pwn 0.5.749 → 0.5.750

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: bf86add502082e14d3299544ffc90b9e86887ada46569fa867d9ce6e0cc8560a
-  data.tar.gz: bb266bdff59a1187f99664eb876dc6b591a43d4f725689025e8af2325b735c00
+  metadata.gz: 60cc9203ea107c4633ec8e0a2ec550fbd4c667c3e8557e8dff360306f786843c
+  data.tar.gz: 637c33f3923027a6d7fbb4fa392e7bee5f56cf2c827fb49bdf5869c64e368c3b
 SHA512:
-  metadata.gz: b5191f1ecd9004cd4152ead0630c3c91df671ecb6f1f7b1c801a8a9983b6d919734b4006dca79740b9dc5be7181880ac5cddd219d343478a4c5cb66c86bad741
-  data.tar.gz: 8ef2550a781242412f81402b29fa5257fcf404fe5a97facd8126544aabe651cd4e26d008cf92bb040b8d675985943a06cac1800a88eb85b2b3c35bb3f79c4692
+  metadata.gz: 608fc78a170ce6fd0c3eb52d41f40c3452fa01b839f9b34f38d42fa65810fdf74954c76c732ae207aa3085522eb42be0af95cd574c9bd96f73e4d53451846bda
+  data.tar.gz: 95549917463adc3b7ac81571915976ed00adc28a6cfafd01e9eefa1ee466be906f07c8a9176d5e058d63597fd40c9a51beded438621fe7e8d1e4575a386d2ba0

data/Gemfile

@@ -49,7 +49,7 @@ gem 'jwt', '3.3.0'
 gem 'libusb', '0.8.0'
 gem 'luhn', '3.0.0'
 gem 'mail', '2.9.1'
-gem 'mcp', '1.6.0'
+gem 'mcp', '1.6.1'
 gem 'meshtastic', '0.0.186'
 gem 'metasm', '1.0.6'
 gem 'mongo', '2.26.0'
@@ -77,7 +77,7 @@ gem 'rbvmomi2', '3.10.0'
 gem 'rdoc', '7.0.4'
 gem 'rest-client', '2.1.0'
 gem 'rex', '2.0.13'
-gem 'rmagick', '7.1.5'
+gem 'rmagick', '7.1.6'
 gem 'rqrcode', '3.2.0'
 gem 'rspec', '3.13.2'
 gem 'rtesseract', '3.1.4'

data/documentation/Blockchain.md

@@ -1,18 +1,116 @@
-# `PWN::Blockchain` - BTC · ETH
+# `PWN::Blockchain` — read-only blockchain intelligence
 
-Lightweight helpers for on-chain recon and wallet interaction.
-Source: `lib/pwn/blockchain/*.rb`.
+Source: `lib/pwn/blockchain/{btc,eth}.rb`. Use each module's `.help` for exact options.
+These modules do not manage keys, sign, broadcast transactions, or modify wallets.
+They report on-chain observations, not real-world identity, ownership, or criminality.
 
-| Module | Purpose |
+## Assessment and changes
+
+Previously, BTC offered chain status, block lookup and an unbounded date-to-transaction-ID
+scan. Its chain status lookup also invoked AI and printed the result. The date scan
+binary-searched block timestamps even though those timestamps are not monotonic.
+ETH offered only two BlockCypher lookups. Both transports disabled TLS verification;
+tests checked only that help/authors existed. Earlier versions of this document
+incorrectly advertised balance, broadcast, and key-management APIs.
+
+Both modules now provide structured, read-only intelligence with validated inputs,
+verified HTTPS, request timeouts, credential-safe errors and deterministic tests.
+BTC no longer invokes AI implicitly. `PWN::AI::Agent::BTC` remains a separate,
+explicit analysis interface; its prose is not independent blockchain evidence.
+
+## Bitcoin Core
+
+BTC uses the existing `PWN::Env[:plugins][:blockchain][:bitcoin]` settings:
+`rpc_host`, `rpc_port`, `rpc_user`, and `rpc_pass`. Optional `rpc_scheme: 'https'`
+enables verified TLS; the existing HTTP default is retained. Use a trusted local
+connection or protected tunnel for HTTP. No prompts or configuration writes occur.
+
+| API | Evidence returned |
 |---|---|
-| `PWN::Blockchain::BTC` | Address balance/UTXO lookup, tx broadcast, key helpers |
-| `PWN::Blockchain::ETH` | Address balance, contract call, event log query |
+| `get_latest_block` | Backward-compatible JSON-RPC chain-status envelope, without stdout or AI |
+| `chain_status` | Chain identity, synchronization, tip and pruning status |
+| `get_block_details` | Block at an explicit height or tip, verbosity 0–3 |
+| `inspect_transaction` | Input/output references, exact integer satoshis, fees only with complete prevouts |
+| `inspect_outpoint` | Current UTXO evidence; null means spent **or unknown**, not proof of spending |
+| `mempool_summary` | Aggregate node-local mempool statistics |
+| `trace_transaction` | Bounded ancestor graph of actual input references, missing evidence and truncation reasons |
+| `scan_transactions` | Paginated explicit-height scan filtered by inclusive UTC header dates |
+| `get_transactions` | Legacy transaction-ID array, only for a complete explicitly bounded scan |
+| `scan_address_activity` | Script-matched received/spent outputs within explicit block heights |
+
+```ruby
+btc = PWN::Blockchain::BTC
+tip = btc.chain_status[:blocks]
+block = btc.get_block_details(height: tip - 6, verbosity: 1)
+tx = btc.inspect_transaction(txid: block[:tx].first, blockhash: block[:hash])
+graph = btc.trace_transaction(txid: tx[:txid], blockhash: block[:hash],
+                              max_depth: 2, max_transactions: 20, max_edges: 100)
+page = btc.scan_transactions(from: '2009-01-12', to: '2009-01-12',
+                             start_height: 160, end_height: 180, max_blocks: 10)
+# Resume with start_height: page[:next_height]; keep end_height fixed.
+```
 
-Also exposed to the agent via `PWN::AI::Agent::BTC` for wallet-aware prompts.
+Migration: `get_transactions(from:, to:)` now also requires `start_height:` and
+`end_height:`. It raises rather than return a silently partial array. For larger
+ranges use `scan_transactions` and inspect `next_height`, `complete`,
+`missing_blocks`, `reorg_detected` and `anchor`. Dates use UTC block-header times,
+not a guaranteed real-world transaction time. Every selected height is examined.
+
+Historical transaction retrieval may require `txindex`; supplying `blockhash`
+helps locate the root transaction but does not locate all its ancestors.
+Pruning, absent undo data and lookup budgets can leave evidence incomplete.
+Address activity is not an address index or a historical/current balance service.
+Ancestor edges prove spending references, not ownership, change, or how mixed
+input value is allocated to outputs. Completeness is limited to the requested
+range or traversal. Scans detect some reorgs, but are not atomic snapshots;
+compare anchors across pages and restart if they change.
+
+## Ethereum
+
+New RPC APIs take an explicit `rpc_url:`; no persistent configuration is required.
+URLs with embedded userinfo are rejected. Endpoint credentials in paths/query
+strings are not included in module errors. Use HTTPS for remote endpoints.
+Existing `get_latest_block(token:)` and `get_block_details(height:, token:)`
+continue to use BlockCypher; they are separate from the JSON-RPC APIs below.
+
+| API | Evidence returned |
+|---|---|
+| `chain_status` | Chain ID, head height, sync state |
+| `block` | Block by number, hash or tag; optional full transactions |
+| `transaction` | Transaction/receipt consistency, pending/success/reverted state, exact wei and gas fees, creation evidence |
+| `account` | Balance, nonce and bytecode pinned to one block hash |
+| `address_activity` | Bounded, whole-block pagination of top-level from/to transactions |
+| `event_logs` | Bounded event query with validated Transfer layouts and explicit provider-completeness caveat |
+| `decode_transfer` | Standard ERC20/ERC721 Transfer layout decoding, not proof of contract conformance |
+| `call` | Read-only ABI-encoded `eth_call`, pinned to a canonical block |
+| `token_metadata` | Optional name, symbol, decimals and total supply, with unavailable ABI fields identified |
 
 ```ruby
-PWN::Blockchain::BTC.balance(address: 'bc1q...')
-PWN::Blockchain::ETH.call(contract: '0x...', method: 'owner()')
+eth = PWN::Blockchain::ETH
+rpc = { rpc_url: 'https://YOUR_ETHEREUM_RPC_ENDPOINT' }
+status = eth.chain_status(rpc)
+block = eth.block(rpc.merge(block: 'finalized', full_transactions: true))[:block]
+tx = eth.transaction(rpc.merge(hash: block[:transactions].first[:hash]))
+state = eth.account(rpc.merge(address: tx[:transaction][:from], block: block[:hash]))
 ```
 
+Pinned account/call/metadata reads require EIP-1898 support. Address scans cover
+top-level transactions, not internal calls, token events, or a full account history;
+their pages are not reorg-safe. Event queries cannot independently establish that
+a provider returned every log, and report `provider_completeness: :unverified`.
+ERC1155 events are not labeled ERC20/ERC721. Token names and event layouts do not
+prove token legitimacy or standards conformance. Fees include execution and blob
+components when supplied, not chain-specific L2 surcharges. Optional metadata RPC
+failures propagate; malformed/unsupported returned ABI layouts are marked unavailable.
+
+## Verification
+
+Default specs use deterministic transport fixtures and need no live credentials.
+Read-only live checks also exercised BTC chain/block/transaction/ancestor/UTXO/date/
+address/mempool paths against the configured unpruned mainnet node, and Ethereum
+chain/block/receipt/pinned-account/activity/Transfer-log/USDC-metadata paths against
+a public mainnet RPC. These checks establish those observed paths, not universal
+provider compatibility or complete attribution. Never put RPC credentials in
+fixtures, documentation, logs, or generated skills.
+
 [← Home](Home.md)

data/documentation/pwn-ai-Agent.md

@@ -10,11 +10,151 @@ so it doesn't repeat it**.
 ## Two ways to run it
 
 ```text
-# 1. Interactive TUI (inside the pwn REPL)
+# 1. Interactive curses console (inside the pwn REPL)
 pwn[CURRENT_VERSION]:001 >>> pwn-ai
-✨ pwn-ai · anthropic · session 20260707_225041_d7f2f3bb
-> Use NmapIt to sweep 10.0.0.0/24, then TransparentBrowser via Burp on any
- host with 443 open, active-scan, and give me a Reports::SAST summary.
+# Or launch directly from the shell:
+pwn-ai
+```
+
+The interactive default is a single-owner curses screen: a provider/model
+and session header, scrollable typed timeline (OPERATOR, TASK, TOOL, RESULT,
+ASSISTANT and WARNING), a persistent multiline composer, and an operational
+sidebar at 100 columns or wider. Narrower screens give the timeline the full width.
+At 100 columns × 26 rows or larger, the header may carve out a bordered retro-game
+ASCII animation on the left (Tetris, Snake, Pong and other 8-bit-style scenes).
+Its framed width in terminal cells equals the complete header's height in rows;
+the interior canvas is `(header height - 2)` cells on each side. This is a
+cell-square, not a pixel-square—terminal glyph cells are usually taller than wide.
+When that interior exceeds the artwork API's 16-cell limit, the complete game
+art is centered and padded inside the larger square rather than clipped or stretched.
+One `PWN::Banner.mini_names` animation is randomly
+selected for the session and retained across redraws, model changes and resize;
+changing the active session selects again. Frames advance at the banner API's
+`MINI_FRAME_SECONDS` cadence using monotonic time in the existing render loop—no extra animation
+thread, input reader or provider call. This is decoration, not progress or
+telemetry. The pane uses the existing border/title/header theme roles. Settings
+wrap in the right-hand region without losing their label colors. A bounded,
+nonrecursive layout pass grows the header and square together. If the terminal
+is narrow/short, or that reduced width would clip any setting, the decoration
+disappears and settings reclaim the full width.
+
+Timestamped entries (`%Y-%m-%d %H:%M:%S%z`), measured request elapsed
+time, completed-tool/event counts, and the last observed tool provide operational
+context. There are no estimated progress bars. Running
+status reflects the request's actual model/tool boundary. Borders are black; the OPERATOR
+label, warnings, and the mission prompt are red. Pane titles (`pwn-ai vN`, `SESSION`,
+`OPERATIONS`, `MISSION`) use `title`, which defaults to red and is separate from
+`border`. The operator's request text is
+white. The header grows to show all wrapped settings without shortening or
+replacing their text. Only when the terminal cannot fit the full header plus
+a one-row timeline, composer and footer does it limit the visible rows; its
+title then points to Ctrl+O (or `/status`) for the complete scrollable settings.
+State remains visible at every supported size. Header setting labels use the
+same `category` color as OPERATIONS labels, including across wrapped lines;
+their values retain the `header` color in both the header and details view.
+Use PgUp/PgDn, arrows or Home/End there; Esc or Ctrl+O restores the unchanged
+draft. The short sidebar prioritizes elapsed time, tools, tokens, cost and last
+tool; Ctrl+O exposes the remaining details even without a sidebar.
+Assistant text is white, tasks green, tools and notices cyan, and results yellow.
+Operations category labels are yellow. Those colors are the default
+`ai.tui.theme` in `~/.pwn/pwn.yaml` (`PWN::Env[:ai][:tui][:theme]`). Named curses
+colors (`black`, `red`, `green`, `yellow`, `blue`, `magenta`, `cyan`, `white`)
+may replace any role; an unknown name keeps that role's default. `NO_COLOR=1`
+keeps labels and Unicode borders without color; `TERM=dumb` or non-TTY input/output
+uses the existing line interface and explicitly reports the fallback.
+
+| Default color | Theme roles |
+| --- | --- |
+| black | `border` |
+| red | `title`, `operator`, `warning`, `prompt`, `selection` |
+| white | `request`, `assistant`, `value`, `composer`, `header`, `footer` |
+| green | `task` |
+| cyan | `tool`, `notice` |
+| yellow | `result`, `category`, `status` |
+
+Fresh configurations and missing roles use these defaults. Existing configured
+colors are preserved by runtime resolution and migration/backfill. This palette
+changes only default values, not the configuration schema; no migration is needed.
+
+| Key or command | Action |
+| --- | --- |
+| Enter | Submit the mission |
+| Shift+Enter, or trailing `\\` + Enter | Insert a newline. Terminals that cannot distinguish Shift+Enter need `tmux set -s extended-keys on` or the backslash fallback. |
+| Up/Down | Move the completion highlight while the Command menu is visible (wrapping at either end). Otherwise recall requests from `~/.pwn/pwn_history`, including prior runs; Down past the newest restores the draft and cursor. Esc closes the menu to resume history recall. |
+| Ctrl+P / Ctrl+N | Alternative completion selection keys; recall history when no menu is visible. Tab still accepts the highlight. |
+| Tab | Accept the highlighted as-you-type parameter |
+| Ctrl+O or `/status` | Inspect full status/settings, including overflowing header text; Esc or Ctrl+O closes |
+| Ctrl+L | Clear only the Session pane, including while a request runs; preserve stored conversation, token totals, and draft |
+| Ctrl+R | Incremental reverse search of request history from `~/.pwn/pwn_history`; repeat for an older match, Enter accepts into the draft without sending, Esc cancels and restores the draft/cursor |
+| `/clear` | Clear the session pane only; stored conversation and token totals remain |
+| `/verbose [on|off]` | Toggle compact or full tool/notice output |
+| Ctrl+G or `/swarm` | Open the draft-preserving swarm workspace; `/swarm dashboard` also opens it |
+| `/` + Enter or `/menu` | Open the slash menu |
+| PgUp/PgDn | Scroll the timeline; new output does not pull a scrolled viewport to the bottom |
+| `/model`, `/sessions resume ID` | Show model settings / change the next request's model or session while idle |
+| `/steer INSTRUCTION` | Redirect the active request at a safe boundary |
+| `/input TEXT` | Send one line to an ordinary tool stdin prompt; not retained in input recall |
+| Ctrl+C | Cancel the active request cooperatively, or clear an idle draft |
+| `back`, `/back`, Ctrl+D | Leave the console; if busy, cancel and wait for safe completion first |
+
+Request recall shares Pry's existing `~/.pwn/pwn_history` (append-only plain-text
+lines), not a separate console history file or only the current transcript.
+Requests are saved through Pry's history owner with its normal duplicate and
+save/ignore settings. Multiline input uses Pry's existing line-oriented format.
+Search and Up/Down never write history; `/input` tool responses are excluded from
+both persistence and recall. Ctrl+L never deletes history or resets token totals.
+
+### Model and reasoning selection
+
+`/model` (also `show` or `status`) displays the current selection. Existing
+`/model list`, `/model list llms`, `/model <engine> [model]` and `/model <model>`
+forms are unchanged. Submitting a supported model opens a **REASONING EFFORT**
+list: Up/Down (or j/k) selects, Enter accepts, Esc/Ctrl+C cancels both changes
+and restores the submitted draft. The current effort is preselected when valid;
+otherwise the catalog default or the existing medium default is used when
+supported. The header updates after acceptance. The legacy interactive line
+interface asks the same question (Enter accepts the default; q/Esc cancels).
+Noninteractive callers use the valid current/default effort without reading stdin.
+
+OpenAI options come from the model catalog's `supported_reasoning_levels` and
+`default_reasoning_level`; catalog discovery happens once on submission, never
+while typing. When metadata is absent, explicit documented GPT-5, GPT-5.1/5.2/
+5.4/5.5 and GPT-6 Astra API contracts provide fallback options. Astra never offers
+`none`. Grok 3 Mini and Grok 4.5/4.6/4.7 use their documented effort levels.
+Unknown models without capability metadata and Anthropic, Gemini, Ollama and
+OpenWebUI do not offer an effort list: those adapters do not consume this effort
+setting (their thinking controls, where present, are different).
+
+Acceptance merges `ai.active`, the selected engine's `model` and its existing
+`reasoning_effort` field into the encrypted vault; other settings are retained.
+Missing decryptor files leave changes session-only. No schema migration is needed.
+Capability references: [OpenAI model pages](https://developers.openai.com/api/docs/models),
+[xAI reasoning](https://docs.x.ai/developers/model-capabilities/text/reasoning).
+
+Settings and new requests are rejected while a request is running, rather than
+mutating its context concurrently. Network-capable local commands (`/mcp`,
+`/cron run`, `/model`) also run off the event thread so `/input` and
+exit remain responsive. `/steer` applies to model requests, not these local
+commands; cancellation waits until a local command returns.
+At less than 48 columns or 14 rows, a compact
+resize notice replaces the panes; cancellation and exit remain available. The
+screen is repainted after a resize. Request history persists through Pry in
+`~/.pwn/pwn_history`. The timeline retains the most recent 2,000 events;
+each displayed event is capped at 16 KiB with an explicit truncation notice.
+Scrollback stays pinned while new output arrives and shows a new-event count.
+The opaque command menu highlights selection even with color disabled. At tiny
+sizes hidden submissions are blocked and the draft is preserved until resize.
+Slash parameters complete after spaces as well as during typing; Up/Down or Ctrl+P/Ctrl+N
+can reach every command, not just the visible menu page. `PWN::` completion lists
+constants without loading each candidate's dependencies, and paths complete
+outside the usual `/tmp` and `/home` roots too. Tab replaces the token at the
+cursor without removing subsequent text. Completion does not fetch provider catalogs.
+
+Describe a concrete outcome in the composer, for example:
+
+```text
+Use NmapIt to sweep 10.0.0.0/24, then TransparentBrowser via Burp on hosts
+with 443 open, active-scan, and give me a Reports::SAST summary.
 ```
 
 ```bash
@@ -22,6 +162,119 @@ pwn[CURRENT_VERSION]:001 >>> pwn-ai
 $ pwn --ai "run bin/pwn_sast against ./src and push findings to DefectDojo"
 ```
 
+## Swarm workspace
+
+Write a mission in MISSION CONTROL, then press **Ctrl+G**. Nothing runs just
+because the workspace opens, an agent is selected, or you navigate. The roster
+shows each persona's role, engine and model (including inherited/default values).
+It uses the existing Swarm registry and backend, not another agent runtime.
+
+| Workspace key | Action |
+|---|---|
+| Tab | Switch roster / session-owned jobs |
+| j / k | Move the highlighted agent or job; outside this overlay, Up/Down select an open Command menu or recall request history |
+| Space | Toggle an agent in the selected set |
+| a | Prepare the current mission draft for the highlighted agent |
+| b | Prepare a broadcast to the explicitly selected agents |
+| d | Prepare a debate among at least two selected agents, in selection order |
+| Enter | Open full agent/job details, or execute a displayed confirmation |
+| s | On a job, compose a separate steering instruction; Enter reviews, Enter again sends |
+| c | On a job, review cancellation; Enter confirms |
+| n | Add a swarm-local persona: name, role, then confirm; no provider call |
+| r | Refresh the local roster; no remote model catalog lookup |
+| PgUp/PgDn, Home/End | Scroll full metadata, results, errors and confirmation text |
+| Esc | Abort a prompt/confirmation, leave details, or return to the mission |
+| Ctrl+G | Close the workspace immediately, leaving the mission draft and cursor unchanged |
+
+Every action is explicit: inspect the action, targets and text before pressing
+Enter on its confirmation. The original draft is never cleared or overwritten
+by a swarm action. A new agent inherits routing defaults; use
+`/swarm spawn NAME ROLE --engine ENGINE --model MODEL --toolsets a,b` for overrides.
+An empty roster offers `n`; an empty jobs tab explains how to launch a mission.
+
+Job details retain observed state, reply/error and the last completed tool for
+this console's lifetime. Result text is redacted for display. Broadcasts visit
+selected personas sequentially in one owned job; debates pass the previous
+speaker's reply into the next turn (one round by default). Separate jobs can
+run concurrently, up to eight active jobs. The workspace does not claim a
+percentage complete or infer success from time elapsed.
+
+Steering and cancellation affect only the selected owned job. Finished jobs
+cannot be steered or revived by cancellation. In the main console Ctrl+C
+cancels its active request and owned swarm jobs; Ctrl+D cancels and waits before
+leaving. Below the minimum terminal size, hidden actions are blocked but Ctrl+C
+and Ctrl+D still work. A running tool finishes cooperatively: neither rollback
+nor remote provider billing cancellation is promised. Swarm jobs should use
+noninteractive tools; their individual stdin prompts are not multiplexed.
+
+Typed `/swarm roster|status|create|use|spawn|retire|ask|broadcast|debate|tail|steer|cancel`
+commands remain available, including parameter completion. These are direct
+operator commands and do not add the workspace's extra confirmation step.
+`/swarm status JOB` includes retained outcomes. Jobs shown here belong to this
+console, not a cross-process durable scheduler.
+
+## Steering a running request
+
+In the interactive native-tool REPL, type a complete line while the agent is
+busy:
+
+```text
+/steer Stop writing the report. Summarize the evidence already collected instead.
+```
+
+This is a **local terminal command**, not Ruby, a model tool, or text appended
+to Pry's busy input buffer. `/help` and TAB include `/steer`. An empty command
+prints usage; at an idle prompt it reports that there is no active request.
+
+* During the protected model call, a scoped cancellation signal stops the
+  local wait and discards the obsolete response. This does not promise that the
+  provider stops remote computation or billing. Setup/planning helpers outside
+  that window finish before the next cooperative checkpoint.
+* During a tool, the notice says **wait until the current tool finishes;
+  already-started work is not undone**. No exception is injected into tool
+  code. Remaining calls in the obsolete batch are marked not executed, with
+  paired tool-result messages, before restarting the loop.
+* Instructions are processed FIFO, only in this request/session. Completed
+  conversation and tool evidence stay available. The latest explicit user
+  instruction takes precedence over conflicting earlier instructions; tool
+  output cannot submit steering.
+* A steer starts a fresh completion scope from the latest instruction, not a
+  keyword-edited reconstruction of the original goal. Include all still-required
+  deliverables in that instruction. Old artifact requirements and a supplied
+  old verification contract are not silently imposed on the revised task.
+  Previously completed work is retained as evidence, not claimed to be undone.
+
+In curses, **only the main event thread draws and reads terminal keys**. One
+request-owned worker runs the real `Loop.run`, inheriting PWN request thread
+locals, and sends output through an event queue. The existing Steering model
+window/checkpoints are reused without starting its canonical stdin reader.
+Ordinary tool stdin is a forwarding pipe: explicitly use `/input TEXT`, never
+send a bare line that could be mistaken for a new request. Each request gets a
+fresh pipe, and cancellation closes it so waiting prompts receive EOF. Ruby
+stdout/stderr and the debug tee are captured for the console lifetime, sanitized,
+and restored on exit; spinner control sequences are not displayed as notices.
+Trace logging remains available, but **ENTER-to-step is disabled** so it cannot
+compete for input. Task/tool/result/final rows remain mirrored into request logs.
+
+`Ctrl+C` stops a model wait or requests cancellation at the next tool boundary.
+It never asynchronously raises into a side-effecting tool. `back` and Ctrl+D do
+the same and keep the UI alive in a closing state until the owned request ends.
+The console joins its own worker; it does not kill unrelated threads. A tool
+that ignores EOF and has no timeout can delay exit until it finishes. Already
+launched durable jobs have their own lifecycle; cancellation does not undo them.
+
+Fullscreen/raw-terminal tools, programs opening `/dev/tty` directly, and code
+writing directly to OS terminal descriptors (rather than captured Ruby output
+or a tool's returned result) are not supported inside curses. Run those outside
+the agent. Stdin temporarily reports non-TTY. Do not expect local cancellation
+to stop remote provider computation or billing.
+
+The non-TTY/dumb-terminal legacy line interface retains its foreground Loop and
+single canonical steering reader: there, non-command lines are forwarded to
+ordinary stdin prompts, and EOF ends the reader. Both paths restore the input
+descriptor, its flags, terminal mode and output on cleanup. Restart the REPL
+after upgrading these files.
+
 ## Anatomy of a turn
 
 1. **PromptBuilder** assembles the system prompt: your request + **engine-budgeted

data/etc/default_skills/pwn/banner/SKILL.md

@@ -12,7 +12,7 @@ metadata:
 
 # PWN::Banner
 
-This file, using the autoload directive loads Banner modules into memory only when they're needed. For more information, see: http://www.rubyinside.com/ruby-techniques-revealed-autoload-1652.html
+Static banners and pure, PWN-branded retro ASCII loops: falling_blocks, snake, and pong. mini_frame returns fresh rows on a borderless canvas; use equal width and height (5..16) for complete square artwork. Dimensions clamp to 0..16, with smaller panes degrading to a wordmark. The caller samples mini_names once per session and advances the explicit frame index every MINI_FRAME_SECONDS (0.1), wrapping at MINI_FRAME_COUNT (60). These are decorative animations, not interactive games or telemetry. Static Banner modules autoload only when needed. For more information, see: http://www.rubyinside.com/ruby-techniques-revealed-autoload-1652.html
 
 ## When to use
 
@@ -28,11 +28,13 @@ Class methods take `(opts = {})` and read `opts`.
 
 ```ruby
 PWN::Banner.help
-PWN::Banner.get(opts)
+PWN::Banner.mini_frame(opts)
 ```
 
 ## Public methods
 
+- `mini_frame`
+- `mini_names`
 - `get`
 - `welcome`
 - `authors`
@@ -48,5 +50,5 @@ PWN::Banner.get(opts)
 
 ## Verification
 
-`PWN::Banner.respond_to?(:get)` after the
+`PWN::Banner.respond_to?(:mini_frame)` after the
 module is loaded. Read the source for parameter names.

data/etc/default_skills/pwn/blockchain/btc/SKILL.md

@@ -12,7 +12,7 @@ metadata:
 
 # PWN::Blockchain::BTC
 
-This plugin interacts with BitCoin's Blockchain API.
+Read-only Bitcoin Core intelligence. No wallet, signing or broadcast RPCs.
 
 ## When to use
 
@@ -34,15 +34,18 @@ PWN::Blockchain::BTC.get_latest_block(opts)
 ## Public methods
 
 - `get_latest_block`
+- `chain_status`
 - `get_block_details`
+- `inspect_transaction`
+- `inspect_outpoint`
+- `mempool_summary`
+- `trace_transaction`
+- `scan_transactions`
 - `get_transactions`
+- `scan_address_activity`
 - `authors`
 - `help`
 
-## References
-
-- `references/urls.md` — URLs from source
-
 ## Source
 
 `pwn/blockchain/btc.rb`

data/etc/default_skills/pwn/blockchain/eth/SKILL.md

@@ -12,7 +12,7 @@ metadata:
 
 # PWN::Blockchain::ETH
 
-This plugin interacts with BitCoin's Blockchain API.
+Read-only Ethereum intelligence via BlockCypher and explicit JSON-RPC endpoints.
 
 ## When to use
 
@@ -28,11 +28,20 @@ Class methods take `(opts = {})` and read `opts`.
 
 ```ruby
 PWN::Blockchain::ETH.help
-PWN::Blockchain::ETH.get_latest_block(opts)
+PWN::Blockchain::ETH.decode_transfer(opts)
 ```
 
 ## Public methods
 
+- `decode_transfer`
+- `event_logs`
+- `address_activity`
+- `call`
+- `token_metadata`
+- `block`
+- `account`
+- `transaction`
+- `chain_status`
 - `get_latest_block`
 - `get_block_details`
 - `authors`
@@ -48,5 +57,5 @@ PWN::Blockchain::ETH.get_latest_block(opts)
 
 ## Verification
 
-`PWN::Blockchain::ETH.respond_to?(:get_latest_block)` after the
+`PWN::Blockchain::ETH.respond_to?(:decode_transfer)` after the
 module is loaded. Read the source for parameter names.

data/etc/default_skills/pwn/blockchain/eth/references/urls.md

@@ -1,3 +1,3 @@
 # PWN::Blockchain::ETH source links
 
-- https://api.blockcypher.com/v1/eth/
+- https://api.blockcypher.com/v1/eth/[redacted]

data/etc/default_skills/pwn/plugins/repl/SKILL.md

@@ -51,6 +51,7 @@ PWN::Plugins::REPL.ready_tty(opts)
 - `persist_ai_selection`
 - `persist_mesh_env`
 - `pwn_ai_activation_session`
+- `pwn_ai_apply_model`
 - `pwn_ai_complete`
 - `pwn_ai_complete_command`
 - `pwn_ai_complete_kind`
@@ -63,7 +64,9 @@ PWN::Plugins::REPL.ready_tty(opts)
 - `pwn_ai_memory_command`
 - `pwn_ai_model_ids`
 - `pwn_ai_profile_command`
+- `pwn_ai_prompt_reasoning`
 - `pwn_ai_provider_class`
+- `pwn_ai_reasoning_selection`
 - `pwn_ai_run_cron`
 - `pwn_ai_run_learning`
 - `pwn_ai_run_mcp`
@@ -71,6 +74,7 @@ PWN::Plugins::REPL.ready_tty(opts)
 - `pwn_ai_run_model`
 - `pwn_ai_run_sessions`
 - `pwn_ai_run_skills`
+- `pwn_ai_run_steerable`
 - `pwn_mesh_complete`
 - `pwn_mesh_dispatch_slash!`
 - `pwn_mesh_menu_rows`

data/etc/default_skills/pwn/plugins/repl/ai/SKILL.md

@@ -37,6 +37,10 @@ PWN::Plugins::REPL::AI.add_commands(opts)
 - `authors`
 - `help`
 
+## References
+
+- `references/urls.md` — URLs from source
+
 ## Source
 
 `pwn/plugins/repl/ai.rb`

data/etc/default_skills/pwn/plugins/repl/ai/references/urls.md

@@ -0,0 +1,5 @@
+# PWN::Plugins::REPL::AI source links
+
+- https://developers.openai.com/api/docs/models/gpt-6-astra
+- https://developers.openai.com/api/docs/models/gpt-5
+- https://docs.x.ai/developers/model-capabilities/text/reasoning

data/etc/default_skills/pwn/plugins/repl/ai_console/SKILL.md

@@ -0,0 +1,48 @@
+---
+name: pwn-plugins-repl-aiconsole
+description: Drive PWN::Plugins::REPL::AIConsole from pwn_eval.
+license: MIT
+allowed-tools: [pwn, pwn_eval]
+metadata:
+  bundled: true
+  generated: true
+  module: PWN::Plugins::REPL::AIConsole
+  source: pwn/plugins/repl/ai_console.rb
+---
+
+# PWN::Plugins::REPL::AIConsole
+
+Single-owner fullscreen agent console. Workers never paint the terminal.
+
+## When to use
+
+Call `PWN::Plugins::REPL::AIConsole` from `pwn_eval` when the task needs this module.
+Do not reimplement it in shell.
+
+## Methodologies
+
+Generated from `pwn/plugins/repl/ai_console.rb`. Prefer the public class methods below.
+Class methods take `(opts = {})` and read `opts`.
+
+## How to call
+
+```ruby
+PWN::Plugins::REPL::AIConsole.help
+PWN::Plugins::REPL::AIConsole.run(opts)
+```
+
+## Public methods
+
+- `run`
+- `theme`
+- `authors`
+- `help`
+
+## Source
+
+`pwn/plugins/repl/ai_console.rb`
+
+## Verification
+
+`PWN::Plugins::REPL::AIConsole.respond_to?(:run)` after the
+module is loaded. Read the source for parameter names.