Tools and safety model
Feature docs index · Repository README
Choose a tool below, then check its path, output, and credential limits before use. Shell guardrails are not an OS sandbox.
| Tool | Contract |
|---|---|
read | Read local files and skills. Local reads supply tags and line anchors for hash_edit. |
bash | Run non-interactive shell commands or bounded background jobs. Aliases: shell, cmd. |
bash_job | Read output, inspect, stop, or list this session's background jobs. |
hash_edit | Edit existing UTF-8 files using current read tags and visible line anchors. Supports block edits, moves, and deletion. |
write | Create or replace a UTF-8 file, up to 1 MiB. Creates parent directories and rejects final symlink targets. |
grep | Search file contents with regex or literal patterns, ranked or raw results, and pagination. |
find | Find file and directory paths with fuzzy matching and ignore rules. |
list_files | List direct children of a directory, including hidden and ignored entries. |
ast_grep | Search source structure, preview rewrites, or list symbols. Does not modify files or require an external CLI. |
view_image | Send one local PNG, JPEG, GIF, or WebP image to the configured vision model. |
subagents | Delegate independent tasks to bounded parallel child sessions. See Parallel subagents. |
web | Search public sources or open a public URL. Does not control a browser. |
code_mode | Run synchronous JavaScript that calls permitted tools and returns selected results. Enabled by default. See Code Mode. |
magi_control | Inspect context usage or compact the current session. Direct calls only. See Sessions and context. |
Availability depends on settings, agent permissions, and approved MCP servers. Direct tools and tools inside Code Mode have separate enable/disable lists.
Every built-in and MCP tool requires a top-level intent: a nonblank string briefly explaining why the call is needed. This includes code_mode, magi_control, and each built-in or MCP call inside Code Mode; an outer batch intent does not replace inner-call intents. Missing, non-string, empty, or whitespace-only intent fails before tool execution. Subagent task objectives remain separate from the batch intent. MCP calls put server inputs in a separate arguments object; only that object reaches the server, preserving any server-owned intent field. Client-registered tools retain their own schemas and do not require harness intent.
Concurrent calls
Consecutive direct calls to read, find, list_files, grep, ast_grep, web, and view_image from one provider response run with at most four scoped workers. Code Mode's synchronous tools.callAll([{name, args}, ...]) uses the same worker path and returns results in input order (at most 32 entries). Calls outside this list, unavailable or rejected calls, and mutation tools form barriers: earlier calls finish before the barrier runs, and later reads observe its changes. Individual tools.call calls remain sequential. MCP and client tools do not run in this parallel path.
Any configured hook keeps the response sequential, including hooks for other phases and disabled hook configurations. Tool permissions, inspection ceilings, per-surface exclusions, and argument validation still apply through normal dispatch.
All calls in a concurrent run are recorded and shown as started before dispatch. Results, execution evidence, read-snapshot coverage, instruction loading, provider context, and finish events are committed in original call order on the lifecycle thread. Cancellation reaches every worker; workers join before ordered result/cancel delivery and turn termination. Completed results remain recorded; unfinished calls receive failed results rather than dangling call records. A slow earlier call can delay display of later results.
Shell commands
bash requires command and a nonblank intent that briefly explains why the command is needed. The transcript card shows intent above command, then result summary. timeout is optional (1–600 seconds, default 600). Omit it for the full ten-minute allowance; set a shorter timeout only when needed. Runtime aliases shell and cmd also require intent.
On Unix, agent shell calls start Bash with pipefail: a failed pipeline stage makes the pipeline fail even when its final filter succeeds. This does not enable errexit and does not change hook shell defaults. Commands can disable pipefail or mask status explicitly, so this is not a sandbox or proof that a build passed. Output truncation and timeouts remain separate from process exit status.
Background jobs
Use bash with background: true for dev servers, watchers, and long builds. Do not add &, nohup, or setsid: the shell must stay in its managed process group. timeout is invalid with background: true; jobs have no lifetime timeout while their owner lives. The call observes up to 200 ms of startup, then returns JSON with job_id, pid (and Unix pgid), status, exit_code, uptime, stdout/stderr, and truncation metadata. Normal foreground commands keep their timeout and cleanup behavior.
Background Bash cards carry a background label. In Mission Control, the Background Jobs header panel (Alt-4, the header Panels menu, or /background-jobs) shows a live list for the primary conversation (not side conversations or subagents), including jobs started through Code Mode. With the panel focused, use Up/Down to select and k to stop its process group; Esc closes it when unpinned. It works during active prompts, never consumes job output, and keeps termination off the UI thread. Finished jobs disappear from this list; bash_job list still includes retained finished jobs.
bash_job requires intent and action. Omit inapplicable optional fields or pass null (including timeout for background Bash calls):
outputrequiresjob_idand consumes only unread stdout/stderr. Optionalwait_seconds(0–30) returns early on output or exit. Cancellation ends only the wait.statusrequiresjob_idand reports state, exit code, uptime, and dropped-byte counts without consuming output.stoprequiresjob_id, terminates the process group, and returns the final bounded output tail plus cleanup warnings. Unread bytes omitted from that tail count as dropped.listtakes no job ID or wait and lists retained jobs owned by this session.
Each owner permits four running jobs and retains at most sixteen jobs, evicting the oldest finished job when needed. A fifth running job fails with the running IDs. Each stream has a 256 KiB ring; old bytes are dropped and counted. Output calls stay within Bash's stdout/stderr limits and expose has_more_output. At most 64 background pipe-reader threads run per process; exhausting that bound fails job setup rather than creating more threads.
Jobs survive turn completion and Esc. Stop them when finished. Session switch, /new, application exit, and local-agent connection close stop owned jobs; subagent completion stops its own jobs, inaccessible to its parent. Unix cleanup targets the whole process group. Processes that deliberately escape it are not managed. Non-Unix cleanup kills only the direct child. SIGKILL of magi-code can orphan jobs; startup does not reap them.
Both tools use normal hooks, approvals, redaction, and permissions. Disabling bash also denies bash_job on that surface, including Code Mode, profiles, and task ceilings; inspection mode denies both. Code Mode jobs belong to the session, not the script. Background file activity is not inferred and never counts as confirmed changed paths. Calls/results are ordinary durable tool records, but live jobs are runtime-only: old, evicted, and foreign IDs return a not-found error rather than being replayed.
AST search and outlines
ast_grep defaults to operation: "search". Search requires pattern and accepts optional rewrite for a read-only diff preview. Use operation: "outline" to inspect symbols before reading source.
Neither operation needs an installed CLI. Embedded AST libraries infer language from file extension unless you supply one. Project configs and custom grammars are not loaded.
- Search returns
path:line:textwith 1-based lines. Rewrite previews identify each match by path, line, and exact end-exclusive byte range, then show removed-lines and replacement+lines. Surrounding text remains unchanged, and files are never written. Overlapping matches produce independent previews, not one combined patch. limitbounds returned lines (default 100, maximum 500); output is capped at 64 KiB. A boundary can cut a preview short, so checktruncatedbefore using it.- Pattern and rewrite each accept up to 4 KiB. Search accepts complete UTF-8 files up to 256 KiB, visits at most 10,000 files, and reads at most 128 MiB per call. Syntax trees are limited to 16,384 nodes and depth 64; patterns to 256 nodes. Replacement allocation estimates are capped at 1 MiB before expansion.
files_skipped,partial, andtruncatedreport oversized, unreadable, non-UTF-8, or overly complex files. - Parsing, including pattern parsing, checks cancellation and a shared 30-second deadline. Traversal checks each AST node before matching. A single bounded library match/replacement or filesystem operation cannot be interrupted mid-call. No detached workers are created. Unknown file types are skipped. In mixed-language searches, a pattern that cannot compile for one language makes results partial. Supply a language for strict pattern validation.
- Outline shares search's complete-file, file-count, scan-byte, and syntax-tree limits. Bundled rules cover Rust, TypeScript/TSX, JavaScript, Python, Go, Kotlin, Java, and Swift. Other languages are ignored.
{"intent":"Inspect Rust symbols before reading implementations","operation":"outline","path":"src","language":"rust","limit":30}
- For outline, omit
patternandrewriteor set them to JSONnull. Other values, including empty strings and placeholders, are rejected. Outline-only fields areitems(structureby default, orexports,imports,all),view(digestby default, ornames,signatures,expanded),name(Rust regex over top-level name/signature/first source line),symbol_type(comma-separated lower camel case types), andpub_members(default false). For search, omit these fields or set them tonull;pub_members: falsedoes not count as absent. Filters do not select members directly. Name/type filters are capped at 512 bytes. - All fields except
intentandoperationtreatnullas omission. Search still requires a nonemptypattern. If supplied,operationmust besearchoroutline. Providers that require every key can use"pattern":null,"rewrite":nullfor outline and null outline-only fields for search. Invalid mode-specific arguments report offending fields with an omit-or-null example. - Explicit
structure/digestdefaults retain internal symbols for directory and file requests.namesomits members and computes only names, though name filtering still computes signatures.signaturesincludes top-level signatures;digestadds member names;expandedadds member signatures. - Each output line is a complete JSON object with
path,language, anditem. Items contain names, types, signatures, export/import flags, ranges, and direct members selected by the view. Lines and columns are 1-based; range ends are exclusive. Byte offsets are omitted. limitcounts top-level entries, not members or source lines (default 100, maximum 500). A parent and its selected members are kept or omitted together. Returned entry text is capped at 64 KiB.- Outline checks cancellation and the shared 30-second deadline before and after each extraction iterator step and before accepting output. One step may traverse the remaining bounded file tree and collect direct members. Individual library calls, including compilation of fixed bundled rules, and filesystem operations cannot stop mid-call. The deadline is cooperative, not a strict timeout. No child processes or detached workers are created.
- Check
symbols_returned,files_skipped,partial, andtruncated. Byte and entry limits retain only complete earlier entries; deadlines return partial failure. Empty outlines succeed. Project or ancestorsgconfig.yml, custom language libraries, and custom outline rules are never loaded. Only approved regular files are read under the same path and read checks as search.
URL routing
Use the narrowest tool that meets the request:
| Need | Tool | Use |
|---|---|---|
| Public unauthenticated plain text, Markdown, JSON, or HTML ordinary GET | web open | Bounded native fetch/readability; no JavaScript or interaction. |
| Public research and page text | web | Search and open; no browser interaction or CSS extraction. |
| Custom or non-public HTTP work | bash | Custom methods, headers, bodies, redirect control, binary downloads, private/local services, or an explicit request. |
URL open applies bounded readability extraction to substantial HTML and converts short HTML to Markdown. Never block or rewrite curl.
web contract
Use one operation per call. Omit unused fields or set them to null:
{"intent":"Find official Rust release notes","operation":"search","query":"Rust official release notes","limit":5,"domainFilter":["rust-lang.org","-old.rust-lang.org"],"recencyFilter":"month"}
{"intent":"Read the Rust project homepage","operation":"open","url":"https://www.rust-lang.org/"}
{"intent":"Read a cached research result","operation":"open","reference":"web:<reference from a result>"}
- With active
openai-codex, search uses the active model and existing Codex OAuth/refresh athttps://chatgpt.com/backend-api/codex/responseswith nativeweb_search. Other providers use Exa/searchand require environment-onlyEXA_API_KEY. URL open fetches directly without credentials or an Exa key. Search backends do not fall back to each other or MCP. - Codex returns a labeled model-generated synthesis with URL citations, not page extracts. It creates no cached references; open cited URLs separately for page text.
limitrequests a source count but does not enforce it. Codex rejectsdomainFilterandrecencyFilter; omit both. The filtered example above applies only to Exa. - Codex search requires
response.completed, rejects failed or incomplete streams, and caps streams at 2 MiB. It reuses the provider's blocking transport, cancellation, byte-idle and semantic-progress watchdogs, and session-expiry refresh. Tool output remains capped at 48 KiB. - Query: 1 to 512 characters after trimming. URL: at most 2,048 UTF-8 bytes. Search results: 1 to 10, default 5.
- Search-only
domainFilteraccepts at most 20 bare ASCII domains (253 characters per domain, 63 per label). Prefix-to exclude. Schemes, ports, paths, whitespace, empty labels, and leading/trailing label hyphens are rejected. Entries are trimmed. Includes/excludes map to ExaincludeDomains/excludeDomains.recencyFilteracceptsday,week,month, oryear, mapped tostartPublishedDateat the current UTC date minus 1, 7, 30, or 365 days. It filters publication date, not crawl freshness. For open, omit filters or set them to null. Aliases and query batches are unsupported. - Only public HTTP(S) URLs are accepted. Credentials and sensitive URL parameters are rejected. URL open uses the DNS-pinned transport in
url_fetch.rs, which rejects private/local destinations and redirects. Exa result URLs are also validated; Exa controls retrieval of search extracts. - Exa search text requests are capped at 10,000 characters per page and responses at 2 MiB, with 5s/20s connect/request timeouts. Direct URL open extracts from at most the first 512 KiB (reads one extra byte to detect truncation) with 5s/15s connect/request timeouts and a 5,000-token text budget. Oversized pages return prefix-extracted text,
truncated: true, and a visiblepage exceeded 512 KiB; content extracted from the first 512 KiBnotice instead of failing. Both check cancellation, cache at most 40,000 bytes per page, and cap output at 48 KiB with truncation metadata. - Exa search shows source URLs, stable references, and bounded excerpts. Cached open returns stored text.
- Runtime stores up to 64 immutable page snapshots and evicts the oldest. Runtime clones share this cache. Sessions do not persist or restore it. An expired or unknown reference fails without network work; reopen the public URL for a new reference.
- URL open makes a direct GET and returns readable text, which may omit page content. Cached open returns the same snapshot without refreshing. Output gives no publication, crawl, or fetch timestamps; do not infer them from a reference.
Visible tool output
Mission Control shows pending and completed tool activity with safe status, metadata, and bounded output. Pending write/hash_edit entries do not include file contents or replacement payloads. web shows query/result/truncation metadata and redacts credential-shaped text. read shows requested paths. Subagent summaries omit raw child output. Shared summaries live in src/output/tool_summary.rs.
Mission Control may show successful hash_edit changes as a bounded changed-line transcript preview and a larger activity-detail diff. Both are display-only, not provider tool results or JSONL session data.
subagents displays a bounded aggregate summary rather than raw child output.
Reading files
Use paths for known local files or skills:
{ "intent": "Read project setup instructions", "paths": ["README.md:1-200"] }
- Provider-facing schema requires
paths, accepting 1 to 8 entries. Runtime validation rejects missing, empty, blank, or unknown-file inputs. - Recognized read targets: local files,
skill://<name>skill definitions, andskill://<name>/<relative-reference>skill reference files. Other resource schemes are unsupported. HTTP(S) targets fail with guidance to usewebopen. - String selector suffixes:
:N,:-M,:N-M,:N+,:N+K,:raw, and compounds such as:raw:50-100. Selectors apply independently to each target. - Hashline and whole-file raw reads accept UTF-8 sources up to 1 MiB. Local files without a selector return up to 400 lines; use line selectors to choose another range. Larger files support explicit
:rawline ranges with a 16 MiB scan-work ceiling and 1 MiB output ceiling. These read-only excerpts do not create edit anchors, report unknown total line counts, and validate UTF-8 only in returned lines. Scan or output exhaustion leaves the requested selection incomplete. - Successful local reads use
<File path="<absolute-path>" tag="<TAG>">around numbered text and end with</File>. Raw reads use<File path="<absolute-path>">, without tag or line numbers. Successful skill reads use<Skill name="<skill-name>" dir="<path/to/SKILL.md>">around the full descriptor; skill reference reads use<Skill name="<skill-name>" file="<path/to/file>">. Both end with</Skill>. No additional file heading surrounds reads. Attribute values are XML-escaped; bodies are not XML-escaped. Structuredfiles[].textuses the same complete wrappers as text output. Errors and omitted resources usestatus="error"orstatus="omitted"; theirpathpreserves the requested target, including selectors. Skill failures usenameandsource="<requested skill URI>", rather than claiming a resolved file path. - Complete multi-file
ToolResult.contentis capped at 1 MiB measured as UTF-8 bytes, including all wrappers, separators, notices, and error text. Space is reserved for a closed omission wrapper for every remaining resource. If requested identifiers alone cannot fit, the read fails before loading resources. - Structured results retain an entry for every requested path, including failed and omitted paths.
outcomeiscomplete,partial, orfailedfor delivery of the request;request_completereports whether the requested selections were delivered.completestill reports whole-source coverage: a successful selected range can haverequest_complete: trueandcomplete: false. Partial success retains usable content and does not imply that every file succeeded. - Output-clipped resources retain their complete opening and closing wrappers and include a
READ INCOMPLETEnotice inside. Aggregate clipping can cut a body mid-line or between UTF-8 characters; only a valid UTF-8 prefix is emitted. Every remaining resource receives a closedstatus="omitted"wrapper in both text output and structured results. Incomplete large raw excerpts also include aREAD INCOMPLETEnotice within the output cap. An empty partial search reportsSEARCH INCOMPLETE, not a confirmed absence. readdoes not search, expand globs, or discover files. Usefindfor fuzzy path discovery orgrepfor content search, thenreadknown paths/resources.- Local file reads default to hashline output:
<File path="<absolute-path>" tag="<TAG>">plus 1-indexedLINE:TEXTrows, followed by</File>. :rawbypasses hashline formatting inside the file wrapper (path:raw,path:raw:20-40); explicit line ranges allow bounded excerpts of larger files, without bypassing output caps or path/skill protections.- Skill reference reads are limited to discovered/enabled skills. References must be relative, cannot contain
.., cannot target symlinks or directories, must remain under the canonical skill directory, and must be UTF-8 under the 1 MiB read cap. - For
hash_edit, use only the latest non-raw read's tag and displayed line anchors. Re-read before another anchored edit, and after an error or unexpected result, unless the rejection supplies sufficient tagged context below. A successful edit response does not replace that read. - Aggregate-clipped reads authorize only fully returned anchor lines; omitted files and partially returned lines do not gain edit authorization.
- Unseen-anchor and stale-tag rejections provide tagged context around requested anchors (up to two surrounding lines per side, 40 complete rows and 16 KiB total). Only displayed complete rows gain anchor authorization; oversized rows and omitted ranges require another read. Review the supplied snapshot and submit a new edit decision. Rejection never automatically reapplies the patch; existing exact stale-edit recovery remains unchanged.
Missing read, list_files, grep, and find paths remain failures but may suggest up to three existing permitted paths. Suggestions compare the next missing component case-insensitively, then by edit distance up to two, scanning at most 1,000 entries in the deepest existing ancestor (up to 32 ancestor steps and a cooperative 100 ms deadline). Symlink targets outside the allowed root are never suggested. Suggestions do not execute a different path.
Finding paths
Discover paths without reading contents:
{ "intent": "Locate library source paths", "query": "lib", "kind": "mixed", "path": "src", "limit": 20, "offset": 0 }
Rules:
queryis required, trimmed, and limited to 512 characters.kinddefaults tofiles;directoriesresults end with/in output, andmixedcan return both files and directories.limitdefaults to 50 and is capped at 200;offsetsupports pagination.pathis optional and must resolve to an existing directory. Relative paths stay within cwd; absolute paths outside cwd requirecapabilities.tools.find.absolute_paths(defaulttrue). Output is slash-normalized, cwd-relative for paths within cwd and absolute for outside paths.findis read-only, uses the shared workspace walker, and does not replacegrepcontent search.- Each scan cooperatively checks cancellation, visits at most 100,000 entries (including the root), and has a 10-second deadline. Filesystem calls can exceed a cooperative deadline. Narrow
pathwhen a scan is partial. scan_completeandtotal_matched_exactidentify exhaustive scans. Partial scans rank only observed matches, settruncated: trueandindex_ready: false, and reporthas_more: null;total_matchedis then a lower bound.entries_scanned,entry_limit_reached,timed_out,walk_errors, andwalk_entries_omittedexplain incomplete traversal. Paging does not resume a partial scan.- Workspace walking honors standard hidden/
.ignore/Git filters plus Perforce defaults and configuration. It prunes ignored directories before scanning and feedsfind,grep, and directory-modeast_grep;list_filesintentionally remains a raw shallow listing. - Perforce resolution follows
P4CONFIG→P4ENVIRO→ process environment → Windows user/system registry.P4IGNOREaccepts ordered semicolon-separated relative or absolute rule files; unset values use.p4ignorefollowed byp4ignore.txt. Non-macOS POSIX systems defaultP4ENVIROto$HOME/.p4enviro; Windows and macOS require it explicitly..p4rootis always excluded;.p4configand configuredP4CONFIGfilename are excluded when configuration discovery is active. P4 rule/config reads are bounded. macOSp4 setpreferences are not read directly; expose those values throughP4CONFIG,P4ENVIRO, or process environment. fs_cachestores directory listings only, keyed by path plus directory mtime with TTL expiry. It does not cache file bytes or provider-visible tool output.
Searching contents
{ "intent": "Locate the run function", "patterns": ["fn run"], "path": "src", "limit": 20, "context": 2 }
grepis fully in-process Rust search; it does not require search credentials.patternsis required, accepts 1 to 8 exact strings, and uses OR semantics.- Each pattern is matched as a case-sensitive regex after common brace/parenthesis repairs; failed compilation falls back to literal matching. The original pattern bytes are preserved for matching diagnostics and display.
- Ranked mode is the default: it emits
grouped_v1path headers,[def-like],[match], and[import]primary labels, and deduplicated context rows. Raw mode emits flat deterministicpath:line:contentrows and uses one extra match as a pagination lookahead. limitandoffsetapply to primary matching lines;contextadds at most 20 lines on either side and does not consume the primary limit. Largercontextvalues are clamped to 20 with a visiblecontext clamped from N to 20notice; the schema still advertises a maximum of 20. Ranked mode keeps at most five primary hits per file and reports omitted hits withtruncatedandper_filemetadata.- Reads are capped at 4 MiB per file and 128 MiB per invocation; traversal is capped at 10,000 files and 10 seconds. The provider-visible output is capped at 65,536 UTF-8 bytes and records are added atomically.
- Hidden, ignored,
.git,target, binary, special, and unreadable files are skipped as applicable. Metadata reports scan byte/file limits, per-file limits, recovery skips, binary/invalid-UTF-8 counts, matcher adjustments/fallbacks, exact-match status, and truncation reasons.
Editing files
Anchor patches to the latest read tags:
{ "intent": "Update README instructions", "input": "[README.md#A1B2]\nSWAP 10.=10:\n+new line" }
Rules:
- Each file section starts
[PATH#TAG]. Form it from the latest current-session non-raw read's decodedpathandtagattributes; only displayed line anchors authorize edits. Patch syntax remains[PATH#TAG], not<File>. - Supported ops:
SWAP,SWAP.BLK,DEL,DEL.BLK,INS.PRE,INS.POST,INS.BLK.POST,INS.HEAD,INS.TAIL,MV, andREM. - Body-taking ops end with
:and accept only+TEXTrows. Literal leading plus/minus content still needs the body prefix:++value,+- item. - A bodyless
DEL N:orDEL N.=M:is accepted with a trailing-colon warning. A following body row remains an error; it is never interpreted as replacement text. - Stale tag or surprising result: re-read file, then reissue patch with fresh tag/line anchors.
hash_editdoes not create new files; usewritefor creation.
Inspecting images
Configure an explicit vision provider/model before use:
{
"capabilities": {
"tools": {
"view_image": {
"vision_model": {
"provider": "local-vision",
"model": "vision-model-id"
},
"absolute_paths": true,
"max_image_bytes": 5242880
}
}
}
}
Rules:
pathmust be an absolute path to one existing regular local image. The path is canonicalized; supported formats are PNG, JPEG, GIF, and WebP, with case-insensitive extensions that must match the file signature. Outside-cwd access followscapabilities.tools.view_image.absolute_paths.promptis required and should name the exact visual details to inspect.- Image bytes go only to the explicitly configured vision model; there is no active-chat-model fallback.
capabilities.tools.view_image.max_image_bytesdefaults to 5 MiB and is configurable from 1 byte through 20 MiB. Returned text is capped at 64 KiB.
Safety boundaries
- Tool calls and effects are represented in the session transcript.
- Relative paths resolve from the active working directory and cannot escape it. Built-ins accepting caller-controlled OS paths have independent, default-allow absolute-path settings; see the authoritative Configuration matrix for keys, aliases, exceptions, and opt-out behavior.
bash/shellallows shell expansion by default. Setcapabilities.tools.bash.shell_expansion: falseto reject$,~, backticks,{, and}in bash commands during preflight.hash_editandwritereject final symlink mutation targets and enforce size/count limits.hash_editvalidates all sections before mutation when semantic/path checks fail, rejects unanchored or unseen-line hunks, and guardsREMto regular files only. Write replacement is atomic, but committed-but-undurable failures are possible.- Cwd/root checks reduce accidental escapes and common symlink escapes, but they are not a full TOCTOU-safe sandbox against a malicious concurrent local process racing the filesystem.
bash/shellruns through the host shell and is not an OS sandbox.webuses Codex OAuth for active Codex search and environment-onlyEXA_API_KEYfor other search, with no backend/MCP fallback. URL open fetches directly without an API key. URL checks, timeouts, bounded snapshots, and output limits are described above.grepoutput is capped at 64 KiB;weboutput is capped at 48 KiB with explicit truncation metadata.- Mission Control has separate transcript-card and Selected Activity Detail projections;
subagentsshows only safe summaries. Local/session raw results remain unchanged by these display projections. Do not ask tools to read or print secrets unless you accept them in terminal output and session artifacts. - File mutations are coordinated by per-file mutation locks.