graff — a fast agentic coding harness in Zig: multi-provider, MCP, workflows, DGM evolution loop, TS/Python SDKs
Python
//! Build-time source for --help and --version. Keep all release notes intact.
/// Shown under `graff --version` — a terse "what's new" for recent releases.
/// Keep it short and current; bump alongside the version each release.
pub const changelog_text =
\\What's new
\\──────────
\\0.302.74.1
\\ • a provider or model switch expands opaque compaction state from its checkpoint instead of replaying it
\\
\\0.302.74
\\ • rejected compaction state restores exact input when a private checkpoint is available
\\ • short TLS-construction outages get a longer bounded recovery window
\\
\\0.302.71
\\ • computer use on macOS: graff ships nanocua, an app that lets agents work in other apps
\\ • graff computer and /computer: status, stop, switches; turning input on takes your click
\\ • Codex's computer-use plugins are no longer loaded
\\
\\0.302.70
\\ • Claude Agent: a reply that starts during a tool-list refresh no longer fails
\\ • simulator tests can't overwrite your clipboard; large uploads aren't cut off
\\ • an empty-reply failure says what happened and how to retry
\\
\\0.302.69
\\ • reading peer mail no longer rewrites history the model was already sent
\\ • graff login kimi works for kimi.ai (international) accounts; sign-in links are copied
\\ • workers can use gateway models; graff harness-tree shows the harness as one tree
\\
\\0.302.68
\\ • Claude Agent: compaction writes a summary instead of continuing the task
\\ • Claude Agent: a prompt after a finished turn keeps its images
\\
\\0.302.67
\\ • Claude Agent: loading a tool no longer drops the call or ends the turn
\\ • a finished task still hears peers who message it directly
\\
\\0.302.66
\\ • a one-shot answer built on hosted search is kept, not bounced and replaced
\\
\\0.302.65
\\ • Claude Agent models use their full 1M context window (was listed as 200k)
\\
\\0.302.64
\\ • an expired Console sign-in is skipped, and the error says to run graff login console
\\
\\0.302.63
\\ • auth errors say which credential was rejected: the sign-in, or the API key (and when it was the fallback)
\\
\\0.302.62
\\ • rlm's llm_query works on Claude Agent, on a throwaway claude process
\\
\\0.302.61
\\ • Claude Agent: Claude models through your own Claude Code login (graff login claude-agent)
\\ • subagent requests keep agent_output in lean mode, so the turn waits for its workers
\\
\\0.302.60
\\ • a resumed session keeps its real context size instead of re-estimating compaction state
\\ • large compaction items work over WebSocket without re-running the turn
\\
\\0.302.59
\\ • steadily arriving partial SSE events no longer trigger false stalls or replay
\\ • genuine stalls still time out and unfinished events remain bounded
\\
\\0.302.58
\\ • ACP startup and model discovery no longer wait on slow model-catalog hosts
\\ • note_constraint states its 240-byte limit up front
\\
\\0.302.57
\\ • images are typed by their bytes, so a mislabelled file no longer wedges a chat
\\ • resending a prompt after a failed request makes a new request
\\ • stream error events surface the provider's error instead of retrying
\\ • withdrawn checklist items can be cancelled with a reason
\\
\\0.302.56
\\ • native file tools follow dependency symlinks that stay inside the workspace
\\ • escaping, dangling, and cyclic links remain blocked
\\
\\0.302.55
\\ • help, release history, and bundled playbooks are compressed until needed
\\ • long home paths keep the macOS background-launch protection
\\
\\0.302.54
\\ • normal terminal startup is quiet; diagnostics stay opt-in
\\ • live thinking shows Ctrl+T collapse/expand beside its header
\\
\\0.302.53
\\ • smaller binaries share prompt text and compress runtime resources
\\ • release builds keep safety checks and leave local learning optional
\\ • workflow synthesis keeps its catalog at the final report boundary
\\
\\0.302.52
\\ • clearer terminal status, Markdown emphasis, and checklist styling
\\ • /workers opens a read-only browser of subagent activity and reports
\\
\\0.302.51
\\ • Kimi API keys and logins work again (no more RstStream)
\\ • a used-up Kimi plan window stops instead of retrying
\\ • retrying a prompt after a connection failure works again
\\ • subagents and workflows can fork the conversation
\\
\\0.302.48
\\ • oMLX is a built-in local provider: --model omlx
\\ • compaction counts opaque state by the provider's own measure
\\ • a stopped local server is named in the failed turn
\\
\\0.302.47
\\ • compaction waits instead of repeating when it can't make room
\\ • the shared-tree checkpoint lets a re-issued call through
\\ • fyi peer mail stays in the inbox without starting a turn
\\
\\0.302.46
\\ • images over ACP reach the hosted default models
\\ • a worker keeps its report when its state can't be saved
\\ • a client's own cancel isn't called the user's
\\
\\0.302.45
\\ • ACP context meter reports what graff kept out of context
\\ • render_html says a page is saved, not displayed
\\ • workspace use: instructions follow the tree you switch to
\\
\\0.302.43
\\ • updated subscription context capacity avoids premature compaction
\\ • model catalog and ACP usage updates report the same window
\\
\\0.302.42
\\ • output-limit recovery keeps later requests and saved history valid
\\ • ACP subagent activity shows the resolved worker model
\\ • current main plus session, MCP, compaction and parked-command fixes
\\
\\0.302.41
\\ • native subscription sign-in across terminal and ACP clients
\\ • MCP structured results, resource content and inbound ping handling
\\ • request-local loopback deadlines and repeatable fixture comparisons
\\
\\0.302.40
\\ • stream recovery preserves partial replies across ACP and resume
\\ • workspace skills in headless runs; actionable claim inventory
\\ • MCP off-lists and JSON configuration listing; atomic companion updates
\\ • stored conversation continuation and typed sign-in recovery
\\
\\0.302.39
\\ • unattended runs recover from output limits and empty replies
\\ • one-shot stream-json reports a resume handle, outcome and usage
\\ • exhausted budgets and account limits stop with distinct exit codes
\\
\\0.302.38
\\ • Windows: faster start, and the companion is found on PATH
\\
\\0.302.37
\\ • ChatGPT sign-in keeps your saved token out of the sign-in URL and always prints it
\\
\\0.302.36
\\ • new worktrees link .agents/linked files and run .agents/prepare
\\
\\0.302.35
\\ • gateway models are priced from the gateway itself; new models show their cost
\\ • graff models shows each gateway model's real rate
\\ • a model without its provider's login names --model codegraff/<model>
\\
\\0.302.34
\\ • Gemini API key sign-in works again
\\ • Gemini 4 Argon shows its cost; /effort minimal runs as low on Gemini
\\ • a gateway timeout mid-answer is retried instead of ending the turn
\\
\\0.302.33
\\ • Claude: a new repo reads graff's tools and instructions from the prompt cache
\\
\\0.302.32
\\ • a yield_turn handoff shows once in ACP clients
\\ • graff login anthropic works from the desktop app, SSH and servers
\\ • graff tui installs its client on first use
\\
\\0.302.31
\\ • tool rows for todo_write, peer_message and schema loads finish in ACP clients
\\ • graff update works on Windows
\\
\\0.302.30
\\ • graff mcp status / reconnect / enable / disable, and /mcp reconnect in a session
\\ • graff login anthropic: use an Anthropic Console account
\\ • Claude Haiku 5.5 replaces Haiku 4.5
\\
\\0.0.302.29
\\ • streamed requests to the Codegraff gateway carry graff's version too
\\
\\0.0.302.28
\\ • the Codegraff usage page shows which graff version and device sent each request
\\
\\0.0.302.27
\\ • choose where compaction starts: /compact-at 70, --compact-at, or the desktop setting
\\
\\0.0.302.26
\\ • compaction replies no longer show up in the desktop chat
\\ • a question no longer restarts an earlier task's checklist
\\ • a model switch across providers keeps the conversation
\\ • computer-use and closed stdio MCP servers recover
\\
\\0.0.302.25
\\ • ChatGPT/OpenAI sessions no longer compact before every tool call
\\ • OpenAI routes compact only with the provider's own compaction
\\ • fewer hosted-model requests per task; empty tool results no longer break a chat
\\
\\0.0.302.24
\\ • Codegraff sign-in: models on the Messages wire use it natively
\\ • goals set from the desktop app start immediately; unfinished work survives a question
\\
\\0.0.302.23
\\ • ChatGPT-plan sign-in caches the prompt again, on every model
\\
\\0.0.302.22
\\ • /import-claude brings this project's Claude Code conversations to /resume
\\ • imported conversations continue on your graff provider and credentials
\\
\\0.0.302.21
\\ • Windows: irm https://codegraff.com/install.ps1 | iex installs graff natively
\\ • graff login prints the link with your code; Windows opens the browser
\\ • an expired ChatGPT sign-in reopens in the browser and the request continues
\\
\\0.0.302.20
\\ • includes everything from 0.0.302.19 (not published on its own)
\\ • one oversized MCP message no longer breaks every later call to that server
\\ • Windows: an empty HOME no longer hides your profile directory
\\ • GRAFF_COMPACT_MIX=1: experimental two-pass compaction
\\
\\0.0.302.19
\\ • tool results inline up to 64 KiB: large files read in one piece
\\ • no false stalls while the model composes a batch of tool calls
\\ • parked work is named above the prompt; graff says it will continue
\\ • PR commands run like any other command (no readiness gate)
\\ • logged in, Jev picks each turn's effort alongside the first request
\\
\\0.0.302.18
\\ • graff tui launches the separate fullscreen client; terminals get the line REPL
\\ • shorter turns: one-line heads-ups, no recomputed answers, fewer false stalls
\\ • instructions cache across repos; the shell names python3's version
\\ • clearer session lists and cross-project resume
\\ • paste warnings, MCP recovery, and catalog-tolerant model selection
\\
\\0.0.302.17
\\ • rlm scripts keep whole MCP results; print() still shows the slim view
\\
\\0.0.302.16
\\ • license: companies that raised over US$500k need a commercial license
\\
\\0.0.302.15
\\ • DeepSeek's default effort thinks at its low level (/effort high for more)
\\
\\0.0.302.14
\\ • MiMo answers without thinking unless effort is high (/effort high)
\\ • MCP servers connect in the background; the first message never waits
\\ • sub-agents start with your prompts and their parent's effort
\\ • the shell tool names the OS; small reads and sed -i pass the guard
\\ • rlm takes JSON literals; todo_write replies with counts
\\
\\0.0.302.13
\\ • fewer round trips: named MCP servers and subagents load up front
\\ • write_file won't replace an unread file and says what it did
\\ • sub-agents call inherited MCP tools from their first request
\\ • piped repl: a bracketed paste is one prompt
\\
\\0.0.302.12
\\ • every ChatGPT login (chatgpt, codex) is the new ChatGPT sign-in
\\ • browser sign-ins survive stray connections and old tabs
\\ • sub-agents get the MCP tools their root loaded
\\ • codex follows the model's default effort and verbosity
\\
\\0.0.302.11
\\ • codex prompt cache is per account: a new repo starts warm
\\ • tool loads keep the cached prefix; rare tools load on demand
\\ • MCP says how it slims results and keeps the full result
\\ • GPT-6 follow-ups steer the running reply; graff login chatgpt
\\ • scripted repl sends the prompt once and waits for background work
\\
\\0.0.302.10
\\ • a tool call cut off mid-response is refused with its cause, not run
\\ • /loop waits for its background work; resends never start a second turn
\\ • slow Anthropic calls keep their prompt cache; OpenRouter caches Claude
\\
\\0.0.302.9
\\ • macOS CLI signed and notarized by the release workflow
\\ • local MCP servers start on first use; "shared": true runs one per machine
\\ • the desktop app is Harness; the in-repo desktop apps are removed
\\ • graff mcp add --everywhere saves to the user-level MCP config
\\ • Harness room messages from other agents arrive framed as advisory
\\
\\0.0.302.8
\\ • instruction-file edits reach the model without busting the prompt cache
\\ • no more /pr-acceptance: a draft PR completes as an unverified handoff
\\ • PR review budget fits small edits in large files; finite jobs stay awaitable
\\ • MCP client: pagination, list_changed, progress, cancellation, MRTR
\\ • graff mcp add infers, checks and hot-joins a server; desktop computer use removed
\\
\\0.0.302.7
\\ • tool shells: provider keys stay out, git and NUL-byte commands are safer
\\ • tool calls left as <tool_call> markup in a reply now run
\\ • ACP: diffs and locations on tool calls, richer approvals, faster session/new offline
\\ • graff mcp serve speaks MCP 2026-07-28
\\
\\0.0.302.6
\\ • ask_user over ACP uses standard elicitation and never freezes a client
\\ • ACP subagents stream their work onto the parent tool call
\\ • ACP v2 draft preview (GRAFF_ACP_V2=1)
\\ • graff peer: message other agents on this device; graff keys: sync logins
\\
\\0.0.302.5
\\ • WebSocket turns stream live, with automatic HTTPS fallback
\\ • ACP sessions use the client's folder and report their worktree
\\ • openai/codex models and ACP image blocks accept images
\\ • the old browser companion is retired; webfetch uses built-in HTTP
\\
\\Full release history: https://github.com/justrach/codegraff/releases
\\
;
/// Release builds leave `graff learn` out (build option `learn`).
const learn_usage = if (@import("build_options").learn)
" graff learn [help] local mutate/evaluate/promote/rollback engine\n"
else
"";
pub const usage_text =
\\graff — a minimal agentic coding harness in Zig (zero deps)
\\
\\usage:
\\ graff [flags] start the line REPL
\\ graff repl line REPL (scripted when stdin is piped)
\\ graff tui [--yes] [args...] launch graff-tui; installs it from the Harness release on first use
\\ graff [-p] "prompt" one-shot: run the prompt, print the answer, exit
\\ graff login get a codegraff key (device-code OAuth)
\\ graff login chatgpt ChatGPT sign-in (also: codex, openai)
\\ graff login anthropic Claude subscription browser sign-in (also: ant)
\\ graff login console metered Console sign-in through the `ant` CLI
\\ graff login claude-agent Claude through your own Claude Code login (`claude`)
\\ graff login kimi Kimi Code OAuth login (kimi.ai; kimi.com for China)
\\ graff login xai Grok/SuperGrok OAuth login (device-code)
\\ graff login zai Z.AI Coding Plan OAuth login
\\ graff key set <provider> <key> store a key (Keychain; POSIX 0600 file; Windows home ACL)
\\ graff key list show which providers have keys
\\ graff refresh pull live catalogs (same as `graff models refresh`)
\\ graff models [refresh] list the live catalog; refresh Codex + provider lists + models.dev
\\ graff route <model>… dry-run which provider/billing a model lands on (no API call)
\\ graff mcp add <name> -- <cmd> add a stdio MCP server to .mcp.json
\\ graff mcp add <name> --url <url> add a Streamable HTTP MCP server
\\ graff mcp login <name> OAuth login for a remote MCP server
\\ graff mcp list configured MCP servers
\\ graff mcp install configure local HTTP service and MCP clients
\\ graff mcp serve [--http] [--port N] expose run_task to MCP clients
\\ graff plugins [load <name>] list Claude/Cursor/Grok/Codex plugin trees (in place)
\\ graff computer [status|stop|off|allow|grant|open] nanocua's switches (macOS)
\\
++ learn_usage ++
\\ graff worktree list list -w tabs and experiment-pool trees (tagged)
\\ graff worktree create <name> [base] mint a task workspace from the remote base (origin/main after fetch; optional base)
\\ graff worktree run <name> run the workspace run script from that checkout
\\ graff worktree archive <name> remove a clean duplicate checkout; dirty or unique-commit trees stay
\\ graff worktree merge <name> squash-land worktree-<name> onto the current branch + clean up
\\ graff worktree land <name> same as merge
\\ graff worktree update <name> merge the remote base into that workspace
\\ graff worktree gc drop dead session trees and clean merged-PR trees
\\ graff worktree remove <name> discard worktree-<name> (drops its scratch work) + delete the branch
\\ graff worktree prune drop git registrations for worktrees whose dirs were deleted
\\ graff servers background servers graff started (this session's or older): pid, port, age
\\ graff servers stop <pid> end one (its whole process tree); `prune` drops records of dead ones
\\ graff peer list|send|inbox talk with other agents on this device (Claude Code, Codex, graff)
\\ graff sandboxes list your gateway sandboxes (what's burning credits)
\\ graff sandboxes stop <id> spin a sandbox down (stops it + settles the meter)
\\ graff cube new spin up a cloud graff (sandbox + serve + preview URL)
\\ graff cube [status|stop] inspect the running cube or spin it down
\\ graff --schema print the machine-readable interface (SDK codegen)
\\ graff serve HTTP/NDJSON bridge over the --json protocol
\\ (--host/--port/--token; sessions are --json children)
\\ graff remote-control [--name n] serve with no listener: dial out to your Codegraff account so
\\ `graff remote` anywhere can drive sessions on this machine
\\ graff remote [new|send|tail|…] list and drive the sessions of machines running remote-control
\\ graff acp Agent Client Protocol agent on stdio (Zed and other ACP editors)
\\ graff update [--force|--check] update graff to the latest stable release
\\ graff update --beta [--check] install the newest beta for the next launch
\\ graff title <prompt> print the AI tab-title for a prompt (test title styles)
\\
\\flags:
\\ --model <name> start on this model (same fuzzy resolution as /model)
\\ --subagent-model <name> pin children/workflows/judges on the root provider
\\ --subagent-provider <id> route pinned workers through this provider
\\ --allow-cross-provider-subagents confirm prompts/code may go to the worker provider
\\ --no-subagent-tier opt out of the default worker tier ladder (inherit the root model)
\\ --resume <name> resume/autosave <name>.session.json
\\ --branch <name> clone --resume into an independent autosave target
\\ --new start a fresh autosaved session (default)
\\ --no-resume ignore --resume and start fresh
\\ --system-prompt <text> replace the built-in system prompt
\\ --append-system-prompt <text> append extra text to the system prompt
\\ --goal <text> seed a standing objective (tracked as a todo checklist) for every turn; persists across completions
\\ --eval <cmd> scoring command for an eval-driven loop (the `eval` tool runs it)
\\ --until <0-100> eval-loop target score; stop when reached (default 90)
\\ --niche <name> fleet niche this eval optimizes (reviewer/researcher/implementer/skeptic or a custom agent); tags submitted scores so the DGM can promote a champion for that role
\\ -w, --worktree <name> isolate this session in a git worktree (.graff/worktrees/<name>) so parallel agents don't collide on files. A second session in a claimed checkout auto-isolates.
\\ --experiment N pre-mint N child worktrees (1-16) under .graff/worktrees/exp-<id>/; next spawns claim a seat
\\ --add-dir <path> extra file-tool root (repeatable, max 16). Not a cwd switch; no skills/sessions from it
\\ --context-limit name=N cap a named prefix: skill_catalog_bytes|mcp_schema_bytes|agents_md_bytes
\\ --no-autocommit with -w, don't auto-commit each turn (default on; land work with `graff worktree merge`)
\\ --yolo skip all permission prompts for the session
\\ --rlm advertise the rlm REPL (default; persistent binds, subagent(), llm_query, mid-stream spec-ptc; GRAFF_RLM=1)
\\ --old, --no-rlm restore the pre-rlm structured-only catalog (GRAFF_OLD=1 or GRAFF_RLM=0)
\\ --lean slim tool surface (8 core tools) + MCP schemas folded behind load_tool_schemas — the DEFAULT for -p one-shots (GRAFF_LEAN=1). `.mcp.json` still connects.
\\ --no-lean opt a one-shot out of the implied --lean: full tool surface + eager MCP schemas, the pre-default -p behavior
\\ --no-local-tools embedder mode: hard-disable the built-in bash/bash_output/bash_kill/read_file/edit_file/write_file/codedb tools for the whole process (subagents included), so graff can run outside the sandbox and get its coding tools from an MCP server instead; webfetch, orchestration and MCP tools still work (GRAFF_NO_LOCAL_TOOLS=1)
\\ -p, --print one-shot print mode (answer on stdout, progress on stderr)
\\ --output-format text|stream-json one-shot output; stream-json emits run_started and result
\\ --timing show per-tool wall-clock on result lines
\\ --cost show running session spend in the prompt
\\ --json structured stdio protocol (JSON in, JSONL events out)
\\ --max-run-tool-calls N aggregate root/descendant tool ceiling for this invocation
\\ --max-tool-calls N reject root tool calls after N per turn (JSON-safe budget)
\\ --compact-at PCT start automatic compaction at PCT% of the context window (10-95, default 80)
\\ --max-model-calls N total provider calls allowed across this run (default 0 = unlimited; includes children/title/judges)
\\ --dedupe-tool-calls reject duplicate root tool name+input calls per turn
\\ --no-telemetry disable anonymous usage telemetry for this run
\\ --learning-privacy <mode> learning egress ceiling: local|aggregate|templates|examples (default aggregate)
\\ -h, --help this help
\\ -V, --version print version
\\
\\keys: <PROVIDER>_API_KEY env vars, `graff key set`, or `graff login`;
\\a Codex CLI login is picked up automatically.
\\inside the REPL: /help lists commands, a bare "/" opens the command menu,
\\"@" opens a fuzzy file picker (a drag-and-dropped image attaches as a
\\native vision block on vision models; other files paste as their path),
\\esc interrupts a streaming response, "always allow" persists to
\\.harness/settings.json.
\\telemetry: /debug is a local content-free HUD (session/turn/tool
\\decisions). Anonymous OTLP usage stats leave the process only when
\\OTEL_EXPORTER_OTLP_ENDPOINT (or GRAFF_OTEL_ENDPOINT) is set; opt out
\\with --no-telemetry or GRAFF_NO_TELEMETRY=1. GRAFF_TELEMETRY_KEY sends
\\an optional x-harness-key token to the configured collector.
\\learning privacy: local learning trials publish prompt-free aggregate grades
\\by default, announced once per machine; /privacy local (or
\\GRAFF_LEARNING_PRIVACY=local, GRAFF_FLEET=off, --no-telemetry) sends nothing.
\\/privacy changes the session ceiling; template text still needs exact approval.
\\
;