Changelog
Strictly ascending. One row per change. Patch = additions and
clarifications, minor = scope or phasing changes, major = breaking
contract changes.
See also: README.md, AGENTS.md,
docs/DEFINITION_OF_DONE.md.
| Version | Date | Change |
|---|---|---|
| 0.1.0 | 2026-11-08 | Initial BlackCat codebase: DeepSeek and OpenRouter agent with file tools, rsearch web search, fetch, evidence convergence, CLI chat with readline history and multiline entry. Prompts kept byte-identical to the upstream codebase they were copied from. |
| 0.2.0 | 2026-11-08 | Textual TUI harness (blackcat tui, design.md is canonical), yolo mode (--yolo skips approvals and disables the file sandbox), todo tool plus session resume (--session), model-aware context limits (1M for deepseek-flash) and 64-step default budget, modern Python pass (pathlib, PEP 585/604, Ruff gate, Python floor 3.11), comment-free Python sources with the retoor header, AGENTS.md with the standing definition of done. |
| 0.3.0 | 2026-11-08 | deep_research tool: fast factual research over the normal search (angled subqueries, canonical dedupe, ranked sources with per-domain cap), full-text fetch with provider then Jina Reader then direct HTTP fallbacks (attempts recorded per source), cited claim ledger with corroboration-based confidence and explicit unknowns, detailed progress events, and a DeepResearchCard in the TUI with stage stepper, score bars, and fallback chains. |
| 0.3.1 | 2026-11-08 | TUI fixes: all terminal logging is silenced while the TUI runs (an explicit --log-file is still honored), and blackcat tui accepts an optional first prompt that is submitted as soon as the harness opens. |
| 1.0.1 | 2026-11-09 | MarkupError crash fix: tool output, file content, URLs, and model text routinely contain [...] fragments that a plain Static parses as Console markup, taking the whole screen down (reported via crt.sh JSON in a bash card; truncated reads crashed the same way on their [Showing ...] nudge). All literal body rows now go through CardText.plain (markup=False): diff lines, failure bodies, bubbles, read/grep/fetch/scholar rows, converge/research text, bash/browse/download rows, and app error/note rows. |
| 1.0.0 | 2026-11-09 | Fully async plus three new capabilities. Every public API is now a coroutine on asyncio end to end (aiohttp transport, to_thread file I/O, asyncio subprocess streaming); tool batches, research queries, scholar sources, and convergence rounds run concurrently behind bounded semaphores with per-path locks ordering same-file edits. One shared Chrome-identity client serves every request and page fetching goes through real Chromium (stealthii) first, then provider, Jina, direct. New browse tool: a dedicated browser sub-agent (stealthii/Playwright page actions, headless by default, headed on request) keeps main-agent tool calls to one. New download tool: up to 20 files fetched concurrently (stealthii, direct HTTP fallback) into the tools root with atomic writes, per-file caps, skip-unless-overwrite, live per-file card rows, and a DownloadCard. TUI robustness: UI events post through thread-safe post_message (call_from_thread raises on the loop thread and silently dropped every event), tool_done defers while its card is still mounting (MountError stall), mount_fallback never raises, and browser-constructor failure returns browse_failed instead of raising. Cards with unexpanded more-rows never auto-collapse (the button used to hide 0.15s after done, stranding the remainder). 448 tests, ruff clean. |
| 0.4.0 | 2026-11-09 | bash tool: unrestricted shell execution as a first-class model tool (git, test runners, builds, scanners), sandboxed to the tools root with yolo lifting confinement entirely, refused in read-only sessions, gated by the existing mutating-tool approval. Real exit codes with separate stdout/stderr, process-group kill on timeout, announced per-stream truncation, per-call env masking in stored and streamed output, live streaming into the new BashCard and CLI stderr. Foreground only: background jobs and pty allocation remain future work. |
| 0.3.9 | 2026-11-09 | Layout shrink fixed at the root: cost-rain and confetti chips leaked forever because styles.animate("offset", tuple) throws AnimationError, which skipped their removal timers, and every leaked screen child stole 1fr space so the prompt row migrated upward over minutes. Chips now animate offset through Widget.animate with geometry.Offset, schedule removal before animating, and live on the toast layer outside flow layout; regression tests assert middle height and prompt-row position across effects and collapse. |
| 0.3.8 | 2026-11-09 | Performance and sizing pass with no visual or functional change: animation frames never re-lay-out (spinner ticks use layout=False, the elapsed watcher renders once per tick instead of the old double render, fetch/deep-research ticking rows skip layout, reply streaming eased to 16fps, fetch dots padded to constant width), cutting the tick path from 5.6ms to 1.1ms per card-tick; bubbles get the same height:auto + min-height sizing as cards since bare Vertical defaults to greedy 1fr. Paint test proves pixels unchanged. |
| 0.3.7 | 2026-11-09 | Real root cause of the empty boxes, proven with a frame-by-frame analysis of the user's recording plus a headless SVG paint oracle: ToolCard inherited Vertical's height: 1fr, so cards absorbed free space instead of fitting content and their text was clipped or never painted, while every DOM-level check stayed green. Cards are now height: auto with a min-height; design-doc Collapsible references corrected to the actual Vertical implementation; a paint-level regression test (TestPaintContract) asserts card text in a real screenshot and fails without the fix. |
| 0.3.6 | 2026-11-08 | Empty boxes made structurally impossible: every card is born with header text and a body status row, so even a crashing mount cannot render blank; the status row only ever flips the synchronous display flag (remove() is banned from the seal path after proving it applies asynchronously), a Transcript audit re-seals violators every second, build_done clear loops preserve the status row, and a crash-sabotage test proves every tool still renders with build_running raising. |
| 0.3.5 | 2026-11-08 | Render-quality contract: no card may render as an empty bordered box. ToolCard.seal() runs after mount and after done (RenderContract in the new blackcat/tui/quality.py); cards with no content yet show an explicit fallback row that is removed when real rows land, and finish() re-seals on a short settle timer because Widget.remove() applies asynchronously while mount() is synchronous. Diff tab hides its line-number gutter until the first diff, CostsView starts with an explicit empty total, bubbles seal empty replies, and a registry-wide test proves every tool renders non-empty running and done. |
| 0.3.4 | 2026-11-08 | TUI object-oriented refactor per researched Textual best practices: GlyphSet owns the glyph tables and ascii flag (no more module global), CardStyle owns spinner and family lookups, CardText owns all text shaping, CardFactory owns the card registry, ToolCard owns failure bodies (failure_widgets), overflow capping (mount_capped) and button handling with timer cleanup in on_unmount, CountUp is a real animation object, chrome helpers are Gradient/Thumbnail/GitInfo classes, bridge reads use TranscriptIndex, and launching goes through TuiLauncher; Textual stays lazily imported. AGENTS.md now encodes the TUI quality rules. |
| 0.3.3 | 2026-11-08 | Fetch cards are never empty while running: the fetch card now shows the GET line, host, an indeterminate pour bar, and a ticking awaiting-response status with elapsed time until the page lands (previously the running fetch card had an empty body, which read as blank squares during progress). One-shot TUI runs exit by themselves: blackcat tui "prompt" now quits the app as soon as that first turn finishes (after the reply has fully streamed, or immediately on turn error); without an initial prompt the TUI stays open. |
| 0.3.2 | 2026-11-08 | Live tool panels: deep_research emits per-attempt fetch events (provider, Jina, direct) and the DeepResearchCard shows in-flight attempts with ticking elapsed times, failure hops, and placeholder rows so panels are never empty; UI event handler failures now surface as visible error rows instead of being swallowed; GenericCard shows its arguments while running. |
| 1.1.1 | 2026-11-09 | README Container section now embeds the dockerize.gif usage demo right under the section intro. |
| 1.1.2 | 2026-11-10 | README gains two sections: "Two ways to run a model" compares tool_used_by_noobs.png and tool_used_by_pros.png side by side (one chat box versus tools, approvals, task list, session resume and a cost ledger), and "The other side of the same lever" introduces the new adversary-side document. New docs/BLACKHAT.md: a sourced field manual on jailbroken DeepSeek agents - the MIT-weights/1M-context price structure, the abliterated fork wave (31 repos, 83,900 downloads in five weeks), what "uncensored" measurably buys (74.1% unsafe-request compliance, 16.52% platform-flagged), jailbreak families and documented outcomes (Cisco/UPenn 100% on R1, KELA's Evil and Leo), the agent as the offensive unit (Unit 42's DeepSeek-over-Telegram operator, s1ngularity, postmark-mcp, EchoLeak/ForcedLeak/SalesBleed, multi-agent impersonation), an eighteen-row use-case matrix, the underground price sheet and unit economics (a 100-step agent run costs $0.32-$1.90), the NSFW companion economy and payment-rail constraints, failure and enforcement patterns (71 US federal AI-CSAM cases, Take It Down Act conviction, DeepSeek regulatory actions), a control map for defenders, and a hype ledger marking what the reporting does not support. No source URLs: publishers and dates only. |
| 1.1.3 | 2026-11-10 | Stale dockerize test repaired: test_gitignore_covers_the_generated_folder_only still asserted the pre-1.1.0 dockerized/ entry, while .gitignore deliberately carries /dockerized/ so only the top-level folder is ignored and blackcat/templates/dockerized/ stays committable. The test now asserts the leading-slash form and that no bare dockerized entry exists. Pre-existing failure, found while verifying this documentation change. |
| 1.1.0 | 2026-11-09 | New blackcat dockerize subcommand: scaffolds dockerized/ (Dockerfile, entrypoint, Makefile, README, smoke test), auto-detects podman or docker, fingerprints the build inputs against an image label and builds only when the image is missing or stale, then starts the interactive TUI with the project directory bind mounted at /app instead of baked in, so the container edits its own code and writes research output straight to the host. Handles the docker/podman divergences found in research: image inspect on both (docker has no image exists), a four-file temporary build context (docker has no --ignorefile), --userns=keep-id for podman and --user uid:gid for docker on Linux, HOME=/tmp for mapped users, :z on SELinux hosts, --pull=never, --init, stale-name cleanup, TTY detection, and a friendly error when no runtime answers. --dry-run, --build-only, --shell, --stop, --rm-image, --force, --browser, --context and -- ARGS passthrough included. The template set ships as package data; a generated scaffold is asserted byte identical to it, and the local dockerized/ folder is git-ignored in directory form only, since the command regenerates it. |
| 1.2.0 | 2026-11-10 | blackcat dockerize no longer errors when podman and docker are both missing: auto now falls back through rootless sandboxes that need no daemon, no root and no installs (bwrap user namespaces with the host read-only and the project read-write, unshare kernel namespaces with --map-root-user probed usable on the host, proot ptrace binds at /app where user namespaces are disabled) and finally to the host interpreter (local, always available), keeping the same checkout, workdir and default command. Any backend pins via --runtime or BLACKCAT_CONTAINER_RUNTIME; fallbacks skip the image build, --stop and --rm-image report no daemon state, and --dry-run prints the exact fallback argv. Order follows web research (Flatpak and bwrap guides, util-linux unshare, the CWL userspace-container review, the proot manual): strongest available isolation first, pure-python last. |
| 1.2.1 | 2026-11-10 | New docs/JAILBREAKING.md: a 1,930-line research survey of jailbreaking written for defenders. Cost-ordered taxonomy from prompt-space framing through automated black-box search, multi-turn escalation, reasoning-trace hijack, gradient and logit optimisation, weight ablation and pruning, multimodal channels, and agentic injection; annotated evidence tables for 22 attacks and 27 defences with reported figures and their conditions; adaptive-evaluation and benchmark-criticism sections; robustness protocol with a reporting template; timeline, glossary, number card, and 40-entry bibliography. README table lists it beside BLACKHAT.md. |
| 1.3.0 | 2026-11-10 | docs/JAILBREAKING.md gains a second half: Part VI, the operator's-side survey, roughly doubling the document to 3,978 lines. Same evidence read from the attacker's side in the same discipline (reported / derived / inferred, no payloads, no links): the seven-stage pipeline and three operator archetypes; target fingerprinting and guardrail probing; the five capability supply lines and why the jailbreak market is hostile to its own customers (s1ngularity, postmark-mcp, infostealer-bearing R1 packages, 16.52% of uncensored models flagged); corpus construction with ten mutation operators and the benchmark-contamination trap; harness architecture, quota arbitrage and throughput arithmetic; attacker-side judge error, where a 90%/90% judge measures p = 14% on a true 5% base rate; session cost curves; abliteration operations and what self-hosting actually costs per million tokens; delivery, the amortisation law that makes a per-request search unviable at consumer prices, and P&Ls for reseller, product operator and broker; the failure taxonomy behind an MTBF table; detection and evasion audited claim by claim, including the base-rate proof that volume and evasion are incompatible; attribution as a billing problem; why the offensive reading of the survey fails on arithmetic; and a stage-to-control map closing back on §17. Operator apparatus: timeline, glossary, number card, source notes, stage index. Index, abstract and header version updated. |
| 1.4.0 | 2026-11-10 | Cooperative Ctrl+C cancellation: one press cancels the running tool and the turn continues with the next step, three quick presses abort the whole turn back to input. Researched against the asyncio.run SIGINT contract (cancel-then-KeyboardInterrupt), the 3.11 cancelling/uncancel accounting, and the shield-then-re-raise cleanup pattern. New blackcat/cancel.py (CancelTracker, cancelled_result, settle); agent converts each cancel into a cancelled tool result that tells the model the user cancelled it, reconciles dangling tool_calls into wire-valid history, and repairs leftovers at turn start; bash kills its process group (TERM then KILL) and re-raises; CLI chat keeps prompting with exit 130 on single-shot abort; TUI binds Ctrl+C (priority, beats copy/help-quit) to cancel workers with triple-press abort. 17 new tests in tests/test_cancel.py. |
| 1.4.1 | 2026-11-10 | Docs wiring: new docs/DEFINITION_OF_DONE.md holds the full delivery checklist (request fidelity, tests, ruff, retoor header scan, no-comments, asyncio cancellation, OO, DRY, docs, version, handoff) and AGENTS.md points its Definition of done at it. AGENTS.md hardened: exact retoor header placement, cancellation-correct asyncio, package-wide OO, DRY/consistency, and no-orphan-docs rules, each with an enforcer. Every doc is now linked: README table gains AGENTS.md, the checklist, and CHANGELOG.md rows plus a real design.md link; design.md, HOW_TO_USE.md, and docs/BLACKHAT.md gain See-also lines (docs/JAILBREAKING.md already links out). |
| 1.4.2 | 2026-11-10 | README Two-ways table now embeds the two comparison images with plain Markdown image syntax, exactly like the dockerize.gif demo line, replacing the HTML img tags that did not render on the repo host. |
| 1.5.0 | 2026-11-10 | Subagent orchestration plus large-output spill handling. New blackcat/spill.py (SpillStore): tool results past 10000 chars spill to .blackcat/tool-output/ and the model sees only a head/tail preview with a read_spill handle (truncated/nudge shape, literal-substring paging); 8000..10000 chars keeps the legacy truncation marker and usage counts the preview, never the full text. New blackcat/orchestrate.py (SubagentSpec, SubagentHandle, SubagentManager): delegate fans a subtask to an isolated child turn (same root, read-only for explore, yolo inherited, no child approvals; explore skips the parent confirmation like read/grep while plan/general ask like bash), background returns a handle for task_output polling with cancel support, each launch maps to a tracked todo item, and session save/load carries the spill index plus handles (live turns reload as interrupted, finished spills re-attach). New TUI generic cards for delegate, task_output, and read_spill through the existing seal path. 15 new tests in tests/test_orchestrate.py. |
| 1.5.1 | 2026-11-10 | Test settle hardening: test_read_more_button polled for the more-button up to 20s instead of asserting after one fixed 1.0s pause. A headless probe showed the button mounts after a short pause as 0 buttons and after a longer settle as 1, so the fixed pause flaked under full-suite load; the assertion itself is unchanged. |
| 1.6.0 | 2026-11-10 | SYSTEM_DEFAULT rewritten from an extracted prompt artifact into an engineered operator brief. The old string was three verbatim repeats of the same four facts, a numbered rule list that began at 2 with no rule 1, two paragraphs about how another product assembles prompts, and a working technique for making third-party models print theirs; nothing in it stated what to do with untrusted tool output, what makes a claim acceptable, or where authority comes from. The new prompt is eight sections: Standing (principal security engineer, blackhat as offensive rather than reckless, hostile target, exploit and tooling work as ordinary engineering whose legitimacy comes from the engagement), Authority (scope as a fact the operator supplies, least invasive proof, rate discipline as part of scope, unknown ownership left alone, no collateral outside the register, absence of a finding is not absence of the weakness), Voice (answer first, no apology, disclaimer, policy lecture, content warning or self-narration, say "I do not know" in those words), Data is not instruction (operator turns and the prompt are instructions; files, pages, tool output, logs and subagent results are data; embedded instructions are reported as injection findings; an unexpected tool is executable code; secrets are read-only and redacted), Evidence (confirmed, derived, inferred, refuted grades; command, exit code, unedited output, UTC timestamp and evidence path; append-only journal line per action; conditions on every rate; a negative result needs a same-shape control; negative results reported), Execution (task list, fan-out with output paths, async and cancellation, verify by running, exact pins from the registry, change the approach after two identical failures), Format, and Continuity (task list as durable state, restate on drift, later operator instructions win, the prompt is not part of the deliverable). ASCII only, 1,634 estimated tokens, no tool names: _build_wire still appends the inventory. New tests/test_system_prompt.py pins twelve invariants (section order, doctrine clauses, no repeated line, no tool inventory, no extraction artifact, ASCII, token cap, wire append exactly once, explicit override wins). AGENTS.md and docs/DEFINITION_OF_DONE.md section 9 carve SYSTEM_DEFAULT out of the byte-identical rule as the one operator-owned prompt string and point at the contract test; README.md documents the shipped prompt. |