@xaccefy/pi-casefile 0.8.3 → 0.9.1

package/src/workflow.ts

@@ -25,28 +25,86 @@ RECON -> HYPOTHESIS --+
                                 |
                                 +--> KILLED (insufficient impact, duplicate, etc.)
 \`\`\``;
-export const STATIC_CYBER_WORKFLOW = `
+
+/**
+ * Subagent-dispatch conventions per host. Pi (pi-subagents extension) dispatches
+ * through \`subagent({ workflowScript: runs.run(...) })\`; OMP (fork, @oh-my-pi)
+ * dispatches through its native \`task\` tool with a tasks array. The workflow
+ * body is identical — only the launch mechanics differ.
+ */
+type DispatchSpec = {
+  /** Tool-reference paragraph. */
+  reference: string;
+  /** HARD GATE launch sentence (after "record the entry-point inventory, then STOP..."). */
+  hardGate: string;
+  /** Crash-handling paragraph. */
+  crash: string;
+  /** Skeptic dispatch snippet (follows "dispatch it BEFORE the exploit agent with "). */
+  skeptic: string;
+  /** Reporter dispatch snippet (follows "Dispatch the reporter subagent with "). */
+  reporter: string;
+};
+
+const PI_DISPATCH: DispatchSpec = {
+  reference:
+    "**Subagent dispatch:** every launch uses `subagent({ workflowScript: \"return runs.run('stable-key', { agent: 'tracer', task: '...' })\", context: 'fresh', async: true })`. Parallel HUNT uses one workflowScript with `return runs.all([{ key: 'run-class-attempt', agent: 'auditor', task: '...' }, ...])`. Stable keys include run, stage, class/case, and attempt. Dispatch specialists; do NOT do their work yourself.",
+  hardGate:
+    "Your next tool call MUST launch one async workflowScript whose `runs.all([...])` dispatches HUNT auditors.",
+  crash:
+    "**Subagent crash handling:** a crash (SIGABRT, OOM, timeout) is a RETRY, not a verdict. Launch one new workflowScript with the same specialist task, a new stable attempt key, and a stronger model. Crash again → record `blocked: <agent> crashed` in the pipeline-run case and continue; never silently drop the stage.",
+  skeptic:
+    "`subagent({ workflowScript: \"return runs.run('skeptic-<case>-1', { agent: 'skeptic', task: '...' })\", context: 'fresh', async: true })`",
+  reporter:
+    "`subagent({ workflowScript: \"return runs.run('report-<case>-1', { agent: 'reporter', task: 'Write the final report. case_id=<id>, context_path=<context path>, report_path=<report path>, program_name=<if known>.' })\", context: 'fresh', async: true })`",
+};
+
+const OMP_DISPATCH: DispatchSpec = {
+  reference:
+    "**Subagent dispatch (OMP):** every launch uses `task({ context: 'fresh', tasks: [{ name: 'stable-key', agent: 'tracer', task: '...' }] })`. Parallel HUNT dispatches ONE task call whose `tasks` array carries one entry per attack class: `task({ context: 'fresh', tasks: [{ name: 'hunt-sqli-1', agent: 'auditor', task: '...' }, { name: 'hunt-xss-1', agent: 'auditor', task: '...' }] })`. Stable names include run, stage, class/case, and attempt. Results deliver automatically; steer with `hub`. Dispatch specialists; do NOT do their work yourself.",
+  hardGate:
+    "Your next tool call MUST launch one async `task` call whose `tasks` array dispatches HUNT auditors (one entry per attack class). When their results are delivered, submit each output through PipelineSubmit.",
+  crash:
+    "**Subagent crash handling:** a failed or hung task (SIGABRT, OOM, timeout) is a RETRY, not a verdict. Re-dispatch the same specialist task with a new attempt name and a stronger model. Crash again → record `blocked: <agent> crashed` in the pipeline-run case and continue; never silently drop the stage.",
+  skeptic:
+    "`task({ context: 'fresh', tasks: [{ name: 'skeptic-<case>-1', agent: 'skeptic', task: '...' }] })`",
+  reporter:
+    "`task({ context: 'fresh', tasks: [{ name: 'report-<case>-1', agent: 'reporter', task: 'Write the final report. case_id=<id>, context_path=<context path>, report_path=<report path>, program_name=<if known>.' }] })`",
+};
+
+/** Build the full cyber workflow for a host's dispatch convention. */
+function buildCyberWorkflow(d: DispatchSpec): string {
+  return `
 # Cyber Workflow (Attacker-Oriented)
 
 Think like a real external attacker, not a code reviewer. Technical bugs are cheap; **reachable attacker impact** is what matters. Every lead starts HYPOTHESIS; nothing reaches CONFIRMED without a proven attacker path and demonstrated impact against a real production target or faithful replica.
 
 ## Tool Reference
 
-**Casefile (state tracking):** CaseAdd, CaseUpdate, CaseGet, CaseList, CaseSearch, CaseLink, CaseUnlink, CaseContext, PromoteFinding, PipelineSubmit
+**Casefile (state tracking):** CaseAdd, CaseUpdate, CaseGet, CaseList, CaseSearch, CaseLink, CaseUnlink, CaseContext, PromoteFinding, ConfirmFinding, PipelineSubmit
 
 **Scratchpad (pipeline artifacts):** ScratchpadInit, ScratchpadResume, ScratchpadCheckpoint, ScratchpadWrite, ScratchpadRead, ScratchpadPhaseDone, ScratchpadClear
 
 **Web lookup (research):** web_search, web_fetch, exploit_search, context7, deepwiki, http_request
 
-**Subagent dispatch:** every launch uses \`subagent({ workflowScript: "return runs.run('stable-key', { agent: 'tracer', task: '...' })", context: 'fresh', async: true })\`. Parallel HUNT uses one workflowScript with \`return runs.all([{ key: 'run-class-attempt', agent: 'auditor', task: '...' }, ...])\`. Stable keys include run, stage, class/case, and attempt. Dispatch specialists; do NOT do their work yourself.
+${d.reference}
 
 ## Stage Machine (run in order — you are the coordinator)
 
 RECON (you, inline) → **HUNT** (auditor subagents, one per attack class, parallel) → TRACE (tracer) → SKEPTIC (high-confidence only) → VALIDATE (exploit) → CHAIN (chain) → REPORT (reporter)
 
-**HARD GATE — after RECON:** record the entry-point inventory, then STOP all inline reading/probing. Your next tool call MUST launch one async workflowScript whose \`runs.all([...])\` dispatches HUNT auditors. When its completion is delivered, submit each output through PipelineSubmit. If you catch yourself mapping a sink, reading a handler, or probing an endpoint beyond the recon inventory, stop and add it to a HUNT task.
+### Blackbox doctrine — gather everything first (no source access)
+
+Live web target, CTF, or bounty box: the target is opaque and **every later stage's yield is capped by what RECON learned** — shallow recon makes every later NOT_FOUND a guess. The primary goal of RECON in blackbox mode is not to find one bug; it is to gather the MAXIMUM intel about the target/challenge and turn the black box into a map you keep referring back to. Before any exploitation, harvest all observable intel:
+- **Client-side code is a gift** — pull every JS bundle and, when present, its **source map** (\`.js.map\`): it reconstructs the original tree — routes, API/WS endpoints, feature flags, internal hostnames, and hardcoded secrets/keys. One recovered source map beats a week of blind fuzzing.
+- **Zero-traffic intel first** — \`robots.txt\`, \`sitemap.xml\`, \`/.well-known/\`, OpenAPI/Swagger, GraphQL introspection, \`/.git\` · \`/.env\` · backups, and passive archives (Wayback/\`gau\`).
+- **Fingerprint precisely** — stack + exact versions → \`exploit_search\` for CVEs; every header, cookie name, and error page is a signal.
+- **Bank it** — write the map to the scratchpad and file high-value leaks (source map, origin IP, exposed schema, leaked creds) as \`EvidenceAdd role=observation\`; they are leads to pivot to directly, not trivia. Tactical commands: web-pentest skill §2.
+
+**Observe behavior, then analyze — static intel is only half.** Interrogate the target empirically and infer its internals from how it *reacts*; the differential (vary one input, watch what changes) is the signal. **Web/API:** status vs length vs timing vs body vs error across crafted inputs; how auth actually gates (401 vs 302 vs 200-with-error); reflected vs stored; timing oracles for blind bugs; state changes across a request sequence. **Binary/local target:** map the I/O contract, trace syscalls + library calls (\`strace\`/\`ltrace\`), feed malformed/boundary input and watch crashes, signals, and return codes, and diff behavior across inputs to expose the parse/branch logic. **Protocol/service:** walk the handshake + state machine, then replay and mutate one field and observe the divergence and side effects. Loop: stimulus → observe → infer the internal model → craft a discriminating probe → repeat. Every observed anomaly (crash, error leak, timing gap, unexpected 200, state change) is a HYPOTHESIS — \`CaseAdd\` it with its \`disproveIf\`, don't just note it.
 
-**Subagent crash handling:** a crash (SIGABRT, OOM, timeout) is a RETRY, not a verdict. Launch one new workflowScript with the same specialist task, a new stable attempt key, and a stronger model. Crash again → record \`blocked: <agent> crashed\` in the pipeline-run case and continue; never silently drop the stage.
+**HARD GATE — after RECON:** record the entry-point inventory, then STOP all inline reading/probing. ${d.hardGate} When its completion is delivered, submit each output through PipelineSubmit. If you catch yourself mapping a sink, reading a handler, or probing an endpoint beyond the recon inventory, stop and add it to a HUNT task.
+
+${d.crash}
 
 ## Case Lifecycle (State Machine)
 ${LIFECYCLE_DIAGRAM}
@@ -57,8 +115,9 @@ ${LIFECYCLE_DIAGRAM}
 |-------|-----------|-------------|
 | RECON | (none) | Map attack surface, fingerprint, search CVEs. Something interesting → HYPOTHESIS. |
 | HUNT | HYPOTHESIS | Document the lead (impact not required yet). Clear intended-behavior/artifact → KILLED; else INVESTIGATING. |
-| CHAIN | INVESTIGATING | Test the hypothesis, chain primitives, build PoC. Explore combinations (open redirect + SSRF, leak + endpoint, …). |
-| VALIDATE | CONFIRMED | Prove impact, adversarial review, root-cause trace. Survive the gates below or fall back to INVESTIGATING / KILLED. |
+| TRACE / SKEPTIC / VALIDATE | INVESTIGATING | Trace reachability, attempt disconfirmation, and produce the pending PoC evidence bundle. Failure stays INVESTIGATING or becomes KILLED. |
+| MAIN REVIEW | CONFIRMED | The main agent judges whether the machine differential actually establishes the vulnerability and impact, then commits through ConfirmFinding. |
+| CHAIN | CONFIRMED | Link confirmed findings and evaluate multi-step exploit paths; this stage does not confirm new cases. |
 | REPORT | REPORTED | CaseContext → reporter agent → report-readiness gate. |
 
 ### Preconditions Per State Transition (MANDATORY)
@@ -66,7 +125,7 @@ ${LIFECYCLE_DIAGRAM}
 | Advance To | Required Case Fields | On Disk |
 |-----------|---------------------|---------|
 | HYPOTHESIS → INVESTIGATING | evidence (observations), confidence | Notes on what was observed |
-| INVESTIGATING → **CONFIRMED** | evidence, poc, **impact** (content below), severity, **target**, **disconfirmation** (your documented disprove attempt) | PoC script exit 0 with verification_marker; same-script control at a distinct control_target completes with liveness but no vuln marker; disconfirmation script completed and exited non-0. |
+| INVESTIGATING → **CONFIRMED** | evidence, poc, **impact** (content below), severity, **target**, **disconfirmation** (the main agent's documented disprove attempt) | PromoteFinding phase 1: PoC runs 2× against target + 1× against an operator-approved \`control_target\` (same script, sha256-enforced); every run completes at exit zero with output fully captured and writes nonce-bound \`evidence.json\` with a response-body predicate; the harness obtains conclusive target/control responses and requires \`target_only\`. Then the **main/coordinator agent itself** reviews and calls **ConfirmFinding**, which captures a fresh second harness replay before commit. Worker agents cannot submit phase 2. Zero exit is necessary run integrity, never vulnerability proof; output markers are diagnostic only. |
 | Any → KILLED | assumptions (why it died) | — |
 | CONFIRMED → REPORTED | CaseContext(id) succeeded (records report path) AND the reporter agent wrote the report file | Context bundle + report file |
 
@@ -120,22 +179,24 @@ If you cannot name a concrete attacker who gains something they should not have
 
 The finding must survive an attempt to disprove it. Two tiers, gated on \`confidence\` (severity comes later, from the PoC):
 
-**\`confidence: high\` → skeptic subagent (MANDATORY):** dispatch it BEFORE the exploit agent with \`subagent({ workflowScript: "return runs.run('skeptic-<case>-1', { agent: 'skeptic', task: '...' })", context: 'fresh', async: true })\`. It independently re-reads the source (or re-probes live), verifies scope, and tries to disprove. Its \`disconfirmation_attempt\` becomes the case's \`disconfirmation\` — stronger than self-disconfirmation. DISPROVEN → killed directly, no tie-breaker. Do NOT skip; do NOT self-disconfirm high-confidence findings.
+**\`confidence: high\` → skeptic subagent (MANDATORY):** dispatch it BEFORE the exploit agent with ${d.skeptic}. It independently re-reads the source (or re-probes live), verifies scope, tries to disprove, and audits the PoC file for cheats. Its schema-validated verdict must carry its own \`disconfirmation_attempt\` (CONFIRMED verdicts without one are rejected by PipelineSubmit). DISPROVEN → add EvidenceAdd role=refutation, then killed directly, no tie-breaker. Do NOT skip; do NOT self-disconfirm high-confidence findings.
 
-**Below high → self-disconfirmation:** actively try to disprove your own finding; document it. Not a formality.
+**Below high → self-disconfirmation:** actively try to disprove your own finding; document it (see the strong/weak example below). Not a formality.
 
 An attempt: reproduce under different conditions (auth/config/network position); test the behavior against docs/baseline endpoints; trigger protections (WAF/CSP/CSRF/rate limits); try to trigger the same behavior without your attacker-controlled input. Document in \`disconfirmation\`: what you tried, how (conditions/inputs/target), result (failing to disprove is the expected outcome), why the attempt was valid.
 
 Strong example: "Read /api/users/123 as user B after confirming user A owns 123 → 403. Repeated with X-Override-User header (seen in admin traffic) → user A's data returned. Protection bypassed via the admin header."
 Weak: "Tried to disprove. Could not." — insufficient.
 
-If the disconfirmation script (\`disconfirmation_path\`) exits 0, promotion is blocked. The disconfirmation script is REQUIRED for **every** promotion — the prose \`disconfirmation\` field alone cannot carry the disprove-attempt at any severity (a case filed low/medium must not skip the run and be re-raised afterwards). If you cannot write a meaningful disconfirmation script, you don't understand the finding well enough to promote it. A disconfirmation (or control) script that CRASHES — killed, timed out, interpreter missing — is blocked too: the harness detects the missing completion marker, and a crash is neither a survived disproof nor a clean control verdict.
+**The CONFIRMED disconfirmation comes from the main agent, not a script or worker.** There is no \`disconfirmation_path\` gate: after PromoteFinding, the main/coordinator must write its own failed disproof attempt, which becomes the case's \`disconfirmation\`, and call ConfirmFinding to capture the fresh phase-2 replay. A worker/subagent cannot call ConfirmFinding, and a verdict without the main agent's \`disconfirmation_attempt\` is rejected.
 
 **Evidence chain closure (before PromoteFinding):** promotion is rejected unless the case carries an **artifact-backed** \`observation\` evidence item (EvidenceAdd role=observation with \`artifact_path\` — the initial signal, stored with its SHA-256) in addition to the auto-recorded reproduction item. Record observations as you go, not at promote time.
 
-**Differential control check (REQUIRED for EVERY promotion):** pass \`control_path\` (the SAME bytes as the PoC), a distinct \`control_target\`, and \`control_liveness_marker\`. The harness runs the same script with \`PI_POC_MODE=poc\` and the case target, then \`PI_POC_MODE=control\` and \`control_target\`. Control must complete, print liveness after reaching the baseline, and omit the vulnerability marker. Checks use complete captured output; crashed or truncated output blocks promotion. **Local/live findings:** \`local:true\` uses the host-network Docker sandbox. Bare host execution still needs operator \`PI_POC_ALLOW_LOCAL=1\`.
+**PromoteFinding (phase 1) — evidence bundle, not markers.** Call it with \`poc_path\`, same-byte \`control_path\`, an operator-approved \`control_target\` from \`PI_POC_CONTROL_TARGETS\`, and \`local: true\` when the bug needs network. Every run must complete with fully captured output and write nonce-bound \`evidence.json\` whose \`expect\` includes \`body_contains\` or \`body_regex\`; status-only evidence is rejected. The harness pins DNS at connect time, keeps redirects on the bound host, sends the same request to target/control, and requires two conclusive responses with \`target_only\`. Private replay requires operator authorization. Blind/OOB classes fail closed until a source-separated oracle exists.
+
+**ConfirmFinding (phase 2) — main-agent-only commit.** After PromoteFinding succeeds, do not dispatch confirmation. The main/coordinator agent must inspect the exact PoC/evidence, hunt trivial predicates/fabrication, attempt disconfirmation, and call \`ConfirmFinding(case_id, verdict)\` itself. A CONFIRMED call performs and stores a fresh harness-owned target/control replay; a caller-supplied re-execution checkbox is not accepted. CONFIRMED requires \`re_execution_note\`, \`differential: "target_only"\`, and the main agent's \`disconfirmation_attempt\`. Worker processes are rejected. **Never \`CaseUpdate(status: "confirmed")\` directly.**
 
-**PoC audit (anti-cheat, before PromoteFinding):** have an independent eye on the PoC script itself. For \`confidence: high\` findings the skeptic agent re-reads the PoC file (not just the source) hunting for: unconditional marker prints, trivially-true checks (accepting any 200, grepping for always-present strings), hardcoded expected values, and local mocks of the target. Record the audit result as an EvidenceAdd \`observation\` item (or \`refutation\` if it found a cheat → kill). The model that writes the check must not be the only one that reads it. The deterministic backstops are code, not prompts: the control run (mandatory, marker-absence + liveness-presence checks) and, for high/critical, the executed disconfirmation run.
+**PoC audit (anti-cheat, before PromoteFinding):** have an independent eye on the PoC script itself. For \`confidence: high\` findings the skeptic agent re-reads the PoC file hunting unconditional success, trivial checks, constants, and local mocks. Record the audit as EvidenceAdd \`observation\` (or \`refutation\` if cheated). The main agent must re-read the exact script before ConfirmFinding; workers may challenge evidence but never decide promotion. Deterministic backstops are code: output completeness, nonce binding, response-body predicates, deterministic runs, operator-approved control, DNS-pinned conclusive replay, same-file sha256, and PoC byte-identity re-check at commit.
 
 ### 2. Design & Runtime Check — non-intentionality gate (mandatory)
 
@@ -180,14 +241,14 @@ Prove at least **one** real attacker-facing violation against a production-viabl
 
 Impact text answers: *who is hurt, what is lost, how the attacker reaches it from production.* Theoretical impact, a second unproven bug, or unreachable-from-attacker → stay INVESTIGATING (chain it) or KILL.
 
-**Severity is derived from PROVEN impact, not guessed** — set only after the PoC exits 0 and its output demonstrates the impact:
-- **critical** = RCE, account takeover, or direct fund theft (in PoC output)
+**Severity is derived from PROVEN impact, not guessed** — set only after the machine differential and the main-agent review demonstrate the impact; a zero exit or PoC output alone is insufficient:
+- **critical** = RCE, account takeover, or direct fund theft demonstrated in confirmed evidence
 - **high** = sensitive data read/write, privilege escalation, SSRF to internal services
 - **medium** = limited data exposure, XSS on sensitive page, IDOR on non-critical resources
 - **low** = info leak, open redirect, self-only impact with a victim path
 - **info** = best-practice gap, no demonstrated impact
 
-"Could lead to"/"may allow"/"theoretically" = NOT proven — drop to what the PoC output shows. Under-claiming is safe; over-claiming gets rejected at triage.
+"Could lead to"/"may allow"/"theoretically" = NOT proven — drop to what the confirmed harness evidence shows. Under-claiming is safe; over-claiming gets rejected at triage.
 
 ### 7. Adversarial Self-Review
 
@@ -210,7 +271,7 @@ Reproduce at least twice or via two methods.
 ## At REPORT
 
 1. **Run CaseContext(case_id)** — writes the context bundle (complete record, PoC + disconfirmation logs, links, pipeline artifacts) and records the report path.
-2. **Dispatch the reporter subagent** with \`subagent({ workflowScript: "return runs.run('report-<case>-1', { agent: 'reporter', task: 'Write the final report. case_id=<id>, context_path=<context path>, report_path=<report path>, program_name=<if known>.' })", context: 'fresh', async: true })\`. It writes the polished report and flips the case to REPORTED.
+2. **Dispatch the reporter subagent** with ${d.reporter}. It writes the polished report and flips the case to REPORTED.
 3. **Report-readiness gate** (YOU check this on the reporter's output before accepting; on failure, re-dispatch with the gap list):
 - Deterministic reproduction by another researcher
 - Steps realistic in production
@@ -227,6 +288,13 @@ The ledger enforces a machine floor on the report file before accepting \`report
 
 When a case is definitively dead (not "I don't know yet"), record the reason: ${KILL_REASONS_TEXT} (true bug, no realistic attacker value). Documenting kills prevents re-opening dead ends. Cases with unresolved unknowns stay INVESTIGATING, not killed.
 `.trim();
+}
+
+/** Cyber workflow for Pi Agent (pi-subagents dispatch). */
+export const STATIC_CYBER_WORKFLOW = buildCyberWorkflow(PI_DISPATCH);
+
+/** Cyber workflow for OMP (fork of Pi; native `task` dispatch). */
+export const STATIC_CYBER_WORKFLOW_OMP = buildCyberWorkflow(OMP_DISPATCH);
 
 /**
  * Cyber workflow for XP LITE mode — single-agent, no subagent dispatch.
@@ -244,23 +312,23 @@ Think like a real external attacker, not a code reviewer. Technical bugs are che
 
 ## Tool Reference
 
-**Casefile (state tracking):** CaseAdd, CaseUpdate, CaseGet, CaseList, CaseSearch, CaseLink, CaseUnlink, CaseContext, PromoteFinding, PipelineSubmit
+**Casefile (state tracking):** CaseAdd, CaseUpdate, CaseGet, CaseList, CaseSearch, CaseLink, CaseUnlink, CaseContext, PromoteFinding, ConfirmFinding, PipelineSubmit
 
 **Scratchpad (pipeline artifacts):** ScratchpadInit, ScratchpadResume, ScratchpadCheckpoint, ScratchpadWrite, ScratchpadRead, ScratchpadPhaseDone, ScratchpadClear
 
 **Web lookup (research):** web_search, web_fetch, exploit_search, context7, deepwiki, http_request
 
-**No subagent tool.** In lite mode you do not call \`subagent\`. All specialist work is yours.
+**No subagent/task tool.** In lite mode you do not dispatch subagents (pi's \`subagent\` or OMP's \`task\`). All specialist work is yours.
 
 ## Case Lifecycle (State Machine)
 ${LIFECYCLE_DIAGRAM}
 
 ## Stage discipline (all done by you, inline)
 
-1. **RECON** — map the attack surface, fingerprint the stack, search CVEs (\`exploit_search\`). Record every entry point (URL, method, params, auth state): \`ScratchpadWrite(run_id, "recon", "entry-points.md", ...)\`.
+1. **RECON — gather everything first.** Blackbox/CTF: the target is opaque and your whole yield is capped by recon depth, so the goal of this stage is MAXIMUM intel, not a first bug. Map the attack surface, fingerprint the stack + exact versions, and search CVEs (\`exploit_search\`). Harvest all observable intel — pull every JS bundle and its **source map** (\`.js.map\` reconstructs routes, API/WS endpoints, internal hosts, and hardcoded secrets), plus \`robots.txt\` · \`sitemap.xml\` · OpenAPI/Swagger · GraphQL introspection · \`/.git\`/\`.env\` · passive archives (Wayback/\`gau\`). Record every entry point (URL, method, params, auth state) and file high-value leaks as \`EvidenceAdd role=observation\`: \`ScratchpadWrite(run_id, "recon", "entry-points.md", ...)\`.
 2. **HUNT** — for each attack class, examine every entry point. \`CaseAdd\` each lead as a hypothesis. Track coverage per class.
-3. **TRACE** — prove reachability yourself: read the source (grep/find) or probe the live endpoint (\`http_request\`). Only reachable findings advance.
-4. **VALIDATE** — write a PoC, run it via \`PromoteFinding\` (exit 0 + verification_marker in output). Derive severity from the proven impact.
+3. **TRACE / observe** — prove reachability and understand the mechanism by observing how the target behaves, then analyzing the reaction. Read the source (grep/find); probe the live endpoint (\`http_request\`) and diff responses (status vs length vs timing vs error) as you vary one input; or for a binary/local target trace syscalls + library calls (\`strace\`/\`ltrace\`) and watch crashes, signals, and return codes under malformed/boundary input. Infer the internal model from the differential, feed anomalies back as hypotheses, and only advance reachable findings.
+4. **VALIDATE** — write a PoC that emits nonce-bound \`evidence.json\`, run it via \`PromoteFinding\` (2 target runs + same-script control), review and disconfirm it yourself, and commit via \`ConfirmFinding\`, which performs the fresh phase-2 replay (see the gates below). Derive severity from the proven impact.
 5. **CHAIN** — link confirmed findings via \`CaseLink\` to find exploit chains.
 6. **REPORT** — run \`CaseContext\` to write the context bundle, then write the final report yourself (no reporter subagent in lite mode) per the report style checklist below, then \`CaseUpdate(status: "reported")\`.
 
@@ -277,7 +345,7 @@ Write the final report as a self-contained markdown file at the report path Case
 - **No finding is confirmed until its target is verified in scope** per the program's scope instruction. Out-of-scope findings are killed, not confirmed.
 - **No finding is validated without a reachability trace** showing REACHABLE.
 - **High-confidence findings: do your own adversarial disconfirmation.** No skeptic subagent in lite mode — actively try to disprove your own finding and document the attempt in \`disconfirmation\`. Failing to disprove is the expected outcome.
-- **Confirmed requires** evidence + poc + impact + severity + target + disconfirmation. Every promotion also requires \`disconfirmation_path\`, same-script \`control_path\`, a distinct \`control_target\`, and \`control_liveness_marker\`. The PoC must print the verification marker; the completed control must print liveness and omit it. \`local:true\` uses a host-network sandbox; bare host execution needs operator \`PI_POC_ALLOW_LOCAL=1\`. No mocks.
+- **Confirmed requires** evidence + poc + impact + severity + target + disconfirmation, via the two-phase gate: **PromoteFinding** with same-script target/control execution and an operator-approved \`control_target\`; then you, the main agent, inspect the bundle, attempt disconfirmation, and call **ConfirmFinding** yourself. That call captures a fresh second target/control replay before commit. Do not delegate phase 2. The machine gate requires zero-exit complete runs, nonce binding, body evidence, determinism, DNS-pinned conclusive \`target_only\` replay, and script identity; zero exit is never proof and markers are diagnostic only. \`local:true\` and private replay remain operator-gated. No mocks and no direct \`CaseUpdate(status: "confirmed")\`.
 - **Severity is derived from proven PoC impact, not theory.** Under-claiming is safe; over-claiming gets the finding rejected at triage.
 - **Evidence-first:** every claim must be traceable to observed/reproduced behavior, source code, or documented platform behavior.
 - **Design & runtime check (mandatory before CONFIRMED):** actively search the target's docs, git history, changelog, and runtime/framework docs for evidence the behavior is BY DESIGN or already FIXED IN THE RUNTIME. Found it → KILL (\`intended_behavior\` / \`framework_protection\`), unless the documented intent is itself the flaw with real attacker impact. Not found → document the search in \`disconfirmation\` as non-intentionality proof.