@kolisachint/hoocode-agent 0.5.49 → 0.5.51

package/dist/core/tools/read-dedup.d.ts

@@ -38,8 +38,19 @@ export declare function rangesOverlap(a: ReadRange, b: ReadRange): boolean;
 export declare const DEDUP_POINTER_PREFIX = "[Already in context:";
 /** Whether a tool-result text is an at-call dedup pointer. */
 export declare function isDedupPointerText(text: string): boolean;
+/** Remember the stamp of the file a read delivered. */
+export declare function recordReadStamp(callId: string, stamp: string): void;
+/**
+ * Whether the file a read delivered still stamps the same. False when no stamp
+ * was recorded, so an unprovable case re-reads rather than asserting freshness.
+ */
+export declare function readStampMatches(callId: string, stamp: string): boolean;
+/** Test seam: drop all recorded stamps. */
+export declare function clearReadStamps(): void;
 /** A covering earlier read, described for the pointer message. */
 export interface CoveringRead {
+    /** Tool call id of the read being pointed at, used to check its stamp. */
+    callId: string;
     /** The path as the earlier read spelled it (for a friendly pointer). */
     display: string;
     /** Delivered range start (1-indexed line). */
@@ -71,6 +82,8 @@ export interface FindCoveringReadOptions {
  * the candidate and the supersession side.
  */
 export declare function findCoveringRead(entries: readonly unknown[], opts: FindCoveringReadOptions): CoveringRead | null;
+/** The parts of a covering read the pointer message renders. */
+export type CoveringReadDisplay = Pick<CoveringRead, "display" | "start" | "end">;
 /** Build the pointer text returned in place of a re-fetch. */
-export declare function buildDedupPointerText(covering: CoveringRead): string;
+export declare function buildDedupPointerText(covering: CoveringReadDisplay): string;
 //# sourceMappingURL=read-dedup.d.ts.map

package/dist/core/tools/read-dedup.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"read-dedup.d.ts","sourceRoot":"","sources":["../../../src/core/tools/read-dedup.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAEH,mFAAmF;AACnF,MAAM,WAAW,SAAS;IACzB,KAAK,EAAE,MAAM,CAAC;IACd,GAAG,EAAE,MAAM,CAAC;CACZ;AAED,sEAAsE;AACtE,eAAO,MAAM,gBAAgB,EAAE,SAAuD,CAAC;AAEvF;;;;GAIG;AACH,wBAAgB,iBAAiB,CAAC,IAAI,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,SAAS,GAAG,SAAS,CAStF;AAED,mDAAmD;AACnD,wBAAgB,aAAa,CAAC,CAAC,EAAE,SAAS,EAAE,CAAC,EAAE,SAAS,GAAG,OAAO,CAEjE;AAOD;;;;GAIG;AACH,eAAO,MAAM,oBAAoB,yBAAyB,CAAC;AAE3D,8DAA8D;AAC9D,wBAAgB,kBAAkB,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAExD;AAED,kEAAkE;AAClE,MAAM,WAAW,YAAY;IAC5B,wEAAwE;IACxE,OAAO,EAAE,MAAM,CAAC;IAChB,8CAA8C;IAC9C,KAAK,EAAE,MAAM,CAAC;IACd,uEAAuE;IACvE,GAAG,EAAE,MAAM,CAAC;CACZ;AAED,MAAM,WAAW,uBAAuB;IACvC,kDAAkD;IAClD,YAAY,EAAE,MAAM,CAAC;IACrB,4CAA4C;IAC5C,cAAc,EAAE,SAAS,CAAC;IAC1B,mFAAmF;IACnF,aAAa,EAAE,MAAM,CAAC;IACtB,mFAAmF;IACnF,WAAW,EAAE,CAAC,OAAO,EAAE,MAAM,KAAK,MAAM,CAAC;CACzC;AAoED;;;;;;;;;;;;GAYG;AACH,wBAAgB,gBAAgB,CAAC,OAAO,EAAE,SAAS,OAAO,EAAE,EAAE,IAAI,EAAE,uBAAuB,GAAG,YAAY,GAAG,IAAI,CA8FhH;AAED,8DAA8D;AAC9D,wBAAgB,qBAAqB,CAAC,QAAQ,EAAE,YAAY,GAAG,MAAM,CAQpE","sourcesContent":["/**\n * Read de-duplication primitives.\n *\n * Two mechanisms share this module:\n *\n * - The post-hoc context GC (`context-gc.ts`), which stubs out a `read` result\n *   once a later overlapping read or an edit/write has superseded it.\n * - The at-call-time guard in the `read` tool, which short-circuits a read whose\n *   requested range is *already fully covered* by an earlier, still-live read in\n *   the current session, returning a pointer instead of re-fetching the file.\n *\n * Both reason about half-open line ranges `[start, end)` (end exclusive), so the\n * range math lives here and cannot drift between them. The guard additionally\n * needs to (a) recognise its own pointer results so they are never treated as\n * content-bearing reads, and (b) walk the session branch to find a covering\n * read — both of which are defined here to keep the tool file lean.\n */\n\n/** Half-open line interval `[start, end)` a read call covers; end is exclusive. */\nexport interface ReadRange {\n\tstart: number;\n\tend: number;\n}\n\n/** A read with no offset/limit covers the whole file (open-ended). */\nexport const WHOLE_FILE_RANGE: ReadRange = { start: 1, end: Number.POSITIVE_INFINITY };\n\n/**\n * Derive the line range a read call covers from its `offset`/`limit` args.\n * Missing offset means \"from line 1\"; missing limit means \"to end of file\"\n * (open-ended, so it overlaps any later read of the same file).\n */\nexport function readRangeFromArgs(args: Record<string, unknown> | undefined): ReadRange {\n\tconst offsetRaw = args?.offset;\n\tconst limitRaw = args?.limit;\n\tconst offset = typeof offsetRaw === \"number\" && Number.isFinite(offsetRaw) && offsetRaw > 0 ? offsetRaw : 1;\n\tconst end =\n\t\ttypeof limitRaw === \"number\" && Number.isFinite(limitRaw)\n\t\t\t? offset + Math.max(0, limitRaw)\n\t\t\t: Number.POSITIVE_INFINITY;\n\treturn { start: offset, end };\n}\n\n/** Whether two half-open line ranges intersect. */\nexport function rangesOverlap(a: ReadRange, b: ReadRange): boolean {\n\treturn a.start < b.end && b.start < a.end;\n}\n\n/** Whether `outer` fully contains `inner` (every line of inner lies within outer). */\nfunction rangeContains(outer: ReadRange, inner: ReadRange): boolean {\n\treturn outer.start <= inner.start && outer.end >= inner.end;\n}\n\n/**\n * Marker prefix for the at-call dedup pointer. Kept stable so the context GC can\n * recognise a pointer result and exclude it from supersession bookkeeping (a\n * pointer fetched no content, so it must not stub the read it points at).\n */\nexport const DEDUP_POINTER_PREFIX = \"[Already in context:\";\n\n/** Whether a tool-result text is an at-call dedup pointer. */\nexport function isDedupPointerText(text: string): boolean {\n\treturn text.trimStart().startsWith(DEDUP_POINTER_PREFIX);\n}\n\n/** A covering earlier read, described for the pointer message. */\nexport interface CoveringRead {\n\t/** The path as the earlier read spelled it (for a friendly pointer). */\n\tdisplay: string;\n\t/** Delivered range start (1-indexed line). */\n\tstart: number;\n\t/** Delivered range end (exclusive; Infinity for a whole-file read). */\n\tend: number;\n}\n\nexport interface FindCoveringReadOptions {\n\t/** Resolved absolute path of the current read. */\n\tresolvedPath: string;\n\t/** Range the current read is asking for. */\n\trequestedRange: ReadRange;\n\t/** Tool call id of the current read, excluded from candidate/supersession sets. */\n\tcurrentCallId: string;\n\t/** Resolve a raw read-arg path the same way the current read resolved its path. */\n\tresolvePath: (rawPath: string) => string;\n}\n\ninterface ContentBlock {\n\ttype?: string;\n\tid?: string;\n\tname?: string;\n\targuments?: Record<string, unknown>;\n\ttext?: string;\n}\n\ninterface MessageLike {\n\trole?: string;\n\tcontent?: ContentBlock[];\n\ttoolCallId?: string;\n\ttoolName?: string;\n\tisError?: boolean;\n}\n\nconst READ_TOOL = \"read\";\nconst MUTATE_TOOLS = new Set([\"edit\", \"write\"]);\n\n/** A session compaction entry — the boundary that trims the live context. */\nfunction isCompactionEntry(entry: unknown): boolean {\n\treturn !!entry && typeof entry === \"object\" && (entry as { type?: unknown }).type === \"compaction\";\n}\n\n/** Accept either raw `AgentMessage`s or session entries wrapping `.message`. */\nfunction toMessage(entry: unknown): MessageLike | null {\n\tif (!entry || typeof entry !== \"object\") return null;\n\tconst e = entry as { message?: unknown; role?: unknown };\n\tif (e.message && typeof e.message === \"object\") return e.message as MessageLike;\n\tif (typeof e.role === \"string\") return e as MessageLike;\n\treturn null;\n}\n\nfunction resultText(m: MessageLike): string {\n\treturn (m.content ?? [])\n\t\t.filter((c) => c.type === \"text\")\n\t\t.map((c) => c.text ?? \"\")\n\t\t.join(\"\");\n}\n\n/**\n * Recover the range a read result actually *delivered* from its text.\n *\n * A cap-truncated read announces `[Showing lines A-B of N ...]`, so it delivered\n * only `[A, B+1)` even though its args declared a wider range. A read whose first\n * line alone exceeded the byte cap delivered nothing. Any other (untruncated)\n * read delivered its full declared range. The user-`limit` early-stop notice\n * (`[N more lines in file ...]`) is *not* a cap truncation — the declared range\n * was delivered in full — so it falls through to `declared`.\n */\nfunction deliveredRange(text: string, declared: ReadRange): ReadRange {\n\t// The truncation notice is always the trailing `\\n\\n[Showing lines A-B of N ...]`\n\t// clause the read tool appends. Anchor to the end so a `[Showing lines ...]`\n\t// string that merely appears *inside* the file's content can't spoof it.\n\tconst showing = text.match(/\\n\\n\\[Showing lines (\\d+)-(\\d+) of \\d+[^\\]]*\\]\\s*$/);\n\tif (showing) {\n\t\tconst a = Number(showing[1]);\n\t\tconst b = Number(showing[2]);\n\t\tif (Number.isFinite(a) && Number.isFinite(b) && b >= a) return { start: a, end: b + 1 };\n\t}\n\t// First line alone exceeded the byte limit: the whole result *is* that notice,\n\t// so it must start the text. Nothing usable was delivered.\n\tif (/^\\[Line \\d+ is .+ exceeds .+ limit\\./.test(text)) return { start: 1, end: 1 };\n\treturn declared;\n}\n\n/**\n * Find the latest earlier read that (a) is for the same resolved path, (b)\n * actually delivered a range containing the requested range, and (c) is still\n * live in the outgoing context — i.e. the post-hoc GC will not have stubbed it,\n * because no later edit/write and no later overlapping content read supersede\n * it. Returns that read's delivered range for the pointer, or null when the\n * current read must actually run.\n *\n * Deliberately conservative: a truncated earlier read only covers what it\n * delivered, a whole-file read must have been delivered untruncated to count as\n * covering, and any pointer results (which fetched nothing) are ignored on both\n * the candidate and the supersession side.\n */\nexport function findCoveringRead(entries: readonly unknown[], opts: FindCoveringReadOptions): CoveringRead | null {\n\t// `declared` mirrors the range the GC uses for supersession (straight from the\n\t// call args); `delivered` is what the result text shows was actually returned,\n\t// used for coverage. The current call has no result yet, so it never appears.\n\tinterface PriorRead {\n\t\tindex: number;\n\t\tdeclared: ReadRange;\n\t\tdelivered: ReadRange;\n\t\tdisplay: string;\n\t}\n\n\t// Resolve each distinct raw path once — the resolver may hit the filesystem.\n\tconst resolveCache = new Map<string, string>();\n\tconst resolvePath = (raw: string): string => {\n\t\tconst hit = resolveCache.get(raw);\n\t\tif (hit !== undefined) return hit;\n\t\tconst resolved = opts.resolvePath(raw);\n\t\tresolveCache.set(raw, resolved);\n\t\treturn resolved;\n\t};\n\n\t// Single ordered pass. `readCall`/`mutateCallPath` map a call id to its path as\n\t// the call is seen (a toolCall always precedes its result); `reads` collects\n\t// the target path's reads and `lastMutateIndex` its last mutate. `order` gives\n\t// live-context position for the supersession/mutate comparisons.\n\tconst readCall = new Map<string, { resolved: string; display: string; declared: ReadRange }>();\n\tconst mutateCallPath = new Map<string, string>();\n\tconst reads: PriorRead[] = [];\n\tlet lastMutateIndex = -1;\n\tlet order = 0;\n\n\tfor (const entry of entries) {\n\t\t// Compaction boundary: everything before it is replaced by a summary in the\n\t\t// live context, so drop the state accumulated so far. Conservative — the\n\t\t// kept tail before the boundary is dropped too — which can only miss a\n\t\t// dedup, never point at content that is no longer in context.\n\t\tif (isCompactionEntry(entry)) {\n\t\t\treadCall.clear();\n\t\t\tmutateCallPath.clear();\n\t\t\treads.length = 0;\n\t\t\tlastMutateIndex = -1;\n\t\t\torder = 0;\n\t\t\tcontinue;\n\t\t}\n\t\tconst m = toMessage(entry);\n\t\tif (!m) continue;\n\t\tconst i = order++;\n\t\tif (m.role === \"assistant\" && Array.isArray(m.content)) {\n\t\t\tfor (const b of m.content) {\n\t\t\t\tif (b.type !== \"toolCall\" || !b.id) continue;\n\t\t\t\tconst raw = b.arguments?.path;\n\t\t\t\tif (typeof raw !== \"string\" || raw.length === 0) continue;\n\t\t\t\tconst resolved = resolvePath(raw);\n\t\t\t\tif (b.name === READ_TOOL) {\n\t\t\t\t\treadCall.set(b.id, { resolved, display: raw, declared: readRangeFromArgs(b.arguments) });\n\t\t\t\t} else if (b.name && MUTATE_TOOLS.has(b.name)) {\n\t\t\t\t\tmutateCallPath.set(b.id, resolved);\n\t\t\t\t}\n\t\t\t}\n\t\t} else if (m.role === \"toolResult\" && !m.isError && m.toolCallId) {\n\t\t\tif (m.toolName === READ_TOOL) {\n\t\t\t\tif (m.toolCallId === opts.currentCallId) continue;\n\t\t\t\tconst info = readCall.get(m.toolCallId);\n\t\t\t\tif (!info || info.resolved !== opts.resolvedPath) continue;\n\t\t\t\tconst text = resultText(m);\n\t\t\t\t// A pointer fetched nothing: the GC excludes it from supersession, so we\n\t\t\t\t// must too (both as a candidate and as a superseder).\n\t\t\t\tif (isDedupPointerText(text)) continue;\n\t\t\t\treads.push({\n\t\t\t\t\tindex: i,\n\t\t\t\t\tdeclared: info.declared,\n\t\t\t\t\tdelivered: deliveredRange(text, info.declared),\n\t\t\t\t\tdisplay: info.display,\n\t\t\t\t});\n\t\t\t} else if (m.toolName && MUTATE_TOOLS.has(m.toolName)) {\n\t\t\t\tif (mutateCallPath.get(m.toolCallId) === opts.resolvedPath) lastMutateIndex = i;\n\t\t\t}\n\t\t}\n\t}\n\n\t// A read survives the GC iff no later mutate and no later read overlaps its\n\t// *declared* range — exactly the GC's own test — so predict it the same way.\n\tconst survivesGc = (r: PriorRead): boolean =>\n\t\tlastMutateIndex <= r.index && !reads.some((o) => o.index > r.index && rangesOverlap(o.declared, r.declared));\n\n\tlet best: PriorRead | null = null;\n\tfor (const r of reads) {\n\t\t// Coverage uses the *delivered* range: a truncated read only holds what it returned.\n\t\tif (!rangeContains(r.delivered, opts.requestedRange)) continue;\n\t\tif (!survivesGc(r)) continue;\n\t\tif (!best || r.index > best.index) best = r;\n\t}\n\tif (!best) return null;\n\treturn { display: best.display, start: best.delivered.start, end: best.delivered.end };\n}\n\n/** Build the pointer text returned in place of a re-fetch. */\nexport function buildDedupPointerText(covering: CoveringRead): string {\n\tconst where =\n\t\tcovering.end === Number.POSITIVE_INFINITY\n\t\t\t? \"the entire file\"\n\t\t\t: covering.end - 1 > covering.start\n\t\t\t\t? `lines ${covering.start}-${covering.end - 1}`\n\t\t\t\t: `line ${covering.start}`;\n\treturn `${DEDUP_POINTER_PREFIX} ${covering.display} (${where}) was already read earlier in this session and has not changed since. Not re-fetched to save tokens — pass a different offset/limit, or edit the file, if you need other or newer content.]`;\n}\n"]}
+{"version":3,"file":"read-dedup.d.ts","sourceRoot":"","sources":["../../../src/core/tools/read-dedup.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAEH,mFAAmF;AACnF,MAAM,WAAW,SAAS;IACzB,KAAK,EAAE,MAAM,CAAC;IACd,GAAG,EAAE,MAAM,CAAC;CACZ;AAED,sEAAsE;AACtE,eAAO,MAAM,gBAAgB,EAAE,SAAuD,CAAC;AAEvF;;;;GAIG;AACH,wBAAgB,iBAAiB,CAAC,IAAI,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,SAAS,GAAG,SAAS,CAStF;AAED,mDAAmD;AACnD,wBAAgB,aAAa,CAAC,CAAC,EAAE,SAAS,EAAE,CAAC,EAAE,SAAS,GAAG,OAAO,CAEjE;AAOD;;;;GAIG;AACH,eAAO,MAAM,oBAAoB,yBAAyB,CAAC;AAE3D,8DAA8D;AAC9D,wBAAgB,kBAAkB,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAExD;AAkBD,uDAAuD;AACvD,wBAAgB,eAAe,CAAC,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,GAAG,IAAI,CAMnE;AAED;;;GAGG;AACH,wBAAgB,gBAAgB,CAAC,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,GAAG,OAAO,CAGvE;AAED,2CAA2C;AAC3C,wBAAgB,eAAe,IAAI,IAAI,CAEtC;AAED,kEAAkE;AAClE,MAAM,WAAW,YAAY;IAC5B,0EAA0E;IAC1E,MAAM,EAAE,MAAM,CAAC;IACf,wEAAwE;IACxE,OAAO,EAAE,MAAM,CAAC;IAChB,8CAA8C;IAC9C,KAAK,EAAE,MAAM,CAAC;IACd,uEAAuE;IACvE,GAAG,EAAE,MAAM,CAAC;CACZ;AAED,MAAM,WAAW,uBAAuB;IACvC,kDAAkD;IAClD,YAAY,EAAE,MAAM,CAAC;IACrB,4CAA4C;IAC5C,cAAc,EAAE,SAAS,CAAC;IAC1B,mFAAmF;IACnF,aAAa,EAAE,MAAM,CAAC;IACtB,mFAAmF;IACnF,WAAW,EAAE,CAAC,OAAO,EAAE,MAAM,KAAK,MAAM,CAAC;CACzC;AAoED;;;;;;;;;;;;GAYG;AACH,wBAAgB,gBAAgB,CAAC,OAAO,EAAE,SAAS,OAAO,EAAE,EAAE,IAAI,EAAE,uBAAuB,GAAG,YAAY,GAAG,IAAI,CA8GhH;AAED,gEAAgE;AAChE,MAAM,MAAM,mBAAmB,GAAG,IAAI,CAAC,YAAY,EAAE,SAAS,GAAG,OAAO,GAAG,KAAK,CAAC,CAAC;AAElF,8DAA8D;AAC9D,wBAAgB,qBAAqB,CAAC,QAAQ,EAAE,mBAAmB,GAAG,MAAM,CAQ3E","sourcesContent":["/**\n * Read de-duplication primitives.\n *\n * Two mechanisms share this module:\n *\n * - The post-hoc context GC (`context-gc.ts`), which stubs out a `read` result\n *   once a later overlapping read or an edit/write has superseded it.\n * - The at-call-time guard in the `read` tool, which short-circuits a read whose\n *   requested range is *already fully covered* by an earlier, still-live read in\n *   the current session, returning a pointer instead of re-fetching the file.\n *\n * Both reason about half-open line ranges `[start, end)` (end exclusive), so the\n * range math lives here and cannot drift between them. The guard additionally\n * needs to (a) recognise its own pointer results so they are never treated as\n * content-bearing reads, and (b) walk the session branch to find a covering\n * read — both of which are defined here to keep the tool file lean.\n */\n\n/** Half-open line interval `[start, end)` a read call covers; end is exclusive. */\nexport interface ReadRange {\n\tstart: number;\n\tend: number;\n}\n\n/** A read with no offset/limit covers the whole file (open-ended). */\nexport const WHOLE_FILE_RANGE: ReadRange = { start: 1, end: Number.POSITIVE_INFINITY };\n\n/**\n * Derive the line range a read call covers from its `offset`/`limit` args.\n * Missing offset means \"from line 1\"; missing limit means \"to end of file\"\n * (open-ended, so it overlaps any later read of the same file).\n */\nexport function readRangeFromArgs(args: Record<string, unknown> | undefined): ReadRange {\n\tconst offsetRaw = args?.offset;\n\tconst limitRaw = args?.limit;\n\tconst offset = typeof offsetRaw === \"number\" && Number.isFinite(offsetRaw) && offsetRaw > 0 ? offsetRaw : 1;\n\tconst end =\n\t\ttypeof limitRaw === \"number\" && Number.isFinite(limitRaw)\n\t\t\t? offset + Math.max(0, limitRaw)\n\t\t\t: Number.POSITIVE_INFINITY;\n\treturn { start: offset, end };\n}\n\n/** Whether two half-open line ranges intersect. */\nexport function rangesOverlap(a: ReadRange, b: ReadRange): boolean {\n\treturn a.start < b.end && b.start < a.end;\n}\n\n/** Whether `outer` fully contains `inner` (every line of inner lies within outer). */\nfunction rangeContains(outer: ReadRange, inner: ReadRange): boolean {\n\treturn outer.start <= inner.start && outer.end >= inner.end;\n}\n\n/**\n * Marker prefix for the at-call dedup pointer. Kept stable so the context GC can\n * recognise a pointer result and exclude it from supersession bookkeeping (a\n * pointer fetched no content, so it must not stub the read it points at).\n */\nexport const DEDUP_POINTER_PREFIX = \"[Already in context:\";\n\n/** Whether a tool-result text is an at-call dedup pointer. */\nexport function isDedupPointerText(text: string): boolean {\n\treturn text.trimStart().startsWith(DEDUP_POINTER_PREFIX);\n}\n\n/**\n * Content stamps for reads that have already been delivered.\n *\n * The dedup pointer tells the model the file \"has not changed since\" the read it\n * names. Nothing in the transcript can establish that: an editor, a formatter, a\n * branch switch or a second agent can rewrite the file between two reads without\n * leaving a trace in it. So each delivered read records a stamp of the file it\n * saw, and the pointer is only served when the file still stamps the same.\n *\n * Keyed by tool call id rather than by path, so two sessions sharing a process\n * cannot overwrite each other's observations. Bounded, and a missing entry\n * counts as \"cannot prove it is unchanged\" - the read then simply runs.\n */\nconst MAX_TRACKED_READS = 500;\nconst stampByCallId = new Map<string, string>();\n\n/** Remember the stamp of the file a read delivered. */\nexport function recordReadStamp(callId: string, stamp: string): void {\n\tif (stampByCallId.size >= MAX_TRACKED_READS) {\n\t\tconst oldest = stampByCallId.keys().next().value;\n\t\tif (oldest !== undefined) stampByCallId.delete(oldest);\n\t}\n\tstampByCallId.set(callId, stamp);\n}\n\n/**\n * Whether the file a read delivered still stamps the same. False when no stamp\n * was recorded, so an unprovable case re-reads rather than asserting freshness.\n */\nexport function readStampMatches(callId: string, stamp: string): boolean {\n\tconst recorded = stampByCallId.get(callId);\n\treturn recorded !== undefined && recorded === stamp;\n}\n\n/** Test seam: drop all recorded stamps. */\nexport function clearReadStamps(): void {\n\tstampByCallId.clear();\n}\n\n/** A covering earlier read, described for the pointer message. */\nexport interface CoveringRead {\n\t/** Tool call id of the read being pointed at, used to check its stamp. */\n\tcallId: string;\n\t/** The path as the earlier read spelled it (for a friendly pointer). */\n\tdisplay: string;\n\t/** Delivered range start (1-indexed line). */\n\tstart: number;\n\t/** Delivered range end (exclusive; Infinity for a whole-file read). */\n\tend: number;\n}\n\nexport interface FindCoveringReadOptions {\n\t/** Resolved absolute path of the current read. */\n\tresolvedPath: string;\n\t/** Range the current read is asking for. */\n\trequestedRange: ReadRange;\n\t/** Tool call id of the current read, excluded from candidate/supersession sets. */\n\tcurrentCallId: string;\n\t/** Resolve a raw read-arg path the same way the current read resolved its path. */\n\tresolvePath: (rawPath: string) => string;\n}\n\ninterface ContentBlock {\n\ttype?: string;\n\tid?: string;\n\tname?: string;\n\targuments?: Record<string, unknown>;\n\ttext?: string;\n}\n\ninterface MessageLike {\n\trole?: string;\n\tcontent?: ContentBlock[];\n\ttoolCallId?: string;\n\ttoolName?: string;\n\tisError?: boolean;\n}\n\nconst READ_TOOL = \"read\";\nconst MUTATE_TOOLS = new Set([\"edit\", \"write\"]);\n\n/** A session compaction entry — the boundary that trims the live context. */\nfunction isCompactionEntry(entry: unknown): boolean {\n\treturn !!entry && typeof entry === \"object\" && (entry as { type?: unknown }).type === \"compaction\";\n}\n\n/** Accept either raw `AgentMessage`s or session entries wrapping `.message`. */\nfunction toMessage(entry: unknown): MessageLike | null {\n\tif (!entry || typeof entry !== \"object\") return null;\n\tconst e = entry as { message?: unknown; role?: unknown };\n\tif (e.message && typeof e.message === \"object\") return e.message as MessageLike;\n\tif (typeof e.role === \"string\") return e as MessageLike;\n\treturn null;\n}\n\nfunction resultText(m: MessageLike): string {\n\treturn (m.content ?? [])\n\t\t.filter((c) => c.type === \"text\")\n\t\t.map((c) => c.text ?? \"\")\n\t\t.join(\"\");\n}\n\n/**\n * Recover the range a read result actually *delivered* from its text.\n *\n * A cap-truncated read announces `[Showing lines A-B of N ...]`, so it delivered\n * only `[A, B+1)` even though its args declared a wider range. A read whose first\n * line alone exceeded the byte cap delivered nothing. Any other (untruncated)\n * read delivered its full declared range. The user-`limit` early-stop notice\n * (`[N more lines in file ...]`) is *not* a cap truncation — the declared range\n * was delivered in full — so it falls through to `declared`.\n */\nfunction deliveredRange(text: string, declared: ReadRange): ReadRange {\n\t// The truncation notice is always the trailing `\\n\\n[Showing lines A-B of N ...]`\n\t// clause the read tool appends. Anchor to the end so a `[Showing lines ...]`\n\t// string that merely appears *inside* the file's content can't spoof it.\n\tconst showing = text.match(/\\n\\n\\[Showing lines (\\d+)-(\\d+) of \\d+[^\\]]*\\]\\s*$/);\n\tif (showing) {\n\t\tconst a = Number(showing[1]);\n\t\tconst b = Number(showing[2]);\n\t\tif (Number.isFinite(a) && Number.isFinite(b) && b >= a) return { start: a, end: b + 1 };\n\t}\n\t// First line alone exceeded the byte limit: the whole result *is* that notice,\n\t// so it must start the text. Nothing usable was delivered.\n\tif (/^\\[Line \\d+ is .+ exceeds .+ limit\\./.test(text)) return { start: 1, end: 1 };\n\treturn declared;\n}\n\n/**\n * Find the latest earlier read that (a) is for the same resolved path, (b)\n * actually delivered a range containing the requested range, and (c) is still\n * live in the outgoing context — i.e. the post-hoc GC will not have stubbed it,\n * because no later edit/write and no later overlapping content read supersede\n * it. Returns that read's delivered range for the pointer, or null when the\n * current read must actually run.\n *\n * Deliberately conservative: a truncated earlier read only covers what it\n * delivered, a whole-file read must have been delivered untruncated to count as\n * covering, and any pointer results (which fetched nothing) are ignored on both\n * the candidate and the supersession side.\n */\nexport function findCoveringRead(entries: readonly unknown[], opts: FindCoveringReadOptions): CoveringRead | null {\n\t// `declared` mirrors the range the GC uses for supersession (straight from the\n\t// call args); `delivered` is what the result text shows was actually returned,\n\t// used for coverage. The current call has no result yet, so it never appears.\n\tinterface PriorRead {\n\t\tindex: number;\n\t\tdeclared: ReadRange;\n\t\tdelivered: ReadRange;\n\t\tdisplay: string;\n\t\tcallId: string;\n\t}\n\n\t// Resolve each distinct raw path once — the resolver may hit the filesystem.\n\tconst resolveCache = new Map<string, string>();\n\tconst resolvePath = (raw: string): string => {\n\t\tconst hit = resolveCache.get(raw);\n\t\tif (hit !== undefined) return hit;\n\t\tconst resolved = opts.resolvePath(raw);\n\t\tresolveCache.set(raw, resolved);\n\t\treturn resolved;\n\t};\n\n\t// Single ordered pass. `readCall`/`mutateCallPath` map a call id to its path as\n\t// the call is seen (a toolCall always precedes its result); `reads` collects\n\t// the target path's reads and `lastMutateIndex` its last mutate. `order` gives\n\t// live-context position for the supersession/mutate comparisons.\n\tconst readCall = new Map<string, { resolved: string; display: string; declared: ReadRange }>();\n\tconst mutateCallPath = new Map<string, string>();\n\tconst reads: PriorRead[] = [];\n\tlet lastMutateIndex = -1;\n\t// A *failed* edit/write is not a mutate - the GC is right to ignore it, since\n\t// the file did not change. It is still the one moment the model most needs the\n\t// real bytes: the failure usually means its copy of the text does not match the\n\t// file. Serving a \"not re-fetched, unchanged\" pointer there leaves it retrying\n\t// the same wrong text with no way to see what is actually on disk.\n\tlet lastFailedMutateIndex = -1;\n\tlet order = 0;\n\n\tfor (const entry of entries) {\n\t\t// Compaction boundary: everything before it is replaced by a summary in the\n\t\t// live context, so drop the state accumulated so far. Conservative — the\n\t\t// kept tail before the boundary is dropped too — which can only miss a\n\t\t// dedup, never point at content that is no longer in context.\n\t\tif (isCompactionEntry(entry)) {\n\t\t\treadCall.clear();\n\t\t\tmutateCallPath.clear();\n\t\t\treads.length = 0;\n\t\t\tlastMutateIndex = -1;\n\t\t\torder = 0;\n\t\t\tcontinue;\n\t\t}\n\t\tconst m = toMessage(entry);\n\t\tif (!m) continue;\n\t\tconst i = order++;\n\t\tif (m.role === \"assistant\" && Array.isArray(m.content)) {\n\t\t\tfor (const b of m.content) {\n\t\t\t\tif (b.type !== \"toolCall\" || !b.id) continue;\n\t\t\t\tconst raw = b.arguments?.path;\n\t\t\t\tif (typeof raw !== \"string\" || raw.length === 0) continue;\n\t\t\t\tconst resolved = resolvePath(raw);\n\t\t\t\tif (b.name === READ_TOOL) {\n\t\t\t\t\treadCall.set(b.id, { resolved, display: raw, declared: readRangeFromArgs(b.arguments) });\n\t\t\t\t} else if (b.name && MUTATE_TOOLS.has(b.name)) {\n\t\t\t\t\tmutateCallPath.set(b.id, resolved);\n\t\t\t\t}\n\t\t\t}\n\t\t} else if (m.role === \"toolResult\" && m.isError && m.toolCallId) {\n\t\t\tif (m.toolName && MUTATE_TOOLS.has(m.toolName) && mutateCallPath.get(m.toolCallId) === opts.resolvedPath) {\n\t\t\t\tlastFailedMutateIndex = i;\n\t\t\t}\n\t\t} else if (m.role === \"toolResult\" && !m.isError && m.toolCallId) {\n\t\t\tif (m.toolName === READ_TOOL) {\n\t\t\t\tif (m.toolCallId === opts.currentCallId) continue;\n\t\t\t\tconst info = readCall.get(m.toolCallId);\n\t\t\t\tif (!info || info.resolved !== opts.resolvedPath) continue;\n\t\t\t\tconst text = resultText(m);\n\t\t\t\t// A pointer fetched nothing: the GC excludes it from supersession, so we\n\t\t\t\t// must too (both as a candidate and as a superseder).\n\t\t\t\tif (isDedupPointerText(text)) continue;\n\t\t\t\treads.push({\n\t\t\t\t\tindex: i,\n\t\t\t\t\tdeclared: info.declared,\n\t\t\t\t\tdelivered: deliveredRange(text, info.declared),\n\t\t\t\t\tdisplay: info.display,\n\t\t\t\t\tcallId: m.toolCallId,\n\t\t\t\t});\n\t\t\t} else if (m.toolName && MUTATE_TOOLS.has(m.toolName)) {\n\t\t\t\tif (mutateCallPath.get(m.toolCallId) === opts.resolvedPath) lastMutateIndex = i;\n\t\t\t}\n\t\t}\n\t}\n\n\t// A read survives the GC iff no later mutate and no later read overlaps its\n\t// *declared* range — exactly the GC's own test — so predict it the same way.\n\tconst survivesGc = (r: PriorRead): boolean =>\n\t\tlastMutateIndex <= r.index && !reads.some((o) => o.index > r.index && rangesOverlap(o.declared, r.declared));\n\n\t// Nothing earlier can be trusted to satisfy this read once an edit against this\n\t// path has failed: let the read run and hand the model the current bytes.\n\tif (lastFailedMutateIndex > -1) return null;\n\n\tlet best: PriorRead | null = null;\n\tfor (const r of reads) {\n\t\t// Coverage uses the *delivered* range: a truncated read only holds what it returned.\n\t\tif (!rangeContains(r.delivered, opts.requestedRange)) continue;\n\t\tif (!survivesGc(r)) continue;\n\t\tif (!best || r.index > best.index) best = r;\n\t}\n\tif (!best) return null;\n\treturn { callId: best.callId, display: best.display, start: best.delivered.start, end: best.delivered.end };\n}\n\n/** The parts of a covering read the pointer message renders. */\nexport type CoveringReadDisplay = Pick<CoveringRead, \"display\" | \"start\" | \"end\">;\n\n/** Build the pointer text returned in place of a re-fetch. */\nexport function buildDedupPointerText(covering: CoveringReadDisplay): string {\n\tconst where =\n\t\tcovering.end === Number.POSITIVE_INFINITY\n\t\t\t? \"the entire file\"\n\t\t\t: covering.end - 1 > covering.start\n\t\t\t\t? `lines ${covering.start}-${covering.end - 1}`\n\t\t\t\t: `line ${covering.start}`;\n\treturn `${DEDUP_POINTER_PREFIX} ${covering.display} (${where}) was already read earlier in this session and has not changed since. Not re-fetched to save tokens — pass a different offset/limit, or edit the file, if you need other or newer content.]`;\n}\n"]}

package/dist/core/tools/read-dedup.js

@@ -49,6 +49,42 @@ export const DEDUP_POINTER_PREFIX = "[Already in context:";
 export function isDedupPointerText(text) {
     return text.trimStart().startsWith(DEDUP_POINTER_PREFIX);
 }
+/**
+ * Content stamps for reads that have already been delivered.
+ *
+ * The dedup pointer tells the model the file "has not changed since" the read it
+ * names. Nothing in the transcript can establish that: an editor, a formatter, a
+ * branch switch or a second agent can rewrite the file between two reads without
+ * leaving a trace in it. So each delivered read records a stamp of the file it
+ * saw, and the pointer is only served when the file still stamps the same.
+ *
+ * Keyed by tool call id rather than by path, so two sessions sharing a process
+ * cannot overwrite each other's observations. Bounded, and a missing entry
+ * counts as "cannot prove it is unchanged" - the read then simply runs.
+ */
+const MAX_TRACKED_READS = 500;
+const stampByCallId = new Map();
+/** Remember the stamp of the file a read delivered. */
+export function recordReadStamp(callId, stamp) {
+    if (stampByCallId.size >= MAX_TRACKED_READS) {
+        const oldest = stampByCallId.keys().next().value;
+        if (oldest !== undefined)
+            stampByCallId.delete(oldest);
+    }
+    stampByCallId.set(callId, stamp);
+}
+/**
+ * Whether the file a read delivered still stamps the same. False when no stamp
+ * was recorded, so an unprovable case re-reads rather than asserting freshness.
+ */
+export function readStampMatches(callId, stamp) {
+    const recorded = stampByCallId.get(callId);
+    return recorded !== undefined && recorded === stamp;
+}
+/** Test seam: drop all recorded stamps. */
+export function clearReadStamps() {
+    stampByCallId.clear();
+}
 const READ_TOOL = "read";
 const MUTATE_TOOLS = new Set(["edit", "write"]);
 /** A session compaction entry — the boundary that trims the live context. */
@@ -131,6 +167,12 @@ export function findCoveringRead(entries, opts) {
     const mutateCallPath = new Map();
     const reads = [];
     let lastMutateIndex = -1;
+    // A *failed* edit/write is not a mutate - the GC is right to ignore it, since
+    // the file did not change. It is still the one moment the model most needs the
+    // real bytes: the failure usually means its copy of the text does not match the
+    // file. Serving a "not re-fetched, unchanged" pointer there leaves it retrying
+    // the same wrong text with no way to see what is actually on disk.
+    let lastFailedMutateIndex = -1;
     let order = 0;
     for (const entry of entries) {
         // Compaction boundary: everything before it is replaced by a summary in the
@@ -165,6 +207,11 @@ export function findCoveringRead(entries, opts) {
                 }
             }
         }
+        else if (m.role === "toolResult" && m.isError && m.toolCallId) {
+            if (m.toolName && MUTATE_TOOLS.has(m.toolName) && mutateCallPath.get(m.toolCallId) === opts.resolvedPath) {
+                lastFailedMutateIndex = i;
+            }
+        }
         else if (m.role === "toolResult" && !m.isError && m.toolCallId) {
             if (m.toolName === READ_TOOL) {
                 if (m.toolCallId === opts.currentCallId)
@@ -182,6 +229,7 @@ export function findCoveringRead(entries, opts) {
                     declared: info.declared,
                     delivered: deliveredRange(text, info.declared),
                     display: info.display,
+                    callId: m.toolCallId,
                 });
             }
             else if (m.toolName && MUTATE_TOOLS.has(m.toolName)) {
@@ -193,6 +241,10 @@ export function findCoveringRead(entries, opts) {
     // A read survives the GC iff no later mutate and no later read overlaps its
     // *declared* range — exactly the GC's own test — so predict it the same way.
     const survivesGc = (r) => lastMutateIndex <= r.index && !reads.some((o) => o.index > r.index && rangesOverlap(o.declared, r.declared));
+    // Nothing earlier can be trusted to satisfy this read once an edit against this
+    // path has failed: let the read run and hand the model the current bytes.
+    if (lastFailedMutateIndex > -1)
+        return null;
     let best = null;
     for (const r of reads) {
         // Coverage uses the *delivered* range: a truncated read only holds what it returned.
@@ -205,7 +257,7 @@ export function findCoveringRead(entries, opts) {
     }
     if (!best)
         return null;
-    return { display: best.display, start: best.delivered.start, end: best.delivered.end };
+    return { callId: best.callId, display: best.display, start: best.delivered.start, end: best.delivered.end };
 }
 /** Build the pointer text returned in place of a re-fetch. */
 export function buildDedupPointerText(covering) {

package/dist/core/tools/read-dedup.js.map

@@ -1 +1 @@
-{"version":3,"file":"read-dedup.js","sourceRoot":"","sources":["../../../src/core/tools/read-dedup.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAQH,sEAAsE;AACtE,MAAM,CAAC,MAAM,gBAAgB,GAAc,EAAE,KAAK,EAAE,CAAC,EAAE,GAAG,EAAE,MAAM,CAAC,iBAAiB,EAAE,CAAC;AAEvF;;;;GAIG;AACH,MAAM,UAAU,iBAAiB,CAAC,IAAyC,EAAa;IACvF,MAAM,SAAS,GAAG,IAAI,EAAE,MAAM,CAAC;IAC/B,MAAM,QAAQ,GAAG,IAAI,EAAE,KAAK,CAAC;IAC7B,MAAM,MAAM,GAAG,OAAO,SAAS,KAAK,QAAQ,IAAI,MAAM,CAAC,QAAQ,CAAC,SAAS,CAAC,IAAI,SAAS,GAAG,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC;IAC5G,MAAM,GAAG,GACR,OAAO,QAAQ,KAAK,QAAQ,IAAI,MAAM,CAAC,QAAQ,CAAC,QAAQ,CAAC;QACxD,CAAC,CAAC,MAAM,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,QAAQ,CAAC;QAChC,CAAC,CAAC,MAAM,CAAC,iBAAiB,CAAC;IAC7B,OAAO,EAAE,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,CAAC;AAAA,CAC9B;AAED,mDAAmD;AACnD,MAAM,UAAU,aAAa,CAAC,CAAY,EAAE,CAAY,EAAW;IAClE,OAAO,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,GAAG,CAAC;AAAA,CAC1C;AAED,sFAAsF;AACtF,SAAS,aAAa,CAAC,KAAgB,EAAE,KAAgB,EAAW;IACnE,OAAO,KAAK,CAAC,KAAK,IAAI,KAAK,CAAC,KAAK,IAAI,KAAK,CAAC,GAAG,IAAI,KAAK,CAAC,GAAG,CAAC;AAAA,CAC5D;AAED;;;;GAIG;AACH,MAAM,CAAC,MAAM,oBAAoB,GAAG,sBAAsB,CAAC;AAE3D,8DAA8D;AAC9D,MAAM,UAAU,kBAAkB,CAAC,IAAY,EAAW;IACzD,OAAO,IAAI,CAAC,SAAS,EAAE,CAAC,UAAU,CAAC,oBAAoB,CAAC,CAAC;AAAA,CACzD;AAuCD,MAAM,SAAS,GAAG,MAAM,CAAC;AACzB,MAAM,YAAY,GAAG,IAAI,GAAG,CAAC,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC;AAEhD,+EAA6E;AAC7E,SAAS,iBAAiB,CAAC,KAAc,EAAW;IACnD,OAAO,CAAC,CAAC,KAAK,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAK,KAA4B,CAAC,IAAI,KAAK,YAAY,CAAC;AAAA,CACnG;AAED,gFAAgF;AAChF,SAAS,SAAS,CAAC,KAAc,EAAsB;IACtD,IAAI,CAAC,KAAK,IAAI,OAAO,KAAK,KAAK,QAAQ;QAAE,OAAO,IAAI,CAAC;IACrD,MAAM,CAAC,GAAG,KAA8C,CAAC;IACzD,IAAI,CAAC,CAAC,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,KAAK,QAAQ;QAAE,OAAO,CAAC,CAAC,OAAsB,CAAC;IAChF,IAAI,OAAO,CAAC,CAAC,IAAI,KAAK,QAAQ;QAAE,OAAO,CAAgB,CAAC;IACxD,OAAO,IAAI,CAAC;AAAA,CACZ;AAED,SAAS,UAAU,CAAC,CAAc,EAAU;IAC3C,OAAO,CAAC,CAAC,CAAC,OAAO,IAAI,EAAE,CAAC;SACtB,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,MAAM,CAAC;SAChC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,IAAI,EAAE,CAAC;SACxB,IAAI,CAAC,EAAE,CAAC,CAAC;AAAA,CACX;AAED;;;;;;;;;GASG;AACH,SAAS,cAAc,CAAC,IAAY,EAAE,QAAmB,EAAa;IACrE,kFAAkF;IAClF,6EAA6E;IAC7E,yEAAyE;IACzE,MAAM,OAAO,GAAG,IAAI,CAAC,KAAK,CAAC,oDAAoD,CAAC,CAAC;IACjF,IAAI,OAAO,EAAE,CAAC;QACb,MAAM,CAAC,GAAG,MAAM,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC;QAC7B,MAAM,CAAC,GAAG,MAAM,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC;QAC7B,IAAI,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,IAAI,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC;YAAE,OAAO,EAAE,KAAK,EAAE,CAAC,EAAE,GAAG,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC;IACzF,CAAC;IACD,+EAA+E;IAC/E,2DAA2D;IAC3D,IAAI,sCAAsC,CAAC,IAAI,CAAC,IAAI,CAAC;QAAE,OAAO,EAAE,KAAK,EAAE,CAAC,EAAE,GAAG,EAAE,CAAC,EAAE,CAAC;IACnF,OAAO,QAAQ,CAAC;AAAA,CAChB;AAED;;;;;;;;;;;;GAYG;AACH,MAAM,UAAU,gBAAgB,CAAC,OAA2B,EAAE,IAA6B,EAAuB;IAWjH,+EAA6E;IAC7E,MAAM,YAAY,GAAG,IAAI,GAAG,EAAkB,CAAC;IAC/C,MAAM,WAAW,GAAG,CAAC,GAAW,EAAU,EAAE,CAAC;QAC5C,MAAM,GAAG,GAAG,YAAY,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;QAClC,IAAI,GAAG,KAAK,SAAS;YAAE,OAAO,GAAG,CAAC;QAClC,MAAM,QAAQ,GAAG,IAAI,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC;QACvC,YAAY,CAAC,GAAG,CAAC,GAAG,EAAE,QAAQ,CAAC,CAAC;QAChC,OAAO,QAAQ,CAAC;IAAA,CAChB,CAAC;IAEF,gFAAgF;IAChF,6EAA6E;IAC7E,+EAA+E;IAC/E,iEAAiE;IACjE,MAAM,QAAQ,GAAG,IAAI,GAAG,EAAsE,CAAC;IAC/F,MAAM,cAAc,GAAG,IAAI,GAAG,EAAkB,CAAC;IACjD,MAAM,KAAK,GAAgB,EAAE,CAAC;IAC9B,IAAI,eAAe,GAAG,CAAC,CAAC,CAAC;IACzB,IAAI,KAAK,GAAG,CAAC,CAAC;IAEd,KAAK,MAAM,KAAK,IAAI,OAAO,EAAE,CAAC;QAC7B,4EAA4E;QAC5E,2EAAyE;QACzE,yEAAuE;QACvE,8DAA8D;QAC9D,IAAI,iBAAiB,CAAC,KAAK,CAAC,EAAE,CAAC;YAC9B,QAAQ,CAAC,KAAK,EAAE,CAAC;YACjB,cAAc,CAAC,KAAK,EAAE,CAAC;YACvB,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC;YACjB,eAAe,GAAG,CAAC,CAAC,CAAC;YACrB,KAAK,GAAG,CAAC,CAAC;YACV,SAAS;QACV,CAAC;QACD,MAAM,CAAC,GAAG,SAAS,CAAC,KAAK,CAAC,CAAC;QAC3B,IAAI,CAAC,CAAC;YAAE,SAAS;QACjB,MAAM,CAAC,GAAG,KAAK,EAAE,CAAC;QAClB,IAAI,CAAC,CAAC,IAAI,KAAK,WAAW,IAAI,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,OAAO,CAAC,EAAE,CAAC;YACxD,KAAK,MAAM,CAAC,IAAI,CAAC,CAAC,OAAO,EAAE,CAAC;gBAC3B,IAAI,CAAC,CAAC,IAAI,KAAK,UAAU,IAAI,CAAC,CAAC,CAAC,EAAE;oBAAE,SAAS;gBAC7C,MAAM,GAAG,GAAG,CAAC,CAAC,SAAS,EAAE,IAAI,CAAC;gBAC9B,IAAI,OAAO,GAAG,KAAK,QAAQ,IAAI,GAAG,CAAC,MAAM,KAAK,CAAC;oBAAE,SAAS;gBAC1D,MAAM,QAAQ,GAAG,WAAW,CAAC,GAAG,CAAC,CAAC;gBAClC,IAAI,CAAC,CAAC,IAAI,KAAK,SAAS,EAAE,CAAC;oBAC1B,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,EAAE,QAAQ,EAAE,OAAO,EAAE,GAAG,EAAE,QAAQ,EAAE,iBAAiB,CAAC,CAAC,CAAC,SAAS,CAAC,EAAE,CAAC,CAAC;gBAC1F,CAAC;qBAAM,IAAI,CAAC,CAAC,IAAI,IAAI,YAAY,CAAC,GAAG,CAAC,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC;oBAC/C,cAAc,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,QAAQ,CAAC,CAAC;gBACpC,CAAC;YACF,CAAC;QACF,CAAC;aAAM,IAAI,CAAC,CAAC,IAAI,KAAK,YAAY,IAAI,CAAC,CAAC,CAAC,OAAO,IAAI,CAAC,CAAC,UAAU,EAAE,CAAC;YAClE,IAAI,CAAC,CAAC,QAAQ,KAAK,SAAS,EAAE,CAAC;gBAC9B,IAAI,CAAC,CAAC,UAAU,KAAK,IAAI,CAAC,aAAa;oBAAE,SAAS;gBAClD,MAAM,IAAI,GAAG,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC;gBACxC,IAAI,CAAC,IAAI,IAAI,IAAI,CAAC,QAAQ,KAAK,IAAI,CAAC,YAAY;oBAAE,SAAS;gBAC3D,MAAM,IAAI,GAAG,UAAU,CAAC,CAAC,CAAC,CAAC;gBAC3B,yEAAyE;gBACzE,sDAAsD;gBACtD,IAAI,kBAAkB,CAAC,IAAI,CAAC;oBAAE,SAAS;gBACvC,KAAK,CAAC,IAAI,CAAC;oBACV,KAAK,EAAE,CAAC;oBACR,QAAQ,EAAE,IAAI,CAAC,QAAQ;oBACvB,SAAS,EAAE,cAAc,CAAC,IAAI,EAAE,IAAI,CAAC,QAAQ,CAAC;oBAC9C,OAAO,EAAE,IAAI,CAAC,OAAO;iBACrB,CAAC,CAAC;YACJ,CAAC;iBAAM,IAAI,CAAC,CAAC,QAAQ,IAAI,YAAY,CAAC,GAAG,CAAC,CAAC,CAAC,QAAQ,CAAC,EAAE,CAAC;gBACvD,IAAI,cAAc,CAAC,GAAG,CAAC,CAAC,CAAC,UAAU,CAAC,KAAK,IAAI,CAAC,YAAY;oBAAE,eAAe,GAAG,CAAC,CAAC;YACjF,CAAC;QACF,CAAC;IACF,CAAC;IAED,4EAA4E;IAC5E,iFAA6E;IAC7E,MAAM,UAAU,GAAG,CAAC,CAAY,EAAW,EAAE,CAC5C,eAAe,IAAI,CAAC,CAAC,KAAK,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,KAAK,IAAI,aAAa,CAAC,CAAC,CAAC,QAAQ,EAAE,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC;IAE9G,IAAI,IAAI,GAAqB,IAAI,CAAC;IAClC,KAAK,MAAM,CAAC,IAAI,KAAK,EAAE,CAAC;QACvB,qFAAqF;QACrF,IAAI,CAAC,aAAa,CAAC,CAAC,CAAC,SAAS,EAAE,IAAI,CAAC,cAAc,CAAC;YAAE,SAAS;QAC/D,IAAI,CAAC,UAAU,CAAC,CAAC,CAAC;YAAE,SAAS;QAC7B,IAAI,CAAC,IAAI,IAAI,CAAC,CAAC,KAAK,GAAG,IAAI,CAAC,KAAK;YAAE,IAAI,GAAG,CAAC,CAAC;IAC7C,CAAC;IACD,IAAI,CAAC,IAAI;QAAE,OAAO,IAAI,CAAC;IACvB,OAAO,EAAE,OAAO,EAAE,IAAI,CAAC,OAAO,EAAE,KAAK,EAAE,IAAI,CAAC,SAAS,CAAC,KAAK,EAAE,GAAG,EAAE,IAAI,CAAC,SAAS,CAAC,GAAG,EAAE,CAAC;AAAA,CACvF;AAED,8DAA8D;AAC9D,MAAM,UAAU,qBAAqB,CAAC,QAAsB,EAAU;IACrE,MAAM,KAAK,GACV,QAAQ,CAAC,GAAG,KAAK,MAAM,CAAC,iBAAiB;QACxC,CAAC,CAAC,iBAAiB;QACnB,CAAC,CAAC,QAAQ,CAAC,GAAG,GAAG,CAAC,GAAG,QAAQ,CAAC,KAAK;YAClC,CAAC,CAAC,SAAS,QAAQ,CAAC,KAAK,IAAI,QAAQ,CAAC,GAAG,GAAG,CAAC,EAAE;YAC/C,CAAC,CAAC,QAAQ,QAAQ,CAAC,KAAK,EAAE,CAAC;IAC9B,OAAO,GAAG,oBAAoB,IAAI,QAAQ,CAAC,OAAO,KAAK,KAAK,+LAA6L,CAAC;AAAA,CAC1P","sourcesContent":["/**\n * Read de-duplication primitives.\n *\n * Two mechanisms share this module:\n *\n * - The post-hoc context GC (`context-gc.ts`), which stubs out a `read` result\n *   once a later overlapping read or an edit/write has superseded it.\n * - The at-call-time guard in the `read` tool, which short-circuits a read whose\n *   requested range is *already fully covered* by an earlier, still-live read in\n *   the current session, returning a pointer instead of re-fetching the file.\n *\n * Both reason about half-open line ranges `[start, end)` (end exclusive), so the\n * range math lives here and cannot drift between them. The guard additionally\n * needs to (a) recognise its own pointer results so they are never treated as\n * content-bearing reads, and (b) walk the session branch to find a covering\n * read — both of which are defined here to keep the tool file lean.\n */\n\n/** Half-open line interval `[start, end)` a read call covers; end is exclusive. */\nexport interface ReadRange {\n\tstart: number;\n\tend: number;\n}\n\n/** A read with no offset/limit covers the whole file (open-ended). */\nexport const WHOLE_FILE_RANGE: ReadRange = { start: 1, end: Number.POSITIVE_INFINITY };\n\n/**\n * Derive the line range a read call covers from its `offset`/`limit` args.\n * Missing offset means \"from line 1\"; missing limit means \"to end of file\"\n * (open-ended, so it overlaps any later read of the same file).\n */\nexport function readRangeFromArgs(args: Record<string, unknown> | undefined): ReadRange {\n\tconst offsetRaw = args?.offset;\n\tconst limitRaw = args?.limit;\n\tconst offset = typeof offsetRaw === \"number\" && Number.isFinite(offsetRaw) && offsetRaw > 0 ? offsetRaw : 1;\n\tconst end =\n\t\ttypeof limitRaw === \"number\" && Number.isFinite(limitRaw)\n\t\t\t? offset + Math.max(0, limitRaw)\n\t\t\t: Number.POSITIVE_INFINITY;\n\treturn { start: offset, end };\n}\n\n/** Whether two half-open line ranges intersect. */\nexport function rangesOverlap(a: ReadRange, b: ReadRange): boolean {\n\treturn a.start < b.end && b.start < a.end;\n}\n\n/** Whether `outer` fully contains `inner` (every line of inner lies within outer). */\nfunction rangeContains(outer: ReadRange, inner: ReadRange): boolean {\n\treturn outer.start <= inner.start && outer.end >= inner.end;\n}\n\n/**\n * Marker prefix for the at-call dedup pointer. Kept stable so the context GC can\n * recognise a pointer result and exclude it from supersession bookkeeping (a\n * pointer fetched no content, so it must not stub the read it points at).\n */\nexport const DEDUP_POINTER_PREFIX = \"[Already in context:\";\n\n/** Whether a tool-result text is an at-call dedup pointer. */\nexport function isDedupPointerText(text: string): boolean {\n\treturn text.trimStart().startsWith(DEDUP_POINTER_PREFIX);\n}\n\n/** A covering earlier read, described for the pointer message. */\nexport interface CoveringRead {\n\t/** The path as the earlier read spelled it (for a friendly pointer). */\n\tdisplay: string;\n\t/** Delivered range start (1-indexed line). */\n\tstart: number;\n\t/** Delivered range end (exclusive; Infinity for a whole-file read). */\n\tend: number;\n}\n\nexport interface FindCoveringReadOptions {\n\t/** Resolved absolute path of the current read. */\n\tresolvedPath: string;\n\t/** Range the current read is asking for. */\n\trequestedRange: ReadRange;\n\t/** Tool call id of the current read, excluded from candidate/supersession sets. */\n\tcurrentCallId: string;\n\t/** Resolve a raw read-arg path the same way the current read resolved its path. */\n\tresolvePath: (rawPath: string) => string;\n}\n\ninterface ContentBlock {\n\ttype?: string;\n\tid?: string;\n\tname?: string;\n\targuments?: Record<string, unknown>;\n\ttext?: string;\n}\n\ninterface MessageLike {\n\trole?: string;\n\tcontent?: ContentBlock[];\n\ttoolCallId?: string;\n\ttoolName?: string;\n\tisError?: boolean;\n}\n\nconst READ_TOOL = \"read\";\nconst MUTATE_TOOLS = new Set([\"edit\", \"write\"]);\n\n/** A session compaction entry — the boundary that trims the live context. */\nfunction isCompactionEntry(entry: unknown): boolean {\n\treturn !!entry && typeof entry === \"object\" && (entry as { type?: unknown }).type === \"compaction\";\n}\n\n/** Accept either raw `AgentMessage`s or session entries wrapping `.message`. */\nfunction toMessage(entry: unknown): MessageLike | null {\n\tif (!entry || typeof entry !== \"object\") return null;\n\tconst e = entry as { message?: unknown; role?: unknown };\n\tif (e.message && typeof e.message === \"object\") return e.message as MessageLike;\n\tif (typeof e.role === \"string\") return e as MessageLike;\n\treturn null;\n}\n\nfunction resultText(m: MessageLike): string {\n\treturn (m.content ?? [])\n\t\t.filter((c) => c.type === \"text\")\n\t\t.map((c) => c.text ?? \"\")\n\t\t.join(\"\");\n}\n\n/**\n * Recover the range a read result actually *delivered* from its text.\n *\n * A cap-truncated read announces `[Showing lines A-B of N ...]`, so it delivered\n * only `[A, B+1)` even though its args declared a wider range. A read whose first\n * line alone exceeded the byte cap delivered nothing. Any other (untruncated)\n * read delivered its full declared range. The user-`limit` early-stop notice\n * (`[N more lines in file ...]`) is *not* a cap truncation — the declared range\n * was delivered in full — so it falls through to `declared`.\n */\nfunction deliveredRange(text: string, declared: ReadRange): ReadRange {\n\t// The truncation notice is always the trailing `\\n\\n[Showing lines A-B of N ...]`\n\t// clause the read tool appends. Anchor to the end so a `[Showing lines ...]`\n\t// string that merely appears *inside* the file's content can't spoof it.\n\tconst showing = text.match(/\\n\\n\\[Showing lines (\\d+)-(\\d+) of \\d+[^\\]]*\\]\\s*$/);\n\tif (showing) {\n\t\tconst a = Number(showing[1]);\n\t\tconst b = Number(showing[2]);\n\t\tif (Number.isFinite(a) && Number.isFinite(b) && b >= a) return { start: a, end: b + 1 };\n\t}\n\t// First line alone exceeded the byte limit: the whole result *is* that notice,\n\t// so it must start the text. Nothing usable was delivered.\n\tif (/^\\[Line \\d+ is .+ exceeds .+ limit\\./.test(text)) return { start: 1, end: 1 };\n\treturn declared;\n}\n\n/**\n * Find the latest earlier read that (a) is for the same resolved path, (b)\n * actually delivered a range containing the requested range, and (c) is still\n * live in the outgoing context — i.e. the post-hoc GC will not have stubbed it,\n * because no later edit/write and no later overlapping content read supersede\n * it. Returns that read's delivered range for the pointer, or null when the\n * current read must actually run.\n *\n * Deliberately conservative: a truncated earlier read only covers what it\n * delivered, a whole-file read must have been delivered untruncated to count as\n * covering, and any pointer results (which fetched nothing) are ignored on both\n * the candidate and the supersession side.\n */\nexport function findCoveringRead(entries: readonly unknown[], opts: FindCoveringReadOptions): CoveringRead | null {\n\t// `declared` mirrors the range the GC uses for supersession (straight from the\n\t// call args); `delivered` is what the result text shows was actually returned,\n\t// used for coverage. The current call has no result yet, so it never appears.\n\tinterface PriorRead {\n\t\tindex: number;\n\t\tdeclared: ReadRange;\n\t\tdelivered: ReadRange;\n\t\tdisplay: string;\n\t}\n\n\t// Resolve each distinct raw path once — the resolver may hit the filesystem.\n\tconst resolveCache = new Map<string, string>();\n\tconst resolvePath = (raw: string): string => {\n\t\tconst hit = resolveCache.get(raw);\n\t\tif (hit !== undefined) return hit;\n\t\tconst resolved = opts.resolvePath(raw);\n\t\tresolveCache.set(raw, resolved);\n\t\treturn resolved;\n\t};\n\n\t// Single ordered pass. `readCall`/`mutateCallPath` map a call id to its path as\n\t// the call is seen (a toolCall always precedes its result); `reads` collects\n\t// the target path's reads and `lastMutateIndex` its last mutate. `order` gives\n\t// live-context position for the supersession/mutate comparisons.\n\tconst readCall = new Map<string, { resolved: string; display: string; declared: ReadRange }>();\n\tconst mutateCallPath = new Map<string, string>();\n\tconst reads: PriorRead[] = [];\n\tlet lastMutateIndex = -1;\n\tlet order = 0;\n\n\tfor (const entry of entries) {\n\t\t// Compaction boundary: everything before it is replaced by a summary in the\n\t\t// live context, so drop the state accumulated so far. Conservative — the\n\t\t// kept tail before the boundary is dropped too — which can only miss a\n\t\t// dedup, never point at content that is no longer in context.\n\t\tif (isCompactionEntry(entry)) {\n\t\t\treadCall.clear();\n\t\t\tmutateCallPath.clear();\n\t\t\treads.length = 0;\n\t\t\tlastMutateIndex = -1;\n\t\t\torder = 0;\n\t\t\tcontinue;\n\t\t}\n\t\tconst m = toMessage(entry);\n\t\tif (!m) continue;\n\t\tconst i = order++;\n\t\tif (m.role === \"assistant\" && Array.isArray(m.content)) {\n\t\t\tfor (const b of m.content) {\n\t\t\t\tif (b.type !== \"toolCall\" || !b.id) continue;\n\t\t\t\tconst raw = b.arguments?.path;\n\t\t\t\tif (typeof raw !== \"string\" || raw.length === 0) continue;\n\t\t\t\tconst resolved = resolvePath(raw);\n\t\t\t\tif (b.name === READ_TOOL) {\n\t\t\t\t\treadCall.set(b.id, { resolved, display: raw, declared: readRangeFromArgs(b.arguments) });\n\t\t\t\t} else if (b.name && MUTATE_TOOLS.has(b.name)) {\n\t\t\t\t\tmutateCallPath.set(b.id, resolved);\n\t\t\t\t}\n\t\t\t}\n\t\t} else if (m.role === \"toolResult\" && !m.isError && m.toolCallId) {\n\t\t\tif (m.toolName === READ_TOOL) {\n\t\t\t\tif (m.toolCallId === opts.currentCallId) continue;\n\t\t\t\tconst info = readCall.get(m.toolCallId);\n\t\t\t\tif (!info || info.resolved !== opts.resolvedPath) continue;\n\t\t\t\tconst text = resultText(m);\n\t\t\t\t// A pointer fetched nothing: the GC excludes it from supersession, so we\n\t\t\t\t// must too (both as a candidate and as a superseder).\n\t\t\t\tif (isDedupPointerText(text)) continue;\n\t\t\t\treads.push({\n\t\t\t\t\tindex: i,\n\t\t\t\t\tdeclared: info.declared,\n\t\t\t\t\tdelivered: deliveredRange(text, info.declared),\n\t\t\t\t\tdisplay: info.display,\n\t\t\t\t});\n\t\t\t} else if (m.toolName && MUTATE_TOOLS.has(m.toolName)) {\n\t\t\t\tif (mutateCallPath.get(m.toolCallId) === opts.resolvedPath) lastMutateIndex = i;\n\t\t\t}\n\t\t}\n\t}\n\n\t// A read survives the GC iff no later mutate and no later read overlaps its\n\t// *declared* range — exactly the GC's own test — so predict it the same way.\n\tconst survivesGc = (r: PriorRead): boolean =>\n\t\tlastMutateIndex <= r.index && !reads.some((o) => o.index > r.index && rangesOverlap(o.declared, r.declared));\n\n\tlet best: PriorRead | null = null;\n\tfor (const r of reads) {\n\t\t// Coverage uses the *delivered* range: a truncated read only holds what it returned.\n\t\tif (!rangeContains(r.delivered, opts.requestedRange)) continue;\n\t\tif (!survivesGc(r)) continue;\n\t\tif (!best || r.index > best.index) best = r;\n\t}\n\tif (!best) return null;\n\treturn { display: best.display, start: best.delivered.start, end: best.delivered.end };\n}\n\n/** Build the pointer text returned in place of a re-fetch. */\nexport function buildDedupPointerText(covering: CoveringRead): string {\n\tconst where =\n\t\tcovering.end === Number.POSITIVE_INFINITY\n\t\t\t? \"the entire file\"\n\t\t\t: covering.end - 1 > covering.start\n\t\t\t\t? `lines ${covering.start}-${covering.end - 1}`\n\t\t\t\t: `line ${covering.start}`;\n\treturn `${DEDUP_POINTER_PREFIX} ${covering.display} (${where}) was already read earlier in this session and has not changed since. Not re-fetched to save tokens — pass a different offset/limit, or edit the file, if you need other or newer content.]`;\n}\n"]}
+{"version":3,"file":"read-dedup.js","sourceRoot":"","sources":["../../../src/core/tools/read-dedup.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAQH,sEAAsE;AACtE,MAAM,CAAC,MAAM,gBAAgB,GAAc,EAAE,KAAK,EAAE,CAAC,EAAE,GAAG,EAAE,MAAM,CAAC,iBAAiB,EAAE,CAAC;AAEvF;;;;GAIG;AACH,MAAM,UAAU,iBAAiB,CAAC,IAAyC,EAAa;IACvF,MAAM,SAAS,GAAG,IAAI,EAAE,MAAM,CAAC;IAC/B,MAAM,QAAQ,GAAG,IAAI,EAAE,KAAK,CAAC;IAC7B,MAAM,MAAM,GAAG,OAAO,SAAS,KAAK,QAAQ,IAAI,MAAM,CAAC,QAAQ,CAAC,SAAS,CAAC,IAAI,SAAS,GAAG,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC;IAC5G,MAAM,GAAG,GACR,OAAO,QAAQ,KAAK,QAAQ,IAAI,MAAM,CAAC,QAAQ,CAAC,QAAQ,CAAC;QACxD,CAAC,CAAC,MAAM,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,QAAQ,CAAC;QAChC,CAAC,CAAC,MAAM,CAAC,iBAAiB,CAAC;IAC7B,OAAO,EAAE,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,CAAC;AAAA,CAC9B;AAED,mDAAmD;AACnD,MAAM,UAAU,aAAa,CAAC,CAAY,EAAE,CAAY,EAAW;IAClE,OAAO,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,GAAG,CAAC;AAAA,CAC1C;AAED,sFAAsF;AACtF,SAAS,aAAa,CAAC,KAAgB,EAAE,KAAgB,EAAW;IACnE,OAAO,KAAK,CAAC,KAAK,IAAI,KAAK,CAAC,KAAK,IAAI,KAAK,CAAC,GAAG,IAAI,KAAK,CAAC,GAAG,CAAC;AAAA,CAC5D;AAED;;;;GAIG;AACH,MAAM,CAAC,MAAM,oBAAoB,GAAG,sBAAsB,CAAC;AAE3D,8DAA8D;AAC9D,MAAM,UAAU,kBAAkB,CAAC,IAAY,EAAW;IACzD,OAAO,IAAI,CAAC,SAAS,EAAE,CAAC,UAAU,CAAC,oBAAoB,CAAC,CAAC;AAAA,CACzD;AAED;;;;;;;;;;;;GAYG;AACH,MAAM,iBAAiB,GAAG,GAAG,CAAC;AAC9B,MAAM,aAAa,GAAG,IAAI,GAAG,EAAkB,CAAC;AAEhD,uDAAuD;AACvD,MAAM,UAAU,eAAe,CAAC,MAAc,EAAE,KAAa,EAAQ;IACpE,IAAI,aAAa,CAAC,IAAI,IAAI,iBAAiB,EAAE,CAAC;QAC7C,MAAM,MAAM,GAAG,aAAa,CAAC,IAAI,EAAE,CAAC,IAAI,EAAE,CAAC,KAAK,CAAC;QACjD,IAAI,MAAM,KAAK,SAAS;YAAE,aAAa,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC;IACxD,CAAC;IACD,aAAa,CAAC,GAAG,CAAC,MAAM,EAAE,KAAK,CAAC,CAAC;AAAA,CACjC;AAED;;;GAGG;AACH,MAAM,UAAU,gBAAgB,CAAC,MAAc,EAAE,KAAa,EAAW;IACxE,MAAM,QAAQ,GAAG,aAAa,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC;IAC3C,OAAO,QAAQ,KAAK,SAAS,IAAI,QAAQ,KAAK,KAAK,CAAC;AAAA,CACpD;AAED,2CAA2C;AAC3C,MAAM,UAAU,eAAe,GAAS;IACvC,aAAa,CAAC,KAAK,EAAE,CAAC;AAAA,CACtB;AAyCD,MAAM,SAAS,GAAG,MAAM,CAAC;AACzB,MAAM,YAAY,GAAG,IAAI,GAAG,CAAC,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC;AAEhD,+EAA6E;AAC7E,SAAS,iBAAiB,CAAC,KAAc,EAAW;IACnD,OAAO,CAAC,CAAC,KAAK,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAK,KAA4B,CAAC,IAAI,KAAK,YAAY,CAAC;AAAA,CACnG;AAED,gFAAgF;AAChF,SAAS,SAAS,CAAC,KAAc,EAAsB;IACtD,IAAI,CAAC,KAAK,IAAI,OAAO,KAAK,KAAK,QAAQ;QAAE,OAAO,IAAI,CAAC;IACrD,MAAM,CAAC,GAAG,KAA8C,CAAC;IACzD,IAAI,CAAC,CAAC,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,KAAK,QAAQ;QAAE,OAAO,CAAC,CAAC,OAAsB,CAAC;IAChF,IAAI,OAAO,CAAC,CAAC,IAAI,KAAK,QAAQ;QAAE,OAAO,CAAgB,CAAC;IACxD,OAAO,IAAI,CAAC;AAAA,CACZ;AAED,SAAS,UAAU,CAAC,CAAc,EAAU;IAC3C,OAAO,CAAC,CAAC,CAAC,OAAO,IAAI,EAAE,CAAC;SACtB,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,MAAM,CAAC;SAChC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,IAAI,EAAE,CAAC;SACxB,IAAI,CAAC,EAAE,CAAC,CAAC;AAAA,CACX;AAED;;;;;;;;;GASG;AACH,SAAS,cAAc,CAAC,IAAY,EAAE,QAAmB,EAAa;IACrE,kFAAkF;IAClF,6EAA6E;IAC7E,yEAAyE;IACzE,MAAM,OAAO,GAAG,IAAI,CAAC,KAAK,CAAC,oDAAoD,CAAC,CAAC;IACjF,IAAI,OAAO,EAAE,CAAC;QACb,MAAM,CAAC,GAAG,MAAM,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC;QAC7B,MAAM,CAAC,GAAG,MAAM,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC;QAC7B,IAAI,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,IAAI,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC;YAAE,OAAO,EAAE,KAAK,EAAE,CAAC,EAAE,GAAG,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC;IACzF,CAAC;IACD,+EAA+E;IAC/E,2DAA2D;IAC3D,IAAI,sCAAsC,CAAC,IAAI,CAAC,IAAI,CAAC;QAAE,OAAO,EAAE,KAAK,EAAE,CAAC,EAAE,GAAG,EAAE,CAAC,EAAE,CAAC;IACnF,OAAO,QAAQ,CAAC;AAAA,CAChB;AAED;;;;;;;;;;;;GAYG;AACH,MAAM,UAAU,gBAAgB,CAAC,OAA2B,EAAE,IAA6B,EAAuB;IAYjH,+EAA6E;IAC7E,MAAM,YAAY,GAAG,IAAI,GAAG,EAAkB,CAAC;IAC/C,MAAM,WAAW,GAAG,CAAC,GAAW,EAAU,EAAE,CAAC;QAC5C,MAAM,GAAG,GAAG,YAAY,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;QAClC,IAAI,GAAG,KAAK,SAAS;YAAE,OAAO,GAAG,CAAC;QAClC,MAAM,QAAQ,GAAG,IAAI,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC;QACvC,YAAY,CAAC,GAAG,CAAC,GAAG,EAAE,QAAQ,CAAC,CAAC;QAChC,OAAO,QAAQ,CAAC;IAAA,CAChB,CAAC;IAEF,gFAAgF;IAChF,6EAA6E;IAC7E,+EAA+E;IAC/E,iEAAiE;IACjE,MAAM,QAAQ,GAAG,IAAI,GAAG,EAAsE,CAAC;IAC/F,MAAM,cAAc,GAAG,IAAI,GAAG,EAAkB,CAAC;IACjD,MAAM,KAAK,GAAgB,EAAE,CAAC;IAC9B,IAAI,eAAe,GAAG,CAAC,CAAC,CAAC;IACzB,8EAA8E;IAC9E,+EAA+E;IAC/E,gFAAgF;IAChF,+EAA+E;IAC/E,mEAAmE;IACnE,IAAI,qBAAqB,GAAG,CAAC,CAAC,CAAC;IAC/B,IAAI,KAAK,GAAG,CAAC,CAAC;IAEd,KAAK,MAAM,KAAK,IAAI,OAAO,EAAE,CAAC;QAC7B,4EAA4E;QAC5E,2EAAyE;QACzE,yEAAuE;QACvE,8DAA8D;QAC9D,IAAI,iBAAiB,CAAC,KAAK,CAAC,EAAE,CAAC;YAC9B,QAAQ,CAAC,KAAK,EAAE,CAAC;YACjB,cAAc,CAAC,KAAK,EAAE,CAAC;YACvB,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC;YACjB,eAAe,GAAG,CAAC,CAAC,CAAC;YACrB,KAAK,GAAG,CAAC,CAAC;YACV,SAAS;QACV,CAAC;QACD,MAAM,CAAC,GAAG,SAAS,CAAC,KAAK,CAAC,CAAC;QAC3B,IAAI,CAAC,CAAC;YAAE,SAAS;QACjB,MAAM,CAAC,GAAG,KAAK,EAAE,CAAC;QAClB,IAAI,CAAC,CAAC,IAAI,KAAK,WAAW,IAAI,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,OAAO,CAAC,EAAE,CAAC;YACxD,KAAK,MAAM,CAAC,IAAI,CAAC,CAAC,OAAO,EAAE,CAAC;gBAC3B,IAAI,CAAC,CAAC,IAAI,KAAK,UAAU,IAAI,CAAC,CAAC,CAAC,EAAE;oBAAE,SAAS;gBAC7C,MAAM,GAAG,GAAG,CAAC,CAAC,SAAS,EAAE,IAAI,CAAC;gBAC9B,IAAI,OAAO,GAAG,KAAK,QAAQ,IAAI,GAAG,CAAC,MAAM,KAAK,CAAC;oBAAE,SAAS;gBAC1D,MAAM,QAAQ,GAAG,WAAW,CAAC,GAAG,CAAC,CAAC;gBAClC,IAAI,CAAC,CAAC,IAAI,KAAK,SAAS,EAAE,CAAC;oBAC1B,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,EAAE,QAAQ,EAAE,OAAO,EAAE,GAAG,EAAE,QAAQ,EAAE,iBAAiB,CAAC,CAAC,CAAC,SAAS,CAAC,EAAE,CAAC,CAAC;gBAC1F,CAAC;qBAAM,IAAI,CAAC,CAAC,IAAI,IAAI,YAAY,CAAC,GAAG,CAAC,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC;oBAC/C,cAAc,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,QAAQ,CAAC,CAAC;gBACpC,CAAC;YACF,CAAC;QACF,CAAC;aAAM,IAAI,CAAC,CAAC,IAAI,KAAK,YAAY,IAAI,CAAC,CAAC,OAAO,IAAI,CAAC,CAAC,UAAU,EAAE,CAAC;YACjE,IAAI,CAAC,CAAC,QAAQ,IAAI,YAAY,CAAC,GAAG,CAAC,CAAC,CAAC,QAAQ,CAAC,IAAI,cAAc,CAAC,GAAG,CAAC,CAAC,CAAC,UAAU,CAAC,KAAK,IAAI,CAAC,YAAY,EAAE,CAAC;gBAC1G,qBAAqB,GAAG,CAAC,CAAC;YAC3B,CAAC;QACF,CAAC;aAAM,IAAI,CAAC,CAAC,IAAI,KAAK,YAAY,IAAI,CAAC,CAAC,CAAC,OAAO,IAAI,CAAC,CAAC,UAAU,EAAE,CAAC;YAClE,IAAI,CAAC,CAAC,QAAQ,KAAK,SAAS,EAAE,CAAC;gBAC9B,IAAI,CAAC,CAAC,UAAU,KAAK,IAAI,CAAC,aAAa;oBAAE,SAAS;gBAClD,MAAM,IAAI,GAAG,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC;gBACxC,IAAI,CAAC,IAAI,IAAI,IAAI,CAAC,QAAQ,KAAK,IAAI,CAAC,YAAY;oBAAE,SAAS;gBAC3D,MAAM,IAAI,GAAG,UAAU,CAAC,CAAC,CAAC,CAAC;gBAC3B,yEAAyE;gBACzE,sDAAsD;gBACtD,IAAI,kBAAkB,CAAC,IAAI,CAAC;oBAAE,SAAS;gBACvC,KAAK,CAAC,IAAI,CAAC;oBACV,KAAK,EAAE,CAAC;oBACR,QAAQ,EAAE,IAAI,CAAC,QAAQ;oBACvB,SAAS,EAAE,cAAc,CAAC,IAAI,EAAE,IAAI,CAAC,QAAQ,CAAC;oBAC9C,OAAO,EAAE,IAAI,CAAC,OAAO;oBACrB,MAAM,EAAE,CAAC,CAAC,UAAU;iBACpB,CAAC,CAAC;YACJ,CAAC;iBAAM,IAAI,CAAC,CAAC,QAAQ,IAAI,YAAY,CAAC,GAAG,CAAC,CAAC,CAAC,QAAQ,CAAC,EAAE,CAAC;gBACvD,IAAI,cAAc,CAAC,GAAG,CAAC,CAAC,CAAC,UAAU,CAAC,KAAK,IAAI,CAAC,YAAY;oBAAE,eAAe,GAAG,CAAC,CAAC;YACjF,CAAC;QACF,CAAC;IACF,CAAC;IAED,4EAA4E;IAC5E,iFAA6E;IAC7E,MAAM,UAAU,GAAG,CAAC,CAAY,EAAW,EAAE,CAC5C,eAAe,IAAI,CAAC,CAAC,KAAK,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,KAAK,IAAI,aAAa,CAAC,CAAC,CAAC,QAAQ,EAAE,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC;IAE9G,gFAAgF;IAChF,0EAA0E;IAC1E,IAAI,qBAAqB,GAAG,CAAC,CAAC;QAAE,OAAO,IAAI,CAAC;IAE5C,IAAI,IAAI,GAAqB,IAAI,CAAC;IAClC,KAAK,MAAM,CAAC,IAAI,KAAK,EAAE,CAAC;QACvB,qFAAqF;QACrF,IAAI,CAAC,aAAa,CAAC,CAAC,CAAC,SAAS,EAAE,IAAI,CAAC,cAAc,CAAC;YAAE,SAAS;QAC/D,IAAI,CAAC,UAAU,CAAC,CAAC,CAAC;YAAE,SAAS;QAC7B,IAAI,CAAC,IAAI,IAAI,CAAC,CAAC,KAAK,GAAG,IAAI,CAAC,KAAK;YAAE,IAAI,GAAG,CAAC,CAAC;IAC7C,CAAC;IACD,IAAI,CAAC,IAAI;QAAE,OAAO,IAAI,CAAC;IACvB,OAAO,EAAE,MAAM,EAAE,IAAI,CAAC,MAAM,EAAE,OAAO,EAAE,IAAI,CAAC,OAAO,EAAE,KAAK,EAAE,IAAI,CAAC,SAAS,CAAC,KAAK,EAAE,GAAG,EAAE,IAAI,CAAC,SAAS,CAAC,GAAG,EAAE,CAAC;AAAA,CAC5G;AAKD,8DAA8D;AAC9D,MAAM,UAAU,qBAAqB,CAAC,QAA6B,EAAU;IAC5E,MAAM,KAAK,GACV,QAAQ,CAAC,GAAG,KAAK,MAAM,CAAC,iBAAiB;QACxC,CAAC,CAAC,iBAAiB;QACnB,CAAC,CAAC,QAAQ,CAAC,GAAG,GAAG,CAAC,GAAG,QAAQ,CAAC,KAAK;YAClC,CAAC,CAAC,SAAS,QAAQ,CAAC,KAAK,IAAI,QAAQ,CAAC,GAAG,GAAG,CAAC,EAAE;YAC/C,CAAC,CAAC,QAAQ,QAAQ,CAAC,KAAK,EAAE,CAAC;IAC9B,OAAO,GAAG,oBAAoB,IAAI,QAAQ,CAAC,OAAO,KAAK,KAAK,+LAA6L,CAAC;AAAA,CAC1P","sourcesContent":["/**\n * Read de-duplication primitives.\n *\n * Two mechanisms share this module:\n *\n * - The post-hoc context GC (`context-gc.ts`), which stubs out a `read` result\n *   once a later overlapping read or an edit/write has superseded it.\n * - The at-call-time guard in the `read` tool, which short-circuits a read whose\n *   requested range is *already fully covered* by an earlier, still-live read in\n *   the current session, returning a pointer instead of re-fetching the file.\n *\n * Both reason about half-open line ranges `[start, end)` (end exclusive), so the\n * range math lives here and cannot drift between them. The guard additionally\n * needs to (a) recognise its own pointer results so they are never treated as\n * content-bearing reads, and (b) walk the session branch to find a covering\n * read — both of which are defined here to keep the tool file lean.\n */\n\n/** Half-open line interval `[start, end)` a read call covers; end is exclusive. */\nexport interface ReadRange {\n\tstart: number;\n\tend: number;\n}\n\n/** A read with no offset/limit covers the whole file (open-ended). */\nexport const WHOLE_FILE_RANGE: ReadRange = { start: 1, end: Number.POSITIVE_INFINITY };\n\n/**\n * Derive the line range a read call covers from its `offset`/`limit` args.\n * Missing offset means \"from line 1\"; missing limit means \"to end of file\"\n * (open-ended, so it overlaps any later read of the same file).\n */\nexport function readRangeFromArgs(args: Record<string, unknown> | undefined): ReadRange {\n\tconst offsetRaw = args?.offset;\n\tconst limitRaw = args?.limit;\n\tconst offset = typeof offsetRaw === \"number\" && Number.isFinite(offsetRaw) && offsetRaw > 0 ? offsetRaw : 1;\n\tconst end =\n\t\ttypeof limitRaw === \"number\" && Number.isFinite(limitRaw)\n\t\t\t? offset + Math.max(0, limitRaw)\n\t\t\t: Number.POSITIVE_INFINITY;\n\treturn { start: offset, end };\n}\n\n/** Whether two half-open line ranges intersect. */\nexport function rangesOverlap(a: ReadRange, b: ReadRange): boolean {\n\treturn a.start < b.end && b.start < a.end;\n}\n\n/** Whether `outer` fully contains `inner` (every line of inner lies within outer). */\nfunction rangeContains(outer: ReadRange, inner: ReadRange): boolean {\n\treturn outer.start <= inner.start && outer.end >= inner.end;\n}\n\n/**\n * Marker prefix for the at-call dedup pointer. Kept stable so the context GC can\n * recognise a pointer result and exclude it from supersession bookkeeping (a\n * pointer fetched no content, so it must not stub the read it points at).\n */\nexport const DEDUP_POINTER_PREFIX = \"[Already in context:\";\n\n/** Whether a tool-result text is an at-call dedup pointer. */\nexport function isDedupPointerText(text: string): boolean {\n\treturn text.trimStart().startsWith(DEDUP_POINTER_PREFIX);\n}\n\n/**\n * Content stamps for reads that have already been delivered.\n *\n * The dedup pointer tells the model the file \"has not changed since\" the read it\n * names. Nothing in the transcript can establish that: an editor, a formatter, a\n * branch switch or a second agent can rewrite the file between two reads without\n * leaving a trace in it. So each delivered read records a stamp of the file it\n * saw, and the pointer is only served when the file still stamps the same.\n *\n * Keyed by tool call id rather than by path, so two sessions sharing a process\n * cannot overwrite each other's observations. Bounded, and a missing entry\n * counts as \"cannot prove it is unchanged\" - the read then simply runs.\n */\nconst MAX_TRACKED_READS = 500;\nconst stampByCallId = new Map<string, string>();\n\n/** Remember the stamp of the file a read delivered. */\nexport function recordReadStamp(callId: string, stamp: string): void {\n\tif (stampByCallId.size >= MAX_TRACKED_READS) {\n\t\tconst oldest = stampByCallId.keys().next().value;\n\t\tif (oldest !== undefined) stampByCallId.delete(oldest);\n\t}\n\tstampByCallId.set(callId, stamp);\n}\n\n/**\n * Whether the file a read delivered still stamps the same. False when no stamp\n * was recorded, so an unprovable case re-reads rather than asserting freshness.\n */\nexport function readStampMatches(callId: string, stamp: string): boolean {\n\tconst recorded = stampByCallId.get(callId);\n\treturn recorded !== undefined && recorded === stamp;\n}\n\n/** Test seam: drop all recorded stamps. */\nexport function clearReadStamps(): void {\n\tstampByCallId.clear();\n}\n\n/** A covering earlier read, described for the pointer message. */\nexport interface CoveringRead {\n\t/** Tool call id of the read being pointed at, used to check its stamp. */\n\tcallId: string;\n\t/** The path as the earlier read spelled it (for a friendly pointer). */\n\tdisplay: string;\n\t/** Delivered range start (1-indexed line). */\n\tstart: number;\n\t/** Delivered range end (exclusive; Infinity for a whole-file read). */\n\tend: number;\n}\n\nexport interface FindCoveringReadOptions {\n\t/** Resolved absolute path of the current read. */\n\tresolvedPath: string;\n\t/** Range the current read is asking for. */\n\trequestedRange: ReadRange;\n\t/** Tool call id of the current read, excluded from candidate/supersession sets. */\n\tcurrentCallId: string;\n\t/** Resolve a raw read-arg path the same way the current read resolved its path. */\n\tresolvePath: (rawPath: string) => string;\n}\n\ninterface ContentBlock {\n\ttype?: string;\n\tid?: string;\n\tname?: string;\n\targuments?: Record<string, unknown>;\n\ttext?: string;\n}\n\ninterface MessageLike {\n\trole?: string;\n\tcontent?: ContentBlock[];\n\ttoolCallId?: string;\n\ttoolName?: string;\n\tisError?: boolean;\n}\n\nconst READ_TOOL = \"read\";\nconst MUTATE_TOOLS = new Set([\"edit\", \"write\"]);\n\n/** A session compaction entry — the boundary that trims the live context. */\nfunction isCompactionEntry(entry: unknown): boolean {\n\treturn !!entry && typeof entry === \"object\" && (entry as { type?: unknown }).type === \"compaction\";\n}\n\n/** Accept either raw `AgentMessage`s or session entries wrapping `.message`. */\nfunction toMessage(entry: unknown): MessageLike | null {\n\tif (!entry || typeof entry !== \"object\") return null;\n\tconst e = entry as { message?: unknown; role?: unknown };\n\tif (e.message && typeof e.message === \"object\") return e.message as MessageLike;\n\tif (typeof e.role === \"string\") return e as MessageLike;\n\treturn null;\n}\n\nfunction resultText(m: MessageLike): string {\n\treturn (m.content ?? [])\n\t\t.filter((c) => c.type === \"text\")\n\t\t.map((c) => c.text ?? \"\")\n\t\t.join(\"\");\n}\n\n/**\n * Recover the range a read result actually *delivered* from its text.\n *\n * A cap-truncated read announces `[Showing lines A-B of N ...]`, so it delivered\n * only `[A, B+1)` even though its args declared a wider range. A read whose first\n * line alone exceeded the byte cap delivered nothing. Any other (untruncated)\n * read delivered its full declared range. The user-`limit` early-stop notice\n * (`[N more lines in file ...]`) is *not* a cap truncation — the declared range\n * was delivered in full — so it falls through to `declared`.\n */\nfunction deliveredRange(text: string, declared: ReadRange): ReadRange {\n\t// The truncation notice is always the trailing `\\n\\n[Showing lines A-B of N ...]`\n\t// clause the read tool appends. Anchor to the end so a `[Showing lines ...]`\n\t// string that merely appears *inside* the file's content can't spoof it.\n\tconst showing = text.match(/\\n\\n\\[Showing lines (\\d+)-(\\d+) of \\d+[^\\]]*\\]\\s*$/);\n\tif (showing) {\n\t\tconst a = Number(showing[1]);\n\t\tconst b = Number(showing[2]);\n\t\tif (Number.isFinite(a) && Number.isFinite(b) && b >= a) return { start: a, end: b + 1 };\n\t}\n\t// First line alone exceeded the byte limit: the whole result *is* that notice,\n\t// so it must start the text. Nothing usable was delivered.\n\tif (/^\\[Line \\d+ is .+ exceeds .+ limit\\./.test(text)) return { start: 1, end: 1 };\n\treturn declared;\n}\n\n/**\n * Find the latest earlier read that (a) is for the same resolved path, (b)\n * actually delivered a range containing the requested range, and (c) is still\n * live in the outgoing context — i.e. the post-hoc GC will not have stubbed it,\n * because no later edit/write and no later overlapping content read supersede\n * it. Returns that read's delivered range for the pointer, or null when the\n * current read must actually run.\n *\n * Deliberately conservative: a truncated earlier read only covers what it\n * delivered, a whole-file read must have been delivered untruncated to count as\n * covering, and any pointer results (which fetched nothing) are ignored on both\n * the candidate and the supersession side.\n */\nexport function findCoveringRead(entries: readonly unknown[], opts: FindCoveringReadOptions): CoveringRead | null {\n\t// `declared` mirrors the range the GC uses for supersession (straight from the\n\t// call args); `delivered` is what the result text shows was actually returned,\n\t// used for coverage. The current call has no result yet, so it never appears.\n\tinterface PriorRead {\n\t\tindex: number;\n\t\tdeclared: ReadRange;\n\t\tdelivered: ReadRange;\n\t\tdisplay: string;\n\t\tcallId: string;\n\t}\n\n\t// Resolve each distinct raw path once — the resolver may hit the filesystem.\n\tconst resolveCache = new Map<string, string>();\n\tconst resolvePath = (raw: string): string => {\n\t\tconst hit = resolveCache.get(raw);\n\t\tif (hit !== undefined) return hit;\n\t\tconst resolved = opts.resolvePath(raw);\n\t\tresolveCache.set(raw, resolved);\n\t\treturn resolved;\n\t};\n\n\t// Single ordered pass. `readCall`/`mutateCallPath` map a call id to its path as\n\t// the call is seen (a toolCall always precedes its result); `reads` collects\n\t// the target path's reads and `lastMutateIndex` its last mutate. `order` gives\n\t// live-context position for the supersession/mutate comparisons.\n\tconst readCall = new Map<string, { resolved: string; display: string; declared: ReadRange }>();\n\tconst mutateCallPath = new Map<string, string>();\n\tconst reads: PriorRead[] = [];\n\tlet lastMutateIndex = -1;\n\t// A *failed* edit/write is not a mutate - the GC is right to ignore it, since\n\t// the file did not change. It is still the one moment the model most needs the\n\t// real bytes: the failure usually means its copy of the text does not match the\n\t// file. Serving a \"not re-fetched, unchanged\" pointer there leaves it retrying\n\t// the same wrong text with no way to see what is actually on disk.\n\tlet lastFailedMutateIndex = -1;\n\tlet order = 0;\n\n\tfor (const entry of entries) {\n\t\t// Compaction boundary: everything before it is replaced by a summary in the\n\t\t// live context, so drop the state accumulated so far. Conservative — the\n\t\t// kept tail before the boundary is dropped too — which can only miss a\n\t\t// dedup, never point at content that is no longer in context.\n\t\tif (isCompactionEntry(entry)) {\n\t\t\treadCall.clear();\n\t\t\tmutateCallPath.clear();\n\t\t\treads.length = 0;\n\t\t\tlastMutateIndex = -1;\n\t\t\torder = 0;\n\t\t\tcontinue;\n\t\t}\n\t\tconst m = toMessage(entry);\n\t\tif (!m) continue;\n\t\tconst i = order++;\n\t\tif (m.role === \"assistant\" && Array.isArray(m.content)) {\n\t\t\tfor (const b of m.content) {\n\t\t\t\tif (b.type !== \"toolCall\" || !b.id) continue;\n\t\t\t\tconst raw = b.arguments?.path;\n\t\t\t\tif (typeof raw !== \"string\" || raw.length === 0) continue;\n\t\t\t\tconst resolved = resolvePath(raw);\n\t\t\t\tif (b.name === READ_TOOL) {\n\t\t\t\t\treadCall.set(b.id, { resolved, display: raw, declared: readRangeFromArgs(b.arguments) });\n\t\t\t\t} else if (b.name && MUTATE_TOOLS.has(b.name)) {\n\t\t\t\t\tmutateCallPath.set(b.id, resolved);\n\t\t\t\t}\n\t\t\t}\n\t\t} else if (m.role === \"toolResult\" && m.isError && m.toolCallId) {\n\t\t\tif (m.toolName && MUTATE_TOOLS.has(m.toolName) && mutateCallPath.get(m.toolCallId) === opts.resolvedPath) {\n\t\t\t\tlastFailedMutateIndex = i;\n\t\t\t}\n\t\t} else if (m.role === \"toolResult\" && !m.isError && m.toolCallId) {\n\t\t\tif (m.toolName === READ_TOOL) {\n\t\t\t\tif (m.toolCallId === opts.currentCallId) continue;\n\t\t\t\tconst info = readCall.get(m.toolCallId);\n\t\t\t\tif (!info || info.resolved !== opts.resolvedPath) continue;\n\t\t\t\tconst text = resultText(m);\n\t\t\t\t// A pointer fetched nothing: the GC excludes it from supersession, so we\n\t\t\t\t// must too (both as a candidate and as a superseder).\n\t\t\t\tif (isDedupPointerText(text)) continue;\n\t\t\t\treads.push({\n\t\t\t\t\tindex: i,\n\t\t\t\t\tdeclared: info.declared,\n\t\t\t\t\tdelivered: deliveredRange(text, info.declared),\n\t\t\t\t\tdisplay: info.display,\n\t\t\t\t\tcallId: m.toolCallId,\n\t\t\t\t});\n\t\t\t} else if (m.toolName && MUTATE_TOOLS.has(m.toolName)) {\n\t\t\t\tif (mutateCallPath.get(m.toolCallId) === opts.resolvedPath) lastMutateIndex = i;\n\t\t\t}\n\t\t}\n\t}\n\n\t// A read survives the GC iff no later mutate and no later read overlaps its\n\t// *declared* range — exactly the GC's own test — so predict it the same way.\n\tconst survivesGc = (r: PriorRead): boolean =>\n\t\tlastMutateIndex <= r.index && !reads.some((o) => o.index > r.index && rangesOverlap(o.declared, r.declared));\n\n\t// Nothing earlier can be trusted to satisfy this read once an edit against this\n\t// path has failed: let the read run and hand the model the current bytes.\n\tif (lastFailedMutateIndex > -1) return null;\n\n\tlet best: PriorRead | null = null;\n\tfor (const r of reads) {\n\t\t// Coverage uses the *delivered* range: a truncated read only holds what it returned.\n\t\tif (!rangeContains(r.delivered, opts.requestedRange)) continue;\n\t\tif (!survivesGc(r)) continue;\n\t\tif (!best || r.index > best.index) best = r;\n\t}\n\tif (!best) return null;\n\treturn { callId: best.callId, display: best.display, start: best.delivered.start, end: best.delivered.end };\n}\n\n/** The parts of a covering read the pointer message renders. */\nexport type CoveringReadDisplay = Pick<CoveringRead, \"display\" | \"start\" | \"end\">;\n\n/** Build the pointer text returned in place of a re-fetch. */\nexport function buildDedupPointerText(covering: CoveringReadDisplay): string {\n\tconst where =\n\t\tcovering.end === Number.POSITIVE_INFINITY\n\t\t\t? \"the entire file\"\n\t\t\t: covering.end - 1 > covering.start\n\t\t\t\t? `lines ${covering.start}-${covering.end - 1}`\n\t\t\t\t: `line ${covering.start}`;\n\treturn `${DEDUP_POINTER_PREFIX} ${covering.display} (${where}) was already read earlier in this session and has not changed since. Not re-fetched to save tokens — pass a different offset/limit, or edit the file, if you need other or newer content.]`;\n}\n"]}

package/dist/core/tools/read.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"read.d.ts","sourceRoot":"","sources":["../../../src/core/tools/read.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,iCAAiC,CAAC;AAKjE,OAAO,EAAE,KAAK,MAAM,EAAE,IAAI,EAAE,MAAM,SAAS,CAAC;AAQ5C,OAAO,KAAK,EAAE,cAAc,EAA2B,MAAM,wBAAwB,CAAC;AAKtF,OAAO,EAAoD,KAAK,gBAAgB,EAAgB,MAAM,eAAe,CAAC;AAEtH,QAAA,MAAM,UAAU;;;;EAId,CAAC;AAEH,MAAM,MAAM,aAAa,GAAG,MAAM,CAAC,OAAO,UAAU,CAAC,CAAC;AAEtD,MAAM,WAAW,eAAe;IAC/B,UAAU,CAAC,EAAE,gBAAgB,CAAC;CAC9B;AA2BD;;;GAGG;AACH,MAAM,WAAW,cAAc;IAC9B,qCAAqC;IACrC,QAAQ,EAAE,CAAC,YAAY,EAAE,MAAM,KAAK,OAAO,CAAC,MAAM,CAAC,CAAC;IACpD,+CAA+C;IAC/C,MAAM,EAAE,CAAC,YAAY,EAAE,MAAM,KAAK,OAAO,CAAC,IAAI,CAAC,CAAC;IAChD,sEAAsE;IACtE,mBAAmB,CAAC,EAAE,CAAC,YAAY,EAAE,MAAM,KAAK,OAAO,CAAC,MAAM,GAAG,IAAI,GAAG,SAAS,CAAC,CAAC;CACnF;AAQD,MAAM,WAAW,eAAe;IAC/B,oEAAoE;IACpE,gBAAgB,CAAC,EAAE,OAAO,CAAC;IAC3B,oEAAoE;IACpE,UAAU,CAAC,EAAE,cAAc,CAAC;IAC5B,mFAAmF;IACnF,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,mFAAmF;IACnF,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB;;;;;OAKG;IACH,UAAU,CAAC,EAAE,OAAO,CAAC;CACrB;AAmJD,wBAAgB,wBAAwB,CACvC,GAAG,EAAE,MAAM,EACX,OAAO,CAAC,EAAE,eAAe,GACvB,cAAc,CAAC,OAAO,UAAU,EAAE,eAAe,GAAG,SAAS,CAAC,CA4LhE;AAED,wBAAgB,cAAc,CAAC,GAAG,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,eAAe,GAAG,SAAS,CAAC,OAAO,UAAU,CAAC,CAEnG","sourcesContent":["import { basename, dirname, isAbsolute, relative, resolve as resolvePath, sep } from \"node:path\";\nimport type { AgentTool } from \"@kolisachint/hoocode-agent-core\";\nimport type { Api, ImageContent, Model, TextContent } from \"@kolisachint/hoocode-ai\";\nimport { Text } from \"@kolisachint/hoocode-tui\";\nimport { constants } from \"fs\";\nimport { access as fsAccess, readFile as fsReadFile } from \"fs/promises\";\nimport { type Static, Type } from \"typebox\";\nimport { getReadmePath } from \"../../config.js\";\nimport { keyHint, keyText } from \"../../modes/interactive/components/keybinding-hints.js\";\nimport { renderReadOutput } from \"../../modes/interactive/components/read-output.js\";\nimport { getLanguageFromPath, highlightCode, messageLabel, type Theme } from \"../../modes/interactive/theme/theme.js\";\nimport { formatDimensionNote, resizeImage } from \"../../utils/image-resize.js\";\nimport { detectSupportedImageMimeTypeFromFile } from \"../../utils/mime.js\";\nimport { formatPathRelativeToCwdOrAbsolute } from \"../../utils/paths.js\";\nimport type { ToolDefinition, ToolRenderResultOptions } from \"../extensions/types.js\";\nimport { resolveReadPath } from \"./path-utils.js\";\nimport { buildDedupPointerText, findCoveringRead, readRangeFromArgs } from \"./read-dedup.js\";\nimport { getTextOutput, invalidArgText, replaceTabs, shortenPath, str } from \"./render-utils.js\";\nimport { wrapToolDefinition } from \"./tool-definition-wrapper.js\";\nimport { DEFAULT_MAX_BYTES, DEFAULT_MAX_LINES, formatSize, type TruncationResult, truncateHead } from \"./truncate.js\";\n\nconst readSchema = Type.Object({\n\tpath: Type.String({ description: \"Path to the file to read (relative or absolute)\" }),\n\toffset: Type.Optional(Type.Number({ description: \"Line number to start reading from (1-indexed)\" })),\n\tlimit: Type.Optional(Type.Number({ description: \"Maximum number of lines to read\" })),\n});\n\nexport type ReadToolInput = Static<typeof readSchema>;\n\nexport interface ReadToolDetails {\n\ttruncation?: TruncationResult;\n}\n\ninterface CompactReadClassification {\n\tkind: \"docs\" | \"resource\" | \"skill\";\n\tlabel: string;\n}\n\nconst COMPACT_RESOURCE_FILE_NAMES = new Set([\"AGENTS.md\", \"AGENTS.MD\", \"CLAUDE.md\", \"CLAUDE.MD\"]);\n\n/**\n * Structured/binary document formats. These are container/binary formats (OOXML\n * is a zip, PDF is binary), so a plain utf-8 read yields garbage — read\n * short-circuits with a note instead of dumping bytes into the context.\n */\nconst DOC_FORMAT_EXTENSIONS = new Map<string, string>([\n\t[\".docx\", \"Word (OOXML)\"],\n\t[\".xlsx\", \"Excel (OOXML)\"],\n\t[\".pptx\", \"PowerPoint (OOXML)\"],\n\t[\".pdf\", \"PDF\"],\n]);\n\nfunction getDocFormatHint(absolutePath: string): string | undefined {\n\tconst dot = absolutePath.lastIndexOf(\".\");\n\tif (dot === -1) return undefined;\n\treturn DOC_FORMAT_EXTENSIONS.get(absolutePath.slice(dot).toLowerCase());\n}\n\n/**\n * Pluggable operations for the read tool.\n * Override these to delegate file reading to remote systems (for example SSH).\n */\nexport interface ReadOperations {\n\t/** Read file contents as a Buffer */\n\treadFile: (absolutePath: string) => Promise<Buffer>;\n\t/** Check if file is readable (throw if not) */\n\taccess: (absolutePath: string) => Promise<void>;\n\t/** Detect image MIME type, return null or undefined for non-images */\n\tdetectImageMimeType?: (absolutePath: string) => Promise<string | null | undefined>;\n}\n\nconst defaultReadOperations: ReadOperations = {\n\treadFile: (path) => fsReadFile(path),\n\taccess: (path) => fsAccess(path, constants.R_OK),\n\tdetectImageMimeType: detectSupportedImageMimeTypeFromFile,\n};\n\nexport interface ReadToolOptions {\n\t/** Whether to auto-resize images to 2000x2000 max. Default: true */\n\tautoResizeImages?: boolean;\n\t/** Custom operations for file reading. Default: local filesystem */\n\toperations?: ReadOperations;\n\t/** Byte cap on the returned text before truncation. Default: DEFAULT_MAX_BYTES. */\n\tmaxOutputBytes?: number;\n\t/** Line cap on the returned text before truncation. Default: DEFAULT_MAX_LINES. */\n\tmaxOutputLines?: number;\n\t/**\n\t * When true, a text read whose requested line range is already fully covered\n\t * by an earlier, still-live read in the current session short-circuits with a\n\t * pointer instead of re-fetching the file. Complements the post-hoc context\n\t * GC; gated on the same setting. Default: false.\n\t */\n\tdedupReads?: boolean;\n}\n\ntype ReadRenderArgs = { path?: string; file_path?: string; offset?: number; limit?: number };\n\nfunction formatReadLineRange(args: ReadRenderArgs | undefined, theme: Theme): string {\n\tif (args?.offset === undefined && args?.limit === undefined) return \"\";\n\tconst startLine = args.offset ?? 1;\n\tconst endLine = args.limit !== undefined ? startLine + args.limit - 1 : \"\";\n\treturn theme.fg(\"warning\", `:${startLine}${endLine ? `-${endLine}` : \"\"}`);\n}\n\nfunction formatReadCall(args: ReadRenderArgs | undefined, theme: Theme): string {\n\tconst rawPath = str(args?.file_path ?? args?.path);\n\tconst path = rawPath !== null ? shortenPath(rawPath) : null;\n\tconst invalidArg = invalidArgText(theme);\n\tconst pathDisplay = path === null ? invalidArg : path ? theme.fg(\"accent\", path) : theme.fg(\"toolOutput\", \"...\");\n\treturn `${theme.fg(\"toolTitle\", theme.bold(\"read\"))} ${pathDisplay}${formatReadLineRange(args, theme)}`;\n}\n\nfunction trimTrailingEmptyLines(lines: string[]): string[] {\n\tlet end = lines.length;\n\twhile (end > 0 && lines[end - 1] === \"\") {\n\t\tend--;\n\t}\n\treturn lines.slice(0, end);\n}\n\nfunction getNonVisionImageNote(model: Model<Api> | undefined): string | undefined {\n\tif (!model || model.input.includes(\"image\")) {\n\t\treturn undefined;\n\t}\n\treturn \"[Current model does not support images. The image will be omitted from this request.]\";\n}\n\nfunction toPosixPath(filePath: string): string {\n\treturn filePath.split(sep).join(\"/\");\n}\n\nfunction getHooCodeDocsClassification(absolutePath: string): CompactReadClassification | undefined {\n\tconst packageRoot = dirname(getReadmePath());\n\tconst relativePath = relative(resolvePath(packageRoot), resolvePath(absolutePath));\n\tif (\n\t\trelativePath === \"\" ||\n\t\trelativePath === \"..\" ||\n\t\trelativePath.startsWith(`..${sep}`) ||\n\t\tisAbsolute(relativePath)\n\t) {\n\t\treturn undefined;\n\t}\n\n\tconst label = toPosixPath(relativePath);\n\tif (label === \"README.md\" || label.startsWith(\"docs/\") || label.startsWith(\"examples/\")) {\n\t\treturn { kind: \"docs\", label };\n\t}\n\treturn undefined;\n}\n\nfunction getCompactReadClassification(\n\targs: ReadRenderArgs | undefined,\n\tcwd: string,\n): CompactReadClassification | undefined {\n\tconst rawPath = str(args?.file_path ?? args?.path);\n\tif (!rawPath) return undefined;\n\n\tconst absolutePath = resolveReadPath(rawPath, cwd);\n\tconst fileName = basename(absolutePath);\n\tif (fileName === \"SKILL.md\") {\n\t\treturn { kind: \"skill\", label: basename(dirname(absolutePath)) || fileName };\n\t}\n\n\tconst docsClassification = getHooCodeDocsClassification(absolutePath);\n\tif (docsClassification) return docsClassification;\n\n\tif (COMPACT_RESOURCE_FILE_NAMES.has(fileName)) {\n\t\treturn { kind: \"resource\", label: formatPathRelativeToCwdOrAbsolute(absolutePath, cwd) };\n\t}\n\n\treturn undefined;\n}\n\nfunction formatCompactReadCall(\n\tclassification: CompactReadClassification,\n\targs: ReadRenderArgs | undefined,\n\ttheme: Theme,\n): string {\n\tconst expandHint = theme.fg(\"dim\", ` (${keyText(\"app.tools.expand\")} to expand)`);\n\tif (classification.kind === \"skill\") {\n\t\treturn (\n\t\t\t`${messageLabel(\"skill\")} ` +\n\t\t\ttheme.fg(\"customMessageText\", classification.label) +\n\t\t\tformatReadLineRange(args, theme) +\n\t\t\texpandHint\n\t\t);\n\t}\n\n\treturn (\n\t\ttheme.fg(\"toolTitle\", theme.bold(`read ${classification.kind}`)) +\n\t\t\" \" +\n\t\ttheme.fg(\"accent\", classification.label) +\n\t\tformatReadLineRange(args, theme) +\n\t\texpandHint\n\t);\n}\n\nfunction formatReadResult(\n\targs: ReadRenderArgs | undefined,\n\tresult: { content: (TextContent | ImageContent)[]; details?: ReadToolDetails },\n\toptions: ToolRenderResultOptions,\n\ttheme: Theme,\n\tshowImages: boolean,\n\tcwd: string,\n\tisError: boolean,\n): string {\n\tif (!options.expanded && !isError && getCompactReadClassification(args, cwd)) {\n\t\treturn \"\";\n\t}\n\n\tconst rawPath = str(args?.file_path ?? args?.path);\n\tconst output = getTextOutput(result, showImages);\n\tconst lang = rawPath ? getLanguageFromPath(rawPath) : undefined;\n\tconst renderedLines = lang ? highlightCode(replaceTabs(output), lang) : output.split(\"\\n\");\n\tconst lines = trimTrailingEmptyLines(renderedLines);\n\tconst maxLines = options.expanded ? lines.length : 10;\n\tconst displayLines = lines.slice(0, maxLines);\n\tconst remaining = lines.length - maxLines;\n\t// Calculate starting line number from args\n\tconst startLine = (args?.offset as number) ?? 1;\n\t// Use the new renderReadOutput function for line numbers\n\tconst numberedLines = renderReadOutput(displayLines.join(\"\\n\"), { startLine });\n\tlet text = `\\n${numberedLines.join(\"\\n\")}`;\n\tif (remaining > 0) {\n\t\ttext += `${theme.fg(\"muted\", `\\n... (${remaining} more lines,`)} ${keyHint(\"app.tools.expand\", \"to expand\")})`;\n\t}\n\n\tconst truncation = result.details?.truncation;\n\tif (truncation?.truncated) {\n\t\tif (truncation.firstLineExceedsLimit) {\n\t\t\ttext += `\\n${theme.fg(\"warning\", `[First line exceeds ${formatSize(truncation.maxBytes ?? DEFAULT_MAX_BYTES)} limit]`)}`;\n\t\t} else if (truncation.truncatedBy === \"lines\") {\n\t\t\ttext += `\\n${theme.fg(\"warning\", `[Truncated: showing ${truncation.outputLines} of ${truncation.totalLines} lines (${truncation.maxLines ?? DEFAULT_MAX_LINES} line limit)]`)}`;\n\t\t} else {\n\t\t\ttext += `\\n${theme.fg(\"warning\", `[Truncated: ${truncation.outputLines} lines shown (${formatSize(truncation.maxBytes ?? DEFAULT_MAX_BYTES)} limit)]`)}`;\n\t\t}\n\t}\n\treturn text;\n}\n\nexport function createReadToolDefinition(\n\tcwd: string,\n\toptions?: ReadToolOptions,\n): ToolDefinition<typeof readSchema, ReadToolDetails | undefined> {\n\tconst autoResizeImages = options?.autoResizeImages ?? true;\n\tconst ops = options?.operations ?? defaultReadOperations;\n\tconst maxBytes = options?.maxOutputBytes ?? DEFAULT_MAX_BYTES;\n\tconst maxLines = options?.maxOutputLines ?? DEFAULT_MAX_LINES;\n\tconst dedupReads = options?.dedupReads ?? false;\n\treturn {\n\t\tname: \"read\",\n\t\tlabel: \"read\",\n\t\tdescription: `Read the contents of a file. Supports text files and images (jpg, png, gif, webp). Images are sent as attachments. For text files, output is truncated to ${maxLines} lines or ${Math.round(maxBytes / 1024)}KB (whichever is hit first). Use offset/limit for large files. When you need the full file, continue with offset until complete.`,\n\t\tpromptSnippet: \"Read file contents\",\n\t\t// No promptGuidelines: \"use read instead of cat/sed\" is already covered by the\n\t\t// file-exploration guideline buildSystemPrompt emits, and again by bash's own\n\t\t// snippet. Three copies of one rule.\n\t\tparameters: readSchema,\n\t\tasync execute(\n\t\t\ttoolCallId,\n\t\t\t{ path, offset, limit }: { path: string; offset?: number; limit?: number },\n\t\t\tsignal?: AbortSignal,\n\t\t\t_onUpdate?,\n\t\t\tctx?,\n\t\t) {\n\t\t\tconst absolutePath = resolveReadPath(path, cwd);\n\t\t\treturn new Promise<{ content: (TextContent | ImageContent)[]; details: ReadToolDetails | undefined }>(\n\t\t\t\t(resolve, reject) => {\n\t\t\t\t\tif (signal?.aborted) {\n\t\t\t\t\t\treject(new Error(\"Operation aborted\"));\n\t\t\t\t\t\treturn;\n\t\t\t\t\t}\n\t\t\t\t\tlet aborted = false;\n\t\t\t\t\tconst onAbort = () => {\n\t\t\t\t\t\taborted = true;\n\t\t\t\t\t\treject(new Error(\"Operation aborted\"));\n\t\t\t\t\t};\n\t\t\t\t\tsignal?.addEventListener(\"abort\", onAbort, { once: true });\n\n\t\t\t\t\t(async () => {\n\t\t\t\t\t\ttry {\n\t\t\t\t\t\t\t// Check if file exists and is readable.\n\t\t\t\t\t\t\tawait ops.access(absolutePath);\n\t\t\t\t\t\t\tif (aborted) return;\n\t\t\t\t\t\t\tconst mimeType = ops.detectImageMimeType ? await ops.detectImageMimeType(absolutePath) : undefined;\n\t\t\t\t\t\t\tlet content: (TextContent | ImageContent)[];\n\t\t\t\t\t\t\tlet details: ReadToolDetails | undefined;\n\t\t\t\t\t\t\tconst nonVisionImageNote = getNonVisionImageNote(ctx?.model);\n\t\t\t\t\t\t\t// At-call dedup: for plain-text reads, if an earlier still-live read\n\t\t\t\t\t\t\t// in this session already covers the requested range, point at it\n\t\t\t\t\t\t\t// instead of re-fetching. Skipped for images and structured docs.\n\t\t\t\t\t\t\tconst dedupCovering =\n\t\t\t\t\t\t\t\tdedupReads && !mimeType && !getDocFormatHint(absolutePath) && ctx?.sessionManager\n\t\t\t\t\t\t\t\t\t? findCoveringRead(ctx.sessionManager.getBranch(), {\n\t\t\t\t\t\t\t\t\t\t\tresolvedPath: absolutePath,\n\t\t\t\t\t\t\t\t\t\t\trequestedRange: readRangeFromArgs({ offset, limit }),\n\t\t\t\t\t\t\t\t\t\t\tcurrentCallId: toolCallId,\n\t\t\t\t\t\t\t\t\t\t\tresolvePath: (raw) => resolveReadPath(raw, cwd),\n\t\t\t\t\t\t\t\t\t\t})\n\t\t\t\t\t\t\t\t\t: null;\n\t\t\t\t\t\t\tif (dedupCovering) {\n\t\t\t\t\t\t\t\tcontent = [{ type: \"text\", text: buildDedupPointerText(dedupCovering) }];\n\t\t\t\t\t\t\t} else if (mimeType) {\n\t\t\t\t\t\t\t\t// Read image as binary.\n\t\t\t\t\t\t\t\tconst buffer = await ops.readFile(absolutePath);\n\t\t\t\t\t\t\t\tconst base64 = buffer.toString(\"base64\");\n\t\t\t\t\t\t\t\tif (autoResizeImages) {\n\t\t\t\t\t\t\t\t\t// Resize image if needed before sending it back to the model.\n\t\t\t\t\t\t\t\t\tconst resized = await resizeImage({ type: \"image\", data: base64, mimeType });\n\t\t\t\t\t\t\t\t\tif (!resized) {\n\t\t\t\t\t\t\t\t\t\tlet textNote = `Read image file [${mimeType}]\\n[Image omitted: could not be resized below the inline image size limit.]`;\n\t\t\t\t\t\t\t\t\t\tif (nonVisionImageNote) textNote += `\\n${nonVisionImageNote}`;\n\t\t\t\t\t\t\t\t\t\tcontent = [{ type: \"text\", text: textNote }];\n\t\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\t\tconst dimensionNote = formatDimensionNote(resized);\n\t\t\t\t\t\t\t\t\t\tlet textNote = `Read image file [${resized.mimeType}]`;\n\t\t\t\t\t\t\t\t\t\tif (dimensionNote) textNote += `\\n${dimensionNote}`;\n\t\t\t\t\t\t\t\t\t\tif (nonVisionImageNote) textNote += `\\n${nonVisionImageNote}`;\n\t\t\t\t\t\t\t\t\t\tcontent = [\n\t\t\t\t\t\t\t\t\t\t\t{ type: \"text\", text: textNote },\n\t\t\t\t\t\t\t\t\t\t\t{ type: \"image\", data: resized.data, mimeType: resized.mimeType },\n\t\t\t\t\t\t\t\t\t\t];\n\t\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\tlet textNote = `Read image file [${mimeType}]`;\n\t\t\t\t\t\t\t\t\tif (nonVisionImageNote) textNote += `\\n${nonVisionImageNote}`;\n\t\t\t\t\t\t\t\t\tcontent = [\n\t\t\t\t\t\t\t\t\t\t{ type: \"text\", text: textNote },\n\t\t\t\t\t\t\t\t\t\t{ type: \"image\", data: base64, mimeType },\n\t\t\t\t\t\t\t\t\t];\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t} else if (getDocFormatHint(absolutePath)) {\n\t\t\t\t\t\t\t\t// Structured/binary document: a utf-8 read is meaningless, and dumping\n\t\t\t\t\t\t\t\t// the bytes would burn context for nothing. Stop here and let the model\n\t\t\t\t\t\t\t\t// choose an extraction path (bash + an appropriate converter).\n\t\t\t\t\t\t\t\tconst label = getDocFormatHint(absolutePath);\n\t\t\t\t\t\t\t\tcontent = [\n\t\t\t\t\t\t\t\t\t{\n\t\t\t\t\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\t\t\t\t\ttext: `[${label} document — not plain text; read cannot show it. Extract it with a converter via bash if you need its contents.]`,\n\t\t\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\t\t];\n\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t// Read text content.\n\t\t\t\t\t\t\t\tconst buffer = await ops.readFile(absolutePath);\n\t\t\t\t\t\t\t\tconst textContent = buffer.toString(\"utf-8\");\n\t\t\t\t\t\t\t\tconst allLines = textContent.split(\"\\n\");\n\t\t\t\t\t\t\t\tconst totalFileLines = allLines.length;\n\t\t\t\t\t\t\t\t// Apply offset if specified. Convert from 1-indexed input to 0-indexed array access.\n\t\t\t\t\t\t\t\tconst startLine = offset ? Math.max(0, offset - 1) : 0;\n\t\t\t\t\t\t\t\tconst startLineDisplay = startLine + 1;\n\t\t\t\t\t\t\t\t// Check if offset is out of bounds.\n\t\t\t\t\t\t\t\tif (startLine >= allLines.length) {\n\t\t\t\t\t\t\t\t\tthrow new Error(`Offset ${offset} is beyond end of file (${allLines.length} lines total)`);\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\tlet selectedContent: string;\n\t\t\t\t\t\t\t\tlet userLimitedLines: number | undefined;\n\t\t\t\t\t\t\t\t// If limit is specified by the user, honor it first. Otherwise truncateHead decides.\n\t\t\t\t\t\t\t\tif (limit !== undefined) {\n\t\t\t\t\t\t\t\t\tconst endLine = Math.min(startLine + limit, allLines.length);\n\t\t\t\t\t\t\t\t\tselectedContent = allLines.slice(startLine, endLine).join(\"\\n\");\n\t\t\t\t\t\t\t\t\tuserLimitedLines = endLine - startLine;\n\t\t\t\t\t\t\t\t} else if (startLine === 0) {\n\t\t\t\t\t\t\t\t\t// Whole-file read (common case): reuse the original string instead of\n\t\t\t\t\t\t\t\t\t// rebuilding it from the split lines.\n\t\t\t\t\t\t\t\t\tselectedContent = textContent;\n\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\tselectedContent = allLines.slice(startLine).join(\"\\n\");\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\t// Apply truncation, respecting both line and byte limits.\n\t\t\t\t\t\t\t\t// The text is handed over byte-for-byte. Whatever the model reads here is\n\t\t\t\t\t\t\t\t// what it sends back as an edit's `oldText`, so altering it (the trailing\n\t\t\t\t\t\t\t\t// whitespace strip that used to run at this point) makes the model's own\n\t\t\t\t\t\t\t\t// faithful copy fail to match the file it came from.\n\t\t\t\t\t\t\t\tconst truncation = truncateHead(selectedContent, { maxBytes, maxLines });\n\t\t\t\t\t\t\t\tlet outputText: string;\n\t\t\t\t\t\t\t\tif (truncation.firstLineExceedsLimit) {\n\t\t\t\t\t\t\t\t\t// First line alone exceeds the byte limit. Point the model at a bash fallback.\n\t\t\t\t\t\t\t\t\tconst firstLineSize = formatSize(Buffer.byteLength(allLines[startLine], \"utf-8\"));\n\t\t\t\t\t\t\t\t\toutputText = `[Line ${startLineDisplay} is ${firstLineSize}, exceeds ${formatSize(maxBytes)} limit. Use bash: sed -n '${startLineDisplay}p' ${path} | head -c ${maxBytes}]`;\n\t\t\t\t\t\t\t\t\tdetails = { truncation };\n\t\t\t\t\t\t\t\t} else if (truncation.truncated) {\n\t\t\t\t\t\t\t\t\t// Truncation occurred. Build an actionable continuation notice.\n\t\t\t\t\t\t\t\t\tconst endLineDisplay = startLineDisplay + truncation.outputLines - 1;\n\t\t\t\t\t\t\t\t\tconst nextOffset = endLineDisplay + 1;\n\t\t\t\t\t\t\t\t\toutputText = truncation.content;\n\t\t\t\t\t\t\t\t\tif (truncation.truncatedBy === \"lines\") {\n\t\t\t\t\t\t\t\t\t\toutputText += `\\n\\n[Showing lines ${startLineDisplay}-${endLineDisplay} of ${totalFileLines}. Use offset=${nextOffset} to continue.]`;\n\t\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\t\toutputText += `\\n\\n[Showing lines ${startLineDisplay}-${endLineDisplay} of ${totalFileLines} (${formatSize(maxBytes)} limit). Use offset=${nextOffset} to continue.]`;\n\t\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\t\tdetails = { truncation };\n\t\t\t\t\t\t\t\t} else if (userLimitedLines !== undefined && startLine + userLimitedLines < allLines.length) {\n\t\t\t\t\t\t\t\t\t// User-specified limit stopped early, but the file still has more content.\n\t\t\t\t\t\t\t\t\tconst remaining = allLines.length - (startLine + userLimitedLines);\n\t\t\t\t\t\t\t\t\tconst nextOffset = startLine + userLimitedLines + 1;\n\t\t\t\t\t\t\t\t\toutputText = `${truncation.content}\\n\\n[${remaining} more lines in file. Use offset=${nextOffset} to continue.]`;\n\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\t// No truncation and no remaining user-limited content.\n\t\t\t\t\t\t\t\t\toutputText = truncation.content;\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\tcontent = [{ type: \"text\", text: outputText }];\n\t\t\t\t\t\t\t}\n\n\t\t\t\t\t\t\tif (aborted) return;\n\t\t\t\t\t\t\tsignal?.removeEventListener(\"abort\", onAbort);\n\t\t\t\t\t\t\tresolve({ content, details });\n\t\t\t\t\t\t} catch (error) {\n\t\t\t\t\t\t\tsignal?.removeEventListener(\"abort\", onAbort);\n\t\t\t\t\t\t\tif (!aborted) reject(error);\n\t\t\t\t\t\t}\n\t\t\t\t\t})();\n\t\t\t\t},\n\t\t\t);\n\t\t},\n\t\trenderCall(args, theme, context) {\n\t\t\tconst text = (context.lastComponent as Text | undefined) ?? new Text(\"\", 0, 0);\n\t\t\tconst classification = !context.expanded ? getCompactReadClassification(args, context.cwd) : undefined;\n\t\t\ttext.setText(\n\t\t\t\tclassification ? formatCompactReadCall(classification, args, theme) : formatReadCall(args, theme),\n\t\t\t);\n\t\t\treturn text;\n\t\t},\n\t\trenderResult(result, options, theme, context) {\n\t\t\tconst text = (context.lastComponent as Text | undefined) ?? new Text(\"\", 0, 0);\n\t\t\ttext.setText(\n\t\t\t\tformatReadResult(context.args, result, options, theme, context.showImages, context.cwd, context.isError),\n\t\t\t);\n\t\t\treturn text;\n\t\t},\n\t};\n}\n\nexport function createReadTool(cwd: string, options?: ReadToolOptions): AgentTool<typeof readSchema> {\n\treturn wrapToolDefinition(createReadToolDefinition(cwd, options));\n}\n"]}
+{"version":3,"file":"read.d.ts","sourceRoot":"","sources":["../../../src/core/tools/read.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,iCAAiC,CAAC;AAMjE,OAAO,EAAE,KAAK,MAAM,EAAE,IAAI,EAAE,MAAM,SAAS,CAAC;AAQ5C,OAAO,KAAK,EAAE,cAAc,EAA2B,MAAM,wBAAwB,CAAC;AAWtF,OAAO,EAAoD,KAAK,gBAAgB,EAAgB,MAAM,eAAe,CAAC;AAEtH,QAAA,MAAM,UAAU;;;;EAId,CAAC;AAEH,MAAM,MAAM,aAAa,GAAG,MAAM,CAAC,OAAO,UAAU,CAAC,CAAC;AAEtD,MAAM,WAAW,eAAe;IAC/B,UAAU,CAAC,EAAE,gBAAgB,CAAC;CAC9B;AA2BD;;;GAGG;AACH,MAAM,WAAW,cAAc;IAC9B,qCAAqC;IACrC,QAAQ,EAAE,CAAC,YAAY,EAAE,MAAM,KAAK,OAAO,CAAC,MAAM,CAAC,CAAC;IACpD,+CAA+C;IAC/C,MAAM,EAAE,CAAC,YAAY,EAAE,MAAM,KAAK,OAAO,CAAC,IAAI,CAAC,CAAC;IAChD,sEAAsE;IACtE,mBAAmB,CAAC,EAAE,CAAC,YAAY,EAAE,MAAM,KAAK,OAAO,CAAC,MAAM,GAAG,IAAI,GAAG,SAAS,CAAC,CAAC;CACnF;AAQD,MAAM,WAAW,eAAe;IAC/B,oEAAoE;IACpE,gBAAgB,CAAC,EAAE,OAAO,CAAC;IAC3B,oEAAoE;IACpE,UAAU,CAAC,EAAE,cAAc,CAAC;IAC5B,mFAAmF;IACnF,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,mFAAmF;IACnF,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB;;;;;OAKG;IACH,UAAU,CAAC,EAAE,OAAO,CAAC;CACrB;AAmJD,wBAAgB,wBAAwB,CACvC,GAAG,EAAE,MAAM,EACX,OAAO,CAAC,EAAE,eAAe,GACvB,cAAc,CAAC,OAAO,UAAU,EAAE,eAAe,GAAG,SAAS,CAAC,CAuMhE;AAED,wBAAgB,cAAc,CAAC,GAAG,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,eAAe,GAAG,SAAS,CAAC,OAAO,UAAU,CAAC,CAEnG","sourcesContent":["import { basename, dirname, isAbsolute, relative, resolve as resolvePath, sep } from \"node:path\";\nimport type { AgentTool } from \"@kolisachint/hoocode-agent-core\";\nimport type { Api, ImageContent, Model, TextContent } from \"@kolisachint/hoocode-ai\";\nimport { Text } from \"@kolisachint/hoocode-tui\";\nimport { createHash } from \"crypto\";\nimport { constants } from \"fs\";\nimport { access as fsAccess, readFile as fsReadFile } from \"fs/promises\";\nimport { type Static, Type } from \"typebox\";\nimport { getReadmePath } from \"../../config.js\";\nimport { keyHint, keyText } from \"../../modes/interactive/components/keybinding-hints.js\";\nimport { renderReadOutput } from \"../../modes/interactive/components/read-output.js\";\nimport { getLanguageFromPath, highlightCode, messageLabel, type Theme } from \"../../modes/interactive/theme/theme.js\";\nimport { formatDimensionNote, resizeImage } from \"../../utils/image-resize.js\";\nimport { detectSupportedImageMimeTypeFromFile } from \"../../utils/mime.js\";\nimport { formatPathRelativeToCwdOrAbsolute } from \"../../utils/paths.js\";\nimport type { ToolDefinition, ToolRenderResultOptions } from \"../extensions/types.js\";\nimport { resolveReadPath } from \"./path-utils.js\";\nimport {\n\tbuildDedupPointerText,\n\tfindCoveringRead,\n\treadRangeFromArgs,\n\treadStampMatches,\n\trecordReadStamp,\n} from \"./read-dedup.js\";\nimport { getTextOutput, invalidArgText, replaceTabs, shortenPath, str } from \"./render-utils.js\";\nimport { wrapToolDefinition } from \"./tool-definition-wrapper.js\";\nimport { DEFAULT_MAX_BYTES, DEFAULT_MAX_LINES, formatSize, type TruncationResult, truncateHead } from \"./truncate.js\";\n\nconst readSchema = Type.Object({\n\tpath: Type.String({ description: \"Path to the file to read (relative or absolute)\" }),\n\toffset: Type.Optional(Type.Number({ description: \"Line number to start reading from (1-indexed)\" })),\n\tlimit: Type.Optional(Type.Number({ description: \"Maximum number of lines to read\" })),\n});\n\nexport type ReadToolInput = Static<typeof readSchema>;\n\nexport interface ReadToolDetails {\n\ttruncation?: TruncationResult;\n}\n\ninterface CompactReadClassification {\n\tkind: \"docs\" | \"resource\" | \"skill\";\n\tlabel: string;\n}\n\nconst COMPACT_RESOURCE_FILE_NAMES = new Set([\"AGENTS.md\", \"AGENTS.MD\", \"CLAUDE.md\", \"CLAUDE.MD\"]);\n\n/**\n * Structured/binary document formats. These are container/binary formats (OOXML\n * is a zip, PDF is binary), so a plain utf-8 read yields garbage — read\n * short-circuits with a note instead of dumping bytes into the context.\n */\nconst DOC_FORMAT_EXTENSIONS = new Map<string, string>([\n\t[\".docx\", \"Word (OOXML)\"],\n\t[\".xlsx\", \"Excel (OOXML)\"],\n\t[\".pptx\", \"PowerPoint (OOXML)\"],\n\t[\".pdf\", \"PDF\"],\n]);\n\nfunction getDocFormatHint(absolutePath: string): string | undefined {\n\tconst dot = absolutePath.lastIndexOf(\".\");\n\tif (dot === -1) return undefined;\n\treturn DOC_FORMAT_EXTENSIONS.get(absolutePath.slice(dot).toLowerCase());\n}\n\n/**\n * Pluggable operations for the read tool.\n * Override these to delegate file reading to remote systems (for example SSH).\n */\nexport interface ReadOperations {\n\t/** Read file contents as a Buffer */\n\treadFile: (absolutePath: string) => Promise<Buffer>;\n\t/** Check if file is readable (throw if not) */\n\taccess: (absolutePath: string) => Promise<void>;\n\t/** Detect image MIME type, return null or undefined for non-images */\n\tdetectImageMimeType?: (absolutePath: string) => Promise<string | null | undefined>;\n}\n\nconst defaultReadOperations: ReadOperations = {\n\treadFile: (path) => fsReadFile(path),\n\taccess: (path) => fsAccess(path, constants.R_OK),\n\tdetectImageMimeType: detectSupportedImageMimeTypeFromFile,\n};\n\nexport interface ReadToolOptions {\n\t/** Whether to auto-resize images to 2000x2000 max. Default: true */\n\tautoResizeImages?: boolean;\n\t/** Custom operations for file reading. Default: local filesystem */\n\toperations?: ReadOperations;\n\t/** Byte cap on the returned text before truncation. Default: DEFAULT_MAX_BYTES. */\n\tmaxOutputBytes?: number;\n\t/** Line cap on the returned text before truncation. Default: DEFAULT_MAX_LINES. */\n\tmaxOutputLines?: number;\n\t/**\n\t * When true, a text read whose requested line range is already fully covered\n\t * by an earlier, still-live read in the current session short-circuits with a\n\t * pointer instead of re-fetching the file. Complements the post-hoc context\n\t * GC; gated on the same setting. Default: false.\n\t */\n\tdedupReads?: boolean;\n}\n\ntype ReadRenderArgs = { path?: string; file_path?: string; offset?: number; limit?: number };\n\nfunction formatReadLineRange(args: ReadRenderArgs | undefined, theme: Theme): string {\n\tif (args?.offset === undefined && args?.limit === undefined) return \"\";\n\tconst startLine = args.offset ?? 1;\n\tconst endLine = args.limit !== undefined ? startLine + args.limit - 1 : \"\";\n\treturn theme.fg(\"warning\", `:${startLine}${endLine ? `-${endLine}` : \"\"}`);\n}\n\nfunction formatReadCall(args: ReadRenderArgs | undefined, theme: Theme): string {\n\tconst rawPath = str(args?.file_path ?? args?.path);\n\tconst path = rawPath !== null ? shortenPath(rawPath) : null;\n\tconst invalidArg = invalidArgText(theme);\n\tconst pathDisplay = path === null ? invalidArg : path ? theme.fg(\"accent\", path) : theme.fg(\"toolOutput\", \"...\");\n\treturn `${theme.fg(\"toolTitle\", theme.bold(\"read\"))} ${pathDisplay}${formatReadLineRange(args, theme)}`;\n}\n\nfunction trimTrailingEmptyLines(lines: string[]): string[] {\n\tlet end = lines.length;\n\twhile (end > 0 && lines[end - 1] === \"\") {\n\t\tend--;\n\t}\n\treturn lines.slice(0, end);\n}\n\nfunction getNonVisionImageNote(model: Model<Api> | undefined): string | undefined {\n\tif (!model || model.input.includes(\"image\")) {\n\t\treturn undefined;\n\t}\n\treturn \"[Current model does not support images. The image will be omitted from this request.]\";\n}\n\nfunction toPosixPath(filePath: string): string {\n\treturn filePath.split(sep).join(\"/\");\n}\n\nfunction getHooCodeDocsClassification(absolutePath: string): CompactReadClassification | undefined {\n\tconst packageRoot = dirname(getReadmePath());\n\tconst relativePath = relative(resolvePath(packageRoot), resolvePath(absolutePath));\n\tif (\n\t\trelativePath === \"\" ||\n\t\trelativePath === \"..\" ||\n\t\trelativePath.startsWith(`..${sep}`) ||\n\t\tisAbsolute(relativePath)\n\t) {\n\t\treturn undefined;\n\t}\n\n\tconst label = toPosixPath(relativePath);\n\tif (label === \"README.md\" || label.startsWith(\"docs/\") || label.startsWith(\"examples/\")) {\n\t\treturn { kind: \"docs\", label };\n\t}\n\treturn undefined;\n}\n\nfunction getCompactReadClassification(\n\targs: ReadRenderArgs | undefined,\n\tcwd: string,\n): CompactReadClassification | undefined {\n\tconst rawPath = str(args?.file_path ?? args?.path);\n\tif (!rawPath) return undefined;\n\n\tconst absolutePath = resolveReadPath(rawPath, cwd);\n\tconst fileName = basename(absolutePath);\n\tif (fileName === \"SKILL.md\") {\n\t\treturn { kind: \"skill\", label: basename(dirname(absolutePath)) || fileName };\n\t}\n\n\tconst docsClassification = getHooCodeDocsClassification(absolutePath);\n\tif (docsClassification) return docsClassification;\n\n\tif (COMPACT_RESOURCE_FILE_NAMES.has(fileName)) {\n\t\treturn { kind: \"resource\", label: formatPathRelativeToCwdOrAbsolute(absolutePath, cwd) };\n\t}\n\n\treturn undefined;\n}\n\nfunction formatCompactReadCall(\n\tclassification: CompactReadClassification,\n\targs: ReadRenderArgs | undefined,\n\ttheme: Theme,\n): string {\n\tconst expandHint = theme.fg(\"dim\", ` (${keyText(\"app.tools.expand\")} to expand)`);\n\tif (classification.kind === \"skill\") {\n\t\treturn (\n\t\t\t`${messageLabel(\"skill\")} ` +\n\t\t\ttheme.fg(\"customMessageText\", classification.label) +\n\t\t\tformatReadLineRange(args, theme) +\n\t\t\texpandHint\n\t\t);\n\t}\n\n\treturn (\n\t\ttheme.fg(\"toolTitle\", theme.bold(`read ${classification.kind}`)) +\n\t\t\" \" +\n\t\ttheme.fg(\"accent\", classification.label) +\n\t\tformatReadLineRange(args, theme) +\n\t\texpandHint\n\t);\n}\n\nfunction formatReadResult(\n\targs: ReadRenderArgs | undefined,\n\tresult: { content: (TextContent | ImageContent)[]; details?: ReadToolDetails },\n\toptions: ToolRenderResultOptions,\n\ttheme: Theme,\n\tshowImages: boolean,\n\tcwd: string,\n\tisError: boolean,\n): string {\n\tif (!options.expanded && !isError && getCompactReadClassification(args, cwd)) {\n\t\treturn \"\";\n\t}\n\n\tconst rawPath = str(args?.file_path ?? args?.path);\n\tconst output = getTextOutput(result, showImages);\n\tconst lang = rawPath ? getLanguageFromPath(rawPath) : undefined;\n\tconst renderedLines = lang ? highlightCode(replaceTabs(output), lang) : output.split(\"\\n\");\n\tconst lines = trimTrailingEmptyLines(renderedLines);\n\tconst maxLines = options.expanded ? lines.length : 10;\n\tconst displayLines = lines.slice(0, maxLines);\n\tconst remaining = lines.length - maxLines;\n\t// Calculate starting line number from args\n\tconst startLine = (args?.offset as number) ?? 1;\n\t// Use the new renderReadOutput function for line numbers\n\tconst numberedLines = renderReadOutput(displayLines.join(\"\\n\"), { startLine });\n\tlet text = `\\n${numberedLines.join(\"\\n\")}`;\n\tif (remaining > 0) {\n\t\ttext += `${theme.fg(\"muted\", `\\n... (${remaining} more lines,`)} ${keyHint(\"app.tools.expand\", \"to expand\")})`;\n\t}\n\n\tconst truncation = result.details?.truncation;\n\tif (truncation?.truncated) {\n\t\tif (truncation.firstLineExceedsLimit) {\n\t\t\ttext += `\\n${theme.fg(\"warning\", `[First line exceeds ${formatSize(truncation.maxBytes ?? DEFAULT_MAX_BYTES)} limit]`)}`;\n\t\t} else if (truncation.truncatedBy === \"lines\") {\n\t\t\ttext += `\\n${theme.fg(\"warning\", `[Truncated: showing ${truncation.outputLines} of ${truncation.totalLines} lines (${truncation.maxLines ?? DEFAULT_MAX_LINES} line limit)]`)}`;\n\t\t} else {\n\t\t\ttext += `\\n${theme.fg(\"warning\", `[Truncated: ${truncation.outputLines} lines shown (${formatSize(truncation.maxBytes ?? DEFAULT_MAX_BYTES)} limit)]`)}`;\n\t\t}\n\t}\n\treturn text;\n}\n\nexport function createReadToolDefinition(\n\tcwd: string,\n\toptions?: ReadToolOptions,\n): ToolDefinition<typeof readSchema, ReadToolDetails | undefined> {\n\tconst autoResizeImages = options?.autoResizeImages ?? true;\n\tconst ops = options?.operations ?? defaultReadOperations;\n\tconst maxBytes = options?.maxOutputBytes ?? DEFAULT_MAX_BYTES;\n\tconst maxLines = options?.maxOutputLines ?? DEFAULT_MAX_LINES;\n\tconst dedupReads = options?.dedupReads ?? false;\n\treturn {\n\t\tname: \"read\",\n\t\tlabel: \"read\",\n\t\tdescription: `Read the contents of a file. Supports text files and images (jpg, png, gif, webp). Images are sent as attachments. For text files, output is truncated to ${maxLines} lines or ${Math.round(maxBytes / 1024)}KB (whichever is hit first). Use offset/limit for large files. When you need the full file, continue with offset until complete.`,\n\t\tpromptSnippet: \"Read file contents\",\n\t\t// No promptGuidelines: \"use read instead of cat/sed\" is already covered by the\n\t\t// file-exploration guideline buildSystemPrompt emits, and again by bash's own\n\t\t// snippet. Three copies of one rule.\n\t\tparameters: readSchema,\n\t\tasync execute(\n\t\t\ttoolCallId,\n\t\t\t{ path, offset, limit }: { path: string; offset?: number; limit?: number },\n\t\t\tsignal?: AbortSignal,\n\t\t\t_onUpdate?,\n\t\t\tctx?,\n\t\t) {\n\t\t\tconst absolutePath = resolveReadPath(path, cwd);\n\t\t\treturn new Promise<{ content: (TextContent | ImageContent)[]; details: ReadToolDetails | undefined }>(\n\t\t\t\t(resolve, reject) => {\n\t\t\t\t\tif (signal?.aborted) {\n\t\t\t\t\t\treject(new Error(\"Operation aborted\"));\n\t\t\t\t\t\treturn;\n\t\t\t\t\t}\n\t\t\t\t\tlet aborted = false;\n\t\t\t\t\tconst onAbort = () => {\n\t\t\t\t\t\taborted = true;\n\t\t\t\t\t\treject(new Error(\"Operation aborted\"));\n\t\t\t\t\t};\n\t\t\t\t\tsignal?.addEventListener(\"abort\", onAbort, { once: true });\n\n\t\t\t\t\t(async () => {\n\t\t\t\t\t\ttry {\n\t\t\t\t\t\t\t// Check if file exists and is readable.\n\t\t\t\t\t\t\tawait ops.access(absolutePath);\n\t\t\t\t\t\t\tif (aborted) return;\n\t\t\t\t\t\t\tconst mimeType = ops.detectImageMimeType ? await ops.detectImageMimeType(absolutePath) : undefined;\n\t\t\t\t\t\t\tlet content: (TextContent | ImageContent)[];\n\t\t\t\t\t\t\tlet details: ReadToolDetails | undefined;\n\t\t\t\t\t\t\tconst nonVisionImageNote = getNonVisionImageNote(ctx?.model);\n\t\t\t\t\t\t\t// At-call dedup: for plain-text reads, if an earlier still-live read\n\t\t\t\t\t\t\t// in this session already covers the requested range, point at it\n\t\t\t\t\t\t\t// instead of re-fetching. Skipped for images and structured docs.\n\t\t\t\t\t\t\tconst dedupCovering =\n\t\t\t\t\t\t\t\tdedupReads && !mimeType && !getDocFormatHint(absolutePath) && ctx?.sessionManager\n\t\t\t\t\t\t\t\t\t? findCoveringRead(ctx.sessionManager.getBranch(), {\n\t\t\t\t\t\t\t\t\t\t\tresolvedPath: absolutePath,\n\t\t\t\t\t\t\t\t\t\t\trequestedRange: readRangeFromArgs({ offset, limit }),\n\t\t\t\t\t\t\t\t\t\t\tcurrentCallId: toolCallId,\n\t\t\t\t\t\t\t\t\t\t\tresolvePath: (raw) => resolveReadPath(raw, cwd),\n\t\t\t\t\t\t\t\t\t\t})\n\t\t\t\t\t\t\t\t\t: null;\n\t\t\t\t\t\t\tif (mimeType) {\n\t\t\t\t\t\t\t\t// Read image as binary.\n\t\t\t\t\t\t\t\tconst buffer = await ops.readFile(absolutePath);\n\t\t\t\t\t\t\t\tconst base64 = buffer.toString(\"base64\");\n\t\t\t\t\t\t\t\tif (autoResizeImages) {\n\t\t\t\t\t\t\t\t\t// Resize image if needed before sending it back to the model.\n\t\t\t\t\t\t\t\t\tconst resized = await resizeImage({ type: \"image\", data: base64, mimeType });\n\t\t\t\t\t\t\t\t\tif (!resized) {\n\t\t\t\t\t\t\t\t\t\tlet textNote = `Read image file [${mimeType}]\\n[Image omitted: could not be resized below the inline image size limit.]`;\n\t\t\t\t\t\t\t\t\t\tif (nonVisionImageNote) textNote += `\\n${nonVisionImageNote}`;\n\t\t\t\t\t\t\t\t\t\tcontent = [{ type: \"text\", text: textNote }];\n\t\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\t\tconst dimensionNote = formatDimensionNote(resized);\n\t\t\t\t\t\t\t\t\t\tlet textNote = `Read image file [${resized.mimeType}]`;\n\t\t\t\t\t\t\t\t\t\tif (dimensionNote) textNote += `\\n${dimensionNote}`;\n\t\t\t\t\t\t\t\t\t\tif (nonVisionImageNote) textNote += `\\n${nonVisionImageNote}`;\n\t\t\t\t\t\t\t\t\t\tcontent = [\n\t\t\t\t\t\t\t\t\t\t\t{ type: \"text\", text: textNote },\n\t\t\t\t\t\t\t\t\t\t\t{ type: \"image\", data: resized.data, mimeType: resized.mimeType },\n\t\t\t\t\t\t\t\t\t\t];\n\t\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\tlet textNote = `Read image file [${mimeType}]`;\n\t\t\t\t\t\t\t\t\tif (nonVisionImageNote) textNote += `\\n${nonVisionImageNote}`;\n\t\t\t\t\t\t\t\t\tcontent = [\n\t\t\t\t\t\t\t\t\t\t{ type: \"text\", text: textNote },\n\t\t\t\t\t\t\t\t\t\t{ type: \"image\", data: base64, mimeType },\n\t\t\t\t\t\t\t\t\t];\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t} else if (getDocFormatHint(absolutePath)) {\n\t\t\t\t\t\t\t\t// Structured/binary document: a utf-8 read is meaningless, and dumping\n\t\t\t\t\t\t\t\t// the bytes would burn context for nothing. Stop here and let the model\n\t\t\t\t\t\t\t\t// choose an extraction path (bash + an appropriate converter).\n\t\t\t\t\t\t\t\tconst label = getDocFormatHint(absolutePath);\n\t\t\t\t\t\t\t\tcontent = [\n\t\t\t\t\t\t\t\t\t{\n\t\t\t\t\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\t\t\t\t\ttext: `[${label} document — not plain text; read cannot show it. Extract it with a converter via bash if you need its contents.]`,\n\t\t\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\t\t];\n\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t// Read text content.\n\t\t\t\t\t\t\t\tconst buffer = await ops.readFile(absolutePath);\n\t\t\t\t\t\t\t\t// Stamp the bytes that are actually on disk. Deduping saves context\n\t\t\t\t\t\t\t\t// tokens, not this read, so the file can be hashed for the price the\n\t\t\t\t\t\t\t\t// call was already paying - and the pointer's \"has not changed since\"\n\t\t\t\t\t\t\t\t// becomes something checked rather than assumed.\n\t\t\t\t\t\t\t\tconst stamp = createHash(\"sha1\").update(buffer).digest(\"hex\");\n\t\t\t\t\t\t\t\tif (dedupCovering && readStampMatches(dedupCovering.callId, stamp)) {\n\t\t\t\t\t\t\t\t\tcontent = [{ type: \"text\", text: buildDedupPointerText(dedupCovering) }];\n\t\t\t\t\t\t\t\t\tif (aborted) return;\n\t\t\t\t\t\t\t\t\tsignal?.removeEventListener(\"abort\", onAbort);\n\t\t\t\t\t\t\t\t\tresolve({ content, details });\n\t\t\t\t\t\t\t\t\treturn;\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\tconst textContent = buffer.toString(\"utf-8\");\n\t\t\t\t\t\t\t\tconst allLines = textContent.split(\"\\n\");\n\t\t\t\t\t\t\t\tconst totalFileLines = allLines.length;\n\t\t\t\t\t\t\t\t// Apply offset if specified. Convert from 1-indexed input to 0-indexed array access.\n\t\t\t\t\t\t\t\tconst startLine = offset ? Math.max(0, offset - 1) : 0;\n\t\t\t\t\t\t\t\tconst startLineDisplay = startLine + 1;\n\t\t\t\t\t\t\t\t// Check if offset is out of bounds.\n\t\t\t\t\t\t\t\tif (startLine >= allLines.length) {\n\t\t\t\t\t\t\t\t\tthrow new Error(`Offset ${offset} is beyond end of file (${allLines.length} lines total)`);\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\tlet selectedContent: string;\n\t\t\t\t\t\t\t\tlet userLimitedLines: number | undefined;\n\t\t\t\t\t\t\t\t// If limit is specified by the user, honor it first. Otherwise truncateHead decides.\n\t\t\t\t\t\t\t\tif (limit !== undefined) {\n\t\t\t\t\t\t\t\t\tconst endLine = Math.min(startLine + limit, allLines.length);\n\t\t\t\t\t\t\t\t\tselectedContent = allLines.slice(startLine, endLine).join(\"\\n\");\n\t\t\t\t\t\t\t\t\tuserLimitedLines = endLine - startLine;\n\t\t\t\t\t\t\t\t} else if (startLine === 0) {\n\t\t\t\t\t\t\t\t\t// Whole-file read (common case): reuse the original string instead of\n\t\t\t\t\t\t\t\t\t// rebuilding it from the split lines.\n\t\t\t\t\t\t\t\t\tselectedContent = textContent;\n\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\tselectedContent = allLines.slice(startLine).join(\"\\n\");\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\t// Apply truncation, respecting both line and byte limits.\n\t\t\t\t\t\t\t\t// The text is handed over byte-for-byte. Whatever the model reads here is\n\t\t\t\t\t\t\t\t// what it sends back as an edit's `oldText`, so altering it (the trailing\n\t\t\t\t\t\t\t\t// whitespace strip that used to run at this point) makes the model's own\n\t\t\t\t\t\t\t\t// faithful copy fail to match the file it came from.\n\t\t\t\t\t\t\t\tconst truncation = truncateHead(selectedContent, { maxBytes, maxLines });\n\t\t\t\t\t\t\t\tlet outputText: string;\n\t\t\t\t\t\t\t\tif (truncation.firstLineExceedsLimit) {\n\t\t\t\t\t\t\t\t\t// First line alone exceeds the byte limit. Point the model at a bash fallback.\n\t\t\t\t\t\t\t\t\tconst firstLineSize = formatSize(Buffer.byteLength(allLines[startLine], \"utf-8\"));\n\t\t\t\t\t\t\t\t\toutputText = `[Line ${startLineDisplay} is ${firstLineSize}, exceeds ${formatSize(maxBytes)} limit. Use bash: sed -n '${startLineDisplay}p' ${path} | head -c ${maxBytes}]`;\n\t\t\t\t\t\t\t\t\tdetails = { truncation };\n\t\t\t\t\t\t\t\t} else if (truncation.truncated) {\n\t\t\t\t\t\t\t\t\t// Truncation occurred. Build an actionable continuation notice.\n\t\t\t\t\t\t\t\t\tconst endLineDisplay = startLineDisplay + truncation.outputLines - 1;\n\t\t\t\t\t\t\t\t\tconst nextOffset = endLineDisplay + 1;\n\t\t\t\t\t\t\t\t\toutputText = truncation.content;\n\t\t\t\t\t\t\t\t\tif (truncation.truncatedBy === \"lines\") {\n\t\t\t\t\t\t\t\t\t\toutputText += `\\n\\n[Showing lines ${startLineDisplay}-${endLineDisplay} of ${totalFileLines}. Use offset=${nextOffset} to continue.]`;\n\t\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\t\toutputText += `\\n\\n[Showing lines ${startLineDisplay}-${endLineDisplay} of ${totalFileLines} (${formatSize(maxBytes)} limit). Use offset=${nextOffset} to continue.]`;\n\t\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\t\tdetails = { truncation };\n\t\t\t\t\t\t\t\t} else if (userLimitedLines !== undefined && startLine + userLimitedLines < allLines.length) {\n\t\t\t\t\t\t\t\t\t// User-specified limit stopped early, but the file still has more content.\n\t\t\t\t\t\t\t\t\tconst remaining = allLines.length - (startLine + userLimitedLines);\n\t\t\t\t\t\t\t\t\tconst nextOffset = startLine + userLimitedLines + 1;\n\t\t\t\t\t\t\t\t\toutputText = `${truncation.content}\\n\\n[${remaining} more lines in file. Use offset=${nextOffset} to continue.]`;\n\t\t\t\t\t\t\t\t} else {\n\t\t\t\t\t\t\t\t\t// No truncation and no remaining user-limited content.\n\t\t\t\t\t\t\t\t\toutputText = truncation.content;\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t\tcontent = [{ type: \"text\", text: outputText }];\n\t\t\t\t\t\t\t\trecordReadStamp(toolCallId, stamp);\n\t\t\t\t\t\t\t}\n\n\t\t\t\t\t\t\tif (aborted) return;\n\t\t\t\t\t\t\tsignal?.removeEventListener(\"abort\", onAbort);\n\t\t\t\t\t\t\tresolve({ content, details });\n\t\t\t\t\t\t} catch (error) {\n\t\t\t\t\t\t\tsignal?.removeEventListener(\"abort\", onAbort);\n\t\t\t\t\t\t\tif (!aborted) reject(error);\n\t\t\t\t\t\t}\n\t\t\t\t\t})();\n\t\t\t\t},\n\t\t\t);\n\t\t},\n\t\trenderCall(args, theme, context) {\n\t\t\tconst text = (context.lastComponent as Text | undefined) ?? new Text(\"\", 0, 0);\n\t\t\tconst classification = !context.expanded ? getCompactReadClassification(args, context.cwd) : undefined;\n\t\t\ttext.setText(\n\t\t\t\tclassification ? formatCompactReadCall(classification, args, theme) : formatReadCall(args, theme),\n\t\t\t);\n\t\t\treturn text;\n\t\t},\n\t\trenderResult(result, options, theme, context) {\n\t\t\tconst text = (context.lastComponent as Text | undefined) ?? new Text(\"\", 0, 0);\n\t\t\ttext.setText(\n\t\t\t\tformatReadResult(context.args, result, options, theme, context.showImages, context.cwd, context.isError),\n\t\t\t);\n\t\t\treturn text;\n\t\t},\n\t};\n}\n\nexport function createReadTool(cwd: string, options?: ReadToolOptions): AgentTool<typeof readSchema> {\n\treturn wrapToolDefinition(createReadToolDefinition(cwd, options));\n}\n"]}

package/dist/core/tools/read.js

@@ -1,5 +1,6 @@
 import { basename, dirname, isAbsolute, relative, resolve as resolvePath, sep } from "node:path";
 import { Text } from "@kolisachint/hoocode-tui";
+import { createHash } from "crypto";
 import { constants } from "fs";
 import { access as fsAccess, readFile as fsReadFile } from "fs/promises";
 import { Type } from "typebox";
@@ -11,7 +12,7 @@ import { formatDimensionNote, resizeImage } from "../../utils/image-resize.js";
 import { detectSupportedImageMimeTypeFromFile } from "../../utils/mime.js";
 import { formatPathRelativeToCwdOrAbsolute } from "../../utils/paths.js";
 import { resolveReadPath } from "./path-utils.js";
-import { buildDedupPointerText, findCoveringRead, readRangeFromArgs } from "./read-dedup.js";
+import { buildDedupPointerText, findCoveringRead, readRangeFromArgs, readStampMatches, recordReadStamp, } from "./read-dedup.js";
 import { getTextOutput, invalidArgText, replaceTabs, shortenPath, str } from "./render-utils.js";
 import { wrapToolDefinition } from "./tool-definition-wrapper.js";
 import { DEFAULT_MAX_BYTES, DEFAULT_MAX_LINES, formatSize, truncateHead } from "./truncate.js";
@@ -202,10 +203,7 @@ export function createReadToolDefinition(cwd, options) {
                                 resolvePath: (raw) => resolveReadPath(raw, cwd),
                             })
                             : null;
-                        if (dedupCovering) {
-                            content = [{ type: "text", text: buildDedupPointerText(dedupCovering) }];
-                        }
-                        else if (mimeType) {
+                        if (mimeType) {
                             // Read image as binary.
                             const buffer = await ops.readFile(absolutePath);
                             const base64 = buffer.toString("base64");
@@ -256,6 +254,19 @@ export function createReadToolDefinition(cwd, options) {
                         else {
                             // Read text content.
                             const buffer = await ops.readFile(absolutePath);
+                            // Stamp the bytes that are actually on disk. Deduping saves context
+                            // tokens, not this read, so the file can be hashed for the price the
+                            // call was already paying - and the pointer's "has not changed since"
+                            // becomes something checked rather than assumed.
+                            const stamp = createHash("sha1").update(buffer).digest("hex");
+                            if (dedupCovering && readStampMatches(dedupCovering.callId, stamp)) {
+                                content = [{ type: "text", text: buildDedupPointerText(dedupCovering) }];
+                                if (aborted)
+                                    return;
+                                signal?.removeEventListener("abort", onAbort);
+                                resolve({ content, details });
+                                return;
+                            }
                             const textContent = buffer.toString("utf-8");
                             const allLines = textContent.split("\n");
                             const totalFileLines = allLines.length;
@@ -319,6 +330,7 @@ export function createReadToolDefinition(cwd, options) {
                                 outputText = truncation.content;
                             }
                             content = [{ type: "text", text: outputText }];
+                            recordReadStamp(toolCallId, stamp);
                         }
                         if (aborted)
                             return;