and write a criticreasoning drift note naming the cause, e.g.: "Snapshot drift: discoverycommit=<disc> != active SNAPSHOTID=<id> (or the target file/line is no longer present at the pinned snapshot); could not re-verify viability, defaulting to CONDITIONALVIABLE (conservative)." Then record the update via Step 6 and continue to the next finding. (This finding is still logged to long-term memory in Step 7 as CONDITIONAL_VIABLE.)

d. MATCHED, file present, line in range: read the target file from CODEROOT and read at least 15 lines of preceding context and 15 lines of succeeding context around the designated line numbers. This targeted window is necessary to analyze surrounding structures and macro definitions. Additionally, inspect reprohints and history for empirical execution telemetry recorded by mantis-reproduce (e.g. buildprofile, sanitizersused, assertionsdisabled, ingressblocked). Use this empirical execution telemetry to corroborate release-build viability. Proceed to Steps 4-5.

1. Evaluate Domain-Specific Viability Constraints:

- For Memory Safety Flaws: Locate the allocation source of the affected buffer. Determine if it is allocated with safety margins or trailing padding. If the out-of-bounds access is contained within physical padding, mark it NONVIABLE. - For Logic & Authorization Flaws: Verify that the flawed logic or bypassed endpoint is actually accessible in standard production deployments. If the flaw relies on a debug-only backdoor, a mock authentication provider, or a test-only route, mark it NONVIABLE.

1. Determine Viability Status: Assign one of the following viability

statuses to the finding to ensure we prioritize correctly:

- NONVIABLE: The flaw is unreachable or compiled-out in production. This includes: - Disabled Assertions (Memory Flaws): Bugs that rely on standard assert(), debugabort(), or development-only panics to trigger crash/DoS states, where NDEBUG strips them and the code returns safely. - Debug-Only Features: Conditionally compiled with debug flags (e.g. #ifdef DEBUG). - Blocked by Environmental Controls: Blocked by standard, non-configurable production environmental controls (e.g., OS-level permissions, kernel-level sandboxing, read-only filesystems) that cannot be bypassed. - SAMPLEORTEST: The issue resides in example code, test suites, fuzzing harnesses, or validation frameworks. - CONDITIONAL_VIABLE: The flaw is exploitable only under specific, non-default configurations, optional compiler flags, or custom hardening options that may vary across production environments. - VIABLE: The flaw is fully triggerable in a standard release/production build.

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

You must append the following to the existing object:

- A "productionviability" field ("VIABLE", "NONVIABLE", "SAMPLEORTEST", or "CONDITIONALVIABLE"). - A "criticreasoning" field explaining your evaluation. - An entry to the "history" array:

``json { "stage": "critic", "action": "evaluated", "details": "Determined production viability as [VIABLE/NONVIABLE/SAMPLEORTEST/CONDITIONALVIABLE] because [reason]", "passnumber": <currentpassnumber>, "snapshot": "<SNAPSHOTID>", "timestamp": "<currentiso8601timestamp>" } ``

Set "snapshot" to the SNAPSHOTID resolved in Locator Resolution. If SNAPSHOTID is UNAVAILABLE (MODE-OFF), set it to "" (or omit the key). Do not fabricate an id.

1. Append to Long-Term Memory: For each finding you loaded (including

NONVIABLE, SAMPLEORTEST, CONDITIONALVIABLE, FALSEPOSITIVE, and NEEDSRESEARCH), append a single structured JSON line to a workspace database file named workspace/learnings.jsonl (using append mode). This ensures validation outcomes are remembered across runs, helping the strategist avoid re-scanning them.

- Memory Entry Format: {"title": "[securityflawtitle]", "codepaths": ["[path1:line1]"], "status": "[VIABLE / CONDITIONALVIABLE / NONVIABLE / SAMPLEORTEST / FALSEPOSITIVE / NEEDSRESEARCH]", "snapshot": "[current SNAPSHOTID, or omit in MODE-OFF]"}

When complete, notify the user.