chore: generate
This commit is contained in:
opencode-agent[bot]
parent
cb93114424
commit
83c638eaac
20 changed files with 2137 additions and 1207 deletions
|
|
@ -39,6 +39,7 @@ package and was **deleted** in Wave 3 (done, see below).
|
|||
From issue #34787 and design discussion. Do not relitigate these casually.
|
||||
|
||||
### Core direction
|
||||
|
||||
- Generic CodeMode lives in its own package: `@opencode-ai/codemode` (repo scope convention;
|
||||
the issue's `@opencode/codemode` name was normalized to the `@opencode-ai/*` convention).
|
||||
- **Keep the hand-rolled interpreter.** No QuickJS/V8/sandbox-engine dependency. We own and
|
||||
|
|
@ -52,6 +53,7 @@ From issue #34787 and design discussion. Do not relitigate these casually.
|
|||
products/blog posts) in code, comments, commit messages, or docs in this repo.
|
||||
|
||||
### MCP / tools
|
||||
|
||||
- The MCP adapter lives in OpenCode, not here. It converts MCP definitions into ordinary
|
||||
`Tool.make(...)` definitions and hands CodeMode a plain tool tree.
|
||||
- Permissions stay in the OpenCode adapter (each tool's `run` wraps the permission ask).
|
||||
|
|
@ -61,8 +63,9 @@ From issue #34787 and design discussion. Do not relitigate these casually.
|
|||
`tools.<server>.<tool>` namespaces before handing them over.
|
||||
|
||||
### Discovery / search
|
||||
|
||||
- **Search only — no separate `describe`.** `tools.$codemode.search({ query?, namespace?,
|
||||
limit? })` over the final tool tree, owned by this package.
|
||||
limit? })` over the final tool tree, owned by this package.
|
||||
- Search result item shape: `{ path, description, signature }` in an `{ items, total }`
|
||||
wrapper. The `signature` string embeds the full input/output TypeScript types — in search
|
||||
results it is the pretty, JSDoc-annotated multiline form (Fix 7), so per-field schema
|
||||
|
|
@ -82,6 +85,7 @@ From issue #34787 and design discussion. Do not relitigate these casually.
|
|||
- Tools without an output schema render `unknown` as their return type.
|
||||
|
||||
### Schemas / Tool.make
|
||||
|
||||
- `Tool.make` carries rich metadata so search can render real signatures.
|
||||
- Support **Effect Schema** (first-class, validating) and **JSON Schema** (initially
|
||||
render-only — used for TypeScript rendering; the adapter may validate on its own). Leave
|
||||
|
|
@ -90,6 +94,7 @@ From issue #34787 and design discussion. Do not relitigate these casually.
|
|||
normalization for plugin authors can come later.
|
||||
|
||||
### Attachments / output
|
||||
|
||||
- **No `output.text/file/image` API in v1.** (Deleted in Wave 2.)
|
||||
- Tool calls return native structured payloads into the sandbox. Files/images emitted by
|
||||
child tools **never enter the sandbox** — the OpenCode adapter strips and accumulates them
|
||||
|
|
@ -100,6 +105,7 @@ From issue #34787 and design discussion. Do not relitigate these casually.
|
|||
image bytes into context or drop attachments.
|
||||
|
||||
### Runtime behavior
|
||||
|
||||
- Limits are EXACTLY the three public knobs: `{ timeoutMs, maxToolCalls, maxOutputBytes }` —
|
||||
matching the original locked spec exactly. NO limit has a default (user direction, Fix 6
|
||||
for the first two; extended to `maxOutputBytes` in the truncation-layering fix below):
|
||||
|
|
@ -129,7 +135,7 @@ From issue #34787 and design discussion. Do not relitigate these casually.
|
|||
- `console.*` is captured into `logs` on the result; the host appends them to model-facing
|
||||
output. Not a tool call; costs no tool budget.
|
||||
- Simple tool-call **start/end hooks** for nested progress: `onToolCallStart({ index, name,
|
||||
input })` and `onToolCallEnd({ index, name, input, durationMs, outcome, message? })`.
|
||||
input })` and `onToolCallEnd({ index, name, input, durationMs, outcome, message? })`.
|
||||
Interrupted calls fire no end event. No `CurrentToolCall` context service (removed in
|
||||
Wave 2).
|
||||
|
||||
|
|
@ -148,8 +154,9 @@ and `bun run typecheck`; from `packages/opencode`, `bun run typecheck` and
|
|||
`test/tool/registry.test.ts`).
|
||||
|
||||
### Wave 0 — scaffold (done)
|
||||
|
||||
- `packages/codemode` created from the experiments implementation: `src/{index,codemode,tool,
|
||||
tool-error,tool-runtime}.ts`, README, AGENTS.md, tests.
|
||||
tool-error,tool-runtime}.ts`, README, AGENTS.md, tests.
|
||||
- `package.json`: name `@opencode-ai/codemode`, deps `acorn@8.15.0`, `typescript: catalog:`,
|
||||
`effect: catalog:` (both repos pin effect `4.0.0-beta.83`; opencode's effect patch only
|
||||
touches `unstable/httpapi`, which this package doesn't use).
|
||||
|
|
@ -157,6 +164,7 @@ and `bun run typecheck`; from `packages/opencode`, `bun run typecheck` and
|
|||
Context.Service key string renamed to `@opencode-ai/codemode/CurrentToolCall`.
|
||||
|
||||
### Wave 1a — forgiving JS semantics (done)
|
||||
|
||||
Ported from the old opencode rune work; `test/parity.test.ts` (24 tests) is the acceptance
|
||||
spec. The seeded interpreter was deliberately strict; these behaviors replaced that:
|
||||
|
||||
|
|
@ -173,6 +181,7 @@ spec. The seeded interpreter was deliberately strict; these behaviors replaced t
|
|||
null/undefined still throws (real JS throws too).
|
||||
|
||||
### Wave 1b-i — stdlib value types: Date, RegExp, Map, Set (done)
|
||||
|
||||
`src/values.ts` holds `SandboxDate/SandboxRegExp/SandboxMap/SandboxSet` (own module so both
|
||||
`codemode.ts` and `tool-runtime.ts` import without a cycle). Design:
|
||||
|
||||
|
|
@ -202,19 +211,20 @@ spec. The seeded interpreter was deliberately strict; these behaviors replaced t
|
|||
interpolation renders `/regex/` and ISO dates directly.
|
||||
|
||||
### Wave 2 — API layer (done)
|
||||
|
||||
The package's public contract, reshaped for the Wave 3 adapter. 101 tests / 0 fail after this
|
||||
wave; both packages typecheck clean.
|
||||
|
||||
- **`Tool.make` schema flexibility** (`src/tool.ts`): `input`/`output` each accept an Effect
|
||||
Schema (validating, decoded both directions as before) OR a raw JSON Schema document
|
||||
(render-only — no validation, values pass through; rendering handles `$defs`/`definitions`
|
||||
+ `$ref`). `output` is **optional** → signature renders `Promise<unknown>` and the host
|
||||
result is exposed as-is. Discrimination via `Schema.isSchema`. New helpers exported from
|
||||
`tool.ts`: `inputTypeScript`/`outputTypeScript`/`decodeInput`/`decodeOutput`/
|
||||
`jsonSchemaToTypeScript`; `tool-runtime.ts` consumes them (no direct `Schema.*` use there
|
||||
anymore). Types `JsonSchema`/`ToolSchema` exported from the index. Note: an empty
|
||||
`Schema.Struct({})` renders as `{ } | Array<unknown>` (effect's JSON Schema emission) —
|
||||
cosmetic, fixed in Wave 4.
|
||||
- `$ref`). `output` is **optional** → signature renders `Promise<unknown>` and the host
|
||||
result is exposed as-is. Discrimination via `Schema.isSchema`. New helpers exported from
|
||||
`tool.ts`: `inputTypeScript`/`outputTypeScript`/`decodeInput`/`decodeOutput`/
|
||||
`jsonSchemaToTypeScript`; `tool-runtime.ts` consumes them (no direct `Schema.*` use there
|
||||
anymore). Types `JsonSchema`/`ToolSchema` exported from the index. Note: an empty
|
||||
`Schema.Struct({})` renders as `{ } | Array<unknown>` (effect's JSON Schema emission) —
|
||||
cosmetic, fixed in Wave 4.
|
||||
- **`output.*` API deleted**: `OutputItem`(+Schema), result `output` fields, the `output`
|
||||
global/namespace dispatch, `invokeOutput`/`outputItem`/helpers, interpreter output fields,
|
||||
instructions line, README section, seeded tests. AGENTS.md keeps a rephrased
|
||||
|
|
@ -228,22 +238,23 @@ wave; both packages typecheck clean.
|
|||
(`ToolError`/`ToolRuntimeError` message, else "Tool execution failed"). Interrupted calls
|
||||
fire no end event (timeout kills the whole execution anyway).
|
||||
- **Limits collapse**: public `ExecutionLimits` = `{ timeoutMs?, maxToolCalls?,
|
||||
maxOutputBytes? }` (defaults 10_000 / 100 / 32_000). This wave kept the other knobs as
|
||||
maxOutputBytes? }` (defaults 10_000 / 100 / 32_000). This wave kept the other knobs as
|
||||
internal defaults reachable through an `@internal` `InternalExecutionLimits` type; Fix 5
|
||||
later deleted that type and the internal limit system entirely.
|
||||
- **`maxOutputBytes` truncation** (CodeMode-owned, never fails): applied via `boundOutput` in
|
||||
a final `Effect.map` over every result path (success/timeout/normalized failure). Oversized
|
||||
serialized values become truncated text + ` [result truncated: N bytes exceeds the M-byte
|
||||
output limit; return a smaller value]`; logs keep leading lines within the remaining budget
|
||||
+ `[logs truncated: showing K of N lines]`; result gains `truncated: true` (also added to
|
||||
`ExecuteResultSchema`). UTF-8-safe truncation (no split code points). (The in-sandbox
|
||||
`maxDataBytes` check that used to throw first on oversized raw values died in Fix 5 —
|
||||
truncation is now the only result-size mechanism.)
|
||||
output limit; return a smaller value]`; logs keep leading lines within the remaining budget
|
||||
- `[logs truncated: showing K of N lines]`; result gains `truncated: true` (also added to
|
||||
`ExecuteResultSchema`). UTF-8-safe truncation (no split code points). (The in-sandbox
|
||||
`maxDataBytes` check that used to throw first on oversized raw values died in Fix 5 —
|
||||
truncation is now the only result-size mechanism.)
|
||||
- **Search polish**: default limit 12 → **10** (`defaultSearchLimit`); exact-path lookup — a
|
||||
trimmed query equal to one tool path (optionally `tools.`-prefixed) returns that tool alone
|
||||
(`total: 1`), bypassing ranking. Tokenization/ranking/shape unchanged.
|
||||
|
||||
### Wave 3 — OpenCode MCP adapter (done)
|
||||
|
||||
`packages/opencode/src/session/code-mode.ts` rewritten as a thin adapter over this package;
|
||||
the vendored rune interpreter is gone. Same `define(mcpTools, mcpDefs, servers)` signature, so
|
||||
`tools.ts` gating (flag on + MCP tools exist → single `execute` tool, early-return suppresses
|
||||
|
|
@ -257,7 +268,7 @@ per-MCP registration; MCP resource tools unaffected) is unchanged.
|
|||
invoked) — so signature rendering, the inline-vs-search switch, and `$codemode.search`
|
||||
availability all come from this package and stay consistent with execution.
|
||||
- **`run` path**: per-child permission ask first (`ctx.ask({ permission: entry.key, patterns:
|
||||
["*"], always: ["*"] })`, exactly the old gating; approving `execute` approves no child).
|
||||
["*"], always: ["*"] })`, exactly the old gating; approving `execute` approves no child).
|
||||
Denials and host failures are mapped to `toolError(message)` so they surface as safe,
|
||||
catchable in-program failures (MCP `isError` text propagates as `e.message`; without this
|
||||
they'd be sanitized to "Tool execution failed"). Dispatch reuses the ai-sdk wrapper from
|
||||
|
|
@ -270,10 +281,10 @@ per-MCP registration; MCP resource tools unaffected) is unchanged.
|
|||
No handles, no `Result<T>` envelope, no base64 in the sandbox, no data-size tuning (the
|
||||
`maxDataBytes` budget that existed at the time was deleted in Fix 5).
|
||||
- **Execute result**: `{ output: formatValue(value) + trailing "Logs:" section (success AND
|
||||
error — logs are plain pre-formatted lines now), attachments: accumulated }` through the
|
||||
error — logs are plain pre-formatted lines now), attachments: accumulated }` through the
|
||||
existing `Tool.ExecuteResult.attachments` → `message-v2.ts` vision plumbing; attachments
|
||||
ride on both success and error results. Diagnostic `suggestions` not already contained in
|
||||
the message are appended to error output. Native outer truncation stays on (adapter never
|
||||
the message are appended to error output. Native outer truncation stays on (adapter never
|
||||
sets `metadata.truncated`); CodeMode's own `maxOutputBytes` (32 KB default at the time)
|
||||
cut first — since the truncation-layering fix, native truncation is the only layer.
|
||||
Limits: `{ timeoutMs: 30_000 }` at the time (matched the default MCP request timeout);
|
||||
|
|
@ -296,6 +307,7 @@ per-MCP registration; MCP resource tools unaffected) is unchanged.
|
|||
describe/`renderType`/`rankTools` tests died with the old design (58+17+24 → 34+16).
|
||||
|
||||
### Wave 4 — instructions/prompting + polish (done)
|
||||
|
||||
Instructions are now the budgeted-catalog + prompting-guidance form; verified e2e against a
|
||||
real MCP config. Package still 101 tests / 0 fail; opencode adapter suites still 34 + 16; both
|
||||
packages typecheck clean.
|
||||
|
|
@ -321,7 +333,7 @@ packages typecheck clean.
|
|||
exported); `CodeMode.execute` (one-shot) passes it too, preserving the
|
||||
`execute`≡`make().execute` law. A speculative `tools.$codemode.search` call on a small
|
||||
catalog now succeeds instead of `UnknownTool`, and unknown-tool suggestions always point at
|
||||
search. Search is *advertised* in the instructions only when the inlined list is PARTIAL,
|
||||
search. Search is _advertised_ in the instructions only when the inlined list is PARTIAL,
|
||||
keeping small-catalog instructions tight.
|
||||
- **Prompting content** in `instructions()`, mapping 1:1 to the §5 transcript failures:
|
||||
parse-string-results-as-JSON, return-small, console-for-intermediates, and
|
||||
|
|
@ -340,7 +352,7 @@ packages typecheck clean.
|
|||
- **E2E (verified, headless)**: from the repo root with `OPENCODE_EXPERIMENTAL_CODE_MODE=1`,
|
||||
the scratch `.opencode/opencode.jsonc` (context7, github, playwright, sentry, memory,
|
||||
sequential-thinking; left uncommitted/as-is), and `bun packages/opencode/src/index.ts run
|
||||
--dangerously-skip-permissions -m opencode/claude-sonnet-4-5 "..."`. Confirmed: a single
|
||||
--dangerously-skip-permissions -m opencode/claude-sonnet-4-5 "..."`. Confirmed: a single
|
||||
`execute` tool registered alongside core tools (per-MCP registration suppressed; MCP
|
||||
resource tools unaffected); the live description read back as "Available tools (PARTIAL —
|
||||
56 of 88 shown; find the rest with tools.$codemode.search):" with correct per-namespace
|
||||
|
|
@ -352,6 +364,7 @@ packages typecheck clean.
|
|||
images, output truncation.
|
||||
|
||||
### Wave 5 — Promise generalization (done)
|
||||
|
||||
First-class promise values in the interpreter; the direct-tool-call-only `Promise.all`
|
||||
restriction (and its bespoke AST checks) is gone. Package suite is 136 tests / 0 fail (35 new
|
||||
in `test/promise.test.ts`); adapter suites and both typechecks unchanged/green; the opencode
|
||||
|
|
@ -495,7 +508,7 @@ adapter needed **no changes**.
|
|||
`tools.*`." (the second line drops the tools clause when the tree is empty).
|
||||
- **`## Workflow`**: numbered steps — find a tool via `tools.$codemode.search` → read
|
||||
the `{ path, description, signature }` matches → call by path → `typeof res ===
|
||||
"string" ? JSON.parse(res) : res` → return only the needed fields. When the catalog is
|
||||
"string" ? JSON.parse(res) : res` → return only the needed fields. When the catalog is
|
||||
COMPLETE the search/read steps collapse into "Pick a tool from the list under
|
||||
`## Available tools`" and the steps renumber (4 instead of 5).
|
||||
- **`## Rules`**: call-by-exact-path; TEXT-is-JSON → JSON.parse; return small (never raw
|
||||
|
|
@ -526,70 +539,72 @@ adapter needed **no changes**.
|
|||
|
||||
**Fix 4 — token-budgeted catalog (was bytes)** (user direction: signatures need a token
|
||||
budget; namespaces must always be present):
|
||||
- `src/token.ts` added: copy of `@opencode-ai/core/util/token` (`round(chars / 4)`), so
|
||||
the package stays dependency-free; keep in sync if the core heuristic changes.
|
||||
- `DiscoveryOptions.maxInlineCatalogBytes` → `maxInlineCatalogTokens` (default 4,000
|
||||
estimated tokens ≈ the old 16,000 bytes at 4 chars/token — behavior parity, not a size
|
||||
reduction). `discoveryPlan` charges `estimate(catalogLine(tool))` per line; cheapest-first
|
||||
+ stop-on-first-miss unchanged at the time (stop-on-first-miss replaced by round-robin in
|
||||
|
||||
- `src/token.ts` added: copy of `@opencode-ai/core/util/token` (`round(chars / 4)`), so
|
||||
the package stays dependency-free; keep in sync if the core heuristic changes.
|
||||
- `DiscoveryOptions.maxInlineCatalogBytes` → `maxInlineCatalogTokens` (default 4,000
|
||||
estimated tokens ≈ the old 16,000 bytes at 4 chars/token — behavior parity, not a size
|
||||
reduction). `discoveryPlan` charges `estimate(catalogLine(tool))` per line; cheapest-first
|
||||
- stop-on-first-miss unchanged at the time (stop-on-first-miss replaced by round-robin in
|
||||
Fix 8). Namespace stub lines were and remain unbudgeted — every
|
||||
namespace always appears with its tool count, even at budget 0 (asserted in package and
|
||||
adapter tests).
|
||||
- Ripple: chars/4 rounding erases small line-length differences, so equal-cost lines fall
|
||||
to the lexicographic path tiebreak; the adapter's PARTIAL test now asserts the
|
||||
lexicographic tail (`op_99`) is excluded instead of `op_149`. Fixed-prose measurements
|
||||
(2026-07): preamble ~44 + Workflow ~146 + Rules ~362 + Syntax ~453 ≈ 1,100 tokens fixed;
|
||||
worst-case net description ≈ fixed + 4,000 ≈ 5,100 estimated tokens.
|
||||
- Ripple: chars/4 rounding erases small line-length differences, so equal-cost lines fall
|
||||
to the lexicographic path tiebreak; the adapter's PARTIAL test now asserts the
|
||||
lexicographic tail (`op_99`) is excluded instead of `op_149`. Fixed-prose measurements
|
||||
(2026-07): preamble ~44 + Workflow ~146 + Rules ~362 + Syntax ~453 ≈ 1,100 tokens fixed;
|
||||
worst-case net description ≈ fixed + 4,000 ≈ 5,100 estimated tokens.
|
||||
|
||||
**Fix 5 — internal limits removed** (user direction: only the three PUBLIC limits survive as
|
||||
configurable knobs; the internal limit system dies):
|
||||
- `ExecutionLimits` (`timeoutMs` 10_000 / `maxToolCalls` 100 / `maxOutputBytes` 32_000 at
|
||||
the time; Fix 6 later removed the first two defaults. Same validation: safe integers,
|
||||
timeoutMs >= 1, others >= 0, RangeError otherwise) is now
|
||||
the ENTIRE limit surface — exactly the shape §2's original locked spec named.
|
||||
`ResolvedExecutionLimits` shrank to those three fields; the `@internal`
|
||||
`InternalExecutionLimits` type is deleted.
|
||||
- **Deleted outright**: `maxOperations` and the whole operation-budget machinery
|
||||
(`recordWork`/`recordOperation`/`budget.operations`, plus the `workUnits`/
|
||||
`cheapArrayMethods` cost helpers); `maxSourceBytes` (the pre-parse source-size check);
|
||||
`maxDataBytes` (every byte-accounting path: `runtimeValueBytes`, `boundedProgramValue`,
|
||||
the container-size caches (`containerSizes`/`objectCounts`), Map/Set incremental `bytes`
|
||||
fields in `values.ts`, string-growth `limitString` checks, tool-argument/result byte
|
||||
checks in `tool-runtime.ts`, and the final-result size check); `maxAuditBytes` (log and
|
||||
audit-trail byte accounting — `toolCalls` records and the start/end hooks are unchanged);
|
||||
`maxCollectionLength` (every array-length/object-field-count check — this knob was
|
||||
actively harmful: an MCP tool returning 20k rows failed). The `OperationLimitExceeded`
|
||||
and `AuditLimitExceeded` diagnostic kinds are gone from the `DiagnosticKind` union and
|
||||
`ExecuteResultSchema` (fine — the package is unreleased).
|
||||
- **Fixed constants, not knobs**: `TOOL_CALL_CONCURRENCY = 8` (codemode.ts; the fork
|
||||
semaphore) and `MAX_VALUE_DEPTH = 32` (tool-runtime.ts; the `copyIn` depth check — kept
|
||||
only because it produces a clearer error than a native stack-overflow RangeError; still
|
||||
`InvalidDataValue`). The `DataLimits` plumbing through `tool-runtime.ts` is gone —
|
||||
`copyIn(value, label)` needs no limits argument, and `ToolRuntime.make` takes just
|
||||
`(tools, maxToolCalls, hooks?, searchIndex?)`.
|
||||
- **Verified fact**: timeout interruption does NOT depend on the operation budget — the
|
||||
Effect fiber runtime auto-yields between interpreter steps, so `timeoutMs` interrupts
|
||||
even a pure `while (true) {}` loop (empirically verified: a 200ms timeout fired at
|
||||
~225ms with maxOperations set to MAX_SAFE_INTEGER before the deletion). A regression
|
||||
test in `codemode.test.ts` asserts exactly this (`while(true){}` + `timeoutMs: 200` →
|
||||
`TimeoutExceeded`, elapsed well under a few seconds).
|
||||
- **Kept (correctness, not budgets)**: circular detection (`copyIn` walks +
|
||||
`rejectCircularInsertion` on mutations), plain-objects-only, blocked properties
|
||||
(`__proto__`/`constructor`/`prototype`), data-only checks, and all three public-limit
|
||||
behaviors unchanged.
|
||||
- Behavior deltas beyond the intended kills: in-sandbox structures deeper than 32 levels
|
||||
now fail at the data boundary (`copyIn`) instead of at construction; array index
|
||||
assignment allows any non-negative integer index (holes permitted, message now "must be
|
||||
a non-negative integer"); interpreter-produced deep/hostile structures that overflow the
|
||||
native stack during a walk still normalize to the existing "Execution exceeded the
|
||||
maximum nesting depth." data diagnostic — failures remain data everywhere.
|
||||
- Tests: deleted the knob-only tests (stdlib Map/Set collection-length growth ×2,
|
||||
enumeration operation-budget, codemode maxDataBytes/maxSourceBytes/maxOperations/
|
||||
maxConcurrency-RangeError assertions, and the adapter's runaway-loop-via-operation-limit
|
||||
test — superseded by the package timeout regression test); rewrote the helpers that used
|
||||
`InternalExecutionLimits` as a convenience to plain `ExecutionLimits`
|
||||
(promise/enumeration/stdlib run helpers). Package suite: 154 pass / 0 fail; adapter
|
||||
suites: 34 + 16.
|
||||
|
||||
- `ExecutionLimits` (`timeoutMs` 10_000 / `maxToolCalls` 100 / `maxOutputBytes` 32_000 at
|
||||
the time; Fix 6 later removed the first two defaults. Same validation: safe integers,
|
||||
timeoutMs >= 1, others >= 0, RangeError otherwise) is now
|
||||
the ENTIRE limit surface — exactly the shape §2's original locked spec named.
|
||||
`ResolvedExecutionLimits` shrank to those three fields; the `@internal`
|
||||
`InternalExecutionLimits` type is deleted.
|
||||
- **Deleted outright**: `maxOperations` and the whole operation-budget machinery
|
||||
(`recordWork`/`recordOperation`/`budget.operations`, plus the `workUnits`/
|
||||
`cheapArrayMethods` cost helpers); `maxSourceBytes` (the pre-parse source-size check);
|
||||
`maxDataBytes` (every byte-accounting path: `runtimeValueBytes`, `boundedProgramValue`,
|
||||
the container-size caches (`containerSizes`/`objectCounts`), Map/Set incremental `bytes`
|
||||
fields in `values.ts`, string-growth `limitString` checks, tool-argument/result byte
|
||||
checks in `tool-runtime.ts`, and the final-result size check); `maxAuditBytes` (log and
|
||||
audit-trail byte accounting — `toolCalls` records and the start/end hooks are unchanged);
|
||||
`maxCollectionLength` (every array-length/object-field-count check — this knob was
|
||||
actively harmful: an MCP tool returning 20k rows failed). The `OperationLimitExceeded`
|
||||
and `AuditLimitExceeded` diagnostic kinds are gone from the `DiagnosticKind` union and
|
||||
`ExecuteResultSchema` (fine — the package is unreleased).
|
||||
- **Fixed constants, not knobs**: `TOOL_CALL_CONCURRENCY = 8` (codemode.ts; the fork
|
||||
semaphore) and `MAX_VALUE_DEPTH = 32` (tool-runtime.ts; the `copyIn` depth check — kept
|
||||
only because it produces a clearer error than a native stack-overflow RangeError; still
|
||||
`InvalidDataValue`). The `DataLimits` plumbing through `tool-runtime.ts` is gone —
|
||||
`copyIn(value, label)` needs no limits argument, and `ToolRuntime.make` takes just
|
||||
`(tools, maxToolCalls, hooks?, searchIndex?)`.
|
||||
- **Verified fact**: timeout interruption does NOT depend on the operation budget — the
|
||||
Effect fiber runtime auto-yields between interpreter steps, so `timeoutMs` interrupts
|
||||
even a pure `while (true) {}` loop (empirically verified: a 200ms timeout fired at
|
||||
~225ms with maxOperations set to MAX_SAFE_INTEGER before the deletion). A regression
|
||||
test in `codemode.test.ts` asserts exactly this (`while(true){}` + `timeoutMs: 200` →
|
||||
`TimeoutExceeded`, elapsed well under a few seconds).
|
||||
- **Kept (correctness, not budgets)**: circular detection (`copyIn` walks +
|
||||
`rejectCircularInsertion` on mutations), plain-objects-only, blocked properties
|
||||
(`__proto__`/`constructor`/`prototype`), data-only checks, and all three public-limit
|
||||
behaviors unchanged.
|
||||
- Behavior deltas beyond the intended kills: in-sandbox structures deeper than 32 levels
|
||||
now fail at the data boundary (`copyIn`) instead of at construction; array index
|
||||
assignment allows any non-negative integer index (holes permitted, message now "must be
|
||||
a non-negative integer"); interpreter-produced deep/hostile structures that overflow the
|
||||
native stack during a walk still normalize to the existing "Execution exceeded the
|
||||
maximum nesting depth." data diagnostic — failures remain data everywhere.
|
||||
- Tests: deleted the knob-only tests (stdlib Map/Set collection-length growth ×2,
|
||||
enumeration operation-budget, codemode maxDataBytes/maxSourceBytes/maxOperations/
|
||||
maxConcurrency-RangeError assertions, and the adapter's runaway-loop-via-operation-limit
|
||||
test — superseded by the package timeout regression test); rewrote the helpers that used
|
||||
`InternalExecutionLimits` as a convenience to plain `ExecutionLimits`
|
||||
(promise/enumeration/stdlib run helpers). Package suite: 154 pass / 0 fail; adapter
|
||||
suites: 34 + 16.
|
||||
|
||||
**Fix 6 — no default timeout / tool-call cap** (user direction): `timeoutMs` and
|
||||
`maxToolCalls` lost their defaults (were 10_000 / 100) — absent now means no timeout /
|
||||
|
|
@ -639,196 +654,200 @@ adapter suites: 34 + 16.
|
|||
**Fix 8 — condensed instructions + round-robin catalog fairness + plural-aware search**
|
||||
(user direction: the fixed instruction prose was too verbose; two discovery fixes ride
|
||||
along). All in `tool-runtime.ts`; no interpreter changes.
|
||||
- **Syntax section inverted**: the three dense allowlist lines (~453 estimated tokens)
|
||||
are replaced by four short lines (~188) built on "models already know JavaScript; name
|
||||
only what is unusual or missing": (1) standard modern JS works — functions/closures,
|
||||
destructuring, template literals, loops, try/catch, spread, optional chaining, the
|
||||
usual Array/String/Object/Math/JSON methods, plus Date/RegExp/Map/Set and
|
||||
Promise.all/allSettled/race/resolve/reject; (2) TypeScript type annotations are
|
||||
stripped before execution, decorators are not supported; (3) NOT supported (each fails
|
||||
with a message naming the alternative): classes, generators, for await...of,
|
||||
.then/.catch/.finally (use await with try/catch), `x instanceof Error` (caught errors
|
||||
are plain `{ name, message }` objects), splice; (4) the data-boundary note (Dates →
|
||||
ISO strings; Map/Set/RegExp → `{}`). Every claim was verified against the interpreter
|
||||
before writing: probed empirically — classes/generators/for-await/.then/.catch/
|
||||
.finally/`instanceof Error`/splice/decorators/BigInt/labeled statements/tagged
|
||||
templates/object getters all fail with clear diagnostics; TS annotations/`as`/
|
||||
interfaces/type aliases are stripped and TS **enums actually work** (transpileModule
|
||||
compiles them to an IIFE the interpreter runs), hence enums deliberately unmentioned.
|
||||
`supportedSyntaxMessage` (the in-diagnostic text in `codemode.ts`) is untouched.
|
||||
- **Workflow/Rules deduped**: the call-by-exact-path, JSON.parse-string-results, and
|
||||
return-small content now lives ONLY in the numbered Workflow steps (with their
|
||||
compliance-driving justifications inline: "most tools return JSON as a string", "raw
|
||||
payloads get truncated and waste context"); Rules keeps only bullets adding new
|
||||
content — filter/aggregate collections in code, console.* intermediates (logs ride
|
||||
back), Promise.all parallelism, Object.keys/for...in enumeration, browse-namespace
|
||||
(PARTIAL only), and the media rule compressed to one line. The no-.then/.catch
|
||||
guidance moved to the Syntax not-supported line. Content upgrades: the PARTIAL search
|
||||
step gained query-style guidance (`— short phrases like "list issues" work best`; a
|
||||
clearly-a-query-string example, not a tool name), and the exact-path guidance is now
|
||||
"call it with the result's `path` as-is (never guess segments)" / COMPLETE: "use it
|
||||
as-is rather than guessing segments".
|
||||
- **Fixed-prose measurements** (instructions split on `"\n## "`, catalog budget 0,
|
||||
bytes/3.7 — same method as Fix 4; chars/4 in parentheses):
|
||||
preamble 44 → 44 (41 → 41), Workflow 146 → 187 (135 → 171), Rules 362 → 191
|
||||
(332 → 176), Syntax 453 → 188 (419 → 174); fixed prose total 1,005 → 610 (927 → 562),
|
||||
≈ 40% reduction with no behavioral content dropped. Workflow grew slightly because it
|
||||
absorbed the deduped parse/return-small justifications.
|
||||
- **Round-robin namespace inlining** (`discoveryPlan`): the ported stop-on-first-miss
|
||||
behavior (alphabetically-late namespaces starved to "none shown" while an early
|
||||
namespace inlines everything) is replaced by round-robin fairness — in each round
|
||||
(namespaces alphabetical), every namespace still holding un-inlined tools attempts to
|
||||
place its next-cheapest line against the shared token budget; a namespace whose next
|
||||
line does not fit is done while the others keep going; stop when all are done. Every
|
||||
namespace gets some representation before any namespace gets everything. Kept:
|
||||
`estimate` (chars/4) budget accounting, unbudgeted namespace stub lines, per-namespace
|
||||
`(N tools)`/`(N tools, K shown)`/`(N tools, none shown)` labels, COMPLETE vs PARTIAL
|
||||
header, alphabetical namespace order in the output, cheapest-first within each
|
||||
namespace's shown set.
|
||||
- **Plural/singular search fix**: `tokenize`d terms matched one-directionally (term must
|
||||
be substring of indexed text), so query "issues" missed a tool whose text only says
|
||||
"issue". Now each term expands to `termForms` — the term plus naive singular variants
|
||||
(trailing "es" stripped when length > 3, trailing "s" when length > 2) — and each of
|
||||
the four field checks passes when ANY form matches. Weights, exact-path lookup, and
|
||||
namespace scoping untouched. A true plural path match still outranks a singular-only
|
||||
description match (path substring 8 + searchable 2 > description 4 + searchable 2).
|
||||
- **Tests**: package instruction/structure assertions updated to the new text; new
|
||||
syntax-section test (leads with "Standard modern JavaScript works", names the
|
||||
verified not-supported list, keeps the data-boundary note); the budget-exhaustion
|
||||
test rewritten to assert the new fairness (alpha.expensive not fitting must NOT
|
||||
prevent beta.cheap from showing: PARTIAL 2 of 3, `- beta (1 tool)` fully shown); new
|
||||
plural/singular test (query "issues" finds a singular-only tool; ranking still
|
||||
prefers the true "issues" path match). Adapter: description assertions updated; the
|
||||
large-catalog PARTIAL test now asserts `zeta_only_tool` IS shown (`- zeta (1 tool)` +
|
||||
its inlined line) — it was "none shown" under starvation. README updated (budgeted
|
||||
catalog paragraph → round-robin; search paragraph → singular variants;
|
||||
instructions-structure paragraph → new section contents). Package suite: 169 pass /
|
||||
0 fail; adapter suites: 34 + 16.
|
||||
|
||||
- **Syntax section inverted**: the three dense allowlist lines (~453 estimated tokens)
|
||||
are replaced by four short lines (~188) built on "models already know JavaScript; name
|
||||
only what is unusual or missing": (1) standard modern JS works — functions/closures,
|
||||
destructuring, template literals, loops, try/catch, spread, optional chaining, the
|
||||
usual Array/String/Object/Math/JSON methods, plus Date/RegExp/Map/Set and
|
||||
Promise.all/allSettled/race/resolve/reject; (2) TypeScript type annotations are
|
||||
stripped before execution, decorators are not supported; (3) NOT supported (each fails
|
||||
with a message naming the alternative): classes, generators, for await...of,
|
||||
.then/.catch/.finally (use await with try/catch), `x instanceof Error` (caught errors
|
||||
are plain `{ name, message }` objects), splice; (4) the data-boundary note (Dates →
|
||||
ISO strings; Map/Set/RegExp → `{}`). Every claim was verified against the interpreter
|
||||
before writing: probed empirically — classes/generators/for-await/.then/.catch/
|
||||
.finally/`instanceof Error`/splice/decorators/BigInt/labeled statements/tagged
|
||||
templates/object getters all fail with clear diagnostics; TS annotations/`as`/
|
||||
interfaces/type aliases are stripped and TS **enums actually work** (transpileModule
|
||||
compiles them to an IIFE the interpreter runs), hence enums deliberately unmentioned.
|
||||
`supportedSyntaxMessage` (the in-diagnostic text in `codemode.ts`) is untouched.
|
||||
- **Workflow/Rules deduped**: the call-by-exact-path, JSON.parse-string-results, and
|
||||
return-small content now lives ONLY in the numbered Workflow steps (with their
|
||||
compliance-driving justifications inline: "most tools return JSON as a string", "raw
|
||||
payloads get truncated and waste context"); Rules keeps only bullets adding new
|
||||
content — filter/aggregate collections in code, console.\* intermediates (logs ride
|
||||
back), Promise.all parallelism, Object.keys/for...in enumeration, browse-namespace
|
||||
(PARTIAL only), and the media rule compressed to one line. The no-.then/.catch
|
||||
guidance moved to the Syntax not-supported line. Content upgrades: the PARTIAL search
|
||||
step gained query-style guidance (`— short phrases like "list issues" work best`; a
|
||||
clearly-a-query-string example, not a tool name), and the exact-path guidance is now
|
||||
"call it with the result's `path` as-is (never guess segments)" / COMPLETE: "use it
|
||||
as-is rather than guessing segments".
|
||||
- **Fixed-prose measurements** (instructions split on `"\n## "`, catalog budget 0,
|
||||
bytes/3.7 — same method as Fix 4; chars/4 in parentheses):
|
||||
preamble 44 → 44 (41 → 41), Workflow 146 → 187 (135 → 171), Rules 362 → 191
|
||||
(332 → 176), Syntax 453 → 188 (419 → 174); fixed prose total 1,005 → 610 (927 → 562),
|
||||
≈ 40% reduction with no behavioral content dropped. Workflow grew slightly because it
|
||||
absorbed the deduped parse/return-small justifications.
|
||||
- **Round-robin namespace inlining** (`discoveryPlan`): the ported stop-on-first-miss
|
||||
behavior (alphabetically-late namespaces starved to "none shown" while an early
|
||||
namespace inlines everything) is replaced by round-robin fairness — in each round
|
||||
(namespaces alphabetical), every namespace still holding un-inlined tools attempts to
|
||||
place its next-cheapest line against the shared token budget; a namespace whose next
|
||||
line does not fit is done while the others keep going; stop when all are done. Every
|
||||
namespace gets some representation before any namespace gets everything. Kept:
|
||||
`estimate` (chars/4) budget accounting, unbudgeted namespace stub lines, per-namespace
|
||||
`(N tools)`/`(N tools, K shown)`/`(N tools, none shown)` labels, COMPLETE vs PARTIAL
|
||||
header, alphabetical namespace order in the output, cheapest-first within each
|
||||
namespace's shown set.
|
||||
- **Plural/singular search fix**: `tokenize`d terms matched one-directionally (term must
|
||||
be substring of indexed text), so query "issues" missed a tool whose text only says
|
||||
"issue". Now each term expands to `termForms` — the term plus naive singular variants
|
||||
(trailing "es" stripped when length > 3, trailing "s" when length > 2) — and each of
|
||||
the four field checks passes when ANY form matches. Weights, exact-path lookup, and
|
||||
namespace scoping untouched. A true plural path match still outranks a singular-only
|
||||
description match (path substring 8 + searchable 2 > description 4 + searchable 2).
|
||||
- **Tests**: package instruction/structure assertions updated to the new text; new
|
||||
syntax-section test (leads with "Standard modern JavaScript works", names the
|
||||
verified not-supported list, keeps the data-boundary note); the budget-exhaustion
|
||||
test rewritten to assert the new fairness (alpha.expensive not fitting must NOT
|
||||
prevent beta.cheap from showing: PARTIAL 2 of 3, `- beta (1 tool)` fully shown); new
|
||||
plural/singular test (query "issues" finds a singular-only tool; ranking still
|
||||
prefers the true "issues" path match). Adapter: description assertions updated; the
|
||||
large-catalog PARTIAL test now asserts `zeta_only_tool` IS shown (`- zeta (1 tool)` +
|
||||
its inlined line) — it was "none shown" under starvation. README updated (budgeted
|
||||
catalog paragraph → round-robin; search paragraph → singular variants;
|
||||
instructions-structure paragraph → new section contents). Package suite: 169 pass /
|
||||
0 fail; adapter suites: 34 + 16.
|
||||
|
||||
**Fix 9 — prompting trims per user review of Fix 8** (user reviewed the condensed
|
||||
instructions and directed further cuts):
|
||||
- Default `maxInlineCatalogTokens` 4,000 → **2,000** (user wants ~2k tokens of signatures
|
||||
auto-inlined; round-robin fairness from Fix 8 spreads it across all namespaces).
|
||||
- Console rule and files/images rule DROPPED from `## Rules`. Replaced by a single
|
||||
`unknown`-treatment warning: "A result typed `Promise<unknown>` has no guaranteed
|
||||
shape — verify what actually came back before relying on its fields." (Deliberately
|
||||
does NOT suggest console.log — user review: naming it there nudges models to log AND
|
||||
return the same data; the prompt stays console-neutral, neither for nor against.)
|
||||
The media-stripping MECHANISM is unchanged and still tested; only the prose about it
|
||||
is gone — the `[N images attached]` marker is self-explanatory in context.
|
||||
- Kept as-is per user: the JSON.parse workflow step (maps to the original motivating
|
||||
transcript failure; NOT copied from prior art — see §5 note), the browse-namespace rule
|
||||
(undecided), no no-fetch/ambient-authority rule added (proposed, not approved).
|
||||
- Explicitly REJECTED for now: auto-parsing JSON-looking text results at the adapter
|
||||
boundary ("could get weird" — type flips, program-sees vs tool-sent divergence). Logged
|
||||
as a next-iteration follow-up below.
|
||||
|
||||
- Default `maxInlineCatalogTokens` 4,000 → **2,000** (user wants ~2k tokens of signatures
|
||||
auto-inlined; round-robin fairness from Fix 8 spreads it across all namespaces).
|
||||
- Console rule and files/images rule DROPPED from `## Rules`. Replaced by a single
|
||||
`unknown`-treatment warning: "A result typed `Promise<unknown>` has no guaranteed
|
||||
shape — verify what actually came back before relying on its fields." (Deliberately
|
||||
does NOT suggest console.log — user review: naming it there nudges models to log AND
|
||||
return the same data; the prompt stays console-neutral, neither for nor against.)
|
||||
The media-stripping MECHANISM is unchanged and still tested; only the prose about it
|
||||
is gone — the `[N images attached]` marker is self-explanatory in context.
|
||||
- Kept as-is per user: the JSON.parse workflow step (maps to the original motivating
|
||||
transcript failure; NOT copied from prior art — see §5 note), the browse-namespace rule
|
||||
(undecided), no no-fetch/ambient-authority rule added (proposed, not approved).
|
||||
- Explicitly REJECTED for now: auto-parsing JSON-looking text results at the adapter
|
||||
boundary ("could get weird" — type flips, program-sees vs tool-sent divergence). Logged
|
||||
as a next-iteration follow-up below.
|
||||
|
||||
**DSL-expansion pass — interpreter-surface batch from §4** (the deferred medium-tier JS
|
||||
parity items, done as one focused pass; no public API or limit changes):
|
||||
- **`instanceof` + real Error values**: the `errorConstructors` names (`Error`,
|
||||
`TypeError`, `RangeError`, `SyntaxError`, `ReferenceError`, `EvalError`, `URIError`) are
|
||||
bound globals (`ErrorConstructorReference`, callable with or without `new`; `typeof` →
|
||||
`"function"`). Error values stay the same plain `{ name, message }` null-prototype
|
||||
objects as before — the constructor name additionally rides on a NON-ENUMERABLE symbol
|
||||
key (`ErrorBrand`), which every `Object.entries`-based walk (copyIn/copyOut, spread,
|
||||
JSON.stringify) is blind to, so serialization is byte-identical to the old shape and the
|
||||
brand is lost on spread/boundary copies exactly like JS loses the prototype.
|
||||
`caughtErrorValue` produces `{ name, message }` wrappers via `createErrorValue`, so
|
||||
caught interpreter AND tool failures are `instanceof Error` and carry the `name` the
|
||||
equivalent real-JS failure would have (follow-up fix, user-directed — "closest to real
|
||||
JS"): `InterpreterRuntimeError` gained an `errorName` field ("Error" default) set
|
||||
fluently at throw sites via `.as(name)` — `JSON.parse` failures are `"SyntaxError"` (and
|
||||
now include the engine's position detail in the message; safe — derived from the
|
||||
program-supplied string), invalid regex patterns/flags `"SyntaxError"`, unknown
|
||||
identifiers and TDZ access `"ReferenceError"`, assignment to a constant `"TypeError"`,
|
||||
a bad `normalize` form `"RangeError"`; a host Error reaching the catch path directly
|
||||
keeps its own name when it is one of the standard seven. Tool failures and everything
|
||||
without a specific analogue stay `"Error"` — internal class names never leak. Specific
|
||||
names satisfy the specific `instanceof` (`e instanceof SyntaxError`), matching JS.
|
||||
The operator is handled in `evaluateBinaryExpression`
|
||||
BEFORE the data-only operand check (like `typeof`, it observes any lhs — promises and
|
||||
functions included); recognized rhs: the error constructors (a specific type matches its
|
||||
own brand or `Error`, never a sibling), `Date`/`RegExp`/`Map`/`Set` (sandbox classes),
|
||||
`Array`, `Object` (any object/function-ish value), `Promise` (`SandboxPromise`), and
|
||||
`Number`/`String`/`Boolean` (always false — no boxed values exist); anything else is a
|
||||
catchable error naming the recognized constructors.
|
||||
- **Array methods**: `splice` (mutating, returns the removed elements; insertions run
|
||||
`rejectCircularInsertion` like push/unshift; one-arg form removes to the end, undefined
|
||||
delete count removes nothing), `fill` (circular-checked value) and `copyWithin`
|
||||
(host-delegated), and `keys`/`values`/`entries` returning **arrays** (the Map/Set
|
||||
convention — for...of and spread work either way). The `retryableArrayMethods`
|
||||
"rewrite using map/filter" hint set emptied out and was deleted with its branch; unknown
|
||||
array properties still read `undefined`.
|
||||
- **String methods**: `localeCompare(that)` (locale/options arguments ignored — host
|
||||
default locale; the dominant use is a sort comparator), `normalize(form?)` (invalid form
|
||||
→ catchable error naming the four valid forms), `trimLeft`/`trimRight` as
|
||||
trimStart/trimEnd aliases.
|
||||
- **Actionable regex failures**: `toHostRegex` and `constructRegExp` now show the
|
||||
offending pattern (or flags) plus the engine reason (deduped "Invalid regular
|
||||
expression:" prefix via `regexFailureReason`) and a shared escaping hint
|
||||
(`escapeRegexHint`); flags failures list the valid flag letters; the
|
||||
replaceAll/matchAll missing-`g` errors spell out the exact `/pattern/g` to write and
|
||||
the single-match alternative.
|
||||
- **copyIn split (the important one)**: `copyIn(value, label, preserveSandboxValues =
|
||||
false)` — recursion moved to a private `copyBounded`; `boundedData` (every intra-sandbox
|
||||
checkpoint: `Object.*` helpers, coercion/Array.from/join inputs, template
|
||||
interpolation, expression-result checkpoints) is now `copyIn(value, label, true)`,
|
||||
which passes `SandboxDate`/`SandboxRegExp`/`SandboxMap`/`SandboxSet` through **by
|
||||
reference as leaves** (contents not walked — Map/Set members are validated at their
|
||||
mutation sites) while keeping the depth (`MAX_VALUE_DEPTH`), circularity,
|
||||
plain-objects-only, blocked-property, and data-only checks; un-awaited promises keep
|
||||
the await-hinting rejection in BOTH modes (deliberate — JS-parity pass-through was
|
||||
considered and skipped to preserve the nudge). The HOST boundary (final result,
|
||||
tool-call arguments, `JSON.stringify`, tool-result intake) uses the default mode and
|
||||
still serializes JSON forms (Date → ISO, RegExp/Map/Set → `{}`); host instances met on
|
||||
the preserving path are defensively wrapped into sandbox equivalents. Ripple: the
|
||||
`Object.*` helpers treat sandbox values as empty objects (`Object.keys(map)` → `[]`,
|
||||
assign sources contribute nothing, hasOwn → false — JS has no own enumerable props
|
||||
there), so interpreter internals (`.map`/`.time`/`.regex`) can never leak; the
|
||||
template-literal sandbox carve-out collapsed into `boundedData`. Object/array spread
|
||||
already preserved instances (reference copies, no checkpoint) — now tested.
|
||||
- **Console formatting**: `formatConsoleArgument` is total and deep
|
||||
(`formatConsoleValue`): numbers render via `String` (`NaN`/`Infinity`/`-Infinity`
|
||||
literally — never the JSON `null`; finite numbers match their JSON form), nested
|
||||
strings are JSON-quoted, sandbox values keep their friendly forms at ANY depth (ISO
|
||||
date, `/regex/flags`, `Map(n) [...]`, `Set(n) [...]`), opaque references become
|
||||
in-place `[CodeMode reference]` markers instead of collapsing the whole argument,
|
||||
cycles render `[Circular]` (reachable via Map/Set members, which mutation never
|
||||
checkpoints), and depth beyond `MAX_CONSOLE_DEPTH = 32` (fixed constant, not a knob)
|
||||
degrades to `…` — console can no longer fail a program. `console.table` guards with
|
||||
`containsOpaqueReference` (sandbox cells render, e.g. ISO dates) and its row/cell
|
||||
walkers treat sandbox values as scalar cells.
|
||||
- **Prose**: the instructions Syntax not-supported line dropped its `instanceof
|
||||
Error`/splice mentions (nothing else reworded); README updated (checkpoint
|
||||
preservation vs boundary serialization, error values/`instanceof`, new array/string
|
||||
methods, regex-failure behavior); `supportedSyntaxMessage` left untouched (it lists
|
||||
supported syntax, was already non-exhaustive, and stays accurate).
|
||||
- **Tests**: package suite 169 → 209 (parity: Error/instanceof + real-JS error-name
|
||||
coverage, splice/fill/copyWithin/keys/values/entries, localeCompare/normalize/trim-alias
|
||||
describes; stdlib: checkpoint survival incl. tool-arg boundary pinning, stdlib
|
||||
`instanceof`, regex-message assertions; codemode: NaN/Infinity + nested/cyclic console
|
||||
rendering, table cells, caught-tool-failure `instanceof`); adapter suites unchanged
|
||||
(34 + 16, green); both packages `tsgo --noEmit` clean.
|
||||
|
||||
- **`instanceof` + real Error values**: the `errorConstructors` names (`Error`,
|
||||
`TypeError`, `RangeError`, `SyntaxError`, `ReferenceError`, `EvalError`, `URIError`) are
|
||||
bound globals (`ErrorConstructorReference`, callable with or without `new`; `typeof` →
|
||||
`"function"`). Error values stay the same plain `{ name, message }` null-prototype
|
||||
objects as before — the constructor name additionally rides on a NON-ENUMERABLE symbol
|
||||
key (`ErrorBrand`), which every `Object.entries`-based walk (copyIn/copyOut, spread,
|
||||
JSON.stringify) is blind to, so serialization is byte-identical to the old shape and the
|
||||
brand is lost on spread/boundary copies exactly like JS loses the prototype.
|
||||
`caughtErrorValue` produces `{ name, message }` wrappers via `createErrorValue`, so
|
||||
caught interpreter AND tool failures are `instanceof Error` and carry the `name` the
|
||||
equivalent real-JS failure would have (follow-up fix, user-directed — "closest to real
|
||||
JS"): `InterpreterRuntimeError` gained an `errorName` field ("Error" default) set
|
||||
fluently at throw sites via `.as(name)` — `JSON.parse` failures are `"SyntaxError"` (and
|
||||
now include the engine's position detail in the message; safe — derived from the
|
||||
program-supplied string), invalid regex patterns/flags `"SyntaxError"`, unknown
|
||||
identifiers and TDZ access `"ReferenceError"`, assignment to a constant `"TypeError"`,
|
||||
a bad `normalize` form `"RangeError"`; a host Error reaching the catch path directly
|
||||
keeps its own name when it is one of the standard seven. Tool failures and everything
|
||||
without a specific analogue stay `"Error"` — internal class names never leak. Specific
|
||||
names satisfy the specific `instanceof` (`e instanceof SyntaxError`), matching JS.
|
||||
The operator is handled in `evaluateBinaryExpression`
|
||||
BEFORE the data-only operand check (like `typeof`, it observes any lhs — promises and
|
||||
functions included); recognized rhs: the error constructors (a specific type matches its
|
||||
own brand or `Error`, never a sibling), `Date`/`RegExp`/`Map`/`Set` (sandbox classes),
|
||||
`Array`, `Object` (any object/function-ish value), `Promise` (`SandboxPromise`), and
|
||||
`Number`/`String`/`Boolean` (always false — no boxed values exist); anything else is a
|
||||
catchable error naming the recognized constructors.
|
||||
- **Array methods**: `splice` (mutating, returns the removed elements; insertions run
|
||||
`rejectCircularInsertion` like push/unshift; one-arg form removes to the end, undefined
|
||||
delete count removes nothing), `fill` (circular-checked value) and `copyWithin`
|
||||
(host-delegated), and `keys`/`values`/`entries` returning **arrays** (the Map/Set
|
||||
convention — for...of and spread work either way). The `retryableArrayMethods`
|
||||
"rewrite using map/filter" hint set emptied out and was deleted with its branch; unknown
|
||||
array properties still read `undefined`.
|
||||
- **String methods**: `localeCompare(that)` (locale/options arguments ignored — host
|
||||
default locale; the dominant use is a sort comparator), `normalize(form?)` (invalid form
|
||||
→ catchable error naming the four valid forms), `trimLeft`/`trimRight` as
|
||||
trimStart/trimEnd aliases.
|
||||
- **Actionable regex failures**: `toHostRegex` and `constructRegExp` now show the
|
||||
offending pattern (or flags) plus the engine reason (deduped "Invalid regular
|
||||
expression:" prefix via `regexFailureReason`) and a shared escaping hint
|
||||
(`escapeRegexHint`); flags failures list the valid flag letters; the
|
||||
replaceAll/matchAll missing-`g` errors spell out the exact `/pattern/g` to write and
|
||||
the single-match alternative.
|
||||
- **copyIn split (the important one)**: `copyIn(value, label, preserveSandboxValues =
|
||||
false)` — recursion moved to a private `copyBounded`; `boundedData` (every intra-sandbox
|
||||
checkpoint: `Object.*` helpers, coercion/Array.from/join inputs, template
|
||||
interpolation, expression-result checkpoints) is now `copyIn(value, label, true)`,
|
||||
which passes `SandboxDate`/`SandboxRegExp`/`SandboxMap`/`SandboxSet` through **by
|
||||
reference as leaves** (contents not walked — Map/Set members are validated at their
|
||||
mutation sites) while keeping the depth (`MAX_VALUE_DEPTH`), circularity,
|
||||
plain-objects-only, blocked-property, and data-only checks; un-awaited promises keep
|
||||
the await-hinting rejection in BOTH modes (deliberate — JS-parity pass-through was
|
||||
considered and skipped to preserve the nudge). The HOST boundary (final result,
|
||||
tool-call arguments, `JSON.stringify`, tool-result intake) uses the default mode and
|
||||
still serializes JSON forms (Date → ISO, RegExp/Map/Set → `{}`); host instances met on
|
||||
the preserving path are defensively wrapped into sandbox equivalents. Ripple: the
|
||||
`Object.*` helpers treat sandbox values as empty objects (`Object.keys(map)` → `[]`,
|
||||
assign sources contribute nothing, hasOwn → false — JS has no own enumerable props
|
||||
there), so interpreter internals (`.map`/`.time`/`.regex`) can never leak; the
|
||||
template-literal sandbox carve-out collapsed into `boundedData`. Object/array spread
|
||||
already preserved instances (reference copies, no checkpoint) — now tested.
|
||||
- **Console formatting**: `formatConsoleArgument` is total and deep
|
||||
(`formatConsoleValue`): numbers render via `String` (`NaN`/`Infinity`/`-Infinity`
|
||||
literally — never the JSON `null`; finite numbers match their JSON form), nested
|
||||
strings are JSON-quoted, sandbox values keep their friendly forms at ANY depth (ISO
|
||||
date, `/regex/flags`, `Map(n) [...]`, `Set(n) [...]`), opaque references become
|
||||
in-place `[CodeMode reference]` markers instead of collapsing the whole argument,
|
||||
cycles render `[Circular]` (reachable via Map/Set members, which mutation never
|
||||
checkpoints), and depth beyond `MAX_CONSOLE_DEPTH = 32` (fixed constant, not a knob)
|
||||
degrades to `…` — console can no longer fail a program. `console.table` guards with
|
||||
`containsOpaqueReference` (sandbox cells render, e.g. ISO dates) and its row/cell
|
||||
walkers treat sandbox values as scalar cells.
|
||||
- **Prose**: the instructions Syntax not-supported line dropped its `instanceof
|
||||
Error`/splice mentions (nothing else reworded); README updated (checkpoint
|
||||
preservation vs boundary serialization, error values/`instanceof`, new array/string
|
||||
methods, regex-failure behavior); `supportedSyntaxMessage` left untouched (it lists
|
||||
supported syntax, was already non-exhaustive, and stays accurate).
|
||||
- **Tests**: package suite 169 → 209 (parity: Error/instanceof + real-JS error-name
|
||||
coverage, splice/fill/copyWithin/keys/values/entries, localeCompare/normalize/trim-alias
|
||||
describes; stdlib: checkpoint survival incl. tool-arg boundary pinning, stdlib
|
||||
`instanceof`, regex-message assertions; codemode: NaN/Infinity + nested/cyclic console
|
||||
rendering, table cells, caught-tool-failure `instanceof`); adapter suites unchanged
|
||||
(34 + 16, green); both packages `tsgo --noEmit` clean.
|
||||
|
||||
**Truncation layering — CodeMode truncation off in OpenCode** (user direction; resolves the
|
||||
§4 outer-truncation item the OPPOSITE way from "kill the outer one"):
|
||||
- `maxOutputBytes` lost its 32,000 default and now behaves exactly like the other two
|
||||
limits: absent = no truncation. All three limits are uniformly no-default — budgets are
|
||||
host policy. `ResolvedExecutionLimits.maxOutputBytes` is `number | undefined`;
|
||||
`boundOutput` only runs when the host set the limit. Explicit values validate as before
|
||||
(safe integer ≥ 0).
|
||||
- OpenCode continues to pass NO limits, which now also means no CodeMode truncation.
|
||||
`execute` is a normal `Tool.define` tool, so OpenCode's native tool-output truncation
|
||||
applies with no special-casing — verified by tracing `wrap()` (`tool.ts:130-144`,
|
||||
50KB/2000-line thresholds in `truncate.ts`, full output dumped to a file under
|
||||
`tool-output/`): the `metadata.truncated` self-truncation exemption never fires for
|
||||
`execute` (its metadata never sets that key). One truncation layer, the host's — and it
|
||||
is the richer one (file dump + explore/grep hint vs an inline marker).
|
||||
- Hosts without their own output bounding set `maxOutputBytes` explicitly; README table
|
||||
and prose updated, adapter comment rewritten. Tests: codemode +1 (absent limit → 100KB
|
||||
value + 50KB log line pass through unbounded, `truncated` undefined); the adapter test
|
||||
that relied on the old default now asserts the oversized result reaches the shared
|
||||
wrapper un-truncated. Suites: 210 + 50, tsgo clean both.
|
||||
|
||||
- `maxOutputBytes` lost its 32,000 default and now behaves exactly like the other two
|
||||
limits: absent = no truncation. All three limits are uniformly no-default — budgets are
|
||||
host policy. `ResolvedExecutionLimits.maxOutputBytes` is `number | undefined`;
|
||||
`boundOutput` only runs when the host set the limit. Explicit values validate as before
|
||||
(safe integer ≥ 0).
|
||||
- OpenCode continues to pass NO limits, which now also means no CodeMode truncation.
|
||||
`execute` is a normal `Tool.define` tool, so OpenCode's native tool-output truncation
|
||||
applies with no special-casing — verified by tracing `wrap()` (`tool.ts:130-144`,
|
||||
50KB/2000-line thresholds in `truncate.ts`, full output dumped to a file under
|
||||
`tool-output/`): the `metadata.truncated` self-truncation exemption never fires for
|
||||
`execute` (its metadata never sets that key). One truncation layer, the host's — and it
|
||||
is the richer one (file dump + explore/grep hint vs an inline marker).
|
||||
- Hosts without their own output bounding set `maxOutputBytes` explicitly; README table
|
||||
and prose updated, adapter comment rewritten. Tests: codemode +1 (absent limit → 100KB
|
||||
value + 50KB log line pass through unbounded, `truncated` undefined); the adapter test
|
||||
that relied on the old default now asserts the oversized result reaches the shared
|
||||
wrapper un-truncated. Suites: 210 + 50, tsgo clean both.
|
||||
|
||||
**Docs polish** (post-API-review): stale `DiscoveryOptions` JSDoc fixed (claimed default
|
||||
4,000 and alphabetical cheapest-first — now 2,000 and round-robin, matching Fix 8/9 reality)
|
||||
|
|
@ -837,132 +856,137 @@ regular dependency; hosts depend on it themselves because the API surface is Eff
|
|||
|
||||
**Registry promotion + permission-aware catalog** (the "promote to a proper tool service"
|
||||
restructure; fixes the §4 permission-advertising bug):
|
||||
- **The adapter moved** `src/session/code-mode.ts` → `src/tool/code-mode.ts` and is now a
|
||||
registry-resident tool service on the TaskTool precedent: `CodeModeTool =
|
||||
Tool.define(CODE_MODE_TOOL, ...)` whose init depends on `MCP.Service`, `Agent.Service`,
|
||||
and `Session.Service`. It is yielded in `ToolRegistry.layer`, gated into `builtin` by
|
||||
`flags.experimentalCodeMode` (like the lsp/plan experiments), and `MCP.node` joined the
|
||||
registry's `node.deps` (`MCP.node` has no ToolRegistry dependency, so no cycle). The
|
||||
session-level special-casing in `session/tools.ts` (ad-hoc `SessionCodeMode.define` +
|
||||
append) is deleted; the early return that suppresses raw per-MCP registration when the
|
||||
flag is on stays session-side, keyed on the same flag+tool-count condition.
|
||||
- **Enablement** lives in `ToolRegistry.tools()` next to the WebSearchTool check: the MCP
|
||||
tool count is consulted once (an Effect) before the synchronous filter, and code mode
|
||||
passes the predicate iff `flags.experimentalCodeMode` && count > 0.
|
||||
- **Description split on the `describeTask` precedent**: the tool's static base
|
||||
description is a two-line summary; `describeCodeMode(agent)` in `registry.tools()`
|
||||
appends the full CodeMode instructions (workflow/rules/syntax + grouped catalog,
|
||||
`catalogInstructions` in the adapter) at the same composition point as task — so
|
||||
`plugin.trigger("tool.definition")` sees the base description first.
|
||||
- **Permission-aware catalog + dispatch** (the bug fix): the visibility predicate from
|
||||
`llm/request.ts` `resolveTools` is hoisted to `Permission.visibleTools(tools, ruleset)`
|
||||
(a record filter over `Permission.disabled` — only a hard `deny` with pattern `"*"`
|
||||
hides a tool; ask-level rules stay fully visible and prompt at call time) and
|
||||
`resolveTools` now uses it, so the two paths cannot drift. `describeCodeMode` filters
|
||||
with the merged agent+session ruleset that `SessionTools.resolve` passes into the
|
||||
registry before building the catalog/search index; `execute` rebuilds the runtime per
|
||||
execution from a fresh, filtered `mcp.tools()` snapshot using the same merged ruleset
|
||||
(`Agent.get(ctx.agent)` + `Session.get(ctx.sessionID)`, matching the merge
|
||||
`SessionTools.context` wires into `ctx.ask`) — a denied tool is not dispatchable
|
||||
even if the model guesses its name and yields the normal unknown-tool diagnostic.
|
||||
Documented gap (out of scope by design): per-message `user.tools[key] === false` arrives
|
||||
at request-prep after descriptions are built and has no child-call equivalent.
|
||||
- **Preserved behavior**: cancellation race + pre-aborted-signal guard, `toSandboxResult`
|
||||
unwrap order, attachment accumulation, `CODE_MODE_TOOL` at all title sites, no execution
|
||||
limits (native truncation only), `displayInput`, per-child `ctx.ask` gating (now wired
|
||||
through `Tool.Context` exactly like every registry tool).
|
||||
- **Explicit non-goal**: memoizing the catalog builder keyed on (ToolsChanged generation,
|
||||
permission ruleset) was considered and deliberately skipped — the per-turn rebuild is
|
||||
cheap (grouping + string rendering); revisit only if profiling shows it matters.
|
||||
- **Tests**: the two adapter suites moved to `test/tool/{code-mode,code-mode-integration}
|
||||
.test.ts` (mocked `MCP.Service`/`Agent.Service`/`Session.Service` replacing the direct
|
||||
`define(...)` construction; description assertions target `catalogInstructions`, the
|
||||
registry's composition input) and gained permission coverage: deny excluded from
|
||||
catalog/search, ask-level stays visible and callable, denied tool undispatchable
|
||||
(unknown-tool diagnostic), `Permission.visibleTools` semantics. `test/tool/
|
||||
registry.test.ts` gained four registry-level tests: registered with flag+MCP tools,
|
||||
excluded without MCP tools, excluded with flag off, and deny/ask catalog filtering
|
||||
through `registry.tools()`. Suites: 43 + 16 adapter tests, 16 registry tests, all green.
|
||||
|
||||
- **The adapter moved** `src/session/code-mode.ts` → `src/tool/code-mode.ts` and is now a
|
||||
registry-resident tool service on the TaskTool precedent: `CodeModeTool =
|
||||
Tool.define(CODE_MODE_TOOL, ...)` whose init depends on `MCP.Service`, `Agent.Service`,
|
||||
and `Session.Service`. It is yielded in `ToolRegistry.layer`, gated into `builtin` by
|
||||
`flags.experimentalCodeMode` (like the lsp/plan experiments), and `MCP.node` joined the
|
||||
registry's `node.deps` (`MCP.node` has no ToolRegistry dependency, so no cycle). The
|
||||
session-level special-casing in `session/tools.ts` (ad-hoc `SessionCodeMode.define` +
|
||||
append) is deleted; the early return that suppresses raw per-MCP registration when the
|
||||
flag is on stays session-side, keyed on the same flag+tool-count condition.
|
||||
- **Enablement** lives in `ToolRegistry.tools()` next to the WebSearchTool check: the MCP
|
||||
tool count is consulted once (an Effect) before the synchronous filter, and code mode
|
||||
passes the predicate iff `flags.experimentalCodeMode` && count > 0.
|
||||
- **Description split on the `describeTask` precedent**: the tool's static base
|
||||
description is a two-line summary; `describeCodeMode(agent)` in `registry.tools()`
|
||||
appends the full CodeMode instructions (workflow/rules/syntax + grouped catalog,
|
||||
`catalogInstructions` in the adapter) at the same composition point as task — so
|
||||
`plugin.trigger("tool.definition")` sees the base description first.
|
||||
- **Permission-aware catalog + dispatch** (the bug fix): the visibility predicate from
|
||||
`llm/request.ts` `resolveTools` is hoisted to `Permission.visibleTools(tools, ruleset)`
|
||||
(a record filter over `Permission.disabled` — only a hard `deny` with pattern `"*"`
|
||||
hides a tool; ask-level rules stay fully visible and prompt at call time) and
|
||||
`resolveTools` now uses it, so the two paths cannot drift. `describeCodeMode` filters
|
||||
with the merged agent+session ruleset that `SessionTools.resolve` passes into the
|
||||
registry before building the catalog/search index; `execute` rebuilds the runtime per
|
||||
execution from a fresh, filtered `mcp.tools()` snapshot using the same merged ruleset
|
||||
(`Agent.get(ctx.agent)` + `Session.get(ctx.sessionID)`, matching the merge
|
||||
`SessionTools.context` wires into `ctx.ask`) — a denied tool is not dispatchable
|
||||
even if the model guesses its name and yields the normal unknown-tool diagnostic.
|
||||
Documented gap (out of scope by design): per-message `user.tools[key] === false` arrives
|
||||
at request-prep after descriptions are built and has no child-call equivalent.
|
||||
- **Preserved behavior**: cancellation race + pre-aborted-signal guard, `toSandboxResult`
|
||||
unwrap order, attachment accumulation, `CODE_MODE_TOOL` at all title sites, no execution
|
||||
limits (native truncation only), `displayInput`, per-child `ctx.ask` gating (now wired
|
||||
through `Tool.Context` exactly like every registry tool).
|
||||
- **Explicit non-goal**: memoizing the catalog builder keyed on (ToolsChanged generation,
|
||||
permission ruleset) was considered and deliberately skipped — the per-turn rebuild is
|
||||
cheap (grouping + string rendering); revisit only if profiling shows it matters.
|
||||
- **Tests**: the two adapter suites moved to `test/tool/{code-mode,code-mode-integration}
|
||||
.test.ts` (mocked `MCP.Service`/`Agent.Service`/`Session.Service` replacing the direct
|
||||
`define(...)` construction; description assertions target `catalogInstructions`, the
|
||||
registry's composition input) and gained permission coverage: deny excluded from
|
||||
catalog/search, ask-level stays visible and callable, denied tool undispatchable
|
||||
(unknown-tool diagnostic), `Permission.visibleTools` semantics. `test/tool/
|
||||
registry.test.ts` gained four registry-level tests: registered with flag+MCP tools,
|
||||
excluded without MCP tools, excluded with flag off, and deny/ask catalog filtering
|
||||
through `registry.tools()`. Suites: 43 + 16 adapter tests, 16 registry tests, all green.
|
||||
|
||||
**Shared MCP invocation middle (`McpInvoke.invoke`)** (closes the §4 "plugin hooks skip
|
||||
child calls" gap):
|
||||
- `packages/opencode/src/mcp/invoke.ts` extracts the duplicated "invoke an MCP tool"
|
||||
middle into one shared `McpInvoke.invoke(input)`: plugin `tool.execute.before` hook →
|
||||
permission ask (`{ permission: key, patterns: ["*"], always: ["*"] }` via the caller's
|
||||
`ctx.ask`) → dispatch through the ai-sdk tool's execute inside the `Tool.execute`
|
||||
tracing span (`tool.name`/`tool.call_id`/`session.id`/`message.id` attributes) →
|
||||
plugin `tool.execute.after` hook. It returns the RAW result the ai-sdk execute
|
||||
resolved with; each caller keeps its own shaping edge — the legacy per-MCP loop in
|
||||
`SessionTools.resolve` applies its existing model-facing shaping/truncation, code
|
||||
mode applies `toSandboxResult`. It lives under `src/mcp/` because both callers
|
||||
already depend on MCP and the function is about invoking an MCP-backed ai-sdk tool,
|
||||
not about sessions or code mode.
|
||||
- **After-hook payload**: fired inside `McpInvoke.invoke` with the raw MCP result —
|
||||
which is exactly what the legacy loop always passed (the raw `CallToolResult`, not
|
||||
the shaped `{title, output, metadata}`), so legacy behavior is preserved bit-for-bit
|
||||
and the hook payload cannot drift between callers. No callback/edge-firing design
|
||||
was needed.
|
||||
- **Synthetic child callID**: code-mode child calls pass `${parentCallID}/${n}` as the
|
||||
hook/span callID (`parentCallID` = the `execute` call's `ctx.callID`, falling back to
|
||||
the entry key; `n` = per-execution counter starting at 1, shared across all child
|
||||
calls in one program). callID is an opaque string — nothing parses it. The ai-sdk
|
||||
`toolCallId` (`options.toolCallId`) stays each caller's existing value
|
||||
(`ctx.callID ?? entry.key` for code mode).
|
||||
- **Child-scoped hook failures**: `CodeModeTool` (which now also yields
|
||||
`Plugin.Service`) wraps the whole child call — hooks, ask, dispatch — in
|
||||
`toCatchable` (the generalization of the old `askPermission` catchCause), so a plugin
|
||||
hook failure fails ONLY that child call as a catchable in-program `toolError`; other
|
||||
calls in the same program keep running and interruption still propagates as
|
||||
interruption. Legacy semantics unchanged: a hook failure fails the tool call.
|
||||
- **Tests**: `test/tool/code-mode.test.ts` +2 (child calls fire before/after with the
|
||||
MCP key and `parent/1`, `parent/2` ids, after hook carries the raw MCP result; a
|
||||
failing before hook is caught in-program, gates dispatch, and leaves the outer
|
||||
execute ok) — both code-mode harnesses gained a `Plugin.Service` mock (pass-through
|
||||
trigger by default, overridable). New `test/session/tools.test.ts` (3 tests) pins
|
||||
`SessionTools.resolve` at the real-registry seam (LayerNode.compile, fake MCP layer):
|
||||
flag on + MCP tools → `execute` present, raw MCP keys suppressed; flag off → raw
|
||||
keys present, `execute` absent; and the legacy raw-MCP execute fires before/after
|
||||
hooks keyed by the ai-sdk toolCallId with the raw result payload. Suites: adapter
|
||||
45 + 16, session/tool/permission all green; this package untouched (211 pass).
|
||||
|
||||
- `packages/opencode/src/mcp/invoke.ts` extracts the duplicated "invoke an MCP tool"
|
||||
middle into one shared `McpInvoke.invoke(input)`: plugin `tool.execute.before` hook →
|
||||
permission ask (`{ permission: key, patterns: ["*"], always: ["*"] }` via the caller's
|
||||
`ctx.ask`) → dispatch through the ai-sdk tool's execute inside the `Tool.execute`
|
||||
tracing span (`tool.name`/`tool.call_id`/`session.id`/`message.id` attributes) →
|
||||
plugin `tool.execute.after` hook. It returns the RAW result the ai-sdk execute
|
||||
resolved with; each caller keeps its own shaping edge — the legacy per-MCP loop in
|
||||
`SessionTools.resolve` applies its existing model-facing shaping/truncation, code
|
||||
mode applies `toSandboxResult`. It lives under `src/mcp/` because both callers
|
||||
already depend on MCP and the function is about invoking an MCP-backed ai-sdk tool,
|
||||
not about sessions or code mode.
|
||||
- **After-hook payload**: fired inside `McpInvoke.invoke` with the raw MCP result —
|
||||
which is exactly what the legacy loop always passed (the raw `CallToolResult`, not
|
||||
the shaped `{title, output, metadata}`), so legacy behavior is preserved bit-for-bit
|
||||
and the hook payload cannot drift between callers. No callback/edge-firing design
|
||||
was needed.
|
||||
- **Synthetic child callID**: code-mode child calls pass `${parentCallID}/${n}` as the
|
||||
hook/span callID (`parentCallID` = the `execute` call's `ctx.callID`, falling back to
|
||||
the entry key; `n` = per-execution counter starting at 1, shared across all child
|
||||
calls in one program). callID is an opaque string — nothing parses it. The ai-sdk
|
||||
`toolCallId` (`options.toolCallId`) stays each caller's existing value
|
||||
(`ctx.callID ?? entry.key` for code mode).
|
||||
- **Child-scoped hook failures**: `CodeModeTool` (which now also yields
|
||||
`Plugin.Service`) wraps the whole child call — hooks, ask, dispatch — in
|
||||
`toCatchable` (the generalization of the old `askPermission` catchCause), so a plugin
|
||||
hook failure fails ONLY that child call as a catchable in-program `toolError`; other
|
||||
calls in the same program keep running and interruption still propagates as
|
||||
interruption. Legacy semantics unchanged: a hook failure fails the tool call.
|
||||
- **Tests**: `test/tool/code-mode.test.ts` +2 (child calls fire before/after with the
|
||||
MCP key and `parent/1`, `parent/2` ids, after hook carries the raw MCP result; a
|
||||
failing before hook is caught in-program, gates dispatch, and leaves the outer
|
||||
execute ok) — both code-mode harnesses gained a `Plugin.Service` mock (pass-through
|
||||
trigger by default, overridable). New `test/session/tools.test.ts` (3 tests) pins
|
||||
`SessionTools.resolve` at the real-registry seam (LayerNode.compile, fake MCP layer):
|
||||
flag on + MCP tools → `execute` present, raw MCP keys suppressed; flag off → raw
|
||||
keys present, `execute` absent; and the legacy raw-MCP execute fires before/after
|
||||
hooks keyed by the ai-sdk toolCallId with the raw result payload. Suites: adapter
|
||||
45 + 16, session/tool/permission all green; this package untouched (211 pass).
|
||||
|
||||
**Signature rendering + compound-assignment parity fixes** (externally reported, both
|
||||
verified real with failing tests before fixing):
|
||||
- **Non-identifier property names in rendered signatures** (`src/tool.ts`): `renderSchema`
|
||||
emitted raw property names, so schema properties like `foo-bar`/`@type`/`x.y`/`123`
|
||||
rendered invalid TypeScript (`{ foo-bar?: string }`). Fixed with a `renderKey` helper —
|
||||
bare identifiers stay bare, everything else is `JSON.stringify`-quoted — applied in the
|
||||
single `field` closure both the compact and pretty renderings share. The
|
||||
`identifierSegment` regex now lives in `tool.ts` (exported) and `tool-runtime.ts`'s
|
||||
bracket-notation `toolExpression` imports it: one source of truth for "is this a bare
|
||||
identifier" across object keys and tool paths. Tests: `signature.test.ts` +4 (compact,
|
||||
pretty with JSDoc on a quoted key, JSON Schema input+output, Effect Schema struct).
|
||||
- **Numeric schema unions keep their real alternatives** (`src/tool.ts`): the old
|
||||
`anyOf`/`oneOf` renderer collapsed any union containing `{ type: "number" }` to just
|
||||
`number`, dropping real JSON Schema alternatives (`string | number`, `number | null`,
|
||||
etc.). The collapse is now restricted to Effect's number-schema artifact
|
||||
(`number | "NaN" | "Infinity" | "-Infinity"`, emitted as single-value string enums),
|
||||
while raw JSON Schema unions render every branch. Tests: `signature.test.ts` +3.
|
||||
- **Compound assignment now matches binary-operator semantics** (`src/codemode.ts`):
|
||||
`applyCompoundAssignment` did raw JS ops on interpreter wrapper objects, so `x += y`
|
||||
diverged from `x = x + y` (sandbox Date `d += 1` produced `"[object Object]1"`;
|
||||
`d -= 400` gave `NaN` instead of epoch arithmetic). The operator table + coercion moved
|
||||
verbatim out of `evaluateBinaryExpression` into a shared `applyBinaryOperator`;
|
||||
compound assignment validates against a `compoundOperators` set (`+=` … `>>>=`) and
|
||||
dispatches through it (`operator.slice(0, -1)`). Logical assignments (`&&=`/`||=`/`??=`)
|
||||
keep their separate short-circuit path (`evaluateLogicalAssignment`), and both
|
||||
assignment call sites still wrap results in `boundedData`. Deliberate side effect:
|
||||
compound assignment now rejects opaque references, consistent with binary operators.
|
||||
Tests: `parity.test.ts` +5 (Date `+=` concat parity, Date `-=`/`/=` epoch parity,
|
||||
string `+=` object/array, member-target compound, 13-case operator sweep vs real JS).
|
||||
Package suite: 220 pass.
|
||||
|
||||
- **Non-identifier property names in rendered signatures** (`src/tool.ts`): `renderSchema`
|
||||
emitted raw property names, so schema properties like `foo-bar`/`@type`/`x.y`/`123`
|
||||
rendered invalid TypeScript (`{ foo-bar?: string }`). Fixed with a `renderKey` helper —
|
||||
bare identifiers stay bare, everything else is `JSON.stringify`-quoted — applied in the
|
||||
single `field` closure both the compact and pretty renderings share. The
|
||||
`identifierSegment` regex now lives in `tool.ts` (exported) and `tool-runtime.ts`'s
|
||||
bracket-notation `toolExpression` imports it: one source of truth for "is this a bare
|
||||
identifier" across object keys and tool paths. Tests: `signature.test.ts` +4 (compact,
|
||||
pretty with JSDoc on a quoted key, JSON Schema input+output, Effect Schema struct).
|
||||
- **Numeric schema unions keep their real alternatives** (`src/tool.ts`): the old
|
||||
`anyOf`/`oneOf` renderer collapsed any union containing `{ type: "number" }` to just
|
||||
`number`, dropping real JSON Schema alternatives (`string | number`, `number | null`,
|
||||
etc.). The collapse is now restricted to Effect's number-schema artifact
|
||||
(`number | "NaN" | "Infinity" | "-Infinity"`, emitted as single-value string enums),
|
||||
while raw JSON Schema unions render every branch. Tests: `signature.test.ts` +3.
|
||||
- **Compound assignment now matches binary-operator semantics** (`src/codemode.ts`):
|
||||
`applyCompoundAssignment` did raw JS ops on interpreter wrapper objects, so `x += y`
|
||||
diverged from `x = x + y` (sandbox Date `d += 1` produced `"[object Object]1"`;
|
||||
`d -= 400` gave `NaN` instead of epoch arithmetic). The operator table + coercion moved
|
||||
verbatim out of `evaluateBinaryExpression` into a shared `applyBinaryOperator`;
|
||||
compound assignment validates against a `compoundOperators` set (`+=` … `>>>=`) and
|
||||
dispatches through it (`operator.slice(0, -1)`). Logical assignments (`&&=`/`||=`/`??=`)
|
||||
keep their separate short-circuit path (`evaluateLogicalAssignment`), and both
|
||||
assignment call sites still wrap results in `boundedData`. Deliberate side effect:
|
||||
compound assignment now rejects opaque references, consistent with binary operators.
|
||||
Tests: `parity.test.ts` +5 (Date `+=` concat parity, Date `-=`/`/=` epoch parity,
|
||||
string `+=` object/array, member-target compound, 13-case operator sweep vs real JS).
|
||||
Package suite: 220 pass.
|
||||
|
||||
---
|
||||
|
||||
## 4. Remaining work (detailed TODO)
|
||||
|
||||
### Next DSL-expansion pass (done — see the DSL-expansion pass entry in §3)
|
||||
|
||||
Batch these together — per user direction: important, but deliberately deferred to one
|
||||
focused interpreter-surface pass rather than picked off piecemeal.
|
||||
|
||||
- [x] Medium-tier JS parity items deferred from the original audit: caught errors are plain
|
||||
`{ name, message }` objects, not `instanceof Error` (and `Error` isn't a value —
|
||||
`x instanceof Error` is unsupported syntax); `splice` (still a
|
||||
|
|
@ -981,6 +1005,7 @@ focused interpreter-surface pass rather than picked off piecemeal.
|
|||
(`console.log({ m: map })`) — could deep-format instead.
|
||||
|
||||
### Next iteration: text-result handling (deliberate follow-up, user-directed)
|
||||
|
||||
- [ ] Revisit how MCP text results reach the program. Today: `structuredContent` when the
|
||||
server sends it, else joined text as a plain string (the program JSON.parses it,
|
||||
guided by a workflow step). Considered and deferred: (a) conservative boundary
|
||||
|
|
@ -992,6 +1017,7 @@ focused interpreter-surface pass rather than picked off piecemeal.
|
|||
revisit once real usage shows which failure modes matter.
|
||||
|
||||
### Next iteration: stdlib surface (prioritized)
|
||||
|
||||
Current instructions say "usual Array/String/Object/Math/JSON methods," but the interpreter is
|
||||
intentionally a subset. Keep CodeMode focused on orchestration and data shaping, not a full host
|
||||
runtime, but close the high-friction gaps models are likely to reach for.
|
||||
|
|
@ -1028,7 +1054,9 @@ Explicit non-goals for now: `structuredClone`, `WeakMap`/`WeakSet`, and timers
|
|||
orchestration use case.
|
||||
|
||||
### Wiring-review findings (subagent code review of the OpenCode integration, triaged)
|
||||
|
||||
Pre-PR fixes (user-approved cut):
|
||||
|
||||
- [x] **Cancellation does not interrupt the interpreter** — the no-limits rationale claimed
|
||||
"user cancel interrupts the execution fiber," but `tools.ts` runs tools via
|
||||
`run.promise` → `Effect.runPromise` (`effect/bridge.ts:64-66`) with NO abort wiring;
|
||||
|
|
@ -1073,6 +1101,7 @@ Pre-PR fixes (user-approved cut):
|
|||
`title: "execute"` sites in `code-mode.ts` now reference `CODE_MODE_TOOL`.
|
||||
|
||||
Post-MVP (logged, not blocking an experimental flag):
|
||||
|
||||
- [x] **Plugin `tool.execute.before/after` hooks skip child calls** — legacy MCP
|
||||
registration fires them per tool (`tools.ts:419-441`); under code mode only the
|
||||
outer `execute` fires them, so auditing/intercepting plugins silently lose MCP
|
||||
|
|
@ -1105,6 +1134,7 @@ Post-MVP (logged, not blocking an experimental flag):
|
|||
names that are no longer directly callable under code mode.
|
||||
|
||||
### Backlog / loose ends (non-blocking, any order)
|
||||
|
||||
- [ ] `evaluateUpdateExpression` (`++`/`--`) still uses raw `Number(current)`, so `d++` on a
|
||||
sandbox Date yields `NaN` where `d += 1` now uses epoch semantics (and real JS `d++`
|
||||
would give epoch+0 numeric). Pre-existing, out of scope of the compound-assignment
|
||||
|
|
@ -1141,7 +1171,7 @@ Post-MVP (logged, not blocking an experimental flag):
|
|||
## 5. Context and gotchas for whoever picks this up
|
||||
|
||||
- **Motivating failure (why forgiving semantics + prompting matter):** in a real transcript,
|
||||
the model wrote `me.result?.login ?? me.result` where the tool result was a JSON *string* —
|
||||
the model wrote `me.result?.login ?? me.result` where the tool result was a JSON _string_ —
|
||||
the old strict interpreter threw (`String property 'login' is not available`); then the
|
||||
model returned a raw 105KB payload, which native truncation dumped to a file, costing a
|
||||
subagent round-trip to extract one number. Interpreter forgiveness stops the crashes;
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue