From 77039540f7ab85373c14e1f3fdb77a1c06e6d361 Mon Sep 17 00:00:00 2001 From: justrach <54503978+justrach@users.noreply.github.com> Date: Fri, 20 Nov 2026 18:04:10 +0800 Subject: [PATCH 1/3] Prefer RLM for supported reads and JSON operations The tool is available on turn one, but availability alone can lead to a shell process for work RLM already handles. Lead with a narrow preference in its catalog entry while retaining direct calls, shell fallback, and the structured-only opt-out. Co-Authored-By: Codegraff --- docs/adr/0351-prefer-rlm-for-supported-work.md | 7 +++++++ docs/adr/README.md | 1 + evals/harness_behavior.jsonl | 2 ++ src/rlm.zig | 5 +++-- 4 files changed, 13 insertions(+), 2 deletions(-) create mode 100644 docs/adr/0351-prefer-rlm-for-supported-work.md diff --git a/docs/adr/0351-prefer-rlm-for-supported-work.md b/docs/adr/0351-prefer-rlm-for-supported-work.md new file mode 100644 index 00000000..00186c65 --- /dev/null +++ b/docs/adr/0351-prefer-rlm-for-supported-work.md @@ -0,0 +1,7 @@ +# 0351. Prefer RLM for supported work + +Status: accepted + +[0140](0140-rlm-is-listed-when-available.md) makes RLM available from the first turn, but availability alone can lead the model to start a shell process for work RLM already handles. + +Lead with a narrow preference in the shared RLM tool description, not in the always-on system prompt. Prefer one RLM call for independent reads, JSON row counts, and field projection. Keep direct tools and the shell available for builds, tests, Git commands, and transforms RLM cannot express. Keep `--old` and `--no-rlm` unchanged. Forcing tool choice or rewriting shell commands would interfere with legitimate fallbacks. diff --git a/docs/adr/README.md b/docs/adr/README.md index 750d0ca8..b6672626 100644 --- a/docs/adr/README.md +++ b/docs/adr/README.md @@ -375,6 +375,7 @@ record only when you need the evidence or the edge cases. | [0348](0348-persistent-python-repl-and-unharnessed-mode.md) | Persistent per-owner Python 3 worker (`python_repl`, folded in `native_fold`) gives sessions cross-turn globals, `_` expression results, fd 1/2 capture, and timeout resilience without turn-1 schema overhead; opt-in `--unharnessed` / `/unharnessed` (`CLI > env > .harness/settings.json`, ACP/JSON parity) exposes only `python` with a single-tool prompt, while `--no-local-tools`, `plan_mode`, and hard policies block both. | | [0349](0349-native-flint-runtime-is-external.md) | Native Flint is a built-in local OpenAI-compatible provider, while its runtime and model files remain separately installed; use only validated loopback endpoints and never cloud-fallback. | | [0350](0350-nonblocking-catalog-refresh-and-slash-refresh.md) | Dynamic provider catalogs (including Codex) stay 0 ms non-blocking at boot while stale-while-revalidating in the background past 5 minutes and activating landed caches in-session; `graff/models` adds bounded `{"force": true}`, and `/refresh` refreshes all signed-in catalogs across REPL and ACP. | +| [0351](0351-prefer-rlm-for-supported-work.md) | The shared RLM catalog entry prefers RLM for independent reads, JSON row counts, and field projection. It does not force tool choice or remove direct tools or the shell fallback. | ## When to write one diff --git a/evals/harness_behavior.jsonl b/evals/harness_behavior.jsonl index b6132251..e767aa78 100644 --- a/evals/harness_behavior.jsonl +++ b/evals/harness_behavior.jsonl @@ -118,6 +118,8 @@ {"args": [], "env": {"GRAFF_NO_CODEDB_GUARD": "1"}, "prompt": "Run the scripted RLM ordering fixture and report only observed results.", "workspace_seed": {"target.txt": "old"}, "id": "rlm-nested-shell-failure-status", "why": "A nested shell failure propagates error status and prevents later nested mutation.", "script": [{"tool": "rlm", "arguments": {"code": "print(bash(\"printf NESTED_FAILURE; exit 7\"), write_file(path=\"marker\", content=\"ran\"))"}}, {"text": "The nested command failed."}], "verify_command": ["python3", "-c", "from pathlib import Path; assert not Path('marker').exists()"], "assert": [{"event": {"type": "tool_call_finished", "name": "rlm", "is_error": true}}, {"request_contains": {"index": 1, "text": "NESTED_FAILURE"}}]} {"args": [], "env": {"GRAFF_NO_CODEDB_GUARD": "1"}, "prompt": "Run the scripted RLM ordering fixture and report only observed results.", "workspace_seed": {"target.txt": "old"}, "id": "rlm-pending-host-blocks-tail-completion", "why": "Pending nested background work preserves pending status, stops later mutation and cannot satisfy same-batch completion. The headless run then waits for the job and continues with its exit instead of ending.", "script": [{"tools": [{"tool": "rlm", "arguments": {"code": "print(bash(command=\"sleep 0.2; exit 0\", run_in_background=true), write_file(path=\"marker\", content=\"ran\"))"}}, {"tool": "attempt_completion", "arguments": {"result": "PREMATURE_COMPLETION"}}]}, {"text": "Verification is still pending."}, {"text": "Verification finished: the check exited 0."}], "verify_command": ["python3", "-c", "from pathlib import Path; assert not Path('marker').exists()"], "assert": [{"event": {"type": "tool_call_finished", "name": "rlm", "is_error": false}}, {"event": {"type": "tool_call_finished", "name": "attempt_completion", "is_error": true}}, {"no_event": {"type": "tool_call_finished", "name": "attempt_completion", "is_error": false}}, {"request_contains": {"index": 2, "text": "exited 0: sleep 0.2; exit 0"}}, {"final_text_contains": "Verification finished: the check exited 0."}]} {"id": "rlm-indexed-print-refuses-and-recovers", "why": "Unsupported Python-style result indexing must report an error without pretending to execute the nested host or executing later side effects; the supported text binding form recovers.", "args": ["--lean"], "prompt": "Run the scripted checks and report only observed results.", "script": [{"tool": "rlm", "arguments": {"code": "print(bash(\"printf ran > nested-marker\")[\"stdout\"])\nx = write_file(path=\"tail-marker\", content=\"ran\")"}}, {"tool": "rlm", "arguments": {"code": "r = bash(\"printf VERIFIED\")\nprint(r)"}}, {"text": "Verified using the supported text result."}], "verify_command": ["python3", "-c", "from pathlib import Path; assert not Path(\"nested-marker\").exists(); assert not Path(\"tail-marker\").exists()"], "assert": [{"event": {"type": "tool_call_finished", "name": "rlm", "is_error": true}}, {"events_at_most": {"match": {"type": "tool_call_finished", "is_error": true}, "count": 1}}, {"request_contains": {"index": 1, "text": "host results are text"}}, {"event": {"type": "tool_result", "name": "rlm", "is_error": false, "text": "VERIFIED"}}, {"final_text_contains": "Verified using the supported text result."}]} +{"id":"rlm-routing-preference","why":"The first request advertises a narrow RLM preference without removing the shell fallback.","args":[],"prompt":"Read the synthetic JSON fixture and return its row count.","workspace_seed":{"rows.json":"[{\"id\":1},{\"id\":2}]"},"script":[{"tool":"rlm","arguments":{"code":"rows = read_file(\"rows.json\")\nprint(len(rows))"}},{"text":"2"}],"assert":[{"request_contains":{"index":0,"text":"Prefer one RLM call for independent reads"}},{"request_contains":{"index":0,"text":"Use shell for builds, tests, git, and unsupported transforms"}},{"event":{"type":"tool_result","name":"rlm","is_error":false,"text":"2"}}]} +{"id":"rlm-routing-structured-opt-out","why":"Structured-only mode does not advertise the RLM preference.","args":["--old"],"prompt":"Read the synthetic fixture.","workspace_seed":{"rows.json":"[]"},"script":[{"tool":"read_file","arguments":{"path":"rows.json"}},{"text":"The array is empty."}],"assert":[{"request_lacks":{"index":0,"text":"Prefer one RLM call for independent reads"}},{"event":{"type":"tool_result","name":"read_file","is_error":false,"text":"[]"}}]} {"id": "gateway-failure-usage-unknown", "why": "A typed gateway failure with no usage must preserve the provider diagnostic and label totals as incomplete instead of implying a free zero-token call.", "args": ["-p", "summarize the fixture"], "prompt": "summarize the fixture", "script": [{"text": "partial", "stream_error": {"type": "gateway_error", "code": "request_failed", "message": "fixture request failed"}}], "expected_exit": 1, "assert": [{"stderr_contains": "gateway_error request_failed"}, {"stderr_contains": "known subtotal: 0 api call(s)"}, {"stderr_contains": "1 failed request attempt(s) without usage (tokens and cost unknown)"}]} {"id": "gateway-retry-usage-remains-incomplete", "why": "A recovered retry retains known successful usage while the failed attempt remains explicitly unknown.", "prompt": "summarize the fixture", "script": [{"text": "partial", "stream_error": {"type": "api_error", "message": "Internal error during token parsing"}}, {"text": "recovered fixture"}], "assert": [{"event": {"type": "turn", "usage_complete": false, "unreported_failed_attempts": 1, "api_calls": 1}}, {"final_text_contains": "recovered fixture"}]} {"id":"terminal-model-unavailable-does-not-retry","why":"#1191: a structured model_unavailable error is terminal when the gateway cannot serve the chosen model; a retry would repeat the same failure. Keep the request id so support can find the failed call.","prompt":"Say hello","script":[{"stream_error":{"type":"service_unavailable","code":"model_unavailable","message":"This model is not available right now. Choose another model or contact Codegraff support with the request ID.","request_id":"cg_req_fixture"}}],"assert":[{"events_at_most":{"match":{"type":"model_call_started"},"count":1}},{"event":{"type":"error","message":"Model lmstudio is currently unavailable. Switch models with /model, or try again later. Technical details — api error: (service_unavailable model_unavailable) This model is not available right now. Choose another model or contact Codegraff support with the request ID. [cg_req_fixture]"}},{"no_event":{"type":"model_call_finished","ok":true}}]} diff --git a/src/rlm.zig b/src/rlm.zig index 6a08ac12..e390b873 100644 --- a/src/rlm.zig +++ b/src/rlm.zig @@ -25,10 +25,11 @@ const mcp_shapes = @import("mcp_shapes.zig"); const read_miss = @import("read_miss.zig"); pub const tool_name = "rlm"; -pub const tool_desc = "Tools as a call-only script, not Python: name = tool(args), args being literals or bound names (a bound name passes its text); no expressions, indexing or comprehensions. To compute over a result, write_file(\"r.json\", name), then shell. Leading read_file/codedb/sleep_ms/llm_query calls overlap; others run in order, stopping on error/cancel/pending. Binds persist. subagent(\"task\") is sidecar-only (critical-path steps stay local). Loaded MCP names are host functions after load_tool_schemas; each(arr, tool, field?) maps a JSON array; len(x)/project(x, field) slim it. print() is the answer."; +const routing_hint = "Prefer one RLM call for independent reads and JSON row counts or field projection. Use shell for builds, tests, git, and unsupported transforms. "; +pub const tool_desc = routing_hint ++ "Call-only, not Python: name = tool(args); args are literals or bound text, no expressions or indexing. Bind results, use len(x)/project(x, field), then print(). Leading read_file/codedb/sleep_ms/llm_query calls overlap; others run in order and stop on error/cancel/pending. Binds persist. subagent() is sidecar-only, not critical-path. Call loaded MCP tools after load_tool_schemas; each(arr, tool, field?) fans out."; /// --lean catalog desc: same contract, no REPL essay. maybeAppend is after /// compactLeanSpecs, so this is the one-shot wire text. -pub const lean_tool_desc = "Batch independent read_file/codedb here. print(read_file(\"p\")) returns the file — do not catalog-read it again. Then edit_file/write_file/shell as catalog tools. Leading reads overlap; other calls run in order, stopping on error/cancel/pending. Binds persist. Loaded MCP names are host functions after load_tool_schemas."; +pub const lean_tool_desc = routing_hint ++ "Batch read_file/codedb; bind arrays for len(x)/project(x, field), then print(). print(read_file(\"p\")) returns the file; do not catalog-read it again. Use direct tools for edits. Leading reads overlap; other calls run in order, stopping on error/cancel/pending. Binds persist. Call loaded MCP tools after load_tool_schemas."; pub const tool_schema = \\{"type": "object", "properties": {"code": {"type": "string", "description": "Call-only script, not Python. One statement per line: name = read_file(\"path\") / codedb(\"command\") / bash(\"cmd\") / sleep_ms(ms) / llm_query(\"prompt\") / subagent(\"task\") / loaded mcp__server__tool() or tools.server.tool(); write_file(\"path\", name) saves a bound result; each(arr, tool[, field]) calls tool per item; len(x) and project(x, field) slim it; name = [..] / {..} / \"text\" binds a literal; print(name, ...) is the result. Arguments are string, number, boolean or JSON array/object literals, or bound names. Assignments persist across rlm calls."}}, "required": ["code"]} ; From 9651809bf7eab9f89e84a80d67ef5e8e8f26afa7 Mon Sep 17 00:00:00 2001 From: justrach <54503978+justrach@users.noreply.github.com> Date: Fri, 20 Nov 2026 20:41:39 +0800 Subject: [PATCH 2/3] Prepare v0.302.74.77 Document the RLM routing preference in the release notes and CLI changelog. Co-Authored-By: Codegraff --- CHANGELOG.md | 4 ++++ src/cli_text_source.zig | 3 +++ 2 files changed, 7 insertions(+) diff --git a/CHANGELOG.md b/CHANGELOG.md index 7a043174..90d2c7ee 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -10,6 +10,10 @@ The release workflow uses a tag's section here as its release notes (a hand-written `docs/releases/.md` wins if present), so keeping this file current is part of cutting a release. +## v0.302.74.77 + +- Prefer `rlm` for independent file reads, JSON row counts, and field extraction. Direct tools, shell fallback, and `--no-rlm` remain available (#1915). + ## v0.302.74.76 - Raise `gpt-6.2-astra` subscription context window to 500k on `codex` and `chatgpt-new` (offline rows, subscription clamp, and live catalog discovery when `max_context_window` advertises a wider window; ADR 0292 addendum). diff --git a/src/cli_text_source.zig b/src/cli_text_source.zig index 704425e7..5dc9ea0c 100644 --- a/src/cli_text_source.zig +++ b/src/cli_text_source.zig @@ -5,6 +5,9 @@ pub const changelog_text = \\What's new \\────────── + \\0.302.74.77 + \\ • Prefer `rlm` for independent reads, JSON row counts, and field extraction (#1915). + \\ \\0.302.74.76 \\ • Raise `gpt-6.2-astra` subscription context window to 500k on `codex`/`chatgpt-new` (ADR 0292 addendum). \\ • Ignore empty optional args and normalize `mcp__` prefixes in MCP schema loading; add `CLA.md` PR check. From 5a2ea8b863e71fbc20e99511be31fbe8e80a0ae6 Mon Sep 17 00:00:00 2001 From: justrach <54503978+justrach@users.noreply.github.com> Date: Fri, 20 Nov 2026 20:46:51 +0800 Subject: [PATCH 3/3] Separate CLI changelog source from help text Keep the accumulated release notes within the Zig source line limit without changing CLI output. Co-Authored-By: Codegraff --- src/cli_changelog_source.zig | 463 ++++++++++++++++++++++++++++++++++ src/cli_text_source.zig | 464 +---------------------------------- 2 files changed, 465 insertions(+), 462 deletions(-) create mode 100644 src/cli_changelog_source.zig diff --git a/src/cli_changelog_source.zig b/src/cli_changelog_source.zig new file mode 100644 index 00000000..301fa99a --- /dev/null +++ b/src/cli_changelog_source.zig @@ -0,0 +1,463 @@ +//! Build-time CLI changelog. 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 + \\────────── + \\0.302.74.77 + \\ • Prefer `rlm` for independent reads, JSON row counts, and field extraction (#1915). + \\ + \\0.302.74.76 + \\ • Raise `gpt-6.2-astra` subscription context window to 500k on `codex`/`chatgpt-new` (ADR 0292 addendum). + \\ • Ignore empty optional args and normalize `mcp__` prefixes in MCP schema loading; add `CLA.md` PR check. + \\ + \\0.302.74.75 + \\ • Add built-in `nanoattention/flint` local provider (#1905, ADR 0349) and persistent `python_repl` + `--unharnessed` single-tool mode (#1904, ADR 0348). + \\ • Add `gpt-6.2-sol` / `gpt-6.2-astra`, `client_version=0.170.0`, 0ms-boot 5-min background catalog refresh + `/refresh` (ADR 0350), and Codex/ChatGPT `ultrafast` service tier. + \\ • Only latch ACP `saw_text` when visible text survives citation stripping (#1912). + \\ + \\0.302.74.74 + \\ • Fix HTTP/2 error-body capture across allocators (#1903), give `webfetch` 64 KiB header/redirect buffers (#1911), and preserve `nanocua` AXError/ScreenCaptureKit failure diagnostics (#1910). + \\ • Add `run_task` boundary metadata, mismatched safe/cwd edit preflight refusal, and exit-only verification caveats while keeping `--safe` intact (#1909). + \\ • Quote Windows `cmd.exe` commands via temp `.cmd` scripts (opt-in Git Bash via `GRAFF_WINDOWS_SHELL=bash`), add Windows UTF-8/non-interactive env defaults, read-only skill sibling reads, `read_file` `contains`+range support, and empty/null optional tool-arg tolerance. + \\ + \\0.302.74.73 + \\ • Keep Orchestrator phase directives turn-scoped and allow quoted `rg` alternations, read-only `gh`, stdout `pdftotext`, and `readOnlyHint` MCP tools in plan mode (#1908). + \\ • Skip Auto Orchestrator planning on informational `?` questions or compacted histories, and settle expanded Phase 2 compaction inside `compactAt()` (#1906). + \\ • Strip Codex `tool_search_*` items on non-Codex Responses failover with local unfolding, and avoid auth-refresh loops on `subscription_sharing_unsupported_capability` (#1907). + \\ + \\0.302.74.71 + \\ • Add Orchestrator Mode (`off | auto | on`) same-chat read-only planning and KV-identical executor handoffs with short follow-up and wake bypasses (ADR 0347). + \\ + \\0.302.74.69 + \\ • Track live rate-limit headers, WebSocket `codex.rate_limits`, and `claude-agent` events, and bypass exhausted `claude-fable-5-2` pools onto `claude-opus-5-5` (ADR 0346). + \\ • Surface signed-in account identities, plan credits, and remaining usage windows across `/usage`, `/cost`, ACP `graff/accounts`, and `_codegraff/usage.planWindows`. + \\ • Advertise `_meta["codegraff/accounts"]` and all 7 terminal `authMethods` in ACP `initialize`. + \\ + \\0.302.74.68 + \\ • Add dual-frontier worker tier selection (`hard_frontier` at `xhigh`/`max`), single-model ladder bounds, and live-catalog rung fallback (ADR 0345). + \\ • Surface separate plan quota pools in `/models` (`plan*`) and ACP `graff/models` (`separateQuota`), with intra-plan frontier stepdown on usage limits. + \\ • Restrict cross-provider flat-rate borrowing to `.plan` subscription logins and keep equal-scored rungs on the active plan. + \\ • Enable the workspace-owning ACP client profile for `zeron` in `graff acp` and read `~/.simple-harness-keys.json` before macOS Keychain lookup. + \\ + \\0.302.74.67 + \\ • Add opt-in runtime extensions (`graff extensions`, `plugins enable|disable`) and worker-inherited `policies` across REPL and ACP (ADR 0343). + \\ • Add opt-in post-write `formatters` with verified-byte restore and multicall symlink preservation. + \\ • Add read-only `lsp` queries and `graff lsp` CLI (`status`, `setup`, diagnostics, hover, definition, references, symbols) with folded root schemas, `~/.graff/lsp/settings.json` defaults, and optional POSIX warm reuse (ADR 0344). + \\ • Accept `provider/model` qualifiers on `--subagent-model` across REPL and ACP. + \\ + \\0.302.74.66 + \\ • Repair paused request history and avoid invalid worker retries (#1895). + \\ • Retain longer prompt-cache TTL across returning interactive turns (#1898). + \\ • Reconcile application targets and strengthen background-safe GUI verification guidance (#1899, #832). + \\ • Add schema-ownership coverage and fix Windows test linking (#1892, #1900). + \\ + \\0.302.74.65 + \\ • Improve startup responsiveness across the REPL, one-shot, and ACP paths. + \\ • Avoid blocking startup on optional companion handshakes and model-catalog refreshes. + \\ • Explain output-length continuation exhaustion in ACP while preserving partial work (#1893). + \\ + \\0.302.74.64 + \\ • mcp: tool schemas are deep-cloned before normalization — mutating the Value copy through shared ArrayHashMap storage corrupted the parsed tools tree, so debug builds segfaulted in mcp_cache.store right after a successful connect (#1892) + \\ + \\0.302.74.63 + \\ • `read_file` honors `offset` and `limit` line-window aliases when `start_line`/`end_line` are omitted (#1885) + \\ • native file tools normalize `/…` absolute paths inside the workspace instead of rejecting them (#1886) + \\ • `tool_balance` treats read-only `shell` inspection commands/pipelines as `.read_explore` without resetting read streaks (#1887) + \\ • standalone `shell` `sleep N` polls run inline alongside background jobs and no longer hold `run_idle` open (#1888) + \\ • headless `graff -p` wires `SIGINT` to `esc_cancel` and flushes the final `result` record (#1889) + \\ • `write_file` and `edit_file` append inline JSON, Zig/ZON, Python, and Go syntax diagnostics (#1890) + \\ + \\0.302.74.62 + \\ • the primary bot is your main bot (ADR 0335): one thread for every project; in a client such as Harness it starts each piece of work as its own chat where it belongs and reports when the client says it stopped; it asks only when the project is unclear + \\ • from a terminal the main bot names the project and gives the command; single-job bots stay as helpers it hands their own jobs to; Bots are still off until `graff bot enable` + \\ + \\0.302.74.61 + \\ • Bots (experimental, off until `graff bot enable`): standing roles with one job and their own memory, work log and routines in `.graff/bots//`; `graff bot run`, `routine add`, `tick` for cron or a client's timer (ADR 0333) + \\ • `graff bot --user`: your own bots in `~/.graff/bots` for any project, one `--primary`; a late check records what it missed instead of running it late (`--stale-after`, `graff bot missed`); `graff bot state --json` for clients (ADR 0334) + \\ + \\0.302.74.60 + \\ • runaway raw reasoning streams stop at a byte bound, run loop detection (`ReplyRunaway` -> `ModelLoop`), and skip the 270s empty-reply backoff (#1878) + \\ • `delta.reasoning` normalizes into `reasoning_content` on replay and `ReplySize` (#1881) + \\ • `#390` final budget-landing calls keep `tools` with `tool_choice: "none"`, settle reasoning-only `length` stops cleanly, and exit `0` when tool work completed (#1882) + \\ • `edit_file` resolves multi-line `old_string` blocks with stripped line-1 indentation and restores the matched prefix (#1880) + \\ • unattended `-p` action turns nudge long read-only streaks and bounce once on zero-mutation unfinished prose (#1879) + \\ • explicit truncation markers on capped instruction/skill/MCP sections (#1871), `CLAUDE_CODE_USER_EMAIL` stripped in `claude-agent` (#1870), and worktree-scoped `codedb` guards + exit-code checks (#54) + \\ + \\0.302.74.59 + \\ • Claude Fable 5.2 in the baked catalog: `--model claude-fable-5-2` resolves by its bare name and is priced + \\ • image reads in one tool batch all reach the model, in call order (#1866): a bounded queue (16 images, 24 MB) replaces the single slot, with a tool error past the limit + \\ • uploads a client leaves as an `Attached images` trailer are staged as native images instead of paths the file tools refuse, also when sent mid-turn (#1865) + \\ • `graff refresh` / `graff models` / `/models` say when a provider rejects the live model-list request instead of silently showing the cached list + \\ • Lean 4 ImageDelivery kernel (3,870 verified cells): every staged image is delivered once or refused, in call order + \\ + \\0.302.74.58 + \\ • prompt-cache stability (#1320): jev_auto off by default (opt-in via /effort auto or GRAFF_JEV_AUTO=1), BustCause.effort HUD tracking, deterministic per-turn prompt_cache_key fallback, and in-session HTTP MCP schema retention + \\ • ACP low-latency path + Lean 4 AcpLatency kernel (3,768 verified cells): 0-subprocess in-process .git worktree resolution, single-parse ACP inbox routing, batch-flushed LineWriter/EventSink with reusable scratch arena, and single-pass startup MCP config load + \\ + \\0.302.74.57 + \\ • adaptive codedb workspace threshold (<8 code files in-process list_dir+read_file; >=8 code files codedb) + ~/.codedb/projects CoW seeding + \\ • pre-tool transcript checkpointing (beforeToolBatch), crash-tail suffix replay, and effect-aware TOOL_INTERRUPTED_READONLY recovery + \\ • deferred, single-pass streaming, cached (.graff/folded_priors.json), and ring-capped trajectory archive folding (#1864) + \\ • Lean 4 resource/durability kernel (3,742 verified cells) + scoped arena and O(1) incremental hashing optimizations + \\ + \\0.302.74.56 + \\ • unified subagent steering (@agent-N, @sa-..., /workers send, ACP graff/agents action: "steer") for foreground and background workers + \\ • named background subagents/tasks in REPL standing bar and ACP session_info_update (_meta["graff/waiting"]) + \\ • cross-turn ACP tool_call_update progress streaming for detached subagents and immediate user follow-up wakeup in agent_output + \\ • monotonic rewind-proof attempt ledger across /undo and /rewind preserved in compact_handoff_note.durableState + \\ • wall-clock grid-aligned recurring schedules (nextGridSlot) with @every and --missed=once|skip policies + \\ • APFS clonefile / Linux reflink Copy-on-Write cache seeding in subagent worktrees and adaptive sub-100ms foreground bash polling + \\ • corrupt-safe playbook JSONL self-healing + storage_warning diagnostics, markdown agent line:col validation, and OAuth child-env scrubbing + \\ + \\0.302.74.55 + \\ • scoped temporary arena parsing in mcp_cache.lookupWith/store/forget and single-server tool entry cloning + \\ • per-request scratch allocation in agent_ws.postResponsesWs and scoped arena cleanup in session_workspaces.remember + \\ • automatic same-plan subscription credential failover in providers.runTurnWithFallback + \\ + \\0.302.74.54 + \\ • Phase-1 deterministic pre-compaction tool-result pruning (compact_prune), crash-tail TOOL_OUTCOME_UNKNOWN vs TOOL_NOT_STARTED repair, and key-sorted repeat hashing + \\ • content-addressed dirty workspace snapshots (workspace_tree_snap) for edit_contract and hook/fsmonitor/LFS-safe subagent worktrees + \\ • Linux bwrap --unshare-pid isolation, Landlock ABI probing, and read-only sandbox denial boundary tagging + \\ + \\0.302.74.53 + \\ • workspace-invariant skill catalog prefix ordering and deepseekChat repo_context split for cross-directory prompt caching + \\ • Kimi global (kimi.ai) vs Mainland China (kimi.com) OAuth region resolution and k3.1-pro catalog + plan routing across CLI and ACP + \\ + \\0.302.74.52 + \\ • read-only trajectory inspection and verified user-facing chat title + chat ID reporting + \\ • background workers preserve independent tool-result sequence counters; synthetic compaction skips mcp_preload scans + \\ • workspace-invariant skill catalog prefix ordering, deepseekChat repo_context split, and Kimi global (kimi.ai) + k3.1-pro catalog routing + \\ + \\0.302.74.51 + \\ • graff sessions reads 64 KiB header prefixes in a reusable scratch arena; session load/resume frees parse trees after load + \\ • single-copy message arena reclaim, scoped trajectory archive folds, and faster remembered-model startup + \\ • -Dminimal and per-subsystem build flags (-Dlearn, -Dserve, -Dacp, -Dextras), codedb symbol ranking, and --eval contrastive reflection + \\ • zero-binary ServerlessHarness SDK runtime with InMemoryWorkspace, pluggable tools, lifecycle hooks, steer/followUp, and state snapshots + \\ + \\0.302.74.5 + \\ • Claude Agent ignores unprefixed tool calls rejected by Claude Code and keeps earlier images across checkpoints + \\ • read_tool_result keeps originating workspace handles across workspace switches + \\ • background shared_cwd workers stay pinned to their spawn workspace when the root session switches trees + \\ + \\0.302.74.4 + \\ • WebSocket Responses turns keep server token counts without re-compacting on ciphertext size + \\ • compaction and session switches reclaim dropped history memory; ACP startup and session/load are faster + \\ • --session-id and -c/--continue launch flags, versioned session files, and upright EXIF photo uploads + \\ + \\0.302.74.3 + \\ • Responses compaction uses server token counts instead of re-compacting on ciphertext size + \\ • cross-format model switches expand compaction checkpoints first and drop tool-only preambles + \\ + \\0.302.74.2 + \\ • Windows installs can recover through GitHub's asset API without skipping checksum verification + \\ • nanocua keeps text and refs by default; coordinates are available with include_frames + \\ • frameless reads skip unused geometry queries; controls-only filtering remains opt-in + \\ + \\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/ + \\ + \\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 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 + \\ +; diff --git a/src/cli_text_source.zig b/src/cli_text_source.zig index 5dc9ea0c..58e190e1 100644 --- a/src/cli_text_source.zig +++ b/src/cli_text_source.zig @@ -1,466 +1,6 @@ -//! Build-time source for --help and --version. Keep all release notes intact. +//! Build-time source for --help and --version. -/// 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 - \\────────── - \\0.302.74.77 - \\ • Prefer `rlm` for independent reads, JSON row counts, and field extraction (#1915). - \\ - \\0.302.74.76 - \\ • Raise `gpt-6.2-astra` subscription context window to 500k on `codex`/`chatgpt-new` (ADR 0292 addendum). - \\ • Ignore empty optional args and normalize `mcp__` prefixes in MCP schema loading; add `CLA.md` PR check. - \\ - \\0.302.74.75 - \\ • Add built-in `nanoattention/flint` local provider (#1905, ADR 0349) and persistent `python_repl` + `--unharnessed` single-tool mode (#1904, ADR 0348). - \\ • Add `gpt-6.2-sol` / `gpt-6.2-astra`, `client_version=0.170.0`, 0ms-boot 5-min background catalog refresh + `/refresh` (ADR 0350), and Codex/ChatGPT `ultrafast` service tier. - \\ • Only latch ACP `saw_text` when visible text survives citation stripping (#1912). - \\ - \\0.302.74.74 - \\ • Fix HTTP/2 error-body capture across allocators (#1903), give `webfetch` 64 KiB header/redirect buffers (#1911), and preserve `nanocua` AXError/ScreenCaptureKit failure diagnostics (#1910). - \\ • Add `run_task` boundary metadata, mismatched safe/cwd edit preflight refusal, and exit-only verification caveats while keeping `--safe` intact (#1909). - \\ • Quote Windows `cmd.exe` commands via temp `.cmd` scripts (opt-in Git Bash via `GRAFF_WINDOWS_SHELL=bash`), add Windows UTF-8/non-interactive env defaults, read-only skill sibling reads, `read_file` `contains`+range support, and empty/null optional tool-arg tolerance. - \\ - \\0.302.74.73 - \\ • Keep Orchestrator phase directives turn-scoped and allow quoted `rg` alternations, read-only `gh`, stdout `pdftotext`, and `readOnlyHint` MCP tools in plan mode (#1908). - \\ • Skip Auto Orchestrator planning on informational `?` questions or compacted histories, and settle expanded Phase 2 compaction inside `compactAt()` (#1906). - \\ • Strip Codex `tool_search_*` items on non-Codex Responses failover with local unfolding, and avoid auth-refresh loops on `subscription_sharing_unsupported_capability` (#1907). - \\ - \\0.302.74.71 - \\ • Add Orchestrator Mode (`off | auto | on`) same-chat read-only planning and KV-identical executor handoffs with short follow-up and wake bypasses (ADR 0347). - \\ - \\0.302.74.69 - \\ • Track live rate-limit headers, WebSocket `codex.rate_limits`, and `claude-agent` events, and bypass exhausted `claude-fable-5-2` pools onto `claude-opus-5-5` (ADR 0346). - \\ • Surface signed-in account identities, plan credits, and remaining usage windows across `/usage`, `/cost`, ACP `graff/accounts`, and `_codegraff/usage.planWindows`. - \\ • Advertise `_meta["codegraff/accounts"]` and all 7 terminal `authMethods` in ACP `initialize`. - \\ - \\0.302.74.68 - \\ • Add dual-frontier worker tier selection (`hard_frontier` at `xhigh`/`max`), single-model ladder bounds, and live-catalog rung fallback (ADR 0345). - \\ • Surface separate plan quota pools in `/models` (`plan*`) and ACP `graff/models` (`separateQuota`), with intra-plan frontier stepdown on usage limits. - \\ • Restrict cross-provider flat-rate borrowing to `.plan` subscription logins and keep equal-scored rungs on the active plan. - \\ • Enable the workspace-owning ACP client profile for `zeron` in `graff acp` and read `~/.simple-harness-keys.json` before macOS Keychain lookup. - \\ - \\0.302.74.67 - \\ • Add opt-in runtime extensions (`graff extensions`, `plugins enable|disable`) and worker-inherited `policies` across REPL and ACP (ADR 0343). - \\ • Add opt-in post-write `formatters` with verified-byte restore and multicall symlink preservation. - \\ • Add read-only `lsp` queries and `graff lsp` CLI (`status`, `setup`, diagnostics, hover, definition, references, symbols) with folded root schemas, `~/.graff/lsp/settings.json` defaults, and optional POSIX warm reuse (ADR 0344). - \\ • Accept `provider/model` qualifiers on `--subagent-model` across REPL and ACP. - \\ - \\0.302.74.66 - \\ • Repair paused request history and avoid invalid worker retries (#1895). - \\ • Retain longer prompt-cache TTL across returning interactive turns (#1898). - \\ • Reconcile application targets and strengthen background-safe GUI verification guidance (#1899, #832). - \\ • Add schema-ownership coverage and fix Windows test linking (#1892, #1900). - \\ - \\0.302.74.65 - \\ • Improve startup responsiveness across the REPL, one-shot, and ACP paths. - \\ • Avoid blocking startup on optional companion handshakes and model-catalog refreshes. - \\ • Explain output-length continuation exhaustion in ACP while preserving partial work (#1893). - \\ - \\0.302.74.64 - \\ • mcp: tool schemas are deep-cloned before normalization — mutating the Value copy through shared ArrayHashMap storage corrupted the parsed tools tree, so debug builds segfaulted in mcp_cache.store right after a successful connect (#1892) - \\ - \\0.302.74.63 - \\ • `read_file` honors `offset` and `limit` line-window aliases when `start_line`/`end_line` are omitted (#1885) - \\ • native file tools normalize `/…` absolute paths inside the workspace instead of rejecting them (#1886) - \\ • `tool_balance` treats read-only `shell` inspection commands/pipelines as `.read_explore` without resetting read streaks (#1887) - \\ • standalone `shell` `sleep N` polls run inline alongside background jobs and no longer hold `run_idle` open (#1888) - \\ • headless `graff -p` wires `SIGINT` to `esc_cancel` and flushes the final `result` record (#1889) - \\ • `write_file` and `edit_file` append inline JSON, Zig/ZON, Python, and Go syntax diagnostics (#1890) - \\ - \\0.302.74.62 - \\ • the primary bot is your main bot (ADR 0335): one thread for every project; in a client such as Harness it starts each piece of work as its own chat where it belongs and reports when the client says it stopped; it asks only when the project is unclear - \\ • from a terminal the main bot names the project and gives the command; single-job bots stay as helpers it hands their own jobs to; Bots are still off until `graff bot enable` - \\ - \\0.302.74.61 - \\ • Bots (experimental, off until `graff bot enable`): standing roles with one job and their own memory, work log and routines in `.graff/bots//`; `graff bot run`, `routine add`, `tick` for cron or a client's timer (ADR 0333) - \\ • `graff bot --user`: your own bots in `~/.graff/bots` for any project, one `--primary`; a late check records what it missed instead of running it late (`--stale-after`, `graff bot missed`); `graff bot state --json` for clients (ADR 0334) - \\ - \\0.302.74.60 - \\ • runaway raw reasoning streams stop at a byte bound, run loop detection (`ReplyRunaway` -> `ModelLoop`), and skip the 270s empty-reply backoff (#1878) - \\ • `delta.reasoning` normalizes into `reasoning_content` on replay and `ReplySize` (#1881) - \\ • `#390` final budget-landing calls keep `tools` with `tool_choice: "none"`, settle reasoning-only `length` stops cleanly, and exit `0` when tool work completed (#1882) - \\ • `edit_file` resolves multi-line `old_string` blocks with stripped line-1 indentation and restores the matched prefix (#1880) - \\ • unattended `-p` action turns nudge long read-only streaks and bounce once on zero-mutation unfinished prose (#1879) - \\ • explicit truncation markers on capped instruction/skill/MCP sections (#1871), `CLAUDE_CODE_USER_EMAIL` stripped in `claude-agent` (#1870), and worktree-scoped `codedb` guards + exit-code checks (#54) - \\ - \\0.302.74.59 - \\ • Claude Fable 5.2 in the baked catalog: `--model claude-fable-5-2` resolves by its bare name and is priced - \\ • image reads in one tool batch all reach the model, in call order (#1866): a bounded queue (16 images, 24 MB) replaces the single slot, with a tool error past the limit - \\ • uploads a client leaves as an `Attached images` trailer are staged as native images instead of paths the file tools refuse, also when sent mid-turn (#1865) - \\ • `graff refresh` / `graff models` / `/models` say when a provider rejects the live model-list request instead of silently showing the cached list - \\ • Lean 4 ImageDelivery kernel (3,870 verified cells): every staged image is delivered once or refused, in call order - \\ - \\0.302.74.58 - \\ • prompt-cache stability (#1320): jev_auto off by default (opt-in via /effort auto or GRAFF_JEV_AUTO=1), BustCause.effort HUD tracking, deterministic per-turn prompt_cache_key fallback, and in-session HTTP MCP schema retention - \\ • ACP low-latency path + Lean 4 AcpLatency kernel (3,768 verified cells): 0-subprocess in-process .git worktree resolution, single-parse ACP inbox routing, batch-flushed LineWriter/EventSink with reusable scratch arena, and single-pass startup MCP config load - \\ - \\0.302.74.57 - \\ • adaptive codedb workspace threshold (<8 code files in-process list_dir+read_file; >=8 code files codedb) + ~/.codedb/projects CoW seeding - \\ • pre-tool transcript checkpointing (beforeToolBatch), crash-tail suffix replay, and effect-aware TOOL_INTERRUPTED_READONLY recovery - \\ • deferred, single-pass streaming, cached (.graff/folded_priors.json), and ring-capped trajectory archive folding (#1864) - \\ • Lean 4 resource/durability kernel (3,742 verified cells) + scoped arena and O(1) incremental hashing optimizations - \\ - \\0.302.74.56 - \\ • unified subagent steering (@agent-N, @sa-..., /workers send, ACP graff/agents action: "steer") for foreground and background workers - \\ • named background subagents/tasks in REPL standing bar and ACP session_info_update (_meta["graff/waiting"]) - \\ • cross-turn ACP tool_call_update progress streaming for detached subagents and immediate user follow-up wakeup in agent_output - \\ • monotonic rewind-proof attempt ledger across /undo and /rewind preserved in compact_handoff_note.durableState - \\ • wall-clock grid-aligned recurring schedules (nextGridSlot) with @every and --missed=once|skip policies - \\ • APFS clonefile / Linux reflink Copy-on-Write cache seeding in subagent worktrees and adaptive sub-100ms foreground bash polling - \\ • corrupt-safe playbook JSONL self-healing + storage_warning diagnostics, markdown agent line:col validation, and OAuth child-env scrubbing - \\ - \\0.302.74.55 - \\ • scoped temporary arena parsing in mcp_cache.lookupWith/store/forget and single-server tool entry cloning - \\ • per-request scratch allocation in agent_ws.postResponsesWs and scoped arena cleanup in session_workspaces.remember - \\ • automatic same-plan subscription credential failover in providers.runTurnWithFallback - \\ - \\0.302.74.54 - \\ • Phase-1 deterministic pre-compaction tool-result pruning (compact_prune), crash-tail TOOL_OUTCOME_UNKNOWN vs TOOL_NOT_STARTED repair, and key-sorted repeat hashing - \\ • content-addressed dirty workspace snapshots (workspace_tree_snap) for edit_contract and hook/fsmonitor/LFS-safe subagent worktrees - \\ • Linux bwrap --unshare-pid isolation, Landlock ABI probing, and read-only sandbox denial boundary tagging - \\ - \\0.302.74.53 - \\ • workspace-invariant skill catalog prefix ordering and deepseekChat repo_context split for cross-directory prompt caching - \\ • Kimi global (kimi.ai) vs Mainland China (kimi.com) OAuth region resolution and k3.1-pro catalog + plan routing across CLI and ACP - \\ - \\0.302.74.52 - \\ • read-only trajectory inspection and verified user-facing chat title + chat ID reporting - \\ • background workers preserve independent tool-result sequence counters; synthetic compaction skips mcp_preload scans - \\ • workspace-invariant skill catalog prefix ordering, deepseekChat repo_context split, and Kimi global (kimi.ai) + k3.1-pro catalog routing - \\ - \\0.302.74.51 - \\ • graff sessions reads 64 KiB header prefixes in a reusable scratch arena; session load/resume frees parse trees after load - \\ • single-copy message arena reclaim, scoped trajectory archive folds, and faster remembered-model startup - \\ • -Dminimal and per-subsystem build flags (-Dlearn, -Dserve, -Dacp, -Dextras), codedb symbol ranking, and --eval contrastive reflection - \\ • zero-binary ServerlessHarness SDK runtime with InMemoryWorkspace, pluggable tools, lifecycle hooks, steer/followUp, and state snapshots - \\ - \\0.302.74.5 - \\ • Claude Agent ignores unprefixed tool calls rejected by Claude Code and keeps earlier images across checkpoints - \\ • read_tool_result keeps originating workspace handles across workspace switches - \\ • background shared_cwd workers stay pinned to their spawn workspace when the root session switches trees - \\ - \\0.302.74.4 - \\ • WebSocket Responses turns keep server token counts without re-compacting on ciphertext size - \\ • compaction and session switches reclaim dropped history memory; ACP startup and session/load are faster - \\ • --session-id and -c/--continue launch flags, versioned session files, and upright EXIF photo uploads - \\ - \\0.302.74.3 - \\ • Responses compaction uses server token counts instead of re-compacting on ciphertext size - \\ • cross-format model switches expand compaction checkpoints first and drop tool-only preambles - \\ - \\0.302.74.2 - \\ • Windows installs can recover through GitHub's asset API without skipping checksum verification - \\ • nanocua keeps text and refs by default; coordinates are available with include_frames - \\ • frameless reads skip unused geometry queries; controls-only filtering remains opt-in - \\ - \\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/ - \\ - \\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 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 - \\ -; +pub const changelog_text = @import("cli_changelog_source.zig").changelog_text; /// Release builds leave `graff learn` out (build option `learn`). const learn_usage = if (@import("build_options").learn)