@xaccefy/pi-casefile 0.9.1 → 0.9.3

package/src/workflow.ts

@@ -1,5 +1,5 @@
 /**
- * Cyber workflow injected into agent context when XP mode is ON.
+ * Cyber workflow injected into agent context when XP mode is SWARM.
  *
  * Skills (cyberwf, web-pentest) cover tool usage and methodology. This file
  * adds the unique attacker discipline: state machine with preconditions,
@@ -39,36 +39,30 @@ type DispatchSpec = {
   hardGate: string;
   /** Crash-handling paragraph. */
   crash: string;
-  /** Skeptic dispatch snippet (follows "dispatch it BEFORE the exploit agent with "). */
+  /** Skeptic dispatch snippet (follows "dispatch it before main-agent validation 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.",
+    "**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([...])`, capped at 3 auditor tasks per round. Batch related classes by surface/family instead of spawning one agent per bug class. Stable keys include run, stage, batch/case, and attempt.",
   hardGate:
-    "Your next tool call MUST launch one async workflowScript whose `runs.all([...])` dispatches HUNT auditors.",
+    "Your next tool call MUST launch one async workflowScript whose `runs.all([...])` dispatches at most 3 batched 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.",
+    "**Subagent failure handling:** a crash, timeout, hung run, unparseable output, or schema-invalid output is a RETRY, not a verdict. Launch one new workflowScript with the same specialist task, a new stable attempt key, and a stronger model. Failure again → record `blocked: <agent> failed` 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.",
+    "**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 at most 3 auditor tasks per round, batching related classes by surface/family instead of spawning one agent per bug class. Stable names include run, stage, batch/case, and attempt. Results deliver automatically; steer with `hub`.",
   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.",
+    "Your next tool call MUST launch one async `task` call whose `tasks` array dispatches at most 3 batched HUNT auditors. 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.",
+    "**Subagent failure handling:** a failed or hung task, timeout, unparseable output, or schema-invalid output is a RETRY, not a verdict. Re-dispatch the same specialist task with a new attempt name and a stronger model. Failure again → record `blocked: <agent> failed` 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. */
@@ -88,17 +82,28 @@ Think like a real external attacker, not a code reviewer. Technical bugs are che
 
 ${d.reference}
 
+**Swarm delegation boundary:** only auditor (HUNT rounds), tracer (TRACE), skeptic (high-confidence challenge), and chain (CHAIN) run as subagents. You, the main coordinator, own RECON, VALIDATE/PoC writing, ConfirmFinding, patching, final reports, state decisions, and all orchestration.
+
+## Dispatch Discipline (fewer calls, batched, verifier-in-the-loop)
+
+This pipeline is SEQUENTIAL-dependent: each stage consumes the previous stage's output, so only HUNT genuinely fans out. Do NOT scatter one async subagent call per finding — that pays full coordination cost for no parallel payoff and turns you into a message router. Two rules keep it cheap:
+
+1. **Two dispatch points, each ONE batched call.** (a) **HUNT** — one call whose \`runs.all\`/\`tasks\` carries ≤3 batched auditors (related classes grouped by surface/family). (b) **TRACE+SKEPTIC** — one call carrying a trace task per prioritized finding, plus a skeptic task for each \`confidence: high\` finding; batch the whole round in a single dispatch, never one dispatch per finding. RECON, VALIDATE/PoC, ConfirmFinding, CHAIN, and REPORT stay INLINE with you.
+2. **Barrier, then submit the whole batch in one pass.** Let the batched call return ALL of its results, then \`PipelineSubmit\` each output back-to-back before choosing the next stage. Do not interleave fresh dispatches with the delivery of a prior batch. A crash / timeout / unparseable / schema-invalid result for one item is a RETRY for THAT item in the next batch — never a verdict, never a reason to drop the stage.
+
+Every stage boundary is a \`PipelineSubmit\` gate (schema + pre-filter, in code): nothing advances on prose. This is the verifier-in-the-loop — the same principle the machine PoC gate applies at CONFIRMED, applied at every stage transition.
+
 ## 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)
+RECON (you, inline) → **HUNT** (2-3 batched auditor subagents) → TRACE (tracer for prioritized findings) → SKEPTIC (bounded high-risk review) → VALIDATE (you, inline) → CHAIN (chain subagent) → REPORT (you, inline)
 
-### Blackbox doctrine — gather everything first (no source access)
+### Blackbox recon — attack-surface mapping (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.
+Live web target, CTF, or bounty box: RECON aggressively gathers high-signal intel and turns it into the map HUNT will use: entry-point inventory, attacker model, auth/role boundaries, trust boundaries, likely vuln-class batches, and known gaps. Aim for the richest useful map, not the largest raw pile. Choose the next recon move from the target and current unknowns. Use JS/source maps, \`robots.txt\`, \`sitemap.xml\`, \`/.well-known/\`, OpenAPI/Swagger, GraphQL introspection, passive archives, and exposed backup/VCS checks when they are likely to change class selection, target selection, or attacker modeling. Stop when additional collection is unlikely to change the HUNT plan; re-enter RECON when HUNT/TRACE exposes missing surface.
+- **Client-side code can be high signal** — pull JS bundles/source maps when the app is SPA/API-heavy or routes are hidden, then bank discovered endpoints, params, and secrets as leads.
+- **Zero-traffic intel is optional, not ritual** — check public metadata, schemas, and passive archives when scope allows and the result can change target selection, auth modeling, or class selection.
+- **Fingerprint for decisions** — stack + version confidence should drive \`exploit_search\` and HUNT class selection; record uncertainty instead of forcing a guess.
+- **Bank useful leads** — write the entry-point map, selected HUNT class batches, and open gaps to the scratchpad; file high-value leaks (source map, origin IP, exposed schema, leaked creds) as \`EvidenceAdd role=observation\`. 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.
 
@@ -106,6 +111,8 @@ Live web target, CTF, or bounty box: the target is opaque and **every later stag
 
 ${d.crash}
 
+**TRACE verdicts:** only schema-valid \`trace_result: "REACHABLE"\` advances toward validation. \`UNREACHABLE\` requires a concrete blocker for the stated attacker model; \`UNDETERMINED\` means missing context/auth/WAF/source ambiguity and blocks or re-dispatches, never kills.
+
 ## Case Lifecycle (State Machine)
 ${LIFECYCLE_DIAGRAM}
 
@@ -118,7 +125,7 @@ ${LIFECYCLE_DIAGRAM}
 | 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. |
+| REPORT | REPORTED | CaseContext → main-agent report writing → report-readiness gate. |
 
 ### Preconditions Per State Transition (MANDATORY)
 
@@ -127,7 +134,7 @@ ${LIFECYCLE_DIAGRAM}
 | HYPOTHESIS → INVESTIGATING | evidence (observations), confidence | Notes on what was observed |
 | 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 |
+| CONFIRMED → REPORTED | CaseContext(id) succeeded (records report path) AND you wrote the report file | Context bundle + report file |
 
 **Empty required field = you cannot advance.** The fields ARE the gates.
 
@@ -179,7 +186,7 @@ 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 ${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.
+**\`confidence: high\` → skeptic subagent (MANDATORY):** dispatch it before main-agent validation with ${d.skeptic}. It independently re-reads the source (or re-probes live), verifies scope, tries to disprove, and audits any PoC file you already have for cheats. Its schema-validated CONFIRMED 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. UNDETERMINED → block/re-dispatch; do not validate yet. Do NOT skip; do NOT self-disconfirm high-confidence findings.
 
 **Below high → self-disconfirmation:** actively try to disprove your own finding; document it (see the strong/weak example below). Not a formality.
 
@@ -188,15 +195,22 @@ An attempt: reproduce under different conditions (auth/config/network position);
 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.
 
-**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.
+**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 PromoteFinding or 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.
 
-**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.
+**Main-agent validation only:** do not dispatch validation. You write the smallest reliable PoC that demonstrates the **maximum reachable impact** of the vulnerability, set the case's poc/evidence/impact/severity/target fields, and run PromoteFinding yourself. "Smallest" means no fragile ceremony, mocks, or unrelated exploit steps — not a weaker impact demonstration. Do not stop at a benign marker if a stronger in-scope, non-destructive primitive is reachable (read/write, privilege change, account takeover path, data exposure, etc.). If the PoC fails, refine it yourself up to the local budget; if proof cannot meet the gate, kill or keep the case investigating with the exact blocker.
+
+**PromoteFinding (phase 1) — evidence bundle, not markers.** Pick the differential \`mode\` that fits the class:
+
+- **\`mode: "inter_host"\` (default)** — body-carried proof that is the same on any host (file read, injection exfil, info leak, reflection). Call with \`poc_path\`, an operator-approved \`control_target\` from \`PI_POC_CONTROL_TARGETS\`, optional same-byte \`control_path\` (defaults to \`poc_path\`), and \`local: true\` when the bug needs network. The harness sends the SAME request to target and control and requires \`target_only\`.
+- **\`mode: "intra_target"\`** — access-control and business-logic classes (IDOR/BOLA, auth bypass, privilege escalation, mass assignment, logic/price tampering) where the discriminating variable is the attacker's IDENTITY or a PARAMETER, not the host. A different host lacks the victim's object/state, so inter-host proves nothing. Instead the evidence declares \`verify.mode: "intra_target"\` and a \`baseline\` (a legitimate SAME-host request — the attacker's own object, a properly-authorized request, the field omitted); the harness sends attack + baseline to the case target and requires the proof on the attack response only. No \`control_target\`/\`control_path\`.
+
+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 and keeps redirects on the bound host. 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 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.
+**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 run validation or 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)
 
@@ -248,7 +262,7 @@ Impact text answers: *who is hurt, what is lost, how the attacker reaches it fro
 - **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 confirmed harness evidence 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. Claim the highest impact the harness and main-agent review actually prove; unsupported escalation gets rejected at triage.
 
 ### 7. Adversarial Self-Review
 
@@ -271,8 +285,8 @@ 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 ${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):
+2. **Write the report yourself** at the returned report path using the context bundle. In XP swarm, reporting stays with the main agent.
+3. **Report-readiness gate** (YOU check this before accepting; on failure, edit the report yourself):
 - Deterministic reproduction by another researcher
 - Steps realistic in production
 - Impact justified without inflation (would the vendor agree?)
@@ -306,7 +320,7 @@ export const STATIC_CYBER_WORKFLOW_OMP = buildCyberWorkflow(OMP_DISPATCH);
 export const STATIC_CYBER_WORKFLOW_LITE = `
 # Cyber Workflow — LITE (Single-Agent)
 
-You are the ONLY agent. Do NOT dispatch subagents (no auditor, tracer, skeptic, exploit, or chain agents). You do every stage yourself, inline: recon, hunt, trace, validate, chain, report — the full attacker discipline without subagent orchestration overhead. Great for CTF and focused single-target engagements.
+You are the ONLY agent. Do NOT dispatch subagents (no auditor, tracer, skeptic, or chain agents). You do every stage yourself, inline: recon, hunt, trace, validate, chain, report — the full attacker discipline without subagent orchestration overhead. Great for CTF and focused single-target engagements.
 
 Think like a real external attacker, not a code reviewer. Technical bugs are cheap; **reachable attacker impact** is what matters.
 
@@ -325,28 +339,28 @@ ${LIFECYCLE_DIAGRAM}
 
 ## Stage discipline (all done by you, inline)
 
-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.
+1. **RECON — attack-surface mapping.** Blackbox/CTF: aggressively gather high-signal intel and turn it into entry points, auth models, trust boundaries, attacker model, vuln-class batches, and gaps. Fingerprint credible stack/version signals and search CVEs (\`exploit_search\`) when the version confidence is useful. Use JS/source maps, \`robots.txt\`, \`sitemap.xml\`, \`/.well-known/\`, OpenAPI/Swagger, GraphQL introspection, exposed backup/VCS checks, and passive archives when they can change class selection, target selection, or attacker modeling. Record discovered entry points (URL, method, params, auth state), selected class targets, and gaps/assumptions: \`ScratchpadWrite(run_id, "recon", "entry-points.md", ...)\`.
+2. **HUNT** — choose attack classes from recon and examine relevant entry points. \`CaseAdd\` each lead as a hypothesis. Track coverage per class.
 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")\`.
+6. **REPORT** — run \`CaseContext\` to write the context bundle, then write the final report yourself per the report style checklist below, then \`CaseUpdate(status: "reported")\`.
 
 ## Report style checklist (lite — you are the writer)
 
 Write the final report as a self-contained markdown file at the report path CaseContext recorded, applying the fixed report format rules:
 
 - **Title:** \`<vuln class>: <exact trigger/location> — <honest impact>\` (e.g. "IDOR: order delivery address of any user", "SQLi: blind boolean-based via GET").
-- **Structure:** Summary (2-3 sentences) → Vulnerability Details (CWE, CVSS 3.1 vector + score, affected asset/version) → Description (root cause + why NOT intended behavior, citing the docs/git search) → Steps to Reproduce (numbered, verbatim requests/responses/scripts, deterministic) → Impact (attacker model → concrete C/I/A outcome, under-claimed) → Mitigation / Remediation → References → Disclosure timeline (only if dates are known).
+- **Structure:** Summary (2-3 sentences) → Vulnerability Details (CWE, CVSS 3.1 vector + score, affected asset/version) → Description (root cause + why NOT intended behavior, citing the docs/git search) → Steps to Reproduce (numbered, verbatim requests/responses/scripts, deterministic) → Impact (attacker model → maximum proven C/I/A outcome, not speculation) → Mitigation / Remediation → References → Disclosure timeline (only if dates are known).
 - **Tone:** factual, calm, evidence-carried. NO case IDs, ledger paths, PoC filenames, local paths, or "I discovered" narratives. Never invent evidence — "version not determined" beats a guess. Severity from proven impact only.
 
 ## Gates (unchanged — these keep findings honest)
 
 - **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.
+- **No finding is validated without a reachability trace** showing REACHABLE. UNREACHABLE requires a concrete blocker; unresolved auth/WAF/source ambiguity stays INVESTIGATING or BLOCKED, not killed.
 - **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, 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.
+- **Confirmed requires** evidence + poc + impact + severity + target + disconfirmation, via the two-phase gate: **PromoteFinding** in the differential \`mode\` that fits the class — \`inter_host\` (default) with an operator-approved \`control_target\` for body-carried proof (file read, injection exfil, info leak, reflection), or \`intra_target\` for access-control/logic classes (IDOR, auth bypass, privilege escalation, logic), where the evidence declares \`verify.mode:"intra_target"\` + a same-host \`baseline\` and the harness requires the proof on the attack response only (no control target). Then you, the main agent, inspect the bundle, attempt disconfirmation, and call **ConfirmFinding** yourself. That call captures a fresh second replay before commit. Do not delegate validation or confirmation. The machine gate requires zero-exit complete runs, nonce binding, body evidence, determinism, a DNS-pinned conclusive \`target_only\` differential, 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.** Demonstrate and claim the highest impact the attacker can actually reach; claiming less than a proven escalation is wrong, and over-claiming an unproven one 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.