pi-supernova

One supernova invocation. Inline code, a workspace program, or an explicit program batch. Four commands inside CodeMode.

const source = await read("validateRefreshToken");
return source;

The model submits JavaScript, not four separately advertised native tools. Ordinary JavaScript control flow remains available; the guest command bindings are only read, edit, write, and bash. Supernova supplies retrieval, transactional file operations, batching, bounded results and the grouped nova UI.

What is new in 0.9.0

• Text clipping does not stop batches: sequential and parallel programs keep running when the combined display text exceeds its allowance. Results disclose truncation; execution, deadline, host-call, log and image limits remain enforced.

• Decoded image validation: PNG/JPEG/GIF/WebP reads and returned attachments require matching formats, canonical base64 and decodable pixels, including all GIF/WebP frames. PNG preflight checks chunk boundaries and CRCs too. Failures occur before shell boundaries or final commit/delivery, with no image attached.

• Isolated image work: Sharp 0.35.4 decodes at most 32 million pixels across all frames, one image at a time, in a Node/Bun subprocess with a 5-second kill deadline. Text-only work never loads the decoder. A bounded 16-entry digest cache avoids decoding unchanged bytes again; neither image data nor file paths are cached. Encoded limits remain 16 attachments / 20 MiB. Sharp's platform-specific optional dependencies must be installed; decoder failure never falls back to unchecked data. Local decoding does not guarantee acceptance under every provider's image policy.

• Data is not a display preview: ordinary text reads return complete data up to 64 MiB; JSON selections remain actual values and directories remain arrays. The former 160-line / 8192-character and 31,744-character restrictions no longer constrain computation inside a program. Only returned/logged text is displayed.

• Large batches stay bounded: small reads share one reply; larger items are delivered with acknowledgements within eight I/O slots, not retained as one giant host-side batch. Explicit arrays and coalesced reads use the same path.

• Output work stays in the worker: bounded formatting avoids expanding large values before clipping. Model-visible source previews retain exact ranges and continuation; a clipped preview cannot be passed back as a complete edit view.

• Cancellation no longer crashes the host: bounded reads and CAS signing use abort-checked file handles instead of aborting streams. A failed read can cancel sibling reads without an uncaught AbortError terminating OMP.

• Memory is charged to the guest: worker-local heap and external buffers replace process-wide RSS accounting. Bun enforcement remains best-effort.

• Focused internals: read adapters, file I/O, transactions, worker lifecycle, tool ownership, source ranking and rendering have separate modules. All read modes, batching, checkpoints and rollback behavior remain supported.

Verified release candidate: 315/315 package tests pass on both Node and Bun; actual Pi/OMP checks and clean tarball installation checks also pass. See verification results and coverage limits below.

Concurrent file operations

Batching and explicit parallel programs remain supported. Within one program, Promise.all now overlaps native edits and writes to different files, up to eight operations at once:

await Promise.all([
  edit("src/a.js", "oldA", "newA"),
  edit("src/b.js", "oldB", "newB"),
]);
return await bash("npm test");

Same-file operations retain submission order. Reads before/after mutations, bash, edit checkpoints, and overridden mutating tools remain ordering barriers. Shell calls inside one program stay sequential; use programs with parallel:true for explicitly independent shell workflows or separate JS workers. Edits still stage until program success (or a shell boundary); transactional disk commits retain their conflict checks. await edit(...) one after another is still sequential, and multiple replacements in one file remain one edit operation.

What is new in 0.8.2

This patch release fixes shell failure handling, cancellation and parallel-batch limits, and makes one-call batching guidance explicit.

• Shell quoting stays intact: quoted executable paths are no longer unwrapped. Shell syntax errors suggest literal bash({command,args}) with data for embedded scripts, or a quoted heredoc. Commands are not rewritten or retried.
• Validation before commit: invalid timeouts and null-byte arguments are rejected before the shell boundary flushes staged files.
• Useful failure output: long command labels are bounded so the original stderr is not crowded out by a repeated script.
• Timeouts retain diagnostics: the outer program deadline stops the worker and gives pending host calls a bounded drain to retain shell output. Explicit cancellation is reported separately from timeout. The outer timeoutMs covers every wait and command, including sleep.
• Parallel budgets fail honestly: exceeding the shared log or image allowance marks the batch failed and stops queued entries. Already-running entries settle; their results and completed commits remain. Aggregate logs stay capped rather than multiplying the allowance per guest. In 0.8.2 this also applied to output text; 0.9.0 makes display-text clipping nonfatal.
• Batch known work in one call: combine independent reads/checks with Promise.all, then sequence edits and verification in the same program. Use another invocation when returned evidence is needed for the next decision.

Verified on macOS / Node 26.7: 267 package tests (384 repository tests), 2,328 stress invocations, actual Pi/OMP host checks, lint and both token-budget checks. This is not a claim of exhaustive platform or formal mutation testing.

0.8.0 features and measurements

• Shared program source: top-level code or file supplies a batch default; entries may override it. A shared program is sent once instead of in every entry, and defaults count once against the 48,000-character admission cap.
• Explicit object defaults: mergeData:true shallowly overlays per-entry data onto common data (entry keys win; nested objects are replaced). Whole-input replacement remains the default.
• Checkpoint failures throw: a failed edit(async () => {...}) rolls back and rethrows its original cause; catch explicitly when rejecting a candidate is intentional. Ignored failures no longer report success.
• Accurate failure cards: the nova card reads the host's error flag, shows the original cause and committed/rolledBack totals, marks writes whose persistence cannot be attributed as attempted, and labels pure JavaScript runs instead of "complete".
• Historical read limits (superseded in 0.9.0): errors stated both limits (160 lines / 8192 characters) with copyable recovery (offset, about, complete:true, and Promise.allSettled for optional siblings). Markdown edits skip code-reference searches; exact-symbol evidence excludes generic matches.
• Fail-closed images: unsupported formats (for example BMP) fail before model delivery with PNG-conversion guidance, and sets over 16 images / 20 MiB report aggregate sizes instead of silently omitting attachments. Pending changes roll back.
• Shell follows the program clock: bash() inherits the program's timeoutMs; explicit per-command limits still win.

Tokens: 0.7.1 to 0.8.0 (js-tiktoken, o200k_base / cl100k_base)

Metric	0.7.1	0.8.0	Change
Standing definition per request	596 / 588	631 / 626	+35 / +38
Frozen 6-call mixed workload, total traffic	9,596 / 9,458	9,841 / 9,724	+2.6% / +2.8%
16-program job with shared source + data (32 files)	17,771 / 17,595	3,587 / 3,533	-79.8% / -79.9%
8 programs sharing a 48-path input	16,309 / 14,649	5,499 / 5,197	-66.3% / -64.5%

Rows 3-4 deliver identical complete outputs and files; only argument placement changes. Row 2 repeats no inputs, so it pays the +35-token guidance and nothing else. Method, gates and limits: TOKEN_COSTS.md.

Speed: engine micro-benchmarks (Apple M5 Max, Node 26.7)

Benchmark	Before	After	Change
Package 200-row report (median, 10k iterations)	0.1665 ms	0.1314 ms	-21%
Package nested source object (median)	0.0366 ms	0.0240 ms	-35%
Idle worker exit	267 ms	17 ms	-94%
8-file read wave p50 / p95 (300 samples)	1.90 / 2.74 ms	1.77 / 2.53 ms	-7% / -8%
Cold unbatched p95 vs coalesced warm p95 (8 reads)	13.85 ms	2.39 ms	-83%

Identical output hashes before and after. Local engine benchmarks, not end-to-end agent latency or provider time.

See the changelog and token measurements.

Install and update

Install the published package in your host:

pi install npm:pi-supernova
omp install npm:pi-supernova

Local checkout installs are for development, not distribution:

pi install /path/to/pi-stack/packages/pi-supernova

0.9.0 adds Sharp for image validation. Keep its platform-specific optional dependencies enabled. For a local checkout, refresh dependencies before starting the host:

npm install --prefix /path/to/pi-stack/packages/pi-supernova --include=optional

Install dependencies on each target machine rather than copying node_modules between operating systems or CPU architectures. Published-package installation resolves these dependencies through the host's package manager.

Git pushes do not update npm installations. Publish the new npm version first, then reinstall it in your host. To pin 0.9.0 once it is published:

pi install npm:pi-supernova@0.9.0
omp install npm:pi-supernova@0.9.0

In Pi, pi list shows the configured package sources. A local path uses that checkout directly; an npm source uses the installed npm copy. Do not assume that pushing a checkout or running /reload updates the copy executing in your host.

Both host manifests use index.js. The old src/bridge/pi-extension.ts path remains a compatibility entrypoint but no longer imports Pi tool factories. After updating JavaScript sources, fully exit Pi and resume in a new process. Pi 0.85.1 can retain native ESM modules across /reload, even after its extension factory cache is cleared; /reload alone is not sufficient in that case. OMP uses the same shared engine; restart it after updating the link as well.

For deferred-context-engine, pin supernova in alwaysActive and neverDefer. Remove the four native names from those pins and restore their previous blocks if you want the wrapper to be the only file/shell surface. Preserve unrelated settings. The runtime does not silently rewrite your tool policy.

Four guest commands

Function	Examples and behavior
read	read(path, offset?, limit?), read({path,offset,limit}); one-based line windows
read	return await read("plot.png"); displays images directly, without a browser
read	read({path,json:".verdict"}); parse full JSON before bounded field selection
read	read(directory), read("symbol or question"), read(path,{about:question}); a path is raw text, a symbol is a view
read	read({query,resolve:true}); same view as read("symbol"): status, path, line, lines, text, complete
read	read({query,evidence:true}); ranked evidence with provenance; optional path scopes discovery
read	read({path,outline:true}); structural declarations
read	read([path1,path2]); up to 64 paths, ordered values; rejects if any path fails
edit	edit(path,oldText,newText), edit({path,edits:[{oldText,newText}]}); unique in the file
edit	edit(view,text) CAS-replaces that span; edit(view,old,new) is unique inside it
edit	edit({path,patch}); unified patch application
edit	edit(async () => {...}); filesystem-only checkpoint, described below
write	write(path,text), write({path,content}); atomic replacement
bash	bash(command,{cwd,timeoutMs}), bash({command,timeoutMs}); bounded output, nonzero exits throw
bash	bash({command,args:[...]}); literal executable argv, without shell expansion of argument strings

bash also accepts timeout in seconds for familiar object arguments. timeoutMs is milliseconds and takes precedence. The owned POSIX adapter launches executable argv directly; use the string form for shell builtins, functions or startup hooks. Windows and delegated/older executors retain quoted-shell compatibility. Argument payloads are not repeated in owned direct-execution errors; stdout/stderr, exit status and source context remain. Session environment variables are taken from the current execution context, not inherited from a different parent session.

Shell strings are executed unchanged, including quoted executable paths. For inline Python/Node scripts, prefer literal argv with data instead of nested shell quotes:

{
  "code": "return await bash({command:\"python3\",args:[\"-c\",data.script]});",
  "data": {"script": "q = {'name': 'example'}\nprint(f\"{q['name']}\")\n"}
}

Shell syntax errors keep the original diagnostic and suggest argv or a quoted heredoc; commands are never automatically rewritten or retried. Invalid timeouts and null-byte arguments fail before flushing staged changes. Failure labels are bounded so a large script cannot crowd out its stderr.

Source questions locate a declaration in one command. An exact declaration match uses one bounded direct ripgrep search, without a prerequisite file listing, persistent index, embeddings or summarization. A transient filename listing is a fallback for unmatched content or unresolved bare filenames. Natural-language questions reuse lexical stemming. Source questions and focused about reads accept at most 16 keywords. Ripgrep must be available on PATH.

read(path) stays raw text. read("symbol") is the same view as read({query, resolve:true}) -- not the file, not a path/range header:

const v = await read("validateRefreshToken");
if (v.status !== "found") return v;
await edit(v, v.text.replace("token.length > 3", "token.length > 5"));

The view contains status, path, the matching line, span lines, unchanged text, complete, and nextOffset when a model-facing preview continues. The internal view is not shortened to fit the display. A declaration snap is that span (complete is false unless the span is the whole file). Uncertain results report ambiguous, not_found or incomplete with no selected path. Use {path: directory, about: question} to narrow the scope. Scoping a query does not relabel a selected span as a complete file. Newly staged files and large staged source participate in discovery before commit. Raw offset/limit windows preserve LF/CRLF endings and the final newline; focused views add line labels. A matching window too large for the output budget is reported as budget-limited, not as an absent match. Do not combine incompatible modes such as outline:true and evidence:true.

Ordinary reads stay self-contained. Outlines and graph evidence remain explicit options, not mandatory stages of source resolution. Ordinary calls also get:

• Bounded fuzzy filename hints when a bare source name has no literal match. Existing frecency/directory ranking orders hints; fuzzy matches never select a path. This reuses filename discovery, examines at most 1024 paths and reports incomplete hints.
• Post-edit source windows and structural warnings for both replacement and patch edits. Body-only edits reuse local declaration spans for lexical reference hints. Up to three names share one bounded, cancellable search; staged callers override disk. These hints do not build an index, are not semantic caller resolution, and report unavailable/truncated searches rather than silently claiming completeness.
• Structural warnings on writes without echoing successful file contents.
• Fresh workspace source around up to four failure locations, including shell timeouts and paths relative to the command cwd. Implicit snippets exclude external symlink targets and files over 1 MiB.

Distant edit regions have separate windows and continuation pointers for omitted lines. Structural warnings and source windows are not substitutes for tests.

Safe read-modify-write

A plain file read returns its complete text up to the 64 MiB UTF-8 I/O limit, independently of model-facing character budgets. Explicit offset/limit reads return exact line windows, including long lines and LF/CRLF endings, or fail at the same byte ceiling. Staged and on-disk data obey the same ceiling. complete:true additionally rejects a window that omits part of the file. For larger files/JSONL, use a bounded parser through bash({command,args}).

Compute on the full value and return only what the model needs. A direct large return is an explicitly truncated display, not a complete artifact to parse or write back. resolve:true source previews retain path/range/continuation metadata; clipped previews are not editable. Prefer edit for replacements. A write after reading that path still requires edit or explicit replace:true. Writes reject Supernova truncation markers, including legacy host-result markers. For intentionally writing literal marker documentation only, opt in with write({path,content,allowReadArtifacts:true}). This is a data-loss guard, not full dataflow tracking or a security sandbox.

Read/modify/write conflict checks retain a signature of the actual disk bytes, including for partial and large-file reads. A fresh explicit text read refreshes that observation; internal receipt reads and body-cache eviction do not. Commits reject changed content and conflicting symlink aliases, including new file paths. These checks do not provide a cross-process lock or make shell/import mutations transactional. Extensionless filenames also support complete:true, for example read({path:"LICENSE",complete:true}).

Directory reads return complete string-entry arrays up to 10,000 unique entries. Larger listings fail with a path and bounded-parser guidance rather than silently returning an incomplete array. Metadata lookups overlap eight at a time.

Explicit read arrays reject missing/failed paths. For typed partial outcomes use Promise.allSettled(paths.map(path => read(path))). Successful arrays remain arrays. For literal file content or scripts, prefer the optional tool-level data parameter:

{
  "code": "await write(data.path,data.content); return await bash({command:\"python3\",args:[data.path]});",
  "data": {
    "path": "probe.py",
    "content": "print(\"literal `backticks` and ${braces}\")\n"
  }
}

data crosses the worker boundary as JSON, never as JavaScript source. For a single program its JSON-encoded length is capped separately at maxCodeChars; batches use the combined admission budget described below. Split larger inputs. The binding exists only when supplied, so older programs declaring their own data remain valid. Syntax errors run no commands and give quoting guidance. For inline source, use String.raw (escaping backtick delimiters) or JSON-quoted strings.

On hosts exposing sessionManager.getArtifactsDir(), bare agent://<id> and artifact://<number> read files from the calling session's artifact directory. They preserve ID casing and support offset/limit and structured continuation. These resources are read-only; ambiguous artifact IDs and escaping symlinks fail. A single ?q=.answer (URL-encoded when needed) selects JSON from either resource using the same bounded projection as read({path,json}). The resource must contain valid JSON; ordinary Markdown is not parsed heuristically. This is not the full OMP URI language: full jq, cross-session search and other schemes are not implemented. Hosts without an artifact directory report that limitation rather than treating the URI as a local path.

For unstructured logs/text, read(path,{about:"STT database"}) returns bounded, line-numbered matching windows, or explicitly reports no matching text. It is not a complete-file read. Write temporary investigation files under the workspace (e.g. .work/probe.py): ordinary write/edit paths cannot escape it, including absolute /tmp paths. Shell execution is a separate trusted boundary, not a sandbox.

Reuse a program without resending its source

Save a trusted JavaScript async body or arrow in the workspace, then invoke it:

{"file":".work/audit.js","data":{"paths":["src/a.js","src/b.js"],"term":"TODO"}}

Supply exactly one of code or file. File programs get the same four commands, optional data, limits, deadlines and transaction semantics. Each invocation rereads the file and starts a fresh guest; there is no implicit last-program state, auto-replay, or retained heap. Paths inside the program still resolve from the calling workspace, not the script directory. This runs a Nova program, not an arbitrary JavaScript module or a Python/shell script.

Program files must be regular UTF-8 files inside the workspace, including symlink targets. Invalid encoding, oversized input and syntax errors fail before commands; no truncated prefix is executed. Review untrusted source before running it. Use ordinary edit to revise saved programs. This is explicit source reuse, not conversation compression: prior calls and read results remain intact. Creation costs an additional call unless combined with other work, so prefer inline code for short one-off operations. See token measurements.

Batch already-known continuations

{
  "programs": [
    {"code": "return await edit(data.path,data.oldText,data.newText);", "data": {"path":"src/config.js","oldText":"limit = 8","newText":"limit = 16"}},
    {"code": "return await bash(\"npm test\");"},
    {"file": ".work/audit.js", "data": {"paths":["src/config.js"],"term":"limit"}}
  ],
  "timeoutMs": 60000
}

Supply 1--32 programs entries. Top-level code OR file supplies an optional source default; an entry's own code OR file replaces it. Without a shared source, every entry still requires its own. A shared file is reread when each entry runs, so earlier commits can create or update it.

Top-level data supplies an optional input default. By default, explicit entry data replaces it entirely, including null, false, 0 and empty strings. With mergeData:true, both the default and every explicit entry input must be objects: own entry fields override default fields in a shallow merge. Nested objects are replaced, not recursively merged. Each guest receives its own copy; mutations cannot leak between guests or back into caller-owned inputs.

The JSON-encoded array plus any shared code/file/data must fit maxCodeChars. Common source/input counts once, before expansion. Every entry is validated before any program runs. Result representations and output limits are unchanged. Entries run sequentially in fresh guests and commit separately. A successful entry can create the file executed by a later entry. No implicit retries, reordering, shared heap or nested batches are introduced.

For independent audits that use the same inputs, send them once:

{
  "data": {"paths": ["src/a.js", "src/b.js"]},
  "programs": [
    {"code": "return await read(data.paths);"},
    {"code": "return await Promise.all(data.paths.map(path => read({path, outline:true})));"}
  ]
}

This avoids repeating literal arguments, without a compression codec or result elision. Mutating data in one guest cannot affect the next. An entry with data:null receives null, not the shared object; there is no implicit object merge.

For the same program with varying inputs, no temporary script or input file is needed. For example, audit both files with a common term:

{
  "code": "const text=await read({path:data.path,complete:true}); return {path:data.path,found:text.includes(data.term),text};",
  "data": {"term":"TODO"},
  "mergeData": true,
  "programs": [
    {"data":{"path":"src/a.js"}},
    {"data":{"path":"src/b.js"}}
  ]
}

The complete source is still returned twice, once for each requested file; shared arguments do not authorize result deduplication or context rewriting. Existing programs that depend on whole-input replacement keep that behavior unless the caller explicitly requests mergeData:true.

The batch stops on the first failed entry, cancellation/deadline, or exhausted log/image budget. Display-text clipping does not stop execution or mark a successful batch failed. Earlier successful commits remain; only the active program's uncommitted writes roll back. Admission errors throw before any program. Execution failures return a typed stop report, rather than throwing away prior results/images: isError and details.ok identify failure, details.programs contains every attempted result, and details.attempted/total identifies unstarted work. Single code/file invocations retain their existing throwing behavior.

Set parallel: true with programs to run independent entries concurrently (up to 8 at once). Each still gets a fresh guest and its own commit; results stay in submission order. A failed entry does not stop siblings. Two entries writing the same file race: the losing commit reports a conflict. Sequential remains the default. parallel and mergeData are invalid on a lone code or file call. Log/image budget overflow marks the batch failed and stops queued entries; already-running entries settle and their completed commits remain. Text overflow only clips the displayed result and sets details.returnTruncated; queued entries still run. Parallel execution does not multiply the aggregate allowances.

The outer deadline, host-call budget, log allowance, text budget and image limits are shared across the batch. Individual read budgets are not reduced. Every attempted program's text is assembled in length-delimited blocks, then the combined display is clipped if necessary. Images retain program/image labels. Return summaries or use focused windows when you need every result to fit on display.

Batch only continuations already chosen by the agent, such as edit then known verification, or create then run known audits. Keep a separate call whenever new source/results are needed to decide the next action. This does not lower reasoning settings, hide observations, or infer a plan on the agent's behalf.

Large inputs and report outputs

The default program limit is 48,000 UTF-16 code units (configurable via maxCodeChars and exposed in the tool schema). The same cap applies separately to serialized JSON data, including quote/newline escaping and object keys. Oversized input fails before commands run and reports its actual serialized size. Split larger documents into separate invocations: first write(path, firstChunk), then write({path,content:nextChunk,append:true}). Append uses the complete internal file buffer, never a bounded model-facing read; it retains conflict checks and per-program rollback. Missing files are created. Multiple invocations are not one atomic transaction: for an all-or-nothing publication, assemble a new staging file and publish it only when complete. External write overrides reject append.

Supernova is a bounded foreground executor, not a durable background-job manager. For long archive scans, use resumable chunks or a host background-job tool and write progress records under .work. Shell commands inherit the current program timeoutMs unless they specify their own; increasing the outer deadline no longer leaves a hidden 60-second shell cap. Set the inner bash timeout shorter than the outer program timeout (for example 10 seconds inside a 20-second program). The outer deadline covers all waits and commands, including sleep; a shell's own timeout command does not extend it. On a deadline or cancellation, the worker stops and pending host calls get a bounded 250ms drain to retain owned-shell diagnostics and finalize process termination. Non-cooperating host executors may still outlive that drain. Cancellation is reported separately from timeout; neither triggers a retry. Progress files survive shell execution but staged VFS writes may roll back.

Large returned objects are bounded previews, not retained artifacts. Select fields and array windows before returning, rather than parsing a truncated preview.

JSON reports and targeted text audits

For JSON, select fields inside the read adapter, before output budgeting:

const [verdict, values] = await read({path:"report.json",json:[".verdict",".values[5000:5003]"]});
return {verdict, values};

Selectors support "." (root), .field, .nested[0], .items[0:10], and .["quoted.key"]. Use json:true for the complete parsed value. Selectors are not full jq: pipes, filters, wildcards and negative indices fail explicitly. Missing keys and indices fail; false, zero and null remain values. Slices use an exclusive end and clamp to array length. Only own JSON properties are traversed; nothing is evaluated.

Inputs are capped at 16 MiB, including staged files. JSON reads require regular files and reject named pipes without waiting for a writer. The entire input must be valid JSON before any selection. Each selector is budgeted before allocating the next slice; sparse selector/path/edit arrays are rejected. Selections have a separate 64 MiB estimated storage limit per read, including aggregate multi-selector expansion. Within it, arrays/objects/scalars retain their actual values, even when larger than the display. No routing object is substituted: .map and .filter work on the selected array. Multi-selector values remain independently mutable. The input/storage safety limits throw with guidance. Plain .json reads are raw text, including malformed JSON; parsing is requested only by json. Explicit line windows are not necessarily JSON documents. Do not combine json with complete, line windows, or source views. External read overrides reject JSON projection rather than silently ignoring the option. Uncaught read errors abort the program, including return {a:await read(...), b:await read(...)}; earlier successful values are not an implicit partial return. For optional sources, explicitly return await Promise.allSettled(paths.map(path => read(path))). This keeps successful text and per-path errors without weakening rollback for uncaught failures.

Other read options, even false-valued flags, do not bypass a captured external read executor; its policy, transforms and failures remain authoritative.

For focused Markdown/log audits, use read(path,{about:"document path"}) or explicit offset/limit. Use complete reads for computation within the 64 MiB ceiling, and return a summary. JSON over 16 MiB needs a streaming parser via bash. Arbitrary returned objects still have bounded previews, not implicit continuation handles.

Execution and automatic batching

Put already-known independent reads and checks in one Supernova program using Promise.all (or Promise.allSettled when failures should remain independent). Sequence edits and their known verification in that same program. Start another invocation only when the returned evidence is needed to decide what to do next; use focused read windows to keep the combined result within its output budget.

Compatible independently started reads coalesce at the worker/host boundary. No additional batching command is required. Individual promises preserve their values, errors and per-read safety limits. Reads and known native mutations to disjoint files overlap up to eight operations; same-file mutations stay ordered. Read/mutation transitions, shells, checkpoints and overrides remain barriers.

This does not reorder sequential awaits or predict future model decisions. Explicit path arrays and automatic coalescing share the same typed delivery. Small batches retain the single-reply fast path; beyond 64 KiB estimated storage, items stream to the guest with acknowledgements. Delivery holds its I/O slot until acknowledged, and the final read waits for its host barrier to settle.

Every program runs in a fresh worker. One pristine worker is prepared for the next invocation, then disposed on session shutdown. Executed workers are never reused, so guest globals cannot leak into a later program. Worker preparation still costs CPU and memory; it is moved off the next invocation's critical path, not eliminated.

maxHeapMb sets Node's native worker heap cap. A worker-local check also limits heap plus external buffers to 1.5 times that value, sampled every 50 ms and before host calls and final results. Process-wide RSS is never charged to an individual guest. These samples are best-effort: Bun does not enforce Node's native heap cap, and non-yielding code can prevent sampling until it reaches a command or returns. The outer deadline still terminates non-yielding workers. This is not a hard process-memory or security boundary.

File changes are staged until program success. A throw before an external-mutation barrier rolls them back. Shell execution flushes preceding changes; external shell side effects cannot be rolled back. Stale commits fail explicitly rather than silently overwriting successful concurrent changes. This is not a cross-process filesystem lock. Outcomes explicitly report committed/rolledBack file versions (counted per flush/checkpoint, not unique paths) and external-call attempts. A successful inner checkpoint merges into the program, not necessarily onto disk. Pending commits or failed recovery are reported as uncertain: inspect disk and recovery backups before retrying. Import-based mutations and shell side effects are outside the VFS counters; this is not a filesystem audit.

edit(async () => {...}) creates a nested filesystem checkpoint. It returns {ok:true,committed:true,value} on success. On failure it rolls back and rethrows the cause, so an ignored failed checkpoint cannot report program success. Use try { await edit(async () => {...}); } catch (error) {...} for deliberate recovery. Shell commands, overlapping/nested checkpoints, and concurrent commands outside the active callback are rejected. Await the checkpoint before proceeding.

Context, caching and failure fidelity

• Plain reads are not replaced with earlier-context references. A local cache hit is not proof the model still retains an earlier result after compaction.
• Text and JSON computation is independent of display limits. Source previews provide exact continuation; raw displayed text can be clipped and must not be parsed as a complete file. I/O/storage limits still fail explicitly.
• Model attachments support PNG, JPEG, GIF and WebP. BMP and other unsupported MIME types fail before attachment or commit, rather than causing a provider HTTP 400 on the next request. Convert those sources to PNG first; Supernova does not silently convert, resize or modify the original image.
• Returned images remain image content blocks, including in arrays/objects. Images not returned by the program stay out of model output. Returned images are limited to 16 attachments / 20 MiB. Overflow fails the program with the aggregate count and byte size, returns no images, and rolls back pending writes. Resize or return fewer images when necessary.
• Dense multiline string arrays can render as verbatim source blocks instead of escaped string literals. Each block gives its array index and exact UTF-16 length; strings and result types are unchanged. This is output framing, not source compression. It is chosen only when shorter in characters than escaped output; it does not guarantee lower billed tokens for every tokenizer or input. Nested multiline strings use the same approach: the complete value structure references raw[n], followed by length-delimited verbatim string blocks. Literal "raw[0]" values stay quoted; duplicate source is never deduplicated. Small values keep their existing format. Machine-facing result values are unchanged.
• Intermediate values stay inside CodeMode unless returned or logged. Final text, errors and logs are bounded with explicit truncation. Details support rendering; they are not a second model-facing transcript.
• Source indexing/caching remains internal. Reads after mutations invalidate stale state. There is no claim of provider-cache or total-task token savings.

Default limits are in src/config/config.default.json. Configuration loads from ~/.pi/agent/supernova.json, the configured host directory, or PI_SUPERNOVA_CONFIG.

Citation elision is disabled by default (seenWindow: 0): each result remains self-contained within the normal output budgets. A positive seenWindow explicitly opts into an experimental ledger with known retention gaps: hidden message details, later context transforms and missing citation targets can invalidate its assumptions. It is not a proven lossless optimization and is not recommended for production. Text limits are character budgets, not tokenizer counts. /supernova reports programs and output characters without labelling characters as tokens.

Optional workspace change notifications

Supernova works without another search package. On hosts exposing pi.events, it provides an advisory workspace:changed event for independent cache/index consumers:

{ version: 1, cwd: "/absolute/workspace", paths: ["/absolute/workspace/file.js"] }

cwd identifies the calling workspace. paths contains absolute file paths after a successful disk flush; paths may contain filesystem aliases. The frozen event and path array contain no source text. Checkpoint merges, restored rollbacks, and read-only programs emit nothing. A shell boundary can flush paths before a later program failure, so notifications are not conditional on overall tool success.

paths: null means the changed paths are unknown: shell or delegated mutation attempts emit this even on failure, as does an incomplete commit recovery. Consumers should invalidate conservatively, not interpret it as an empty change list or a guarantee that shell effects stayed inside cwd.

Subscribe with pi.events.on("workspace:changed", handler). Mark cached state dirty synchronously, then refresh when queried; asynchronous listeners are not awaited. Observer failures cannot roll back writes. This is an optional extension convention, not a built-in host standard, durable event log, or cross-process filesystem watcher. There is no dependency on or automatic routing to any consumer package.

Security and host boundary

CodeMode executes trusted JavaScript in a terminable worker, not a security sandbox. The four adapters constrain writes/edits to the workspace and allow explicit external reads. Absolute temporary-directory paths outside the workspace are not write destinations: use a workspace path such as .work/verification.log, or a separately authorized external command. Errors identify the rejected path and workspace, including symlink escapes. JavaScript imports and shell commands still have process privileges. Do not run untrusted programs as though these adapters isolate them.

Pi preflights the outer supernova call. Internal primitives do not emit ordinary native tool_call events, so third-party guards that only recognize top-level edit or bash need CodeMode-aware handling. Configured exclusions and supported host-session execution safeguards remain enforced. Guards inspecting code/file inputs must also understand the programs array; its entries do not emit separate top-level tool_call events. Actual-host smoke checks are not a claim that every third-party permission extension has been validated.

Development and evidence

Implementation is grouped under src/; all replacement tests are under tests/. Read-budget assertions now test the deliberate 0.9.0 contract: complete internal values and bounded model-facing output. Frozen token snapshots are unchanged. Strict regressions cover batching fidelity, image/context retention, checkpoints, mutation ordering, external symlinks, deadlines, worker isolation, bounded read delivery, and execution-context environment.

Module boundaries

Responsibility	Modules
Registration and session lifecycle	index.js
Program admission and batching	runtime/program.js, runtime/batch-input.js, runtime/program-batch.js
Worker ownership and guest RPC	runtime/runtime.js, runtime/worker-pool.js, runtime/guest-worker.js, runtime/guest-api.js
Host permissions, ordering and trace	bridge/host-bridge.js, bridge/tool-registry.js, bridge/trace.js
Read routing and typed readers	adapters/read.js, adapters/read-{text,json,image,focus}.js
Image validation and isolated decoding	shared/png.js, shared/image.js, shared/image-worker.js
Bounded I/O and atomic transactions	fs/file-io.js, fs/read-window.js, fs/vfs.js, fs/commit.js
Source search and ranking	context/query.js, context/snap-search.js, context/source-entry.js, context/evidence-{graph,rank}.js
Model output and host rendering	output/final.js, output/outcome.js, ui/host-render.js, ui/trace.js, ui/render.js

Paths above are relative to src/, except index.js. Layer tests require an acyclic import graph and prohibit context modules from importing runtime code.

Verification

The final 0.9.0 release candidate was verified on macOS:

Check	Result
Node 26.7 package suite	315 passed; zero failures or skips
Bun 1.4 package suite	315 passed; zero failures or skips
Pi 0.86.1 / OMP 18.2.6	Actual host execution checks passed; Pi TUI checks passed
Clean production-only tarball installation	Node/Bun and actual Pi/OMP smoke checks passed
Lint, frozen token gates, repository release check and publish preflight	Passed

Failure-first regressions cover corrupt image data, malformed base64, MIME mismatches, decoded-pixel limits and nonfatal batch-text clipping. Additional checks cover animated images, content-based validation reuse, cancellation, decoder-watchdog recovery, and real failures after clipped output. The compiled OMP native-module issue is covered by the actual-host checks, not only unit tests.

Coverage limits: a live Codex round-trip and Linux/Spark smoke checks were not rerun after the final image-validation changes. These results are not a claim of exhaustive platform coverage or guaranteed acceptance by every model provider.

Test user-visible contracts through registered programs: exact source, on-disk results, failure/rollback, isolation, bounded output, and usable host rendering. Inject filesystem faults only to exercise real failure paths; do not prescribe private helper layouts, staging filenames, or syscall counts. New regressions must fail before the fix; for existing behavior, verify that a named deliberate defect makes the intended test fail before accepting it. Preserve existing assertions unless an explicitly requested contract changes; then test both the new behavior and the retained safety boundary. Cost gates cover avoidable search processes, per-read safety limits and progress flooding; latency claims belong in the explicit measurement lane, not arbitrary wall-clock assertions.

npm test --prefix packages/pi-supernova
npm run lint:supernova
npm run measure --prefix packages/pi-supernova
npm run test:tokens --prefix packages/pi-supernova

PI_SUPERNOVA_PI_ROOT=/path/to/pi-coding-agent \
PI_SUPERNOVA_OMP=/path/to/omp \
npm run test:hosts --prefix packages/pi-supernova

The explicit host runner requires macOS network sandboxing and fails, rather than skips, when prerequisites are absent. Verified locally against Pi 0.86.1 and OMP 18.2.6: CodeMode execution, four primitives, automatic read coalescing, checkpoints, images and failed execution. Pi's actual loader/runner and TUI are exercised; OMP runs in a disposable process through its actual session registry, with networking denied. The Pi runner supplies a minimal tool registry, not a full provider session.

The local measurement compares identical eight-file programs with coalescing off, coalescing on, and a pristine-ready worker. It reports latency percentiles, observed maxima, raw samples and bridge calls. Set SUPERNOVA_MEASURE_SAMPLES (20 to 10000; default 200) for longer runs. Maxima describe the measured sample, not hard real-time guarantees. It excludes model latency, provider tokens and prewarm time; it is not a universal comparison against every CodeMode implementation.

See the changelog for changes and compatibility notes.

Research and prior art

These references describe internal algorithms, not additional guest commands.

The existing implementation references are retained below. Mechanisms requiring an additional model call are not silently invoked by the tools.

Work	What we use it for	Where
Zero-Mem: Zero-Token Memory Operations for LLM Agents, Xiao, Zhu, Zhang, Chen, Hong, Zhuang, Zhang, Chen, Ouyang, Ren, Huang (arXiv:2607.29377)	Evidence selection: entity–context graph with co-occurrence weights (eq. 3–4), turn/window/episode hierarchy as line/span/file (eq. 5, 11), query profile and relational/local routing (eq. 6–7), lexical entity alignment and one propagation step (eq. 8–9), personalized PageRank over spans (eq. 10), per-view normalisation and ρ-weighted fusion (eq. 12–13), closure with bridges and neighbours (eq. 14), deterministic calibration (eq. 15). Top-K = 5 follows the paper's Top-5 ≈ Top-10 finding.	src/context/evidence.js
Agent Zero Memory: Provenance-Aware Long-Term Memory for LLM Agents, Zhu, Wu (arXiv:2608.29606)	Every returned unit carries provenance (path, line range, verbatim text); the L0→L1→L2 read discipline (read(query) → read(path, {about}) → read(path, offset, limit)); the citation-lock idea that a model should only cite what it actually opened.	src/context/evidence.js, src/context/outline.js, tool guidance
Harness-of-Harness: Multi-Day Autonomous Software Development with Continual Improvement, Yan, Su, et al. (arXiv:2609.01481)	Progressive disclosure (index first, detail on demand) and carrying evidence forward instead of reconstructing it from code.	outline / result shaping
Act More, Decide Less: Skill-Guided Adaptive Action Chunking for Long-Horizon LLM Agents, Yang, Jin, Zhao, et al. (arXiv:2609.02042)	Framing: one supernova program is an action chunk (one model decision, many primitive actions, stop at the first failing one).	runtime design
fff, Dmitriy Kovalenko, MIT, https://github.com/dmtrKovalenko/fff	File search. We reimplemented fff's ranking in plain JavaScript after reading its Rust sources (crates/fff-core/src/score.rs, dbs/frecency.rs, path_utils.rs); the formulas and constants are fff's, the code is ours, and nothing runs out of process. Bounded fuzzy filename hints now run automatically for unmatched bare source names, using in-memory frecency and directory distance without extra filesystem probes. The full internal search implementation also retains typo-tolerant fuzzy path matching with boundary/consecutive/case bonuses; smart-case; exact-filename +40% and filename +20% bonuses; frecency boost base·f/100 with fff's AI-mode decay (3-day half-life, 7-day window) and modification-recency steps (30s/5m/15m/1h/4h); git-modified +15%; directory-distance penalty from the current file (−1 per hop, floor −20); definition-first result hinting; fuzzy fallback on zero literal matches; weak-match cutoff; watcher-driven index refresh. Git/mtime boosts and full indexed grep are not mandatory stages of ordinary reads. Not ported: fff's SIMD/frizbee matcher (ours is an fzf-style greedy match with backward tightening), LMDB persistence (frecency is per session), and the MCP/Neovim surfaces.	src/context/fuzzy.js, src/context/repo-index.js, src/bridge/host-bridge.js

License

MIT. fff is © Dmitriy Kovalenko and contributors, also MIT.