coding-agents
tool-selection
bialy / release.json
paintedwolf-ai's picture
Release open1-b5-e4-bialy
39012d0 verified
Raw History Blame Contribute Delete
26.5 kB
{
"version": 1,
"release": "open1-b5-e4",
"backbone": {
"model": "convaiinnovations/laya-multilingual",
"revision": "e4e9ddf21a7b1903b7acffd8814ad4307bf63a67"
},
"heads": {
"turn-load": {
"sha256": "550f30b94d5c6a8680d2cb2a607ebc9ee33e709c7679f72606d1e9fda17f2e90",
"label": "open1-turn-load-B5-release-independent"
},
"unit-rank": {
"sha256": "c62674993eabba94660eccfafbb6e04d1549adc7335656c0ddefa7f068fa58f7",
"label": "open1-unit-rank-e4-dense1"
},
"code-rank": {
"sha256": "abe235c347a223e957a0f8b7b979e3a584aebab76ee239e5f62117e839b9c40a",
"label": "open1-code-rank"
}
},
"preload": {
"catalog_revision": "dc3bf9215aac8877",
"encoding": "independent",
"option_words": 60,
"max_len": 1024,
"head_tokens": 512,
"options": {
"answer_decision": "Answer a worker's needs_decision: pick one of its offered options; the worker resumes on its same session with your choice.",
"ask_user": "Structured question for the user \u2014 this tool is the only way to ask one, and assistant prose is not an answer. Host parks until the user resolves; next wake returns this tool result. Prefer choices[] when present. Parallel while pending \u2192 ASK_USER_ALREADY_PENDING. Fill args per field docs below.",
"capture_page": "One-shot open\u2192act\u2192snapshot\u2192close; prefer page_* for live/authenticated state. Use project_dir for a static tree or url plus process_handle for a loopback app. Targets are exclusive. Returns structure, requests, errors, console, and a screenshot; use filmstrip for settled sequence claims and timeline for anything that happens while the page is painting.",
"chmod": "Change approval bits on repo-relative paths.",
"chown": "Change file ownership within write scope (never use command chown).",
"code_rewrite": "Rewrite a code shape across path or paths[]. $VAR matches one node; $$$VAR a sequence. Narrow bare `$A | $B` patterns (CODE_REWRITE_PATTERN_TOO_BROAD). Parse-breaking rewrites are refused (MUTATION_BROKE_PARSE). For search, use grep structural: true. Track changed, blocked (code/details), failed, unobserved_files and truncated separately: blocked is not no-match, and successful siblings may already be saved. Address each cause without repeating saved mutations",
"command": "Needs `command` or `pipeline` (not both). Runs as argv without a shell: `&&`/`||`/`;` sequence commands, `pipeline` carries `|`, `>`/`>>`/`2>` redirect, globs and `@scratch/<path>` expand, `$()`/backticks reject. Leading `NAME=value` assignments scope to their command; shared variables belong in `env`. Pass cwd; `cd` is rejected. Omit output paths to return output tails.",
"command_output": "Read command job output since cursor and current host-observed network facts. Inspect the handle of the process making a failed connection before attributing an application error to the sandbox. `network` records observed destinations and allow/deny decisions; `boundary_refusal` reports host attribution. Allowed access does not imply a successful upstream response.",
"command_stop": "Request termination of a command job process tree. stop_requested acknowledges the request; running stays true until exit is observed. Request once, then wait on the handle or inspect command_output rather than repeating stop.",
"complete_leg": "Finish this worker leg. Call once when the assignment is done or blocked. The host builds the parent report from these fields plus disk proof.",
"copy": "Copy files within write scope (files only; never use command cp).",
"delegate_decompose": "Decompose an approved plan into delegation legs",
"delegate_dispatch": "Dispatch a delegation leg to the worker queue",
"delegate_init": "Create a delegation departure for the current coordinator session",
"delegate_status": "Return delegation departure status and legs",
"delete": "Remove files and empty directories within write scope; the person can undo each removal.",
"diff": "Unified diff between two repo-relative file paths (read-scope only; never use command diff).",
"edit": "Replace an exact span in a write-scope file. Batch independent edits to distinct files. Read in an earlier response (same-batch read+edit rejects: EDITOR_DOCUMENT_READ_REQUIRED). Copy old_string after line-number tab; unique unless replace_all. Parse-breaking splices: MUTATION_BROKE_PARSE. New files: write.",
"extend_worker_budget": "Grant a running worker a higher tool-round ceiling, usually in answer to its request_budget; the worker is told on its next round and its request closes.",
"extract_archive": "Unpack .zip or .tar.gz into a write-scoped directory (replaces unzip/tar xzf; HITL ask by default).",
"fanout_plan": "Stamp a planned parallel fan-out (1\u2013N read-only legs). Call on fanout_planned phases, then workflow_advance. Stamping records the plan only; it does not enqueue workers. In the worker phase, dispatch each leg with task and its host-issued workflow_work_id. Recover incomplete legs within the plan's attempt allowance.",
"fetch_url": "Read a web page, documentation site, or other online resource. mode=text (default) extracts readable page text (HTML\u2192markdown); large bodies return a head + symbol map \u2014 page with offset+limit (cached). mode=raw returns the verbatim resource: textual MIME (HTML/CSS/JS/JSON/SVG/\u2026) as text with the same paging; image MIME types without dest automatically ingest into the session visual store and emit visual captures",
"find": "Bounded recursive directory walk \u2014 canonical for ** name_glob. Default type=any returns files and directories together. Wide unscoped walk at path=\".\" (2000+ paths) returns view digest (selected=0) with distribution + FIND_OVERFLOW_NARROW \u2014 pass name_glob when you can name a filename pattern. A concrete path (not \".\") or name_glob returns literal results[] (same as list_dir targeted). Explicit max_results or offset also",
"git_blame": "Per-line blame JSON (`available`, `lines` with commit/author/date/content).",
"git_branches": "List local branches JSON (`available`, `branches` with name/current).",
"git_checkout": "Switch to an existing local branch, or explicitly create and switch to a new one. Rehearses and reviews file changes before applying. Preserves unrelated local edits; refuses conflicting edits. Reports the actual branch and HEAD even after failure.",
"git_commit": "Stage and commit paths with a message; amend=true replaces HEAD with the selected paths, including new files, leaving unrelated staged changes alone. Success JSON: available, hash, paths, and uncommitted_paths/uncommitted_count for dirty files this commit left alone. Failure JSON: available=false, error, optional hook_failed, output_tail (last hook/CLI lines), and hint. Empty paths stages only files this session wrote; name paths explicitly to",
"git_compare": "Compare frozen commit IDs: ahead/behind counts for head_ref relative to base_ref, plus all merge bases.",
"git_diff": "Git diff JSON, one page of files in git's order (`available`, `files` with per-file insertions/deletions and, unless `stat`, that file's hunks in `diff`; `available:false` on non-git projects). Pages hold up to 80 files and 5 KiB of hunks by default: read `files_total`, use `next_offset` for more files when needed; exhaust the selection for tasks requiring complete coverage. `wire_spill_path` retains the",
"git_log": "Read paged commit history from a branch, tag, commit, or revision range. Use ref for revisions and path only for a file. main..feature lists commits unique to feature; main...feature lists both sides. Follow next_offset to continue.",
"git_merge": "Merge a revision into the current branch. Rehearses and reviews file effects. Conflicts are explicit and retain the merge state: edit resolutions, then continue with explicit paths; abort only when asked to discard merge work. Check status and after state, not diagnostics, for success.",
"git_ref": "Resolve refs to SHAs JSON (`available`, `refs` with ref/sha/peeled).",
"git_restore": "Restore explicit paths in the Git worktree, index, or both. Defaults to Git's ordinary worktree-from-index behavior. This discards selected state \u2014 use only when the user explicitly asked to revert named paths.",
"git_show": "Show commit or file at ref JSON (`available`, `show` with kind diff/blob).",
"git_stash": "Save explicit repository paths temporarily, apply a saved stash, or explicitly drop one. Save preserves unrelated edits and index entries. Apply retains the stash even on success; use drop separately after verifying restoration. Conflicts and partial effects are reported in structured state.",
"git_stash_list": "List stashes with immutable IDs and current selectors for apply/drop.",
"git_status": "Git working tree status JSON (`available`, `branch`, `head_short`, `upstream`, `ahead`/`behind`, `dirty`, staged/unstaged/untracked counts, `files` with porcelain XY status one entry per file \u2014 untracked directories are expanded; `available:false` on non-git projects). Pages of 80 files: read `files_total`, use `next_offset` to continue when more coverage is needed; exhaust pages when the task requires the complete selection. `paths` narrows the selection to",
"grep": "Search files with RE2 text patterns or tree-sitter structural patterns. Recurses under path with optional path_glob, searches gitignored files, and skips engine/VCS metadata. Wide results may return a view digest.",
"handoff_health": "Report reservation counts for a handoff session plane.",
"handoff_init": "Bind the parent session handoff plane for sibling worker coordination.",
"handoff_release": "Release one or more path reservations held by the caller agent.",
"handoff_release_all": "Release every path reservation held by the caller agent on the session plane.",
"handoff_reserve": "Record repo-relative path reservations on the parent session plane (informational; merge resolves overlaps).",
"held_result": "Read a held call by its handle, running or settled with its original outcome, codes, and result. wait_ms waits up to 30 s for a running call.",
"held_stop": "Cancel a running held call. Request once, then wait on the handle or read call_result.",
"http_request": "Send one bounded request to an API, webhook, or service endpoint and return its status, headers, and raw body.",
"jq": "Query JSON, YAML, or TOML via sandboxed jq. Pass limit/offset for literal values[].",
"jq_edit": "Update a JSON, YAML, or TOML document with a jq program and write the result in the same format, in place or to dest. The program returns the whole document (`.a.b = 1`, `del(.x)`, `.items |= map(...)`). Unchanged values keep their literals, key order, and YAML comments; syntax a rewrite would lose is refused (JQ_EDIT_LOSSY).",
"list_dir": "Locate repository areas with {\"path\":\".\"}. A named directory or explicit listing controls return entries instead. Use summarize to explain the area.",
"measure_page": "Measure selector geometry in CSS pixels, and whether the pointer can reach each element or what covers or clips it. Use id for an existing live page, or one exclusive project_dir/loopback url target. annotate adds boxes; page_geometry remains the numeric evidence.",
"mkdir": "Create directories within write scope (never use command mkdir).",
"move": "Rename files within write scope (never use command mv); a replaced destination can be undone.",
"pack_board": "When the board is stale, workers changed, or you need a scan/layout pulse \u2014 return the board snapshot. Not a substitute for summarize or a known-file read.",
"page_act": "Drive a held page with trusted input. Each step aims where its target receives the pointer; the result reports that point plus the requests, errors, and console output the drive caused. With record, the drive is also a timeline seen as a labeled contact sheet.",
"page_close": "Tear down a held page; optional final snapshot.",
"page_open": "Open a live page; returns an id held across turns. The page has its own cookies and storage, which persist until page_close. Up to max_pages stay open per session; re-opening a held target reloads it with the given viewport and routes. Prefer capture_page for a single check. Bindings match capture_page (project_dir or url).",
"page_snapshot": "Universal {state, snapshot, log, artifact} for a held page (does not close it).",
"parse_decomposition": "Validate decomposition.",
"parse_delegation_plan": "Validate delegation plan.",
"parse_evaluation": "Validate evaluation.",
"parse_extract_json": "Extract JSON.",
"parse_validate": "Validate JSON.",
"plan_append_review_evidence": "Append advisory plan review evidence to critic.jsonl",
"preview_overlay": "Merge preview \u2014 same plan as promote_overlay without landing. overlay_id == job_id. Returns path_status and ready_resolutions; spill_path is host-data-relative (promote-spills/<id>.json). Example: {\"overlay_id\":\"<job_id>\"}.",
"process_list": "Inspect host processes after approval. Returns names, executable paths, UID, parent PID, and opaque instance references; no argv or environment. Use pid for a known target. Use next_after_pid for another page of the changing snapshot. Unavailable metadata does not mean exited. Use command_output and command_stop for jobs you launched with command.",
"process_signal": "Send a signal to exact host process instances after approval. Pass opaque references returned by process_list in this task. PID reuse or changed identity rejects; names, bare PIDs, groups, and descendants are not targets. TERM requests graceful termination; KILL can discard unsaved work. Delivered means the signal was accepted, not that exit was observed. Inspect the result for each target.",
"promote_overlay": "Atomically promote a worker overlay (overlay_id=job_id), considering all observed changes. Conflicts return a menu without writes; resolve all via resolutions/drop.",
"read": "Read a file and return its content or structure in JSON; unscoped large files return a structure digest. For literal content, set mode / offset / limit / ranges / symbol.",
"recall": "Look up what was already observed instead of observing it again. Reach for it when a finished worker leg's envelope omits a detail the leg saw, when a continuation record or a `[compacted \u2026]` banner names a handle whose body is gone from context, or \u2014 from a root session \u2014 when an earlier session on this project may already",
"record_finding": "Post peer intent for parallel siblings \u2014 shared paths, registry slots, merge-relevant facts. Cite a real path/handle.",
"reject_overlay": "Abort overlay without landing. overlay_id == job_id. Wrong/superseded only \u2014 not landable conflicts. Example: {\"overlay_id\":\"<job_id>\",\"reason\":\"Superseded by corrected dispatch\"}.",
"render_view": "Rasterize SVG or HTML into an authored visual: UI/product mockups (especially before building), diagrams, comparisons, timelines, or charts. Use when spatial layout improves clarity; never as runtime evidence \u2014 use capture_page for observed UI. Host injects an offline design kit (OFL fonts, Lucide-adapted icons, theme tokens, spacing/type scales, media placeholders). No CDN links, scripts, or external https assets \u2014 use",
"replace_lines": "Atomically replace or reindent line ranges. Use direct fields once, or operations for coupled non-overlapping seams in one snapshot. Read existing editable files in an earlier response before applying changes; concurrent edits rebase against that read. EDITOR_DOCUMENT_READ_REQUIRED or EDITOR_DOCUMENT_CHANGING requires read and recomputation.",
"request_budget": "Ask your coordinator for more tool rounds when this leg's goal needs more than remain. Name the rounds and the concrete work they cover. You keep working while the coordinator decides; a granted ceiling arrives as a host notice on your next round. If none arrives, finish with an honest partial report at your current ceiling. One request can be",
"request_decision": "Escalate an ambiguity you can't resolve in scope \u2014 question + 2+ options, the accepted call suspends your run. The coordinator answers and you resume with the same workspace. Do not also call complete_leg.",
"request_tools": "Load more tool schemas for the next model call by describing what the work needs",
"restore_version": "Restore a file to a previously retained source version as a forward mutation in the source ledger. Use source_history to inspect available versions or diffs.",
"scan_compare": "Diff two complete scans \u2014 returns the findings introduced, resolved, and unchanged since the baseline. After a fix, pass the post-fix scan as new_scan_id and omit old_scan_id: the host resolves the previous complete scan in the same category. Use it to prove a fix landed; a board Scan: tail is not proof.",
"scan_list": "List recent code scans for the active project (read-only). Returns a slim id roster with no guidance bodies \u2014 use it to resolve a scan_id for scan_summary, scan_query, or scan_compare when the board does not already show one.",
"scan_pack": "Run the host's configured code scanners over the active project and return one scan_pack. Use it when the ask is security or code-quality work and the board shows no complete scan at the current head. Returns per_scan[].scan_id \u2014 pass that exact array to scan_summary or scan_query for a cross-engine aggregate, or set completion=summary to receive the aggregate when this call",
"scan_query": "Page findings for one or more scans. One id queries one scan; the exact id array returned by scan_pack queries that pack as a cross-engine aggregate. Narrow with rule_id, path, code, or level rather than reading every finding. When a result is truncated, page it with offset \u2014 do not re-run the scan.",
"scan_summary": "Read status and guidance for one or more scans. One id reads one scan; the exact id array returned by scan_pack reads that pack as a single cross-engine aggregate, and a pass_id reads a full pass, including scanners that have not started yet. This is the first call after a scan completes \u2014 start here rather than grepping the repo",
"secret_generate": "Generate random secret material and return an opaque {{paintedwolf-secret:\u2026}} reference for tool execution.",
"secret_list": "List visible managed secret references and metadata, never values.",
"secret_revoke": "Revoke a managed secret reference while retaining provider-screening evidence.",
"skills_read": "Read a skill or resource by name, find skills by describing the work, or omit arguments to list every loaded skill.",
"source_history": "Query the source ledger for file effects (mode=effects), current line authorship (mode=lines), this session's paths (mode=mine), historical version content (mode=version), or version diffs (mode=diff). Includes uncommitted state. Actors: you / agent / user / external / mixed / unknown. Missing history is reported explicitly.",
"stat": "File metadata (mode, size, mtime, entry_count for directories)",
"state_close": "Exit the active workflow run for the session",
"state_query": "Query active workflow run and scaffold vars",
"state_start": "Propose a workflow run on the current session. The host records the proposal; an explicit human action starts it.",
"state_update": "Update model-authored scaffold data on the active workflow run",
"submit_verdict": "Record the phase's review_loop verdict per verdict_schema. Terminal verdicts satisfy its evidence_passed gate; other gates and advancement rules still apply. Non-terminal verdicts re-loop within the cap. Terminal verdicts must cite observed evidence paths/handles in cited_evidence or fetched URLs in cited_urls. Invented references reject; non-terminal verdicts need no citations.",
"summarize": "Explain a named file or area with a bounded, source-cited briefing. Path sets scope; task ranks within it. Use next_actions only when more evidence is needed.",
"surface_note": "Post a concise, grounded mid-turn note for the user, then keep working \u2014 this does not end the turn. Use it for meaningful progress or expectation changes. Say what happened and what comes next. Cite at least one cited_evidence or cited_urls entry. Carries artifact_ids, which is how a still reaches the user mid-run. For a question use ask_user; the final",
"survey_repo": "One catalog map (Tier 1 digest, selected=0), not a briefing. Call when you still cannot name a destination and the ask is unknown layout (layout_overview), duplicated env/config/stores (ssot_drift), or where HTTP routes/handlers live (api_routes). Then summarize the destination. For how/what on a named file/package/dir, use summarize(path=\u2026) \u2014 never survey_repo for depth on a path you can already say.",
"task": "Start background workers for parallel execution across concurrent legs and return control immediately. Dispatch ready legs together in the same turn or launch companion legs while work runs. For single-thread, localized changes, execute inline directly. Match the worker profile and scope to the assignment.",
"terminal_close": "Tear down a held pty (graceful then kill). It returns lifecycle acknowledgement only; use terminal_read or terminal_snapshot before closing when output matters.",
"terminal_open": "Live host terminal (pty): run a command-class argv command. Prefer command when you only need exit code/tail for yourself. Use this tool when the session must be held or the user should see the screen \u2014 then terminal_snapshot. Returns an id held across turns plus the requested observation (screen by default). Never supply known plaintext via later terminal_send \u2014 use",
"terminal_read": "Read incremental pty output since the last read (idle/timeout bounded).",
"terminal_send": "Write literal text and/or named control tokens ({Enter}, {Ctrl-C}, \u2026) to a held pty, then return the settled vt grid. Never send known plaintext passwords, tokens, or passphrases. An opaque {{paintedwolf-secret:\u2026}} reference is allowed when the interactive prompt is the consuming boundary; the host resolves and screens it.",
"terminal_snapshot": "User-visible photograph of the held terminal session (CLI output or TUI), parallel to capture_page for pages. Returns one bounded VT screen and a receipt, never a duplicate main-stream log. Use for any held session \u2014 not only full-screen TUIs. Settles on output quiescence; does not advance terminal_read's cursor.",
"update_progress": "Set session progress checklist (markdown). Replaces the whole checklist \u2014 see content for markers and caps.",
"verify": "Run a verification check and report its terminal outcome. Bare verify runs the selected project command. With no selected command, supply the documented check appropriate to the task. A different command or pipeline is rejected when a project command is selected; use command(verification: true) for targeted checks. Selecting a command does not require it after every edit. outcome is passed",
"view_image": "Inspect an image file (.svg, .png, .jpg, .jpeg, .webp, .gif) or a session visual and perceive its rendered pixels and extracted text. Returns format metadata and extracted text, and emits a visual capture for multimodal models.",
"view_video": "See moments of a screen recording or other video (.mp4, .mov, .webm), including an attached one. Returns one sheet of frames, each labeled with its timestamp, plus the video's duration and size. Without times_ms it spreads count frames across start_ms to end_ms (default: the whole video); name exact times_ms to look closer at a moment, and crop to read a",
"wait": "Park without a worker slot until a condition is satisfied or timeout expires. Use until_complete for exact process handles without periodic timer wakes; omit conditions for a timer.",
"wc": "Count lines, bytes, and words in files or recursively in directories. bytes is the true file size; truncated marks counts taken from a capped scan.",
"web_search": "Search the web by time window. The host selects Direct search or custom providers from Settings \u2192 Web research.",
"worker_cancel": "Cancel a pending or running worker job; running workers get a prose-only cancellation report before the leg ends",
"workflow_advance": "Leave the current coordinator-controlled phase when gates pass. Returns phase_gate_unmet with failed_leaves when blocked; already_advanced if host auto-advanced first. Unavailable on host-auto phases.",
"workflow_catalog_summaries": "List bundled, session-composed, and template workflow ids.",
"workflow_compose": "Compose a session workflow from YAML. Returns effective_summary; host replays coordinator_brief each Prompt.",
"workflow_compose_from_template": "Compose from a bundled template. Returns effective_summary; read prepended coordinator_brief before proposing with state_start.",
"workflow_persist": "Save a session-composed workflow to .paintedwolf/workflows/<workflow_id>/workflow.yaml (requires user confirmation).",
"workflow_transition": "Fire choice leave by transition_id (Phase exit). Pending input latches block; human_approval_awaiting does not.",
"workflow_user_feedback": "Pending user feedback for the active run. Wait for the host card \u2014 do not infer answers.",
"write": "Create/replace files in write scope (append: true appends). Batch independent writes to distinct files; stage same-file or dependent writes. New or previously clean supported source must parse (MUTATION_BROKE_PARSE); broken files must improve and finish clean. Scope violation: WRITE_SCOPE_DENIED. Read existing editable files in an earlier response; concurrent edits rebase against it. EDITOR_DOCUMENT_READ_REQUIRED or EDITOR_DOCUMENT_CHANGING requires rereading and recomputatio"
}
}
}