@bastani/atomic 0.9.5-alpha.8 → 0.9.5

package/dist/core/agent-session-auto-compaction.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"agent-session-auto-compaction.d.ts","sourceRoot":"","sources":["../../src/core/agent-session-auto-compaction.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,8BAA8B,CAAC;AAMrE,OAAO,KAAK,EAAE,2BAA2B,IAAI,YAAY,EAAE,MAAM,4BAA4B,CAAC;AAE9F,wBAAsB,gBAAgB,CAAC,IAAI,EAAE,YAAY,EAAE,gBAAgB,EAAE,gBAAgB,EAAE,gBAAgB,UAAO,GAAG,OAAO,CAAC,IAAI,CAAC,CAgGrI;AAGD,wBAAgB,6CAA6C,CAAC,IAAI,EAAE,YAAY,EAAE,gBAAgB,EAAE,gBAAgB,GAAG,OAAO,CAQ7H;AAED,wBAAgB,mCAAmC,CAAC,gBAAgB,EAAE,gBAAgB,GAAG,OAAO,CAE/F;AAED;;GAEG;AAEH,wBAAgB,kDAAkD,CAAC,IAAI,EAAE,YAAY,GAAG,IAAI,CAQ3F;AAED;;GAEG;AAEH,wBAAgB,4CAA4C,CAAC,IAAI,EAAE,YAAY,EAC9E,MAAM,EAAE,UAAU,GAAG,WAAW,EAChC,SAAS,EAAE,OAAO,GAChB,IAAI,CAyCN;AAED,wBAAsB,+CAA+C,CAAC,IAAI,EAAE,YAAY,GAAG,OAAO,CAAC,IAAI,CAAC,CAIvG;AAED;;GAEG;AAEH,wBAAsB,0BAA0B,CAAC,IAAI,EAAE,YAAY,GAAG,OAAO,CAAC,IAAI,CAAC,CAWlF;AAOD,wBAAsB,kBAAkB,CAAC,IAAI,EAAE,YAAY,EAAE,MAAM,EAAE,UAAU,GAAG,WAAW,EAAE,SAAS,EAAE,OAAO,GAAG,OAAO,CAAC,IAAI,CAAC,CAyEhI;AAED;;GAEG;AAEH,eAAO,MAAM,iCAAiC;;;;;;;;CAQ7C,CAAC","sourcesContent":["import type { AssistantMessage } from \"@earendil-works/pi-ai/compat\";\nimport { isContextOverflow } from \"@earendil-works/pi-ai/compat\";\nimport { getEffectiveInputBudget } from \"./context-window.ts\";\nimport { parseCopilotPromptLimitError } from \"./copilot-errors.ts\";\nimport { calculateContextTokens, estimateContextTokens, shouldCompact } from \"./compaction/index.ts\";\nimport { getLatestCompactionBoundaryEntry } from \"./session-manager.ts\";\nimport type { AgentSessionInternalSurface as AgentSession } from \"./agent-session-methods.ts\";\n\nexport async function _checkCompaction(this: AgentSession, assistantMessage: AssistantMessage, skipAbortedCheck = true): Promise<void> {\n\tconst settings = this.settingsManager.getCompactionSettings();\n\tif (!settings.enabled) return;\n\n\t// Skip if message was aborted (user cancelled) - unless skipAbortedCheck is false\n\tif (skipAbortedCheck && assistantMessage.stopReason === \"aborted\") return;\n\n\tconst contextWindow = this.model?.contextWindow ?? 0;\n\n\t// Skip overflow check if the message came from a different model.\n\t// This handles the case where user switched from a smaller-context model (e.g. opus)\n\t// to a larger-context model (e.g. codex) - the overflow error from the old model\n\t// shouldn't trigger compaction for the new model.\n\tconst sameModel =\n\t\tthis.model && assistantMessage.provider === this.model.provider && assistantMessage.model === this.model.id;\n\n\t// Skip compaction checks if this assistant message is older than the latest\n\t// compaction boundary. This prevents a stale pre-compaction usage/error\n\t// from retriggering compaction on the first prompt after compaction.\n\tconst compactionBoundaryEntry = getLatestCompactionBoundaryEntry(this.sessionManager.getBranch());\n\tconst assistantIsFromBeforeCompactionBoundary =\n\t\tcompactionBoundaryEntry !== null &&\n\t\tassistantMessage.timestamp <= new Date(compactionBoundaryEntry.timestamp).getTime();\n\tif (assistantIsFromBeforeCompactionBoundary) {\n\t\treturn;\n\t}\n\n\t// Case 1: Overflow - LLM returned context overflow error\n\t// When Copilot rejects a 1m client-budget prompt at a lower server cap (for example\n\t// because long-context/usage-based billing entitlement is missing), leave the friendly\n\t// error visible instead of auto-compacting down to a smaller server tier silently.\n\tif (sameModel && this._isCopilotServerCapBelowSelectedContextWindow(assistantMessage)) {\n\t\treturn;\n\t}\n\tif (sameModel && isContextOverflow(assistantMessage, contextWindow)) {\n\t\tconst willRetry = assistantMessage.stopReason !== \"stop\";\n\t\tif (!willRetry) {\n\t\t\tawait this._runAutoCompaction(\"overflow\", false);\n\t\t\treturn;\n\t\t}\n\n\t\tif (this._overflowRecoveryAttempted) {\n\t\t\tthis._emit({\n\t\t\t\ttype: \"compaction_end\",\n\t\t\t\treason: \"overflow\",\n\t\t\t\tresult: undefined,\n\t\t\t\taborted: false,\n\t\t\t\twillRetry: false,\n\t\t\t\tunresolvedOverflow: true,\n\t\t\t\terrorMessage:\n\t\t\t\t\t\"Context overflow recovery failed after one compact-and-retry attempt. Try reducing context or switching to a larger-context model.\",\n\t\t\t});\n\t\t\treturn;\n\t\t}\n\n\t\tthis._overflowRecoveryAttempted = true;\n\t\t// Remove the error message from agent state (it IS saved to session for history,\n\t\t// but we don't want it in context for the retry)\n\t\tconst messages = this.agent.state.messages;\n\t\tif (messages.length > 0 && messages[messages.length - 1].role === \"assistant\") {\n\t\t\tthis.agent.state.messages = messages.slice(0, -1);\n\t\t}\n\t\tawait this._runAutoCompaction(\"overflow\", willRetry);\n\t\treturn;\n\t}\n\n\t// Case 2: Threshold - context is getting large\n\t// For error messages (no usage data), estimate from last successful response.\n\t// This ensures sessions that hit persistent API errors (e.g. 529) can still compact.\n\tlet contextTokens: number;\n\tif (assistantMessage.stopReason === \"error\") {\n\t\tconst messages = this.agent.state.messages;\n\t\tconst estimate = estimateContextTokens(messages);\n\t\tif (estimate.lastUsageIndex === null) return; // No usage data at all\n\t\t// Verify the usage source is post-compaction. Kept pre-compaction messages\n\t\t// have stale usage reflecting the old (larger) context and would falsely\n\t\t// trigger compaction right after one just finished.\n\t\tconst usageMsg = messages[estimate.lastUsageIndex];\n\t\tif (\n\t\t\tcompactionBoundaryEntry &&\n\t\t\tusageMsg.role === \"assistant\" &&\n\t\t\t(usageMsg as AssistantMessage).timestamp <= new Date(compactionBoundaryEntry.timestamp).getTime()\n\t\t) {\n\t\t\treturn;\n\t\t}\n\t\tcontextTokens = estimate.tokens;\n\t} else {\n\t\tcontextTokens = calculateContextTokens(assistantMessage.usage);\n\t}\n\t// Compact against the effective input budget (the hard prompt cap for providers like Copilot\n\t// that advertise a larger total window) so we compact before overrunning the server-side limit\n\t// rather than relying on reactive overflow recovery near the cap.\n\tconst compactionBudget = this.model ? getEffectiveInputBudget(this.model) : contextWindow;\n\tif (shouldCompact(contextTokens, compactionBudget, settings)) {\n\t\tawait this._runAutoCompaction(\"threshold\", shouldRetryAfterThresholdCompaction(assistantMessage));\n\t}\n}\n\n\nexport function _isCopilotServerCapBelowSelectedContextWindow(this: AgentSession, assistantMessage: AssistantMessage): boolean {\n\tif (!this.model || this.model.provider !== \"github-copilot\" || !assistantMessage.errorMessage) return false;\n\tconst promptLimitError = parseCopilotPromptLimitError(assistantMessage.errorMessage);\n\t// Compare against the effective input budget (the model's real prompt cap), not the displayed\n\t// total window. A rejection at the prompt cap is a normal overflow we should compact-and-retry;\n\t// only a rejection *below* the cap (e.g. a missing long-context entitlement dropping the account\n\t// to a lower server tier) keeps the friendly error visible instead of silently compacting down.\n\treturn promptLimitError !== undefined && getEffectiveInputBudget(this.model) > promptLimitError.limitTokens;\n}\n\nexport function shouldRetryAfterThresholdCompaction(assistantMessage: AssistantMessage): boolean {\n\treturn assistantMessage.stopReason === \"length\" && assistantMessage.usage.output > 0;\n}\n\n/**\n * Internal: remove an incomplete assistant from retry context before auto-continuing after compaction.\n */\n\nexport function _dropTrailingAutoCompactionRetryAssistantIfPresent(this: AgentSession): void {\n\tconst messages = this.agent.state.messages;\n\tconst lastMsg = messages[messages.length - 1];\n\tif (lastMsg?.role !== \"assistant\") return;\n\tconst stopReason = (lastMsg as AssistantMessage).stopReason;\n\tif (stopReason === \"error\" || stopReason === \"length\") {\n\t\tthis.agent.state.messages = messages.slice(0, -1);\n\t}\n}\n\n/**\n * Internal: schedule a live post-event continuation probe after compaction_end listeners can flush queues.\n */\n\nexport function _schedulePostAutoCompactionContinuationProbe(this: AgentSession,\n\treason: \"overflow\" | \"threshold\",\n\twillRetry: boolean,\n): void {\n\tif (reason === \"overflow\" && willRetry) {\n\t\tconst token = this._overflowPostCompactionContinuationToken + 1;\n\t\tthis._overflowPostCompactionContinuationToken = token;\n\t\tlet pending: Promise<void>;\n\t\tpending = new Promise<void>((resolve) => {\n\t\t\tsetTimeout(() => {\n\t\t\t\tvoid (async () => {\n\t\t\t\t\ttry {\n\t\t\t\t\t\tif (this._overflowPostCompactionContinuationToken !== token) return;\n\t\t\t\t\t\tif (this.isCompacting || this.isStreaming) return;\n\t\t\t\t\t\tawait this._resumeAfterAutoCompaction();\n\t\t\t\t\t} finally {\n\t\t\t\t\t\tif (this._pendingOverflowPostCompactionContinuation === pending) {\n\t\t\t\t\t\t\tthis._pendingOverflowPostCompactionContinuation = undefined;\n\t\t\t\t\t\t}\n\t\t\t\t\t\tresolve();\n\t\t\t\t\t}\n\t\t\t\t})();\n\t\t\t}, 100);\n\t\t});\n\t\tthis._pendingOverflowPostCompactionContinuation = pending;\n\t\treturn;\n\t}\n\n\tsetTimeout(() => {\n\t\tif (this.isCompacting || this.isStreaming) {\n\t\t\treturn;\n\t\t}\n\n\t\tif (willRetry) {\n\t\t\tvoid this._resumeAfterAutoCompaction();\n\t\t\treturn;\n\t\t}\n\n\t\tif (!this.agent.hasQueuedMessages()) {\n\t\t\treturn;\n\t\t}\n\n\t\tvoid this._resumeAfterAutoCompaction();\n\t}, 100);\n}\n\nexport async function _awaitPendingOverflowPostCompactionContinuation(this: AgentSession): Promise<void> {\n\tconst pending = this._pendingOverflowPostCompactionContinuation;\n\tif (pending === undefined) return;\n\tawait pending;\n}\n\n/**\n * Internal: resume generation after successful auto-compaction only when active work remains.\n */\n\nexport async function _resumeAfterAutoCompaction(this: AgentSession): Promise<void> {\n\ttry {\n\t\tawait this._runAgentContinue();\n\t} catch (error) {\n\t\tconst message = error instanceof Error ? error.message : String(error);\n\t\tthis._emit({\n\t\t\ttype: \"agent_continue_error\",\n\t\t\tsource: \"post_compaction\",\n\t\t\terrorMessage: `Post-compaction continuation failed: ${message}`,\n\t\t});\n\t}\n}\n\n\nfunction overflowUnresolved(reason: \"overflow\" | \"threshold\", aborted = false): boolean | undefined {\n\treturn reason === \"overflow\" && !aborted ? true : undefined;\n}\n\nexport async function _runAutoCompaction(this: AgentSession, reason: \"overflow\" | \"threshold\", willRetry: boolean): Promise<void> {\n\tthis._emit({ type: \"compaction_start\", reason });\n\tthis._autoCompactionAbortController = new AbortController();\n\n\ttry {\n\t\tif (!this.model) {\n\t\t\tthis._emit({\n\t\t\t\ttype: \"compaction_end\",\n\t\t\t\treason,\n\t\t\t\tresult: undefined,\n\t\t\t\taborted: false,\n\t\t\t\twillRetry: false,\n\t\t\t\tunresolvedOverflow: overflowUnresolved(reason),\n\t\t\t});\n\t\t\treturn;\n\t\t}\n\n\t\t// Auth is resolved lazily: only called when the planner fallback is needed.\n\t\t// This allows extension-provided deletion requests to run before auth is checked,\n\t\t// enabling local extension compaction even when API credentials are unavailable.\n\t\t// Auto-mode resolver returns undefined (rather than throwing) when auth is missing;\n\t\t// overflow recovery then falls back to deterministic no-auth eviction, while\n\t\t// threshold compaction keeps the previous no-op behavior.\n\t\tconst model = this.model;\n\t\tconst result = await this._applyContextVerbatimCompaction({\n\t\t\tresolvePlannerAuth: async () => {\n\t\t\t\tconst authResult = await this._modelRegistry.getApiKeyAndHeaders(model);\n\t\t\t\tif (!authResult.ok || !authResult.apiKey) {\n\t\t\t\t\treturn undefined;\n\t\t\t\t}\n\t\t\t\treturn { apiKey: authResult.apiKey, headers: authResult.headers };\n\t\t\t},\n\t\t\tabortController: this._autoCompactionAbortController,\n\t\t\tbackupLabel: reason === \"overflow\" ? \"overflow-auto-compact\" : \"auto-compact\",\n\t\t\treason,\n\t\t});\n\t\tif (!result) {\n\t\t\tthis._emit({\n\t\t\t\ttype: \"compaction_end\",\n\t\t\t\treason,\n\t\t\t\tresult: undefined,\n\t\t\t\taborted: false,\n\t\t\t\twillRetry: false,\n\t\t\t\tunresolvedOverflow: overflowUnresolved(reason),\n\t\t\t});\n\t\t\treturn;\n\t\t}\n\n\t\tif (willRetry) {\n\t\t\tthis._dropTrailingAutoCompactionRetryAssistantIfPresent();\n\t\t}\n\n\t\tthis._emit({ type: \"compaction_end\", reason, result, aborted: false, willRetry });\n\t\tthis._schedulePostAutoCompactionContinuationProbe(reason, willRetry);\n\t} catch (error) {\n\t\tconst errorMessage = error instanceof Error ? error.message : \"compaction failed\";\n\t\tconst aborted = errorMessage === \"Compaction cancelled\" || (error instanceof Error && error.name === \"AbortError\");\n\t\tthis._emit({\n\t\t\ttype: \"compaction_end\",\n\t\t\treason,\n\t\t\tresult: undefined,\n\t\t\taborted,\n\t\t\twillRetry: false,\n\t\t\tunresolvedOverflow: overflowUnresolved(reason, aborted),\n\t\t\terrorMessage: aborted\n\t\t\t\t? undefined\n\t\t\t\t: reason === \"overflow\"\n\t\t\t\t\t? `Context overflow recovery failed: ${errorMessage}`\n\t\t\t\t\t: `Auto-compaction failed: ${errorMessage}`,\n\t\t});\n\t} finally {\n\t\tthis._autoCompactionAbortController = undefined;\n\t}\n}\n\n/**\n * Toggle auto-compaction setting.\n */\n\nexport const agentSessionAutoCompactionMethods = {\n\t_awaitPendingOverflowPostCompactionContinuation,\n\t_checkCompaction,\n\t_isCopilotServerCapBelowSelectedContextWindow,\n\t_dropTrailingAutoCompactionRetryAssistantIfPresent,\n\t_schedulePostAutoCompactionContinuationProbe,\n\t_resumeAfterAutoCompaction,\n\t_runAutoCompaction,\n};\n"]}
+{"version":3,"file":"agent-session-auto-compaction.d.ts","sourceRoot":"","sources":["../../src/core/agent-session-auto-compaction.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,8BAA8B,CAAC;AAOrE,OAAO,KAAK,EAAE,2BAA2B,IAAI,YAAY,EAAE,MAAM,4BAA4B,CAAC;AAE9F;;;;;;GAMG;AACH,eAAO,MAAM,gCAAgC,IAAI,CAAC;AAElD,eAAO,MAAM,6CAA6C,IAAI,CAAC;AAc/D,wBAAsB,gBAAgB,CAAC,IAAI,EAAE,YAAY,EAAE,gBAAgB,EAAE,gBAAgB,EAAE,gBAAgB,UAAO,GAAG,OAAO,CAAC,IAAI,CAAC,CA8HrI;AAGD,wBAAgB,6CAA6C,CAAC,IAAI,EAAE,YAAY,EAAE,gBAAgB,EAAE,gBAAgB,GAAG,OAAO,CAQ7H;AAED,wBAAgB,uBAAuB,CAAC,gBAAgB,EAAE,gBAAgB,GAAG,OAAO,CAEnF;AAmDD,wBAAgB,8BAA8B,CAAC,gBAAgB,EAAE,gBAAgB,GAAG,OAAO,CAQ1F;AAED,wBAAgB,mCAAmC,CAAC,gBAAgB,EAAE,gBAAgB,GAAG,OAAO,CAE/F;AAED;;GAEG;AAEH,wBAAgB,kDAAkD,CAAC,IAAI,EAAE,YAAY,GAAG,IAAI,CAQ3F;AAED;;GAEG;AAEH,wBAAgB,4CAA4C,CAAC,IAAI,EAAE,YAAY,EAC9E,OAAO,EAAE,UAAU,GAAG,WAAW,EACjC,SAAS,EAAE,OAAO,GAChB,IAAI,CAoCN;AAED,wBAAsB,uCAAuC,CAAC,IAAI,EAAE,YAAY,GAAG,OAAO,CAAC,IAAI,CAAC,CAI/F;AAED;;GAEG;AAEH,wBAAsB,0BAA0B,CAAC,IAAI,EAAE,YAAY,GAAG,OAAO,CAAC,IAAI,CAAC,CAWlF;AAED;;;;;;GAMG;AAEH,wBAAgB,4BAA4B,CAAC,IAAI,EAAE,YAAY,GAAG,IAAI,CAQrE;AAOD,wBAAsB,kBAAkB,CAAC,IAAI,EAAE,YAAY,EAAE,MAAM,EAAE,UAAU,GAAG,WAAW,EAAE,SAAS,EAAE,OAAO,GAAG,OAAO,CAAC,IAAI,CAAC,CAyEhI;AAED;;GAEG;AAEH,eAAO,MAAM,iCAAiC;;;;;;;;;CAS7C,CAAC","sourcesContent":["import type { AssistantMessage } from \"@earendil-works/pi-ai/compat\";\nimport { isContextOverflow } from \"@earendil-works/pi-ai/compat\";\nimport { getEffectiveInputBudget } from \"./context-window.ts\";\nimport { parseCopilotPromptLimitError } from \"./copilot-errors.ts\";\nimport { calculateContextTokens, estimateContextTokens, shouldCompact } from \"./compaction/index.ts\";\nimport { getLatestCompactionBoundaryEntry } from \"./session-manager.ts\";\nimport { MIN_RESPONSES_MAX_OUTPUT_TOKENS } from \"./openai-responses-payload-sanitizer.ts\";\nimport type { AgentSessionInternalSurface as AgentSession } from \"./agent-session-methods.ts\";\n\n/**\n * Upper bound on consecutive automatic continuations of a response that was\n * truncated at the output-token cap (\"length\") while the context is still\n * below the compaction budget. Each continuation regenerates the cut-off turn,\n * so a model that insists on emitting more than its per-turn output cap can\n * still terminate instead of looping forever.\n */\nexport const MAX_LENGTH_CONTINUATION_ATTEMPTS = 3;\n\nexport const MAX_OUTPUT_BUDGET_ERROR_CONTINUATION_ATTEMPTS = 1;\n\ntype JsonValue = string | number | boolean | null | JsonValue[] | { [key: string]: JsonValue };\ntype ProviderErrorDetails = {\n\tmessage?: string;\n\tcode?: string;\n\tparam?: string;\n};\n\nconst OUTPUT_BUDGET_PARAMETER_PATTERN = /\\bmax_output_tokens\\b/;\nconst OUTPUT_BUDGET_UNDERFLOW_PATTERN = new RegExp(\n\t`(?:integer\\\\s+below\\\\s+minimum\\\\s+value|expected\\\\s+(?:a\\\\s+)?value\\\\s*>=\\\\s*${MIN_RESPONSES_MAX_OUTPUT_TOKENS}|got\\\\s+1\\\\s+instead)`,\n);\n\nexport async function _checkCompaction(this: AgentSession, assistantMessage: AssistantMessage, skipAbortedCheck = true): Promise<void> {\n\t// The agent_end path passes skipAbortedCheck=true; the pre-prompt path passes\n\t// false. Only the live turn-completion path may auto-continue a truncated\n\t// response — before a fresh user prompt we must not resume the old turn.\n\tconst isLiveTurnCompletion = skipAbortedCheck;\n\tconst settings = this.settingsManager.getCompactionSettings();\n\tif (!settings.enabled) return;\n\n\t// Skip if message was aborted (user cancelled) - unless skipAbortedCheck is false\n\tif (skipAbortedCheck && assistantMessage.stopReason === \"aborted\") return;\n\n\tconst contextWindow = this.model?.contextWindow ?? 0;\n\n\t// Skip overflow check if the message came from a different model.\n\t// This handles the case where user switched from a smaller-context model (e.g. opus)\n\t// to a larger-context model (e.g. codex) - the overflow error from the old model\n\t// shouldn't trigger compaction for the new model.\n\tconst sameModel =\n\t\tthis.model && assistantMessage.provider === this.model.provider && assistantMessage.model === this.model.id;\n\n\t// Skip compaction checks if this assistant message is older than the latest\n\t// compaction boundary. This prevents a stale pre-compaction usage/error\n\t// from retriggering compaction on the first prompt after compaction.\n\tconst compactionBoundaryEntry = getLatestCompactionBoundaryEntry(this.sessionManager.getBranch());\n\tconst assistantIsFromBeforeCompactionBoundary =\n\t\tcompactionBoundaryEntry !== null &&\n\t\tassistantMessage.timestamp <= new Date(compactionBoundaryEntry.timestamp).getTime();\n\tif (assistantIsFromBeforeCompactionBoundary) {\n\t\treturn;\n\t}\n\n\t// Case 1: Overflow - LLM returned context overflow error\n\t// When Copilot rejects a 1m client-budget prompt at a lower server cap (for example\n\t// because long-context/usage-based billing entitlement is missing), leave the friendly\n\t// error visible instead of auto-compacting down to a smaller server tier silently.\n\tif (sameModel && this._isCopilotServerCapBelowSelectedContextWindow(assistantMessage)) {\n\t\treturn;\n\t}\n\tif (sameModel && isContextOverflow(assistantMessage, contextWindow)) {\n\t\tconst willRetry = assistantMessage.stopReason !== \"stop\";\n\t\tif (!willRetry) {\n\t\t\tawait this._runAutoCompaction(\"overflow\", false);\n\t\t\treturn;\n\t\t}\n\n\t\tif (this._overflowRecoveryAttempted) {\n\t\t\tthis._emit({\n\t\t\t\ttype: \"compaction_end\",\n\t\t\t\treason: \"overflow\",\n\t\t\t\tresult: undefined,\n\t\t\t\taborted: false,\n\t\t\t\twillRetry: false,\n\t\t\t\tunresolvedOverflow: true,\n\t\t\t\terrorMessage:\n\t\t\t\t\t\"Context overflow recovery failed after one compact-and-retry attempt. Try reducing context or switching to a larger-context model.\",\n\t\t\t});\n\t\t\treturn;\n\t\t}\n\n\t\tthis._overflowRecoveryAttempted = true;\n\t\t// Remove the error message from agent state (it IS saved to session for history,\n\t\t// but we don't want it in context for the retry)\n\t\tconst messages = this.agent.state.messages;\n\t\tif (messages.length > 0 && messages[messages.length - 1].role === \"assistant\") {\n\t\t\tthis.agent.state.messages = messages.slice(0, -1);\n\t\t}\n\t\tawait this._runAutoCompaction(\"overflow\", willRetry);\n\t\treturn;\n\t}\n\n\t// Case 2: Threshold - context is getting large\n\t// For error messages (no usage data), estimate from last successful response.\n\t// This ensures sessions that hit persistent API errors (e.g. 529) can still compact.\n\tlet contextTokens: number;\n\tif (assistantMessage.stopReason === \"error\") {\n\t\tconst messages = this.agent.state.messages;\n\t\tconst estimate = estimateContextTokens(messages);\n\t\tif (estimate.lastUsageIndex === null) return; // No usage data at all\n\t\t// Verify the usage source is post-compaction. Kept pre-compaction messages\n\t\t// have stale usage reflecting the old (larger) context and would falsely\n\t\t// trigger compaction right after one just finished.\n\t\tconst usageMsg = messages[estimate.lastUsageIndex];\n\t\tif (\n\t\t\tcompactionBoundaryEntry &&\n\t\t\tusageMsg.role === \"assistant\" &&\n\t\t\t(usageMsg as AssistantMessage).timestamp <= new Date(compactionBoundaryEntry.timestamp).getTime()\n\t\t) {\n\t\t\treturn;\n\t\t}\n\t\tcontextTokens = estimate.tokens;\n\t} else {\n\t\tcontextTokens = calculateContextTokens(assistantMessage.usage);\n\t}\n\t// Compact against the effective input budget (the hard prompt cap for providers like Copilot\n\t// that advertise a larger total window) so we compact before overrunning the server-side limit\n\t// rather than relying on reactive overflow recovery near the cap.\n\tconst compactionBudget = this.model ? getEffectiveInputBudget(this.model) : contextWindow;\n\tif (shouldCompact(contextTokens, compactionBudget, settings)) {\n\t\tconst willRetry = shouldRetryAfterThresholdCompaction(assistantMessage);\n\t\tif (willRetry && isRetryWorthyOutputBudgetError(assistantMessage)) {\n\t\t\tif (this._outputBudgetErrorContinuationAttempts >= MAX_OUTPUT_BUDGET_ERROR_CONTINUATION_ATTEMPTS) {\n\t\t\t\tthis._emit({\n\t\t\t\t\ttype: \"compaction_end\",\n\t\t\t\t\treason: \"threshold\",\n\t\t\t\t\tresult: undefined,\n\t\t\t\t\taborted: false,\n\t\t\t\t\twillRetry: false,\n\t\t\t\t\terrorMessage:\n\t\t\t\t\t\t\"Output-budget recovery stopped after a compact-and-retry attempt. Try reducing context or switching to a larger-context model.\",\n\t\t\t\t});\n\t\t\t\treturn;\n\t\t\t}\n\t\t\tthis._outputBudgetErrorContinuationAttempts += 1;\n\t\t}\n\t\tawait this._runAutoCompaction(\"threshold\", willRetry);\n\t\treturn;\n\t}\n\n\t// A response truncated at the output-token cap (\"length\") with the context\n\t// still below the compaction budget is genuine work cut off mid-flight, not a\n\t// context overflow. Compaction would not free any room, so continue the\n\t// generation directly instead of dead-ending on the truncation and leaving\n\t// the task half-finished.\n\tif (isLiveTurnCompletion && isRetryWorthyLengthStop(assistantMessage)) {\n\t\tthis._resumeAfterLengthTruncation();\n\t}\n}\n\n\nexport function _isCopilotServerCapBelowSelectedContextWindow(this: AgentSession, assistantMessage: AssistantMessage): boolean {\n\tif (!this.model || this.model.provider !== \"github-copilot\" || !assistantMessage.errorMessage) return false;\n\tconst promptLimitError = parseCopilotPromptLimitError(assistantMessage.errorMessage);\n\t// Compare against the effective input budget (the model's real prompt cap), not the displayed\n\t// total window. A rejection at the prompt cap is a normal overflow we should compact-and-retry;\n\t// only a rejection *below* the cap (e.g. a missing long-context entitlement dropping the account\n\t// to a lower server tier) keeps the friendly error visible instead of silently compacting down.\n\treturn promptLimitError !== undefined && getEffectiveInputBudget(this.model) > promptLimitError.limitTokens;\n}\n\nexport function isRetryWorthyLengthStop(assistantMessage: AssistantMessage): boolean {\n\treturn assistantMessage.stopReason === \"length\" && assistantMessage.usage.output > 0;\n}\n\nfunction isJsonRecord(value: JsonValue): value is { [key: string]: JsonValue } {\n\treturn typeof value === \"object\" && value !== null && !Array.isArray(value);\n}\n\nfunction stringField(record: { [key: string]: JsonValue }, key: string): string | undefined {\n\tconst value = record[key];\n\treturn typeof value === \"string\" ? value : undefined;\n}\n\nfunction parseProviderErrorDetails(errorMessage: string): ProviderErrorDetails | undefined {\n\tconst start = errorMessage.indexOf(\"{\");\n\tconst end = errorMessage.lastIndexOf(\"}\");\n\tif (start === -1 || end <= start) return undefined;\n\n\ttry {\n\t\tconst parsed = JSON.parse(errorMessage.slice(start, end + 1)) as JsonValue;\n\t\tif (!isJsonRecord(parsed)) return undefined;\n\n\t\tconst nestedError = parsed.error;\n\t\tif (nestedError !== undefined && isJsonRecord(nestedError)) {\n\t\t\treturn {\n\t\t\t\tmessage: stringField(nestedError, \"message\"),\n\t\t\t\tcode: stringField(nestedError, \"code\") ?? stringField(parsed, \"code\"),\n\t\t\t\tparam: stringField(nestedError, \"param\") ?? stringField(parsed, \"param\"),\n\t\t\t};\n\t\t}\n\n\t\treturn {\n\t\t\tmessage: stringField(parsed, \"message\"),\n\t\t\tcode: stringField(parsed, \"code\"),\n\t\t\tparam: stringField(parsed, \"param\"),\n\t\t};\n\t} catch {\n\t\treturn undefined;\n\t}\n}\n\nfunction isOutputBudgetUnderflowText(text: string): boolean {\n\tconst message = text.toLowerCase();\n\treturn OUTPUT_BUDGET_PARAMETER_PATTERN.test(message) && OUTPUT_BUDGET_UNDERFLOW_PATTERN.test(message);\n}\n\nfunction isStructuredOutputBudgetUnderflow(details: ProviderErrorDetails): boolean {\n\tconst message = details.message?.toLowerCase() ?? \"\";\n\tconst param = details.param?.toLowerCase();\n\tif (!OUTPUT_BUDGET_UNDERFLOW_PATTERN.test(message)) return false;\n\treturn param !== undefined ? OUTPUT_BUDGET_PARAMETER_PATTERN.test(param) : OUTPUT_BUDGET_PARAMETER_PATTERN.test(message);\n}\n\nexport function isRetryWorthyOutputBudgetError(assistantMessage: AssistantMessage): boolean {\n\tif (assistantMessage.stopReason !== \"error\" || !assistantMessage.errorMessage) return false;\n\tif (assistantMessage.api !== \"openai-responses\") return false;\n\n\tconst structuredDetails = parseProviderErrorDetails(assistantMessage.errorMessage);\n\tif (structuredDetails && isStructuredOutputBudgetUnderflow(structuredDetails)) return true;\n\n\treturn isOutputBudgetUnderflowText(assistantMessage.errorMessage);\n}\n\nexport function shouldRetryAfterThresholdCompaction(assistantMessage: AssistantMessage): boolean {\n\treturn isRetryWorthyLengthStop(assistantMessage) || isRetryWorthyOutputBudgetError(assistantMessage);\n}\n\n/**\n * Internal: remove an incomplete assistant from retry context before auto-continuing after compaction.\n */\n\nexport function _dropTrailingAutoCompactionRetryAssistantIfPresent(this: AgentSession): void {\n\tconst messages = this.agent.state.messages;\n\tconst lastMsg = messages[messages.length - 1];\n\tif (lastMsg?.role !== \"assistant\") return;\n\tconst stopReason = (lastMsg as AssistantMessage).stopReason;\n\tif (stopReason === \"error\" || stopReason === \"length\") {\n\t\tthis.agent.state.messages = messages.slice(0, -1);\n\t}\n}\n\n/**\n * Internal: schedule a live post-event continuation probe after compaction_end listeners can flush queues.\n */\n\nexport function _schedulePostAutoCompactionContinuationProbe(this: AgentSession,\n\t_reason: \"overflow\" | \"threshold\",\n\twillRetry: boolean,\n): void {\n\tif (willRetry) {\n\t\tconst token = this._postCompactionContinuationToken + 1;\n\t\tthis._postCompactionContinuationToken = token;\n\t\tlet pending: Promise<void>;\n\t\tpending = new Promise<void>((resolve) => {\n\t\t\tsetTimeout(() => {\n\t\t\t\tvoid (async () => {\n\t\t\t\t\ttry {\n\t\t\t\t\t\tif (this._postCompactionContinuationToken !== token) return;\n\t\t\t\t\t\tif (this.isCompacting || this.isStreaming) return;\n\t\t\t\t\t\tawait this._resumeAfterAutoCompaction();\n\t\t\t\t\t} finally {\n\t\t\t\t\t\tif (this._pendingPostCompactionContinuation === pending) {\n\t\t\t\t\t\t\tthis._pendingPostCompactionContinuation = undefined;\n\t\t\t\t\t\t}\n\t\t\t\t\t\tresolve();\n\t\t\t\t\t}\n\t\t\t\t})();\n\t\t\t}, 100);\n\t\t});\n\t\tthis._pendingPostCompactionContinuation = pending;\n\t\treturn;\n\t}\n\n\tsetTimeout(() => {\n\t\tif (this.isCompacting || this.isStreaming) {\n\t\t\treturn;\n\t\t}\n\n\t\tif (!this.agent.hasQueuedMessages()) {\n\t\t\treturn;\n\t\t}\n\n\t\tvoid this._resumeAfterAutoCompaction();\n\t}, 100);\n}\n\nexport async function _awaitPendingPostCompactionContinuation(this: AgentSession): Promise<void> {\n\tconst pending = this._pendingPostCompactionContinuation;\n\tif (pending === undefined) return;\n\tawait pending;\n}\n\n/**\n * Internal: resume generation after successful auto-compaction only when active work remains.\n */\n\nexport async function _resumeAfterAutoCompaction(this: AgentSession): Promise<void> {\n\ttry {\n\t\tawait this._runAgentContinue();\n\t} catch (error) {\n\t\tconst message = error instanceof Error ? error.message : String(error);\n\t\tthis._emit({\n\t\t\ttype: \"agent_continue_error\",\n\t\t\tsource: \"post_compaction\",\n\t\t\terrorMessage: `Post-compaction continuation failed: ${message}`,\n\t\t});\n\t}\n}\n\n/**\n * Internal: resume a response that was truncated at the output-token cap\n * (\"length\") when the context does not warrant compaction. The generation is\n * continued directly so the model finishes the work it was cut off from, rather\n * than dead-ending on the truncation. Bounded by MAX_LENGTH_CONTINUATION_ATTEMPTS\n * so a turn that keeps exceeding the per-turn output cap can still terminate.\n */\n\nexport function _resumeAfterLengthTruncation(this: AgentSession): void {\n\tif (this._lengthContinuationAttempts >= MAX_LENGTH_CONTINUATION_ATTEMPTS) return;\n\tthis._lengthContinuationAttempts += 1;\n\t// agent.continue() rejects an assistant tail; drop the incomplete\n\t// length-stopped message so the preceding user/tool-result anchors the\n\t// continuation. It remains persisted in session history.\n\tthis._dropTrailingAutoCompactionRetryAssistantIfPresent();\n\tthis._schedulePostAutoCompactionContinuationProbe(\"threshold\", true);\n}\n\n\nfunction overflowUnresolved(reason: \"overflow\" | \"threshold\", aborted = false): boolean | undefined {\n\treturn reason === \"overflow\" && !aborted ? true : undefined;\n}\n\nexport async function _runAutoCompaction(this: AgentSession, reason: \"overflow\" | \"threshold\", willRetry: boolean): Promise<void> {\n\tthis._emit({ type: \"compaction_start\", reason });\n\tthis._autoCompactionAbortController = new AbortController();\n\n\ttry {\n\t\tif (!this.model) {\n\t\t\tthis._emit({\n\t\t\t\ttype: \"compaction_end\",\n\t\t\t\treason,\n\t\t\t\tresult: undefined,\n\t\t\t\taborted: false,\n\t\t\t\twillRetry: false,\n\t\t\t\tunresolvedOverflow: overflowUnresolved(reason),\n\t\t\t});\n\t\t\treturn;\n\t\t}\n\n\t\t// Auth is resolved lazily: only called when the planner fallback is needed.\n\t\t// This allows extension-provided deletion requests to run before auth is checked,\n\t\t// enabling local extension compaction even when API credentials are unavailable.\n\t\t// Auto-mode resolver returns undefined (rather than throwing) when auth is missing;\n\t\t// overflow recovery then falls back to deterministic no-auth eviction, while\n\t\t// threshold compaction keeps the previous no-op behavior.\n\t\tconst model = this.model;\n\t\tconst result = await this._applyContextVerbatimCompaction({\n\t\t\tresolvePlannerAuth: async () => {\n\t\t\t\tconst authResult = await this._modelRegistry.getApiKeyAndHeaders(model);\n\t\t\t\tif (!authResult.ok || !authResult.apiKey) {\n\t\t\t\t\treturn undefined;\n\t\t\t\t}\n\t\t\t\treturn { apiKey: authResult.apiKey, headers: authResult.headers };\n\t\t\t},\n\t\t\tabortController: this._autoCompactionAbortController,\n\t\t\tbackupLabel: reason === \"overflow\" ? \"overflow-auto-compact\" : \"auto-compact\",\n\t\t\treason,\n\t\t});\n\t\tif (!result) {\n\t\t\tthis._emit({\n\t\t\t\ttype: \"compaction_end\",\n\t\t\t\treason,\n\t\t\t\tresult: undefined,\n\t\t\t\taborted: false,\n\t\t\t\twillRetry: false,\n\t\t\t\tunresolvedOverflow: overflowUnresolved(reason),\n\t\t\t});\n\t\t\treturn;\n\t\t}\n\n\t\tif (willRetry) {\n\t\t\tthis._dropTrailingAutoCompactionRetryAssistantIfPresent();\n\t\t}\n\n\t\tthis._emit({ type: \"compaction_end\", reason, result, aborted: false, willRetry });\n\t\tthis._schedulePostAutoCompactionContinuationProbe(reason, willRetry);\n\t} catch (error) {\n\t\tconst errorMessage = error instanceof Error ? error.message : \"compaction failed\";\n\t\tconst aborted = errorMessage === \"Compaction cancelled\" || (error instanceof Error && error.name === \"AbortError\");\n\t\tthis._emit({\n\t\t\ttype: \"compaction_end\",\n\t\t\treason,\n\t\t\tresult: undefined,\n\t\t\taborted,\n\t\t\twillRetry: false,\n\t\t\tunresolvedOverflow: overflowUnresolved(reason, aborted),\n\t\t\terrorMessage: aborted\n\t\t\t\t? undefined\n\t\t\t\t: reason === \"overflow\"\n\t\t\t\t\t? `Context overflow recovery failed: ${errorMessage}`\n\t\t\t\t\t: `Auto-compaction failed: ${errorMessage}`,\n\t\t});\n\t} finally {\n\t\tthis._autoCompactionAbortController = undefined;\n\t}\n}\n\n/**\n * Toggle auto-compaction setting.\n */\n\nexport const agentSessionAutoCompactionMethods = {\n\t_awaitPendingPostCompactionContinuation,\n\t_checkCompaction,\n\t_isCopilotServerCapBelowSelectedContextWindow,\n\t_dropTrailingAutoCompactionRetryAssistantIfPresent,\n\t_schedulePostAutoCompactionContinuationProbe,\n\t_resumeAfterAutoCompaction,\n\t_resumeAfterLengthTruncation,\n\t_runAutoCompaction,\n};\n"]}

package/dist/core/agent-session-auto-compaction.js

@@ -3,7 +3,23 @@ import { getEffectiveInputBudget } from "./context-window.js";
 import { parseCopilotPromptLimitError } from "./copilot-errors.js";
 import { calculateContextTokens, estimateContextTokens, shouldCompact } from "./compaction/index.js";
 import { getLatestCompactionBoundaryEntry } from "./session-manager.js";
+import { MIN_RESPONSES_MAX_OUTPUT_TOKENS } from "./openai-responses-payload-sanitizer.js";
+/**
+ * Upper bound on consecutive automatic continuations of a response that was
+ * truncated at the output-token cap ("length") while the context is still
+ * below the compaction budget. Each continuation regenerates the cut-off turn,
+ * so a model that insists on emitting more than its per-turn output cap can
+ * still terminate instead of looping forever.
+ */
+export const MAX_LENGTH_CONTINUATION_ATTEMPTS = 3;
+export const MAX_OUTPUT_BUDGET_ERROR_CONTINUATION_ATTEMPTS = 1;
+const OUTPUT_BUDGET_PARAMETER_PATTERN = /\bmax_output_tokens\b/;
+const OUTPUT_BUDGET_UNDERFLOW_PATTERN = new RegExp(`(?:integer\\s+below\\s+minimum\\s+value|expected\\s+(?:a\\s+)?value\\s*>=\\s*${MIN_RESPONSES_MAX_OUTPUT_TOKENS}|got\\s+1\\s+instead)`);
 export async function _checkCompaction(assistantMessage, skipAbortedCheck = true) {
+    // The agent_end path passes skipAbortedCheck=true; the pre-prompt path passes
+    // false. Only the live turn-completion path may auto-continue a truncated
+    // response — before a fresh user prompt we must not resume the old turn.
+    const isLiveTurnCompletion = skipAbortedCheck;
     const settings = this.settingsManager.getCompactionSettings();
     if (!settings.enabled)
         return;
@@ -88,7 +104,31 @@ export async function _checkCompaction(assistantMessage, skipAbortedCheck = true
     // rather than relying on reactive overflow recovery near the cap.
     const compactionBudget = this.model ? getEffectiveInputBudget(this.model) : contextWindow;
     if (shouldCompact(contextTokens, compactionBudget, settings)) {
-        await this._runAutoCompaction("threshold", shouldRetryAfterThresholdCompaction(assistantMessage));
+        const willRetry = shouldRetryAfterThresholdCompaction(assistantMessage);
+        if (willRetry && isRetryWorthyOutputBudgetError(assistantMessage)) {
+            if (this._outputBudgetErrorContinuationAttempts >= MAX_OUTPUT_BUDGET_ERROR_CONTINUATION_ATTEMPTS) {
+                this._emit({
+                    type: "compaction_end",
+                    reason: "threshold",
+                    result: undefined,
+                    aborted: false,
+                    willRetry: false,
+                    errorMessage: "Output-budget recovery stopped after a compact-and-retry attempt. Try reducing context or switching to a larger-context model.",
+                });
+                return;
+            }
+            this._outputBudgetErrorContinuationAttempts += 1;
+        }
+        await this._runAutoCompaction("threshold", willRetry);
+        return;
+    }
+    // A response truncated at the output-token cap ("length") with the context
+    // still below the compaction budget is genuine work cut off mid-flight, not a
+    // context overflow. Compaction would not free any room, so continue the
+    // generation directly instead of dead-ending on the truncation and leaving
+    // the task half-finished.
+    if (isLiveTurnCompletion && isRetryWorthyLengthStop(assistantMessage)) {
+        this._resumeAfterLengthTruncation();
     }
 }
 export function _isCopilotServerCapBelowSelectedContextWindow(assistantMessage) {
@@ -101,9 +141,67 @@ export function _isCopilotServerCapBelowSelectedContextWindow(assistantMessage)
     // to a lower server tier) keeps the friendly error visible instead of silently compacting down.
     return promptLimitError !== undefined && getEffectiveInputBudget(this.model) > promptLimitError.limitTokens;
 }
-export function shouldRetryAfterThresholdCompaction(assistantMessage) {
+export function isRetryWorthyLengthStop(assistantMessage) {
     return assistantMessage.stopReason === "length" && assistantMessage.usage.output > 0;
 }
+function isJsonRecord(value) {
+    return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+function stringField(record, key) {
+    const value = record[key];
+    return typeof value === "string" ? value : undefined;
+}
+function parseProviderErrorDetails(errorMessage) {
+    const start = errorMessage.indexOf("{");
+    const end = errorMessage.lastIndexOf("}");
+    if (start === -1 || end <= start)
+        return undefined;
+    try {
+        const parsed = JSON.parse(errorMessage.slice(start, end + 1));
+        if (!isJsonRecord(parsed))
+            return undefined;
+        const nestedError = parsed.error;
+        if (nestedError !== undefined && isJsonRecord(nestedError)) {
+            return {
+                message: stringField(nestedError, "message"),
+                code: stringField(nestedError, "code") ?? stringField(parsed, "code"),
+                param: stringField(nestedError, "param") ?? stringField(parsed, "param"),
+            };
+        }
+        return {
+            message: stringField(parsed, "message"),
+            code: stringField(parsed, "code"),
+            param: stringField(parsed, "param"),
+        };
+    }
+    catch {
+        return undefined;
+    }
+}
+function isOutputBudgetUnderflowText(text) {
+    const message = text.toLowerCase();
+    return OUTPUT_BUDGET_PARAMETER_PATTERN.test(message) && OUTPUT_BUDGET_UNDERFLOW_PATTERN.test(message);
+}
+function isStructuredOutputBudgetUnderflow(details) {
+    const message = details.message?.toLowerCase() ?? "";
+    const param = details.param?.toLowerCase();
+    if (!OUTPUT_BUDGET_UNDERFLOW_PATTERN.test(message))
+        return false;
+    return param !== undefined ? OUTPUT_BUDGET_PARAMETER_PATTERN.test(param) : OUTPUT_BUDGET_PARAMETER_PATTERN.test(message);
+}
+export function isRetryWorthyOutputBudgetError(assistantMessage) {
+    if (assistantMessage.stopReason !== "error" || !assistantMessage.errorMessage)
+        return false;
+    if (assistantMessage.api !== "openai-responses")
+        return false;
+    const structuredDetails = parseProviderErrorDetails(assistantMessage.errorMessage);
+    if (structuredDetails && isStructuredOutputBudgetUnderflow(structuredDetails))
+        return true;
+    return isOutputBudgetUnderflowText(assistantMessage.errorMessage);
+}
+export function shouldRetryAfterThresholdCompaction(assistantMessage) {
+    return isRetryWorthyLengthStop(assistantMessage) || isRetryWorthyOutputBudgetError(assistantMessage);
+}
 /**
  * Internal: remove an incomplete assistant from retry context before auto-continuing after compaction.
  */
@@ -120,49 +218,45 @@ export function _dropTrailingAutoCompactionRetryAssistantIfPresent() {
 /**
  * Internal: schedule a live post-event continuation probe after compaction_end listeners can flush queues.
  */
-export function _schedulePostAutoCompactionContinuationProbe(reason, willRetry) {
-    if (reason === "overflow" && willRetry) {
-        const token = this._overflowPostCompactionContinuationToken + 1;
-        this._overflowPostCompactionContinuationToken = token;
+export function _schedulePostAutoCompactionContinuationProbe(_reason, willRetry) {
+    if (willRetry) {
+        const token = this._postCompactionContinuationToken + 1;
+        this._postCompactionContinuationToken = token;
         let pending;
         pending = new Promise((resolve) => {
             setTimeout(() => {
                 void (async () => {
                     try {
-                        if (this._overflowPostCompactionContinuationToken !== token)
+                        if (this._postCompactionContinuationToken !== token)
                             return;
                         if (this.isCompacting || this.isStreaming)
                             return;
                         await this._resumeAfterAutoCompaction();
                     }
                     finally {
-                        if (this._pendingOverflowPostCompactionContinuation === pending) {
-                            this._pendingOverflowPostCompactionContinuation = undefined;
+                        if (this._pendingPostCompactionContinuation === pending) {
+                            this._pendingPostCompactionContinuation = undefined;
                         }
                         resolve();
                     }
                 })();
             }, 100);
         });
-        this._pendingOverflowPostCompactionContinuation = pending;
+        this._pendingPostCompactionContinuation = pending;
         return;
     }
     setTimeout(() => {
         if (this.isCompacting || this.isStreaming) {
             return;
         }
-        if (willRetry) {
-            void this._resumeAfterAutoCompaction();
-            return;
-        }
         if (!this.agent.hasQueuedMessages()) {
             return;
         }
         void this._resumeAfterAutoCompaction();
     }, 100);
 }
-export async function _awaitPendingOverflowPostCompactionContinuation() {
-    const pending = this._pendingOverflowPostCompactionContinuation;
+export async function _awaitPendingPostCompactionContinuation() {
+    const pending = this._pendingPostCompactionContinuation;
     if (pending === undefined)
         return;
     await pending;
@@ -183,6 +277,23 @@ export async function _resumeAfterAutoCompaction() {
         });
     }
 }
+/**
+ * Internal: resume a response that was truncated at the output-token cap
+ * ("length") when the context does not warrant compaction. The generation is
+ * continued directly so the model finishes the work it was cut off from, rather
+ * than dead-ending on the truncation. Bounded by MAX_LENGTH_CONTINUATION_ATTEMPTS
+ * so a turn that keeps exceeding the per-turn output cap can still terminate.
+ */
+export function _resumeAfterLengthTruncation() {
+    if (this._lengthContinuationAttempts >= MAX_LENGTH_CONTINUATION_ATTEMPTS)
+        return;
+    this._lengthContinuationAttempts += 1;
+    // agent.continue() rejects an assistant tail; drop the incomplete
+    // length-stopped message so the preceding user/tool-result anchors the
+    // continuation. It remains persisted in session history.
+    this._dropTrailingAutoCompactionRetryAssistantIfPresent();
+    this._schedulePostAutoCompactionContinuationProbe("threshold", true);
+}
 function overflowUnresolved(reason, aborted = false) {
     return reason === "overflow" && !aborted ? true : undefined;
 }
@@ -262,12 +373,13 @@ export async function _runAutoCompaction(reason, willRetry) {
  * Toggle auto-compaction setting.
  */
 export const agentSessionAutoCompactionMethods = {
-    _awaitPendingOverflowPostCompactionContinuation,
+    _awaitPendingPostCompactionContinuation,
     _checkCompaction,
     _isCopilotServerCapBelowSelectedContextWindow,
     _dropTrailingAutoCompactionRetryAssistantIfPresent,
     _schedulePostAutoCompactionContinuationProbe,
     _resumeAfterAutoCompaction,
+    _resumeAfterLengthTruncation,
     _runAutoCompaction,
 };
 //# sourceMappingURL=agent-session-auto-compaction.js.map

package/dist/core/agent-session-auto-compaction.js.map

@@ -1 +1 @@
-{"version":3,"file":"agent-session-auto-compaction.js","sourceRoot":"","sources":["../../src/core/agent-session-auto-compaction.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,iBAAiB,EAAE,MAAM,8BAA8B,CAAC;AACjE,OAAO,EAAE,uBAAuB,EAAE,MAAM,qBAAqB,CAAC;AAC9D,OAAO,EAAE,4BAA4B,EAAE,MAAM,qBAAqB,CAAC;AACnE,OAAO,EAAE,sBAAsB,EAAE,qBAAqB,EAAE,aAAa,EAAE,MAAM,uBAAuB,CAAC;AACrG,OAAO,EAAE,gCAAgC,EAAE,MAAM,sBAAsB,CAAC;AAGxE,MAAM,CAAC,KAAK,UAAU,gBAAgB,CAAqB,gBAAkC,EAAE,gBAAgB,GAAG,IAAI;IACrH,MAAM,QAAQ,GAAG,IAAI,CAAC,eAAe,CAAC,qBAAqB,EAAE,CAAC;IAC9D,IAAI,CAAC,QAAQ,CAAC,OAAO;QAAE,OAAO;IAE9B,kFAAkF;IAClF,IAAI,gBAAgB,IAAI,gBAAgB,CAAC,UAAU,KAAK,SAAS;QAAE,OAAO;IAE1E,MAAM,aAAa,GAAG,IAAI,CAAC,KAAK,EAAE,aAAa,IAAI,CAAC,CAAC;IAErD,kEAAkE;IAClE,qFAAqF;IACrF,iFAAiF;IACjF,kDAAkD;IAClD,MAAM,SAAS,GACd,IAAI,CAAC,KAAK,IAAI,gBAAgB,CAAC,QAAQ,KAAK,IAAI,CAAC,KAAK,CAAC,QAAQ,IAAI,gBAAgB,CAAC,KAAK,KAAK,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC;IAE7G,4EAA4E;IAC5E,wEAAwE;IACxE,qEAAqE;IACrE,MAAM,uBAAuB,GAAG,gCAAgC,CAAC,IAAI,CAAC,cAAc,CAAC,SAAS,EAAE,CAAC,CAAC;IAClG,MAAM,uCAAuC,GAC5C,uBAAuB,KAAK,IAAI;QAChC,gBAAgB,CAAC,SAAS,IAAI,IAAI,IAAI,CAAC,uBAAuB,CAAC,SAAS,CAAC,CAAC,OAAO,EAAE,CAAC;IACrF,IAAI,uCAAuC,EAAE,CAAC;QAC7C,OAAO;IACR,CAAC;IAED,yDAAyD;IACzD,oFAAoF;IACpF,uFAAuF;IACvF,mFAAmF;IACnF,IAAI,SAAS,IAAI,IAAI,CAAC,6CAA6C,CAAC,gBAAgB,CAAC,EAAE,CAAC;QACvF,OAAO;IACR,CAAC;IACD,IAAI,SAAS,IAAI,iBAAiB,CAAC,gBAAgB,EAAE,aAAa,CAAC,EAAE,CAAC;QACrE,MAAM,SAAS,GAAG,gBAAgB,CAAC,UAAU,KAAK,MAAM,CAAC;QACzD,IAAI,CAAC,SAAS,EAAE,CAAC;YAChB,MAAM,IAAI,CAAC,kBAAkB,CAAC,UAAU,EAAE,KAAK,CAAC,CAAC;YACjD,OAAO;QACR,CAAC;QAED,IAAI,IAAI,CAAC,0BAA0B,EAAE,CAAC;YACrC,IAAI,CAAC,KAAK,CAAC;gBACV,IAAI,EAAE,gBAAgB;gBACtB,MAAM,EAAE,UAAU;gBAClB,MAAM,EAAE,SAAS;gBACjB,OAAO,EAAE,KAAK;gBACd,SAAS,EAAE,KAAK;gBAChB,kBAAkB,EAAE,IAAI;gBACxB,YAAY,EACX,oIAAoI;aACrI,CAAC,CAAC;YACH,OAAO;QACR,CAAC;QAED,IAAI,CAAC,0BAA0B,GAAG,IAAI,CAAC;QACvC,iFAAiF;QACjF,iDAAiD;QACjD,MAAM,QAAQ,GAAG,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,QAAQ,CAAC;QAC3C,IAAI,QAAQ,CAAC,MAAM,GAAG,CAAC,IAAI,QAAQ,CAAC,QAAQ,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;YAC/E,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,QAAQ,GAAG,QAAQ,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;QACnD,CAAC;QACD,MAAM,IAAI,CAAC,kBAAkB,CAAC,UAAU,EAAE,SAAS,CAAC,CAAC;QACrD,OAAO;IACR,CAAC;IAED,+CAA+C;IAC/C,8EAA8E;IAC9E,qFAAqF;IACrF,IAAI,aAAqB,CAAC;IAC1B,IAAI,gBAAgB,CAAC,UAAU,KAAK,OAAO,EAAE,CAAC;QAC7C,MAAM,QAAQ,GAAG,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,QAAQ,CAAC;QAC3C,MAAM,QAAQ,GAAG,qBAAqB,CAAC,QAAQ,CAAC,CAAC;QACjD,IAAI,QAAQ,CAAC,cAAc,KAAK,IAAI;YAAE,OAAO,CAAC,uBAAuB;QACrE,2EAA2E;QAC3E,yEAAyE;QACzE,oDAAoD;QACpD,MAAM,QAAQ,GAAG,QAAQ,CAAC,QAAQ,CAAC,cAAc,CAAC,CAAC;QACnD,IACC,uBAAuB;YACvB,QAAQ,CAAC,IAAI,KAAK,WAAW;YAC5B,QAA6B,CAAC,SAAS,IAAI,IAAI,IAAI,CAAC,uBAAuB,CAAC,SAAS,CAAC,CAAC,OAAO,EAAE,EAChG,CAAC;YACF,OAAO;QACR,CAAC;QACD,aAAa,GAAG,QAAQ,CAAC,MAAM,CAAC;IACjC,CAAC;SAAM,CAAC;QACP,aAAa,GAAG,sBAAsB,CAAC,gBAAgB,CAAC,KAAK,CAAC,CAAC;IAChE,CAAC;IACD,6FAA6F;IAC7F,+FAA+F;IAC/F,kEAAkE;IAClE,MAAM,gBAAgB,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,uBAAuB,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,aAAa,CAAC;IAC1F,IAAI,aAAa,CAAC,aAAa,EAAE,gBAAgB,EAAE,QAAQ,CAAC,EAAE,CAAC;QAC9D,MAAM,IAAI,CAAC,kBAAkB,CAAC,WAAW,EAAE,mCAAmC,CAAC,gBAAgB,CAAC,CAAC,CAAC;IACnG,CAAC;AACF,CAAC;AAGD,MAAM,UAAU,6CAA6C,CAAqB,gBAAkC;IACnH,IAAI,CAAC,IAAI,CAAC,KAAK,IAAI,IAAI,CAAC,KAAK,CAAC,QAAQ,KAAK,gBAAgB,IAAI,CAAC,gBAAgB,CAAC,YAAY;QAAE,OAAO,KAAK,CAAC;IAC5G,MAAM,gBAAgB,GAAG,4BAA4B,CAAC,gBAAgB,CAAC,YAAY,CAAC,CAAC;IACrF,8FAA8F;IAC9F,gGAAgG;IAChG,iGAAiG;IACjG,gGAAgG;IAChG,OAAO,gBAAgB,KAAK,SAAS,IAAI,uBAAuB,CAAC,IAAI,CAAC,KAAK,CAAC,GAAG,gBAAgB,CAAC,WAAW,CAAC;AAC7G,CAAC;AAED,MAAM,UAAU,mCAAmC,CAAC,gBAAkC;IACrF,OAAO,gBAAgB,CAAC,UAAU,KAAK,QAAQ,IAAI,gBAAgB,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC;AACtF,CAAC;AAED;;GAEG;AAEH,MAAM,UAAU,kDAAkD;IACjE,MAAM,QAAQ,GAAG,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,QAAQ,CAAC;IAC3C,MAAM,OAAO,GAAG,QAAQ,CAAC,QAAQ,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;IAC9C,IAAI,OAAO,EAAE,IAAI,KAAK,WAAW;QAAE,OAAO;IAC1C,MAAM,UAAU,GAAI,OAA4B,CAAC,UAAU,CAAC;IAC5D,IAAI,UAAU,KAAK,OAAO,IAAI,UAAU,KAAK,QAAQ,EAAE,CAAC;QACvD,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,QAAQ,GAAG,QAAQ,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;IACnD,CAAC;AACF,CAAC;AAED;;GAEG;AAEH,MAAM,UAAU,4CAA4C,CAC3D,MAAgC,EAChC,SAAkB;IAElB,IAAI,MAAM,KAAK,UAAU,IAAI,SAAS,EAAE,CAAC;QACxC,MAAM,KAAK,GAAG,IAAI,CAAC,wCAAwC,GAAG,CAAC,CAAC;QAChE,IAAI,CAAC,wCAAwC,GAAG,KAAK,CAAC;QACtD,IAAI,OAAsB,CAAC;QAC3B,OAAO,GAAG,IAAI,OAAO,CAAO,CAAC,OAAO,EAAE,EAAE;YACvC,UAAU,CAAC,GAAG,EAAE;gBACf,KAAK,CAAC,KAAK,IAAI,EAAE;oBAChB,IAAI,CAAC;wBACJ,IAAI,IAAI,CAAC,wCAAwC,KAAK,KAAK;4BAAE,OAAO;wBACpE,IAAI,IAAI,CAAC,YAAY,IAAI,IAAI,CAAC,WAAW;4BAAE,OAAO;wBAClD,MAAM,IAAI,CAAC,0BAA0B,EAAE,CAAC;oBACzC,CAAC;4BAAS,CAAC;wBACV,IAAI,IAAI,CAAC,0CAA0C,KAAK,OAAO,EAAE,CAAC;4BACjE,IAAI,CAAC,0CAA0C,GAAG,SAAS,CAAC;wBAC7D,CAAC;wBACD,OAAO,EAAE,CAAC;oBACX,CAAC;gBACF,CAAC,CAAC,EAAE,CAAC;YACN,CAAC,EAAE,GAAG,CAAC,CAAC;QACT,CAAC,CAAC,CAAC;QACH,IAAI,CAAC,0CAA0C,GAAG,OAAO,CAAC;QAC1D,OAAO;IACR,CAAC;IAED,UAAU,CAAC,GAAG,EAAE;QACf,IAAI,IAAI,CAAC,YAAY,IAAI,IAAI,CAAC,WAAW,EAAE,CAAC;YAC3C,OAAO;QACR,CAAC;QAED,IAAI,SAAS,EAAE,CAAC;YACf,KAAK,IAAI,CAAC,0BAA0B,EAAE,CAAC;YACvC,OAAO;QACR,CAAC;QAED,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,iBAAiB,EAAE,EAAE,CAAC;YACrC,OAAO;QACR,CAAC;QAED,KAAK,IAAI,CAAC,0BAA0B,EAAE,CAAC;IACxC,CAAC,EAAE,GAAG,CAAC,CAAC;AACT,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,+CAA+C;IACpE,MAAM,OAAO,GAAG,IAAI,CAAC,0CAA0C,CAAC;IAChE,IAAI,OAAO,KAAK,SAAS;QAAE,OAAO;IAClC,MAAM,OAAO,CAAC;AACf,CAAC;AAED;;GAEG;AAEH,MAAM,CAAC,KAAK,UAAU,0BAA0B;IAC/C,IAAI,CAAC;QACJ,MAAM,IAAI,CAAC,iBAAiB,EAAE,CAAC;IAChC,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QAChB,MAAM,OAAO,GAAG,KAAK,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC;QACvE,IAAI,CAAC,KAAK,CAAC;YACV,IAAI,EAAE,sBAAsB;YAC5B,MAAM,EAAE,iBAAiB;YACzB,YAAY,EAAE,wCAAwC,OAAO,EAAE;SAC/D,CAAC,CAAC;IACJ,CAAC;AACF,CAAC;AAGD,SAAS,kBAAkB,CAAC,MAAgC,EAAE,OAAO,GAAG,KAAK;IAC5E,OAAO,MAAM,KAAK,UAAU,IAAI,CAAC,OAAO,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,SAAS,CAAC;AAC7D,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,kBAAkB,CAAqB,MAAgC,EAAE,SAAkB;IAChH,IAAI,CAAC,KAAK,CAAC,EAAE,IAAI,EAAE,kBAAkB,EAAE,MAAM,EAAE,CAAC,CAAC;IACjD,IAAI,CAAC,8BAA8B,GAAG,IAAI,eAAe,EAAE,CAAC;IAE5D,IAAI,CAAC;QACJ,IAAI,CAAC,IAAI,CAAC,KAAK,EAAE,CAAC;YACjB,IAAI,CAAC,KAAK,CAAC;gBACV,IAAI,EAAE,gBAAgB;gBACtB,MAAM;gBACN,MAAM,EAAE,SAAS;gBACjB,OAAO,EAAE,KAAK;gBACd,SAAS,EAAE,KAAK;gBAChB,kBAAkB,EAAE,kBAAkB,CAAC,MAAM,CAAC;aAC9C,CAAC,CAAC;YACH,OAAO;QACR,CAAC;QAED,4EAA4E;QAC5E,kFAAkF;QAClF,iFAAiF;QACjF,oFAAoF;QACpF,6EAA6E;QAC7E,0DAA0D;QAC1D,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC;QACzB,MAAM,MAAM,GAAG,MAAM,IAAI,CAAC,+BAA+B,CAAC;YACzD,kBAAkB,EAAE,KAAK,IAAI,EAAE;gBAC9B,MAAM,UAAU,GAAG,MAAM,IAAI,CAAC,cAAc,CAAC,mBAAmB,CAAC,KAAK,CAAC,CAAC;gBACxE,IAAI,CAAC,UAAU,CAAC,EAAE,IAAI,CAAC,UAAU,CAAC,MAAM,EAAE,CAAC;oBAC1C,OAAO,SAAS,CAAC;gBAClB,CAAC;gBACD,OAAO,EAAE,MAAM,EAAE,UAAU,CAAC,MAAM,EAAE,OAAO,EAAE,UAAU,CAAC,OAAO,EAAE,CAAC;YACnE,CAAC;YACD,eAAe,EAAE,IAAI,CAAC,8BAA8B;YACpD,WAAW,EAAE,MAAM,KAAK,UAAU,CAAC,CAAC,CAAC,uBAAuB,CAAC,CAAC,CAAC,cAAc;YAC7E,MAAM;SACN,CAAC,CAAC;QACH,IAAI,CAAC,MAAM,EAAE,CAAC;YACb,IAAI,CAAC,KAAK,CAAC;gBACV,IAAI,EAAE,gBAAgB;gBACtB,MAAM;gBACN,MAAM,EAAE,SAAS;gBACjB,OAAO,EAAE,KAAK;gBACd,SAAS,EAAE,KAAK;gBAChB,kBAAkB,EAAE,kBAAkB,CAAC,MAAM,CAAC;aAC9C,CAAC,CAAC;YACH,OAAO;QACR,CAAC;QAED,IAAI,SAAS,EAAE,CAAC;YACf,IAAI,CAAC,kDAAkD,EAAE,CAAC;QAC3D,CAAC;QAED,IAAI,CAAC,KAAK,CAAC,EAAE,IAAI,EAAE,gBAAgB,EAAE,MAAM,EAAE,MAAM,EAAE,OAAO,EAAE,KAAK,EAAE,SAAS,EAAE,CAAC,CAAC;QAClF,IAAI,CAAC,4CAA4C,CAAC,MAAM,EAAE,SAAS,CAAC,CAAC;IACtE,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QAChB,MAAM,YAAY,GAAG,KAAK,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,mBAAmB,CAAC;QAClF,MAAM,OAAO,GAAG,YAAY,KAAK,sBAAsB,IAAI,CAAC,KAAK,YAAY,KAAK,IAAI,KAAK,CAAC,IAAI,KAAK,YAAY,CAAC,CAAC;QACnH,IAAI,CAAC,KAAK,CAAC;YACV,IAAI,EAAE,gBAAgB;YACtB,MAAM;YACN,MAAM,EAAE,SAAS;YACjB,OAAO;YACP,SAAS,EAAE,KAAK;YAChB,kBAAkB,EAAE,kBAAkB,CAAC,MAAM,EAAE,OAAO,CAAC;YACvD,YAAY,EAAE,OAAO;gBACpB,CAAC,CAAC,SAAS;gBACX,CAAC,CAAC,MAAM,KAAK,UAAU;oBACtB,CAAC,CAAC,qCAAqC,YAAY,EAAE;oBACrD,CAAC,CAAC,2BAA2B,YAAY,EAAE;SAC7C,CAAC,CAAC;IACJ,CAAC;YAAS,CAAC;QACV,IAAI,CAAC,8BAA8B,GAAG,SAAS,CAAC;IACjD,CAAC;AACF,CAAC;AAED;;GAEG;AAEH,MAAM,CAAC,MAAM,iCAAiC,GAAG;IAChD,+CAA+C;IAC/C,gBAAgB;IAChB,6CAA6C;IAC7C,kDAAkD;IAClD,4CAA4C;IAC5C,0BAA0B;IAC1B,kBAAkB;CAClB,CAAC","sourcesContent":["import type { AssistantMessage } from \"@earendil-works/pi-ai/compat\";\nimport { isContextOverflow } from \"@earendil-works/pi-ai/compat\";\nimport { getEffectiveInputBudget } from \"./context-window.ts\";\nimport { parseCopilotPromptLimitError } from \"./copilot-errors.ts\";\nimport { calculateContextTokens, estimateContextTokens, shouldCompact } from \"./compaction/index.ts\";\nimport { getLatestCompactionBoundaryEntry } from \"./session-manager.ts\";\nimport type { AgentSessionInternalSurface as AgentSession } from \"./agent-session-methods.ts\";\n\nexport async function _checkCompaction(this: AgentSession, assistantMessage: AssistantMessage, skipAbortedCheck = true): Promise<void> {\n\tconst settings = this.settingsManager.getCompactionSettings();\n\tif (!settings.enabled) return;\n\n\t// Skip if message was aborted (user cancelled) - unless skipAbortedCheck is false\n\tif (skipAbortedCheck && assistantMessage.stopReason === \"aborted\") return;\n\n\tconst contextWindow = this.model?.contextWindow ?? 0;\n\n\t// Skip overflow check if the message came from a different model.\n\t// This handles the case where user switched from a smaller-context model (e.g. opus)\n\t// to a larger-context model (e.g. codex) - the overflow error from the old model\n\t// shouldn't trigger compaction for the new model.\n\tconst sameModel =\n\t\tthis.model && assistantMessage.provider === this.model.provider && assistantMessage.model === this.model.id;\n\n\t// Skip compaction checks if this assistant message is older than the latest\n\t// compaction boundary. This prevents a stale pre-compaction usage/error\n\t// from retriggering compaction on the first prompt after compaction.\n\tconst compactionBoundaryEntry = getLatestCompactionBoundaryEntry(this.sessionManager.getBranch());\n\tconst assistantIsFromBeforeCompactionBoundary =\n\t\tcompactionBoundaryEntry !== null &&\n\t\tassistantMessage.timestamp <= new Date(compactionBoundaryEntry.timestamp).getTime();\n\tif (assistantIsFromBeforeCompactionBoundary) {\n\t\treturn;\n\t}\n\n\t// Case 1: Overflow - LLM returned context overflow error\n\t// When Copilot rejects a 1m client-budget prompt at a lower server cap (for example\n\t// because long-context/usage-based billing entitlement is missing), leave the friendly\n\t// error visible instead of auto-compacting down to a smaller server tier silently.\n\tif (sameModel && this._isCopilotServerCapBelowSelectedContextWindow(assistantMessage)) {\n\t\treturn;\n\t}\n\tif (sameModel && isContextOverflow(assistantMessage, contextWindow)) {\n\t\tconst willRetry = assistantMessage.stopReason !== \"stop\";\n\t\tif (!willRetry) {\n\t\t\tawait this._runAutoCompaction(\"overflow\", false);\n\t\t\treturn;\n\t\t}\n\n\t\tif (this._overflowRecoveryAttempted) {\n\t\t\tthis._emit({\n\t\t\t\ttype: \"compaction_end\",\n\t\t\t\treason: \"overflow\",\n\t\t\t\tresult: undefined,\n\t\t\t\taborted: false,\n\t\t\t\twillRetry: false,\n\t\t\t\tunresolvedOverflow: true,\n\t\t\t\terrorMessage:\n\t\t\t\t\t\"Context overflow recovery failed after one compact-and-retry attempt. Try reducing context or switching to a larger-context model.\",\n\t\t\t});\n\t\t\treturn;\n\t\t}\n\n\t\tthis._overflowRecoveryAttempted = true;\n\t\t// Remove the error message from agent state (it IS saved to session for history,\n\t\t// but we don't want it in context for the retry)\n\t\tconst messages = this.agent.state.messages;\n\t\tif (messages.length > 0 && messages[messages.length - 1].role === \"assistant\") {\n\t\t\tthis.agent.state.messages = messages.slice(0, -1);\n\t\t}\n\t\tawait this._runAutoCompaction(\"overflow\", willRetry);\n\t\treturn;\n\t}\n\n\t// Case 2: Threshold - context is getting large\n\t// For error messages (no usage data), estimate from last successful response.\n\t// This ensures sessions that hit persistent API errors (e.g. 529) can still compact.\n\tlet contextTokens: number;\n\tif (assistantMessage.stopReason === \"error\") {\n\t\tconst messages = this.agent.state.messages;\n\t\tconst estimate = estimateContextTokens(messages);\n\t\tif (estimate.lastUsageIndex === null) return; // No usage data at all\n\t\t// Verify the usage source is post-compaction. Kept pre-compaction messages\n\t\t// have stale usage reflecting the old (larger) context and would falsely\n\t\t// trigger compaction right after one just finished.\n\t\tconst usageMsg = messages[estimate.lastUsageIndex];\n\t\tif (\n\t\t\tcompactionBoundaryEntry &&\n\t\t\tusageMsg.role === \"assistant\" &&\n\t\t\t(usageMsg as AssistantMessage).timestamp <= new Date(compactionBoundaryEntry.timestamp).getTime()\n\t\t) {\n\t\t\treturn;\n\t\t}\n\t\tcontextTokens = estimate.tokens;\n\t} else {\n\t\tcontextTokens = calculateContextTokens(assistantMessage.usage);\n\t}\n\t// Compact against the effective input budget (the hard prompt cap for providers like Copilot\n\t// that advertise a larger total window) so we compact before overrunning the server-side limit\n\t// rather than relying on reactive overflow recovery near the cap.\n\tconst compactionBudget = this.model ? getEffectiveInputBudget(this.model) : contextWindow;\n\tif (shouldCompact(contextTokens, compactionBudget, settings)) {\n\t\tawait this._runAutoCompaction(\"threshold\", shouldRetryAfterThresholdCompaction(assistantMessage));\n\t}\n}\n\n\nexport function _isCopilotServerCapBelowSelectedContextWindow(this: AgentSession, assistantMessage: AssistantMessage): boolean {\n\tif (!this.model || this.model.provider !== \"github-copilot\" || !assistantMessage.errorMessage) return false;\n\tconst promptLimitError = parseCopilotPromptLimitError(assistantMessage.errorMessage);\n\t// Compare against the effective input budget (the model's real prompt cap), not the displayed\n\t// total window. A rejection at the prompt cap is a normal overflow we should compact-and-retry;\n\t// only a rejection *below* the cap (e.g. a missing long-context entitlement dropping the account\n\t// to a lower server tier) keeps the friendly error visible instead of silently compacting down.\n\treturn promptLimitError !== undefined && getEffectiveInputBudget(this.model) > promptLimitError.limitTokens;\n}\n\nexport function shouldRetryAfterThresholdCompaction(assistantMessage: AssistantMessage): boolean {\n\treturn assistantMessage.stopReason === \"length\" && assistantMessage.usage.output > 0;\n}\n\n/**\n * Internal: remove an incomplete assistant from retry context before auto-continuing after compaction.\n */\n\nexport function _dropTrailingAutoCompactionRetryAssistantIfPresent(this: AgentSession): void {\n\tconst messages = this.agent.state.messages;\n\tconst lastMsg = messages[messages.length - 1];\n\tif (lastMsg?.role !== \"assistant\") return;\n\tconst stopReason = (lastMsg as AssistantMessage).stopReason;\n\tif (stopReason === \"error\" || stopReason === \"length\") {\n\t\tthis.agent.state.messages = messages.slice(0, -1);\n\t}\n}\n\n/**\n * Internal: schedule a live post-event continuation probe after compaction_end listeners can flush queues.\n */\n\nexport function _schedulePostAutoCompactionContinuationProbe(this: AgentSession,\n\treason: \"overflow\" | \"threshold\",\n\twillRetry: boolean,\n): void {\n\tif (reason === \"overflow\" && willRetry) {\n\t\tconst token = this._overflowPostCompactionContinuationToken + 1;\n\t\tthis._overflowPostCompactionContinuationToken = token;\n\t\tlet pending: Promise<void>;\n\t\tpending = new Promise<void>((resolve) => {\n\t\t\tsetTimeout(() => {\n\t\t\t\tvoid (async () => {\n\t\t\t\t\ttry {\n\t\t\t\t\t\tif (this._overflowPostCompactionContinuationToken !== token) return;\n\t\t\t\t\t\tif (this.isCompacting || this.isStreaming) return;\n\t\t\t\t\t\tawait this._resumeAfterAutoCompaction();\n\t\t\t\t\t} finally {\n\t\t\t\t\t\tif (this._pendingOverflowPostCompactionContinuation === pending) {\n\t\t\t\t\t\t\tthis._pendingOverflowPostCompactionContinuation = undefined;\n\t\t\t\t\t\t}\n\t\t\t\t\t\tresolve();\n\t\t\t\t\t}\n\t\t\t\t})();\n\t\t\t}, 100);\n\t\t});\n\t\tthis._pendingOverflowPostCompactionContinuation = pending;\n\t\treturn;\n\t}\n\n\tsetTimeout(() => {\n\t\tif (this.isCompacting || this.isStreaming) {\n\t\t\treturn;\n\t\t}\n\n\t\tif (willRetry) {\n\t\t\tvoid this._resumeAfterAutoCompaction();\n\t\t\treturn;\n\t\t}\n\n\t\tif (!this.agent.hasQueuedMessages()) {\n\t\t\treturn;\n\t\t}\n\n\t\tvoid this._resumeAfterAutoCompaction();\n\t}, 100);\n}\n\nexport async function _awaitPendingOverflowPostCompactionContinuation(this: AgentSession): Promise<void> {\n\tconst pending = this._pendingOverflowPostCompactionContinuation;\n\tif (pending === undefined) return;\n\tawait pending;\n}\n\n/**\n * Internal: resume generation after successful auto-compaction only when active work remains.\n */\n\nexport async function _resumeAfterAutoCompaction(this: AgentSession): Promise<void> {\n\ttry {\n\t\tawait this._runAgentContinue();\n\t} catch (error) {\n\t\tconst message = error instanceof Error ? error.message : String(error);\n\t\tthis._emit({\n\t\t\ttype: \"agent_continue_error\",\n\t\t\tsource: \"post_compaction\",\n\t\t\terrorMessage: `Post-compaction continuation failed: ${message}`,\n\t\t});\n\t}\n}\n\n\nfunction overflowUnresolved(reason: \"overflow\" | \"threshold\", aborted = false): boolean | undefined {\n\treturn reason === \"overflow\" && !aborted ? true : undefined;\n}\n\nexport async function _runAutoCompaction(this: AgentSession, reason: \"overflow\" | \"threshold\", willRetry: boolean): Promise<void> {\n\tthis._emit({ type: \"compaction_start\", reason });\n\tthis._autoCompactionAbortController = new AbortController();\n\n\ttry {\n\t\tif (!this.model) {\n\t\t\tthis._emit({\n\t\t\t\ttype: \"compaction_end\",\n\t\t\t\treason,\n\t\t\t\tresult: undefined,\n\t\t\t\taborted: false,\n\t\t\t\twillRetry: false,\n\t\t\t\tunresolvedOverflow: overflowUnresolved(reason),\n\t\t\t});\n\t\t\treturn;\n\t\t}\n\n\t\t// Auth is resolved lazily: only called when the planner fallback is needed.\n\t\t// This allows extension-provided deletion requests to run before auth is checked,\n\t\t// enabling local extension compaction even when API credentials are unavailable.\n\t\t// Auto-mode resolver returns undefined (rather than throwing) when auth is missing;\n\t\t// overflow recovery then falls back to deterministic no-auth eviction, while\n\t\t// threshold compaction keeps the previous no-op behavior.\n\t\tconst model = this.model;\n\t\tconst result = await this._applyContextVerbatimCompaction({\n\t\t\tresolvePlannerAuth: async () => {\n\t\t\t\tconst authResult = await this._modelRegistry.getApiKeyAndHeaders(model);\n\t\t\t\tif (!authResult.ok || !authResult.apiKey) {\n\t\t\t\t\treturn undefined;\n\t\t\t\t}\n\t\t\t\treturn { apiKey: authResult.apiKey, headers: authResult.headers };\n\t\t\t},\n\t\t\tabortController: this._autoCompactionAbortController,\n\t\t\tbackupLabel: reason === \"overflow\" ? \"overflow-auto-compact\" : \"auto-compact\",\n\t\t\treason,\n\t\t});\n\t\tif (!result) {\n\t\t\tthis._emit({\n\t\t\t\ttype: \"compaction_end\",\n\t\t\t\treason,\n\t\t\t\tresult: undefined,\n\t\t\t\taborted: false,\n\t\t\t\twillRetry: false,\n\t\t\t\tunresolvedOverflow: overflowUnresolved(reason),\n\t\t\t});\n\t\t\treturn;\n\t\t}\n\n\t\tif (willRetry) {\n\t\t\tthis._dropTrailingAutoCompactionRetryAssistantIfPresent();\n\t\t}\n\n\t\tthis._emit({ type: \"compaction_end\", reason, result, aborted: false, willRetry });\n\t\tthis._schedulePostAutoCompactionContinuationProbe(reason, willRetry);\n\t} catch (error) {\n\t\tconst errorMessage = error instanceof Error ? error.message : \"compaction failed\";\n\t\tconst aborted = errorMessage === \"Compaction cancelled\" || (error instanceof Error && error.name === \"AbortError\");\n\t\tthis._emit({\n\t\t\ttype: \"compaction_end\",\n\t\t\treason,\n\t\t\tresult: undefined,\n\t\t\taborted,\n\t\t\twillRetry: false,\n\t\t\tunresolvedOverflow: overflowUnresolved(reason, aborted),\n\t\t\terrorMessage: aborted\n\t\t\t\t? undefined\n\t\t\t\t: reason === \"overflow\"\n\t\t\t\t\t? `Context overflow recovery failed: ${errorMessage}`\n\t\t\t\t\t: `Auto-compaction failed: ${errorMessage}`,\n\t\t});\n\t} finally {\n\t\tthis._autoCompactionAbortController = undefined;\n\t}\n}\n\n/**\n * Toggle auto-compaction setting.\n */\n\nexport const agentSessionAutoCompactionMethods = {\n\t_awaitPendingOverflowPostCompactionContinuation,\n\t_checkCompaction,\n\t_isCopilotServerCapBelowSelectedContextWindow,\n\t_dropTrailingAutoCompactionRetryAssistantIfPresent,\n\t_schedulePostAutoCompactionContinuationProbe,\n\t_resumeAfterAutoCompaction,\n\t_runAutoCompaction,\n};\n"]}
+{"version":3,"file":"agent-session-auto-compaction.js","sourceRoot":"","sources":["../../src/core/agent-session-auto-compaction.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,iBAAiB,EAAE,MAAM,8BAA8B,CAAC;AACjE,OAAO,EAAE,uBAAuB,EAAE,MAAM,qBAAqB,CAAC;AAC9D,OAAO,EAAE,4BAA4B,EAAE,MAAM,qBAAqB,CAAC;AACnE,OAAO,EAAE,sBAAsB,EAAE,qBAAqB,EAAE,aAAa,EAAE,MAAM,uBAAuB,CAAC;AACrG,OAAO,EAAE,gCAAgC,EAAE,MAAM,sBAAsB,CAAC;AACxE,OAAO,EAAE,+BAA+B,EAAE,MAAM,yCAAyC,CAAC;AAG1F;;;;;;GAMG;AACH,MAAM,CAAC,MAAM,gCAAgC,GAAG,CAAC,CAAC;AAElD,MAAM,CAAC,MAAM,6CAA6C,GAAG,CAAC,CAAC;AAS/D,MAAM,+BAA+B,GAAG,uBAAuB,CAAC;AAChE,MAAM,+BAA+B,GAAG,IAAI,MAAM,CACjD,gFAAgF,+BAA+B,uBAAuB,CACtI,CAAC;AAEF,MAAM,CAAC,KAAK,UAAU,gBAAgB,CAAqB,gBAAkC,EAAE,gBAAgB,GAAG,IAAI;IACrH,8EAA8E;IAC9E,0EAA0E;IAC1E,yEAAyE;IACzE,MAAM,oBAAoB,GAAG,gBAAgB,CAAC;IAC9C,MAAM,QAAQ,GAAG,IAAI,CAAC,eAAe,CAAC,qBAAqB,EAAE,CAAC;IAC9D,IAAI,CAAC,QAAQ,CAAC,OAAO;QAAE,OAAO;IAE9B,kFAAkF;IAClF,IAAI,gBAAgB,IAAI,gBAAgB,CAAC,UAAU,KAAK,SAAS;QAAE,OAAO;IAE1E,MAAM,aAAa,GAAG,IAAI,CAAC,KAAK,EAAE,aAAa,IAAI,CAAC,CAAC;IAErD,kEAAkE;IAClE,qFAAqF;IACrF,iFAAiF;IACjF,kDAAkD;IAClD,MAAM,SAAS,GACd,IAAI,CAAC,KAAK,IAAI,gBAAgB,CAAC,QAAQ,KAAK,IAAI,CAAC,KAAK,CAAC,QAAQ,IAAI,gBAAgB,CAAC,KAAK,KAAK,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC;IAE7G,4EAA4E;IAC5E,wEAAwE;IACxE,qEAAqE;IACrE,MAAM,uBAAuB,GAAG,gCAAgC,CAAC,IAAI,CAAC,cAAc,CAAC,SAAS,EAAE,CAAC,CAAC;IAClG,MAAM,uCAAuC,GAC5C,uBAAuB,KAAK,IAAI;QAChC,gBAAgB,CAAC,SAAS,IAAI,IAAI,IAAI,CAAC,uBAAuB,CAAC,SAAS,CAAC,CAAC,OAAO,EAAE,CAAC;IACrF,IAAI,uCAAuC,EAAE,CAAC;QAC7C,OAAO;IACR,CAAC;IAED,yDAAyD;IACzD,oFAAoF;IACpF,uFAAuF;IACvF,mFAAmF;IACnF,IAAI,SAAS,IAAI,IAAI,CAAC,6CAA6C,CAAC,gBAAgB,CAAC,EAAE,CAAC;QACvF,OAAO;IACR,CAAC;IACD,IAAI,SAAS,IAAI,iBAAiB,CAAC,gBAAgB,EAAE,aAAa,CAAC,EAAE,CAAC;QACrE,MAAM,SAAS,GAAG,gBAAgB,CAAC,UAAU,KAAK,MAAM,CAAC;QACzD,IAAI,CAAC,SAAS,EAAE,CAAC;YAChB,MAAM,IAAI,CAAC,kBAAkB,CAAC,UAAU,EAAE,KAAK,CAAC,CAAC;YACjD,OAAO;QACR,CAAC;QAED,IAAI,IAAI,CAAC,0BAA0B,EAAE,CAAC;YACrC,IAAI,CAAC,KAAK,CAAC;gBACV,IAAI,EAAE,gBAAgB;gBACtB,MAAM,EAAE,UAAU;gBAClB,MAAM,EAAE,SAAS;gBACjB,OAAO,EAAE,KAAK;gBACd,SAAS,EAAE,KAAK;gBAChB,kBAAkB,EAAE,IAAI;gBACxB,YAAY,EACX,oIAAoI;aACrI,CAAC,CAAC;YACH,OAAO;QACR,CAAC;QAED,IAAI,CAAC,0BAA0B,GAAG,IAAI,CAAC;QACvC,iFAAiF;QACjF,iDAAiD;QACjD,MAAM,QAAQ,GAAG,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,QAAQ,CAAC;QAC3C,IAAI,QAAQ,CAAC,MAAM,GAAG,CAAC,IAAI,QAAQ,CAAC,QAAQ,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;YAC/E,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,QAAQ,GAAG,QAAQ,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;QACnD,CAAC;QACD,MAAM,IAAI,CAAC,kBAAkB,CAAC,UAAU,EAAE,SAAS,CAAC,CAAC;QACrD,OAAO;IACR,CAAC;IAED,+CAA+C;IAC/C,8EAA8E;IAC9E,qFAAqF;IACrF,IAAI,aAAqB,CAAC;IAC1B,IAAI,gBAAgB,CAAC,UAAU,KAAK,OAAO,EAAE,CAAC;QAC7C,MAAM,QAAQ,GAAG,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,QAAQ,CAAC;QAC3C,MAAM,QAAQ,GAAG,qBAAqB,CAAC,QAAQ,CAAC,CAAC;QACjD,IAAI,QAAQ,CAAC,cAAc,KAAK,IAAI;YAAE,OAAO,CAAC,uBAAuB;QACrE,2EAA2E;QAC3E,yEAAyE;QACzE,oDAAoD;QACpD,MAAM,QAAQ,GAAG,QAAQ,CAAC,QAAQ,CAAC,cAAc,CAAC,CAAC;QACnD,IACC,uBAAuB;YACvB,QAAQ,CAAC,IAAI,KAAK,WAAW;YAC5B,QAA6B,CAAC,SAAS,IAAI,IAAI,IAAI,CAAC,uBAAuB,CAAC,SAAS,CAAC,CAAC,OAAO,EAAE,EAChG,CAAC;YACF,OAAO;QACR,CAAC;QACD,aAAa,GAAG,QAAQ,CAAC,MAAM,CAAC;IACjC,CAAC;SAAM,CAAC;QACP,aAAa,GAAG,sBAAsB,CAAC,gBAAgB,CAAC,KAAK,CAAC,CAAC;IAChE,CAAC;IACD,6FAA6F;IAC7F,+FAA+F;IAC/F,kEAAkE;IAClE,MAAM,gBAAgB,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,uBAAuB,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,aAAa,CAAC;IAC1F,IAAI,aAAa,CAAC,aAAa,EAAE,gBAAgB,EAAE,QAAQ,CAAC,EAAE,CAAC;QAC9D,MAAM,SAAS,GAAG,mCAAmC,CAAC,gBAAgB,CAAC,CAAC;QACxE,IAAI,SAAS,IAAI,8BAA8B,CAAC,gBAAgB,CAAC,EAAE,CAAC;YACnE,IAAI,IAAI,CAAC,sCAAsC,IAAI,6CAA6C,EAAE,CAAC;gBAClG,IAAI,CAAC,KAAK,CAAC;oBACV,IAAI,EAAE,gBAAgB;oBACtB,MAAM,EAAE,WAAW;oBACnB,MAAM,EAAE,SAAS;oBACjB,OAAO,EAAE,KAAK;oBACd,SAAS,EAAE,KAAK;oBAChB,YAAY,EACX,gIAAgI;iBACjI,CAAC,CAAC;gBACH,OAAO;YACR,CAAC;YACD,IAAI,CAAC,sCAAsC,IAAI,CAAC,CAAC;QAClD,CAAC;QACD,MAAM,IAAI,CAAC,kBAAkB,CAAC,WAAW,EAAE,SAAS,CAAC,CAAC;QACtD,OAAO;IACR,CAAC;IAED,2EAA2E;IAC3E,8EAA8E;IAC9E,wEAAwE;IACxE,2EAA2E;IAC3E,0BAA0B;IAC1B,IAAI,oBAAoB,IAAI,uBAAuB,CAAC,gBAAgB,CAAC,EAAE,CAAC;QACvE,IAAI,CAAC,4BAA4B,EAAE,CAAC;IACrC,CAAC;AACF,CAAC;AAGD,MAAM,UAAU,6CAA6C,CAAqB,gBAAkC;IACnH,IAAI,CAAC,IAAI,CAAC,KAAK,IAAI,IAAI,CAAC,KAAK,CAAC,QAAQ,KAAK,gBAAgB,IAAI,CAAC,gBAAgB,CAAC,YAAY;QAAE,OAAO,KAAK,CAAC;IAC5G,MAAM,gBAAgB,GAAG,4BAA4B,CAAC,gBAAgB,CAAC,YAAY,CAAC,CAAC;IACrF,8FAA8F;IAC9F,gGAAgG;IAChG,iGAAiG;IACjG,gGAAgG;IAChG,OAAO,gBAAgB,KAAK,SAAS,IAAI,uBAAuB,CAAC,IAAI,CAAC,KAAK,CAAC,GAAG,gBAAgB,CAAC,WAAW,CAAC;AAC7G,CAAC;AAED,MAAM,UAAU,uBAAuB,CAAC,gBAAkC;IACzE,OAAO,gBAAgB,CAAC,UAAU,KAAK,QAAQ,IAAI,gBAAgB,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC;AACtF,CAAC;AAED,SAAS,YAAY,CAAC,KAAgB;IACrC,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,IAAI,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;AAC7E,CAAC;AAED,SAAS,WAAW,CAAC,MAAoC,EAAE,GAAW;IACrE,MAAM,KAAK,GAAG,MAAM,CAAC,GAAG,CAAC,CAAC;IAC1B,OAAO,OAAO,KAAK,KAAK,QAAQ,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,SAAS,CAAC;AACtD,CAAC;AAED,SAAS,yBAAyB,CAAC,YAAoB;IACtD,MAAM,KAAK,GAAG,YAAY,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC;IACxC,MAAM,GAAG,GAAG,YAAY,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC;IAC1C,IAAI,KAAK,KAAK,CAAC,CAAC,IAAI,GAAG,IAAI,KAAK;QAAE,OAAO,SAAS,CAAC;IAEnD,IAAI,CAAC;QACJ,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,YAAY,CAAC,KAAK,CAAC,KAAK,EAAE,GAAG,GAAG,CAAC,CAAC,CAAc,CAAC;QAC3E,IAAI,CAAC,YAAY,CAAC,MAAM,CAAC;YAAE,OAAO,SAAS,CAAC;QAE5C,MAAM,WAAW,GAAG,MAAM,CAAC,KAAK,CAAC;QACjC,IAAI,WAAW,KAAK,SAAS,IAAI,YAAY,CAAC,WAAW,CAAC,EAAE,CAAC;YAC5D,OAAO;gBACN,OAAO,EAAE,WAAW,CAAC,WAAW,EAAE,SAAS,CAAC;gBAC5C,IAAI,EAAE,WAAW,CAAC,WAAW,EAAE,MAAM,CAAC,IAAI,WAAW,CAAC,MAAM,EAAE,MAAM,CAAC;gBACrE,KAAK,EAAE,WAAW,CAAC,WAAW,EAAE,OAAO,CAAC,IAAI,WAAW,CAAC,MAAM,EAAE,OAAO,CAAC;aACxE,CAAC;QACH,CAAC;QAED,OAAO;YACN,OAAO,EAAE,WAAW,CAAC,MAAM,EAAE,SAAS,CAAC;YACvC,IAAI,EAAE,WAAW,CAAC,MAAM,EAAE,MAAM,CAAC;YACjC,KAAK,EAAE,WAAW,CAAC,MAAM,EAAE,OAAO,CAAC;SACnC,CAAC;IACH,CAAC;IAAC,MAAM,CAAC;QACR,OAAO,SAAS,CAAC;IAClB,CAAC;AACF,CAAC;AAED,SAAS,2BAA2B,CAAC,IAAY;IAChD,MAAM,OAAO,GAAG,IAAI,CAAC,WAAW,EAAE,CAAC;IACnC,OAAO,+BAA+B,CAAC,IAAI,CAAC,OAAO,CAAC,IAAI,+BAA+B,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC;AACvG,CAAC;AAED,SAAS,iCAAiC,CAAC,OAA6B;IACvE,MAAM,OAAO,GAAG,OAAO,CAAC,OAAO,EAAE,WAAW,EAAE,IAAI,EAAE,CAAC;IACrD,MAAM,KAAK,GAAG,OAAO,CAAC,KAAK,EAAE,WAAW,EAAE,CAAC;IAC3C,IAAI,CAAC,+BAA+B,CAAC,IAAI,CAAC,OAAO,CAAC;QAAE,OAAO,KAAK,CAAC;IACjE,OAAO,KAAK,KAAK,SAAS,CAAC,CAAC,CAAC,+BAA+B,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,+BAA+B,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC;AAC1H,CAAC;AAED,MAAM,UAAU,8BAA8B,CAAC,gBAAkC;IAChF,IAAI,gBAAgB,CAAC,UAAU,KAAK,OAAO,IAAI,CAAC,gBAAgB,CAAC,YAAY;QAAE,OAAO,KAAK,CAAC;IAC5F,IAAI,gBAAgB,CAAC,GAAG,KAAK,kBAAkB;QAAE,OAAO,KAAK,CAAC;IAE9D,MAAM,iBAAiB,GAAG,yBAAyB,CAAC,gBAAgB,CAAC,YAAY,CAAC,CAAC;IACnF,IAAI,iBAAiB,IAAI,iCAAiC,CAAC,iBAAiB,CAAC;QAAE,OAAO,IAAI,CAAC;IAE3F,OAAO,2BAA2B,CAAC,gBAAgB,CAAC,YAAY,CAAC,CAAC;AACnE,CAAC;AAED,MAAM,UAAU,mCAAmC,CAAC,gBAAkC;IACrF,OAAO,uBAAuB,CAAC,gBAAgB,CAAC,IAAI,8BAA8B,CAAC,gBAAgB,CAAC,CAAC;AACtG,CAAC;AAED;;GAEG;AAEH,MAAM,UAAU,kDAAkD;IACjE,MAAM,QAAQ,GAAG,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,QAAQ,CAAC;IAC3C,MAAM,OAAO,GAAG,QAAQ,CAAC,QAAQ,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;IAC9C,IAAI,OAAO,EAAE,IAAI,KAAK,WAAW;QAAE,OAAO;IAC1C,MAAM,UAAU,GAAI,OAA4B,CAAC,UAAU,CAAC;IAC5D,IAAI,UAAU,KAAK,OAAO,IAAI,UAAU,KAAK,QAAQ,EAAE,CAAC;QACvD,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,QAAQ,GAAG,QAAQ,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;IACnD,CAAC;AACF,CAAC;AAED;;GAEG;AAEH,MAAM,UAAU,4CAA4C,CAC3D,OAAiC,EACjC,SAAkB;IAElB,IAAI,SAAS,EAAE,CAAC;QACf,MAAM,KAAK,GAAG,IAAI,CAAC,gCAAgC,GAAG,CAAC,CAAC;QACxD,IAAI,CAAC,gCAAgC,GAAG,KAAK,CAAC;QAC9C,IAAI,OAAsB,CAAC;QAC3B,OAAO,GAAG,IAAI,OAAO,CAAO,CAAC,OAAO,EAAE,EAAE;YACvC,UAAU,CAAC,GAAG,EAAE;gBACf,KAAK,CAAC,KAAK,IAAI,EAAE;oBAChB,IAAI,CAAC;wBACJ,IAAI,IAAI,CAAC,gCAAgC,KAAK,KAAK;4BAAE,OAAO;wBAC5D,IAAI,IAAI,CAAC,YAAY,IAAI,IAAI,CAAC,WAAW;4BAAE,OAAO;wBAClD,MAAM,IAAI,CAAC,0BAA0B,EAAE,CAAC;oBACzC,CAAC;4BAAS,CAAC;wBACV,IAAI,IAAI,CAAC,kCAAkC,KAAK,OAAO,EAAE,CAAC;4BACzD,IAAI,CAAC,kCAAkC,GAAG,SAAS,CAAC;wBACrD,CAAC;wBACD,OAAO,EAAE,CAAC;oBACX,CAAC;gBACF,CAAC,CAAC,EAAE,CAAC;YACN,CAAC,EAAE,GAAG,CAAC,CAAC;QACT,CAAC,CAAC,CAAC;QACH,IAAI,CAAC,kCAAkC,GAAG,OAAO,CAAC;QAClD,OAAO;IACR,CAAC;IAED,UAAU,CAAC,GAAG,EAAE;QACf,IAAI,IAAI,CAAC,YAAY,IAAI,IAAI,CAAC,WAAW,EAAE,CAAC;YAC3C,OAAO;QACR,CAAC;QAED,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,iBAAiB,EAAE,EAAE,CAAC;YACrC,OAAO;QACR,CAAC;QAED,KAAK,IAAI,CAAC,0BAA0B,EAAE,CAAC;IACxC,CAAC,EAAE,GAAG,CAAC,CAAC;AACT,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,uCAAuC;IAC5D,MAAM,OAAO,GAAG,IAAI,CAAC,kCAAkC,CAAC;IACxD,IAAI,OAAO,KAAK,SAAS;QAAE,OAAO;IAClC,MAAM,OAAO,CAAC;AACf,CAAC;AAED;;GAEG;AAEH,MAAM,CAAC,KAAK,UAAU,0BAA0B;IAC/C,IAAI,CAAC;QACJ,MAAM,IAAI,CAAC,iBAAiB,EAAE,CAAC;IAChC,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QAChB,MAAM,OAAO,GAAG,KAAK,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC;QACvE,IAAI,CAAC,KAAK,CAAC;YACV,IAAI,EAAE,sBAAsB;YAC5B,MAAM,EAAE,iBAAiB;YACzB,YAAY,EAAE,wCAAwC,OAAO,EAAE;SAC/D,CAAC,CAAC;IACJ,CAAC;AACF,CAAC;AAED;;;;;;GAMG;AAEH,MAAM,UAAU,4BAA4B;IAC3C,IAAI,IAAI,CAAC,2BAA2B,IAAI,gCAAgC;QAAE,OAAO;IACjF,IAAI,CAAC,2BAA2B,IAAI,CAAC,CAAC;IACtC,kEAAkE;IAClE,uEAAuE;IACvE,yDAAyD;IACzD,IAAI,CAAC,kDAAkD,EAAE,CAAC;IAC1D,IAAI,CAAC,4CAA4C,CAAC,WAAW,EAAE,IAAI,CAAC,CAAC;AACtE,CAAC;AAGD,SAAS,kBAAkB,CAAC,MAAgC,EAAE,OAAO,GAAG,KAAK;IAC5E,OAAO,MAAM,KAAK,UAAU,IAAI,CAAC,OAAO,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,SAAS,CAAC;AAC7D,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,kBAAkB,CAAqB,MAAgC,EAAE,SAAkB;IAChH,IAAI,CAAC,KAAK,CAAC,EAAE,IAAI,EAAE,kBAAkB,EAAE,MAAM,EAAE,CAAC,CAAC;IACjD,IAAI,CAAC,8BAA8B,GAAG,IAAI,eAAe,EAAE,CAAC;IAE5D,IAAI,CAAC;QACJ,IAAI,CAAC,IAAI,CAAC,KAAK,EAAE,CAAC;YACjB,IAAI,CAAC,KAAK,CAAC;gBACV,IAAI,EAAE,gBAAgB;gBACtB,MAAM;gBACN,MAAM,EAAE,SAAS;gBACjB,OAAO,EAAE,KAAK;gBACd,SAAS,EAAE,KAAK;gBAChB,kBAAkB,EAAE,kBAAkB,CAAC,MAAM,CAAC;aAC9C,CAAC,CAAC;YACH,OAAO;QACR,CAAC;QAED,4EAA4E;QAC5E,kFAAkF;QAClF,iFAAiF;QACjF,oFAAoF;QACpF,6EAA6E;QAC7E,0DAA0D;QAC1D,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC;QACzB,MAAM,MAAM,GAAG,MAAM,IAAI,CAAC,+BAA+B,CAAC;YACzD,kBAAkB,EAAE,KAAK,IAAI,EAAE;gBAC9B,MAAM,UAAU,GAAG,MAAM,IAAI,CAAC,cAAc,CAAC,mBAAmB,CAAC,KAAK,CAAC,CAAC;gBACxE,IAAI,CAAC,UAAU,CAAC,EAAE,IAAI,CAAC,UAAU,CAAC,MAAM,EAAE,CAAC;oBAC1C,OAAO,SAAS,CAAC;gBAClB,CAAC;gBACD,OAAO,EAAE,MAAM,EAAE,UAAU,CAAC,MAAM,EAAE,OAAO,EAAE,UAAU,CAAC,OAAO,EAAE,CAAC;YACnE,CAAC;YACD,eAAe,EAAE,IAAI,CAAC,8BAA8B;YACpD,WAAW,EAAE,MAAM,KAAK,UAAU,CAAC,CAAC,CAAC,uBAAuB,CAAC,CAAC,CAAC,cAAc;YAC7E,MAAM;SACN,CAAC,CAAC;QACH,IAAI,CAAC,MAAM,EAAE,CAAC;YACb,IAAI,CAAC,KAAK,CAAC;gBACV,IAAI,EAAE,gBAAgB;gBACtB,MAAM;gBACN,MAAM,EAAE,SAAS;gBACjB,OAAO,EAAE,KAAK;gBACd,SAAS,EAAE,KAAK;gBAChB,kBAAkB,EAAE,kBAAkB,CAAC,MAAM,CAAC;aAC9C,CAAC,CAAC;YACH,OAAO;QACR,CAAC;QAED,IAAI,SAAS,EAAE,CAAC;YACf,IAAI,CAAC,kDAAkD,EAAE,CAAC;QAC3D,CAAC;QAED,IAAI,CAAC,KAAK,CAAC,EAAE,IAAI,EAAE,gBAAgB,EAAE,MAAM,EAAE,MAAM,EAAE,OAAO,EAAE,KAAK,EAAE,SAAS,EAAE,CAAC,CAAC;QAClF,IAAI,CAAC,4CAA4C,CAAC,MAAM,EAAE,SAAS,CAAC,CAAC;IACtE,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QAChB,MAAM,YAAY,GAAG,KAAK,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,mBAAmB,CAAC;QAClF,MAAM,OAAO,GAAG,YAAY,KAAK,sBAAsB,IAAI,CAAC,KAAK,YAAY,KAAK,IAAI,KAAK,CAAC,IAAI,KAAK,YAAY,CAAC,CAAC;QACnH,IAAI,CAAC,KAAK,CAAC;YACV,IAAI,EAAE,gBAAgB;YACtB,MAAM;YACN,MAAM,EAAE,SAAS;YACjB,OAAO;YACP,SAAS,EAAE,KAAK;YAChB,kBAAkB,EAAE,kBAAkB,CAAC,MAAM,EAAE,OAAO,CAAC;YACvD,YAAY,EAAE,OAAO;gBACpB,CAAC,CAAC,SAAS;gBACX,CAAC,CAAC,MAAM,KAAK,UAAU;oBACtB,CAAC,CAAC,qCAAqC,YAAY,EAAE;oBACrD,CAAC,CAAC,2BAA2B,YAAY,EAAE;SAC7C,CAAC,CAAC;IACJ,CAAC;YAAS,CAAC;QACV,IAAI,CAAC,8BAA8B,GAAG,SAAS,CAAC;IACjD,CAAC;AACF,CAAC;AAED;;GAEG;AAEH,MAAM,CAAC,MAAM,iCAAiC,GAAG;IAChD,uCAAuC;IACvC,gBAAgB;IAChB,6CAA6C;IAC7C,kDAAkD;IAClD,4CAA4C;IAC5C,0BAA0B;IAC1B,4BAA4B;IAC5B,kBAAkB;CAClB,CAAC","sourcesContent":["import type { AssistantMessage } from \"@earendil-works/pi-ai/compat\";\nimport { isContextOverflow } from \"@earendil-works/pi-ai/compat\";\nimport { getEffectiveInputBudget } from \"./context-window.ts\";\nimport { parseCopilotPromptLimitError } from \"./copilot-errors.ts\";\nimport { calculateContextTokens, estimateContextTokens, shouldCompact } from \"./compaction/index.ts\";\nimport { getLatestCompactionBoundaryEntry } from \"./session-manager.ts\";\nimport { MIN_RESPONSES_MAX_OUTPUT_TOKENS } from \"./openai-responses-payload-sanitizer.ts\";\nimport type { AgentSessionInternalSurface as AgentSession } from \"./agent-session-methods.ts\";\n\n/**\n * Upper bound on consecutive automatic continuations of a response that was\n * truncated at the output-token cap (\"length\") while the context is still\n * below the compaction budget. Each continuation regenerates the cut-off turn,\n * so a model that insists on emitting more than its per-turn output cap can\n * still terminate instead of looping forever.\n */\nexport const MAX_LENGTH_CONTINUATION_ATTEMPTS = 3;\n\nexport const MAX_OUTPUT_BUDGET_ERROR_CONTINUATION_ATTEMPTS = 1;\n\ntype JsonValue = string | number | boolean | null | JsonValue[] | { [key: string]: JsonValue };\ntype ProviderErrorDetails = {\n\tmessage?: string;\n\tcode?: string;\n\tparam?: string;\n};\n\nconst OUTPUT_BUDGET_PARAMETER_PATTERN = /\\bmax_output_tokens\\b/;\nconst OUTPUT_BUDGET_UNDERFLOW_PATTERN = new RegExp(\n\t`(?:integer\\\\s+below\\\\s+minimum\\\\s+value|expected\\\\s+(?:a\\\\s+)?value\\\\s*>=\\\\s*${MIN_RESPONSES_MAX_OUTPUT_TOKENS}|got\\\\s+1\\\\s+instead)`,\n);\n\nexport async function _checkCompaction(this: AgentSession, assistantMessage: AssistantMessage, skipAbortedCheck = true): Promise<void> {\n\t// The agent_end path passes skipAbortedCheck=true; the pre-prompt path passes\n\t// false. Only the live turn-completion path may auto-continue a truncated\n\t// response — before a fresh user prompt we must not resume the old turn.\n\tconst isLiveTurnCompletion = skipAbortedCheck;\n\tconst settings = this.settingsManager.getCompactionSettings();\n\tif (!settings.enabled) return;\n\n\t// Skip if message was aborted (user cancelled) - unless skipAbortedCheck is false\n\tif (skipAbortedCheck && assistantMessage.stopReason === \"aborted\") return;\n\n\tconst contextWindow = this.model?.contextWindow ?? 0;\n\n\t// Skip overflow check if the message came from a different model.\n\t// This handles the case where user switched from a smaller-context model (e.g. opus)\n\t// to a larger-context model (e.g. codex) - the overflow error from the old model\n\t// shouldn't trigger compaction for the new model.\n\tconst sameModel =\n\t\tthis.model && assistantMessage.provider === this.model.provider && assistantMessage.model === this.model.id;\n\n\t// Skip compaction checks if this assistant message is older than the latest\n\t// compaction boundary. This prevents a stale pre-compaction usage/error\n\t// from retriggering compaction on the first prompt after compaction.\n\tconst compactionBoundaryEntry = getLatestCompactionBoundaryEntry(this.sessionManager.getBranch());\n\tconst assistantIsFromBeforeCompactionBoundary =\n\t\tcompactionBoundaryEntry !== null &&\n\t\tassistantMessage.timestamp <= new Date(compactionBoundaryEntry.timestamp).getTime();\n\tif (assistantIsFromBeforeCompactionBoundary) {\n\t\treturn;\n\t}\n\n\t// Case 1: Overflow - LLM returned context overflow error\n\t// When Copilot rejects a 1m client-budget prompt at a lower server cap (for example\n\t// because long-context/usage-based billing entitlement is missing), leave the friendly\n\t// error visible instead of auto-compacting down to a smaller server tier silently.\n\tif (sameModel && this._isCopilotServerCapBelowSelectedContextWindow(assistantMessage)) {\n\t\treturn;\n\t}\n\tif (sameModel && isContextOverflow(assistantMessage, contextWindow)) {\n\t\tconst willRetry = assistantMessage.stopReason !== \"stop\";\n\t\tif (!willRetry) {\n\t\t\tawait this._runAutoCompaction(\"overflow\", false);\n\t\t\treturn;\n\t\t}\n\n\t\tif (this._overflowRecoveryAttempted) {\n\t\t\tthis._emit({\n\t\t\t\ttype: \"compaction_end\",\n\t\t\t\treason: \"overflow\",\n\t\t\t\tresult: undefined,\n\t\t\t\taborted: false,\n\t\t\t\twillRetry: false,\n\t\t\t\tunresolvedOverflow: true,\n\t\t\t\terrorMessage:\n\t\t\t\t\t\"Context overflow recovery failed after one compact-and-retry attempt. Try reducing context or switching to a larger-context model.\",\n\t\t\t});\n\t\t\treturn;\n\t\t}\n\n\t\tthis._overflowRecoveryAttempted = true;\n\t\t// Remove the error message from agent state (it IS saved to session for history,\n\t\t// but we don't want it in context for the retry)\n\t\tconst messages = this.agent.state.messages;\n\t\tif (messages.length > 0 && messages[messages.length - 1].role === \"assistant\") {\n\t\t\tthis.agent.state.messages = messages.slice(0, -1);\n\t\t}\n\t\tawait this._runAutoCompaction(\"overflow\", willRetry);\n\t\treturn;\n\t}\n\n\t// Case 2: Threshold - context is getting large\n\t// For error messages (no usage data), estimate from last successful response.\n\t// This ensures sessions that hit persistent API errors (e.g. 529) can still compact.\n\tlet contextTokens: number;\n\tif (assistantMessage.stopReason === \"error\") {\n\t\tconst messages = this.agent.state.messages;\n\t\tconst estimate = estimateContextTokens(messages);\n\t\tif (estimate.lastUsageIndex === null) return; // No usage data at all\n\t\t// Verify the usage source is post-compaction. Kept pre-compaction messages\n\t\t// have stale usage reflecting the old (larger) context and would falsely\n\t\t// trigger compaction right after one just finished.\n\t\tconst usageMsg = messages[estimate.lastUsageIndex];\n\t\tif (\n\t\t\tcompactionBoundaryEntry &&\n\t\t\tusageMsg.role === \"assistant\" &&\n\t\t\t(usageMsg as AssistantMessage).timestamp <= new Date(compactionBoundaryEntry.timestamp).getTime()\n\t\t) {\n\t\t\treturn;\n\t\t}\n\t\tcontextTokens = estimate.tokens;\n\t} else {\n\t\tcontextTokens = calculateContextTokens(assistantMessage.usage);\n\t}\n\t// Compact against the effective input budget (the hard prompt cap for providers like Copilot\n\t// that advertise a larger total window) so we compact before overrunning the server-side limit\n\t// rather than relying on reactive overflow recovery near the cap.\n\tconst compactionBudget = this.model ? getEffectiveInputBudget(this.model) : contextWindow;\n\tif (shouldCompact(contextTokens, compactionBudget, settings)) {\n\t\tconst willRetry = shouldRetryAfterThresholdCompaction(assistantMessage);\n\t\tif (willRetry && isRetryWorthyOutputBudgetError(assistantMessage)) {\n\t\t\tif (this._outputBudgetErrorContinuationAttempts >= MAX_OUTPUT_BUDGET_ERROR_CONTINUATION_ATTEMPTS) {\n\t\t\t\tthis._emit({\n\t\t\t\t\ttype: \"compaction_end\",\n\t\t\t\t\treason: \"threshold\",\n\t\t\t\t\tresult: undefined,\n\t\t\t\t\taborted: false,\n\t\t\t\t\twillRetry: false,\n\t\t\t\t\terrorMessage:\n\t\t\t\t\t\t\"Output-budget recovery stopped after a compact-and-retry attempt. Try reducing context or switching to a larger-context model.\",\n\t\t\t\t});\n\t\t\t\treturn;\n\t\t\t}\n\t\t\tthis._outputBudgetErrorContinuationAttempts += 1;\n\t\t}\n\t\tawait this._runAutoCompaction(\"threshold\", willRetry);\n\t\treturn;\n\t}\n\n\t// A response truncated at the output-token cap (\"length\") with the context\n\t// still below the compaction budget is genuine work cut off mid-flight, not a\n\t// context overflow. Compaction would not free any room, so continue the\n\t// generation directly instead of dead-ending on the truncation and leaving\n\t// the task half-finished.\n\tif (isLiveTurnCompletion && isRetryWorthyLengthStop(assistantMessage)) {\n\t\tthis._resumeAfterLengthTruncation();\n\t}\n}\n\n\nexport function _isCopilotServerCapBelowSelectedContextWindow(this: AgentSession, assistantMessage: AssistantMessage): boolean {\n\tif (!this.model || this.model.provider !== \"github-copilot\" || !assistantMessage.errorMessage) return false;\n\tconst promptLimitError = parseCopilotPromptLimitError(assistantMessage.errorMessage);\n\t// Compare against the effective input budget (the model's real prompt cap), not the displayed\n\t// total window. A rejection at the prompt cap is a normal overflow we should compact-and-retry;\n\t// only a rejection *below* the cap (e.g. a missing long-context entitlement dropping the account\n\t// to a lower server tier) keeps the friendly error visible instead of silently compacting down.\n\treturn promptLimitError !== undefined && getEffectiveInputBudget(this.model) > promptLimitError.limitTokens;\n}\n\nexport function isRetryWorthyLengthStop(assistantMessage: AssistantMessage): boolean {\n\treturn assistantMessage.stopReason === \"length\" && assistantMessage.usage.output > 0;\n}\n\nfunction isJsonRecord(value: JsonValue): value is { [key: string]: JsonValue } {\n\treturn typeof value === \"object\" && value !== null && !Array.isArray(value);\n}\n\nfunction stringField(record: { [key: string]: JsonValue }, key: string): string | undefined {\n\tconst value = record[key];\n\treturn typeof value === \"string\" ? value : undefined;\n}\n\nfunction parseProviderErrorDetails(errorMessage: string): ProviderErrorDetails | undefined {\n\tconst start = errorMessage.indexOf(\"{\");\n\tconst end = errorMessage.lastIndexOf(\"}\");\n\tif (start === -1 || end <= start) return undefined;\n\n\ttry {\n\t\tconst parsed = JSON.parse(errorMessage.slice(start, end + 1)) as JsonValue;\n\t\tif (!isJsonRecord(parsed)) return undefined;\n\n\t\tconst nestedError = parsed.error;\n\t\tif (nestedError !== undefined && isJsonRecord(nestedError)) {\n\t\t\treturn {\n\t\t\t\tmessage: stringField(nestedError, \"message\"),\n\t\t\t\tcode: stringField(nestedError, \"code\") ?? stringField(parsed, \"code\"),\n\t\t\t\tparam: stringField(nestedError, \"param\") ?? stringField(parsed, \"param\"),\n\t\t\t};\n\t\t}\n\n\t\treturn {\n\t\t\tmessage: stringField(parsed, \"message\"),\n\t\t\tcode: stringField(parsed, \"code\"),\n\t\t\tparam: stringField(parsed, \"param\"),\n\t\t};\n\t} catch {\n\t\treturn undefined;\n\t}\n}\n\nfunction isOutputBudgetUnderflowText(text: string): boolean {\n\tconst message = text.toLowerCase();\n\treturn OUTPUT_BUDGET_PARAMETER_PATTERN.test(message) && OUTPUT_BUDGET_UNDERFLOW_PATTERN.test(message);\n}\n\nfunction isStructuredOutputBudgetUnderflow(details: ProviderErrorDetails): boolean {\n\tconst message = details.message?.toLowerCase() ?? \"\";\n\tconst param = details.param?.toLowerCase();\n\tif (!OUTPUT_BUDGET_UNDERFLOW_PATTERN.test(message)) return false;\n\treturn param !== undefined ? OUTPUT_BUDGET_PARAMETER_PATTERN.test(param) : OUTPUT_BUDGET_PARAMETER_PATTERN.test(message);\n}\n\nexport function isRetryWorthyOutputBudgetError(assistantMessage: AssistantMessage): boolean {\n\tif (assistantMessage.stopReason !== \"error\" || !assistantMessage.errorMessage) return false;\n\tif (assistantMessage.api !== \"openai-responses\") return false;\n\n\tconst structuredDetails = parseProviderErrorDetails(assistantMessage.errorMessage);\n\tif (structuredDetails && isStructuredOutputBudgetUnderflow(structuredDetails)) return true;\n\n\treturn isOutputBudgetUnderflowText(assistantMessage.errorMessage);\n}\n\nexport function shouldRetryAfterThresholdCompaction(assistantMessage: AssistantMessage): boolean {\n\treturn isRetryWorthyLengthStop(assistantMessage) || isRetryWorthyOutputBudgetError(assistantMessage);\n}\n\n/**\n * Internal: remove an incomplete assistant from retry context before auto-continuing after compaction.\n */\n\nexport function _dropTrailingAutoCompactionRetryAssistantIfPresent(this: AgentSession): void {\n\tconst messages = this.agent.state.messages;\n\tconst lastMsg = messages[messages.length - 1];\n\tif (lastMsg?.role !== \"assistant\") return;\n\tconst stopReason = (lastMsg as AssistantMessage).stopReason;\n\tif (stopReason === \"error\" || stopReason === \"length\") {\n\t\tthis.agent.state.messages = messages.slice(0, -1);\n\t}\n}\n\n/**\n * Internal: schedule a live post-event continuation probe after compaction_end listeners can flush queues.\n */\n\nexport function _schedulePostAutoCompactionContinuationProbe(this: AgentSession,\n\t_reason: \"overflow\" | \"threshold\",\n\twillRetry: boolean,\n): void {\n\tif (willRetry) {\n\t\tconst token = this._postCompactionContinuationToken + 1;\n\t\tthis._postCompactionContinuationToken = token;\n\t\tlet pending: Promise<void>;\n\t\tpending = new Promise<void>((resolve) => {\n\t\t\tsetTimeout(() => {\n\t\t\t\tvoid (async () => {\n\t\t\t\t\ttry {\n\t\t\t\t\t\tif (this._postCompactionContinuationToken !== token) return;\n\t\t\t\t\t\tif (this.isCompacting || this.isStreaming) return;\n\t\t\t\t\t\tawait this._resumeAfterAutoCompaction();\n\t\t\t\t\t} finally {\n\t\t\t\t\t\tif (this._pendingPostCompactionContinuation === pending) {\n\t\t\t\t\t\t\tthis._pendingPostCompactionContinuation = undefined;\n\t\t\t\t\t\t}\n\t\t\t\t\t\tresolve();\n\t\t\t\t\t}\n\t\t\t\t})();\n\t\t\t}, 100);\n\t\t});\n\t\tthis._pendingPostCompactionContinuation = pending;\n\t\treturn;\n\t}\n\n\tsetTimeout(() => {\n\t\tif (this.isCompacting || this.isStreaming) {\n\t\t\treturn;\n\t\t}\n\n\t\tif (!this.agent.hasQueuedMessages()) {\n\t\t\treturn;\n\t\t}\n\n\t\tvoid this._resumeAfterAutoCompaction();\n\t}, 100);\n}\n\nexport async function _awaitPendingPostCompactionContinuation(this: AgentSession): Promise<void> {\n\tconst pending = this._pendingPostCompactionContinuation;\n\tif (pending === undefined) return;\n\tawait pending;\n}\n\n/**\n * Internal: resume generation after successful auto-compaction only when active work remains.\n */\n\nexport async function _resumeAfterAutoCompaction(this: AgentSession): Promise<void> {\n\ttry {\n\t\tawait this._runAgentContinue();\n\t} catch (error) {\n\t\tconst message = error instanceof Error ? error.message : String(error);\n\t\tthis._emit({\n\t\t\ttype: \"agent_continue_error\",\n\t\t\tsource: \"post_compaction\",\n\t\t\terrorMessage: `Post-compaction continuation failed: ${message}`,\n\t\t});\n\t}\n}\n\n/**\n * Internal: resume a response that was truncated at the output-token cap\n * (\"length\") when the context does not warrant compaction. The generation is\n * continued directly so the model finishes the work it was cut off from, rather\n * than dead-ending on the truncation. Bounded by MAX_LENGTH_CONTINUATION_ATTEMPTS\n * so a turn that keeps exceeding the per-turn output cap can still terminate.\n */\n\nexport function _resumeAfterLengthTruncation(this: AgentSession): void {\n\tif (this._lengthContinuationAttempts >= MAX_LENGTH_CONTINUATION_ATTEMPTS) return;\n\tthis._lengthContinuationAttempts += 1;\n\t// agent.continue() rejects an assistant tail; drop the incomplete\n\t// length-stopped message so the preceding user/tool-result anchors the\n\t// continuation. It remains persisted in session history.\n\tthis._dropTrailingAutoCompactionRetryAssistantIfPresent();\n\tthis._schedulePostAutoCompactionContinuationProbe(\"threshold\", true);\n}\n\n\nfunction overflowUnresolved(reason: \"overflow\" | \"threshold\", aborted = false): boolean | undefined {\n\treturn reason === \"overflow\" && !aborted ? true : undefined;\n}\n\nexport async function _runAutoCompaction(this: AgentSession, reason: \"overflow\" | \"threshold\", willRetry: boolean): Promise<void> {\n\tthis._emit({ type: \"compaction_start\", reason });\n\tthis._autoCompactionAbortController = new AbortController();\n\n\ttry {\n\t\tif (!this.model) {\n\t\t\tthis._emit({\n\t\t\t\ttype: \"compaction_end\",\n\t\t\t\treason,\n\t\t\t\tresult: undefined,\n\t\t\t\taborted: false,\n\t\t\t\twillRetry: false,\n\t\t\t\tunresolvedOverflow: overflowUnresolved(reason),\n\t\t\t});\n\t\t\treturn;\n\t\t}\n\n\t\t// Auth is resolved lazily: only called when the planner fallback is needed.\n\t\t// This allows extension-provided deletion requests to run before auth is checked,\n\t\t// enabling local extension compaction even when API credentials are unavailable.\n\t\t// Auto-mode resolver returns undefined (rather than throwing) when auth is missing;\n\t\t// overflow recovery then falls back to deterministic no-auth eviction, while\n\t\t// threshold compaction keeps the previous no-op behavior.\n\t\tconst model = this.model;\n\t\tconst result = await this._applyContextVerbatimCompaction({\n\t\t\tresolvePlannerAuth: async () => {\n\t\t\t\tconst authResult = await this._modelRegistry.getApiKeyAndHeaders(model);\n\t\t\t\tif (!authResult.ok || !authResult.apiKey) {\n\t\t\t\t\treturn undefined;\n\t\t\t\t}\n\t\t\t\treturn { apiKey: authResult.apiKey, headers: authResult.headers };\n\t\t\t},\n\t\t\tabortController: this._autoCompactionAbortController,\n\t\t\tbackupLabel: reason === \"overflow\" ? \"overflow-auto-compact\" : \"auto-compact\",\n\t\t\treason,\n\t\t});\n\t\tif (!result) {\n\t\t\tthis._emit({\n\t\t\t\ttype: \"compaction_end\",\n\t\t\t\treason,\n\t\t\t\tresult: undefined,\n\t\t\t\taborted: false,\n\t\t\t\twillRetry: false,\n\t\t\t\tunresolvedOverflow: overflowUnresolved(reason),\n\t\t\t});\n\t\t\treturn;\n\t\t}\n\n\t\tif (willRetry) {\n\t\t\tthis._dropTrailingAutoCompactionRetryAssistantIfPresent();\n\t\t}\n\n\t\tthis._emit({ type: \"compaction_end\", reason, result, aborted: false, willRetry });\n\t\tthis._schedulePostAutoCompactionContinuationProbe(reason, willRetry);\n\t} catch (error) {\n\t\tconst errorMessage = error instanceof Error ? error.message : \"compaction failed\";\n\t\tconst aborted = errorMessage === \"Compaction cancelled\" || (error instanceof Error && error.name === \"AbortError\");\n\t\tthis._emit({\n\t\t\ttype: \"compaction_end\",\n\t\t\treason,\n\t\t\tresult: undefined,\n\t\t\taborted,\n\t\t\twillRetry: false,\n\t\t\tunresolvedOverflow: overflowUnresolved(reason, aborted),\n\t\t\terrorMessage: aborted\n\t\t\t\t? undefined\n\t\t\t\t: reason === \"overflow\"\n\t\t\t\t\t? `Context overflow recovery failed: ${errorMessage}`\n\t\t\t\t\t: `Auto-compaction failed: ${errorMessage}`,\n\t\t});\n\t} finally {\n\t\tthis._autoCompactionAbortController = undefined;\n\t}\n}\n\n/**\n * Toggle auto-compaction setting.\n */\n\nexport const agentSessionAutoCompactionMethods = {\n\t_awaitPendingPostCompactionContinuation,\n\t_checkCompaction,\n\t_isCopilotServerCapBelowSelectedContextWindow,\n\t_dropTrailingAutoCompactionRetryAssistantIfPresent,\n\t_schedulePostAutoCompactionContinuationProbe,\n\t_resumeAfterAutoCompaction,\n\t_resumeAfterLengthTruncation,\n\t_runAutoCompaction,\n};\n"]}

package/dist/core/agent-session-events.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"agent-session-events.d.ts","sourceRoot":"","sources":["../../src/core/agent-session-events.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,UAAU,EAAE,YAAY,EAAE,MAAM,+BAA+B,CAAC;AAC9E,OAAO,KAAK,EAAE,gBAAgB,EAAE,OAAO,EAAe,MAAM,8BAA8B,CAAC;AAG3F,OAAO,KAAK,EAAE,2BAA2B,IAAI,YAAY,EAAE,MAAM,4BAA4B,CAAC;AAC9F,OAAO,EAA0F,KAAK,iBAAiB,EAAE,KAAK,yBAAyB,EAAE,MAAM,0BAA0B,CAAC;AAG1L,wBAAgB,KAAK,CAAC,IAAI,EAAE,YAAY,EAAE,KAAK,EAAE,iBAAiB,GAAG,IAAI,CAIxE;AAGD,wBAAgB,gBAAgB,CAAC,IAAI,EAAE,YAAY,GAAG,IAAI,CAMzD;AAED,4EAA4E;AAE5E,wBAAgB,iBAAiB,CAAC,IAAI,EAAE,YAAY,EAAE,KAAK,EAAE,UAAU,GAAG,IAAI,CAe7E;AAGD,wBAAgB,8BAA8B,CAAC,IAAI,EAAE,YAAY,EAAE,KAAK,EAAE,UAAU,GAAG,IAAI,CAuB1F;AAGD,wBAAgB,4BAA4B,CAAC,IAAI,EAAE,YAAY,EAAE,QAAQ,EAAE,YAAY,EAAE,GAAG,gBAAgB,GAAG,SAAS,CAQvH;AAGD,wBAAsB,kBAAkB,CAAC,IAAI,EAAE,YAAY,EAAE,KAAK,EAAE,UAAU,GAAG,OAAO,CAAC,IAAI,CAAC,CAiH7F;AAGD,wBAAgB,2BAA2B,CAAC,IAAI,EAAE,YAAY,EAAE,KAAK,EAAE,UAAU,GAAG,IAAI,CAsBvF;AAGD,wBAAgB,2BAA2B,CAAC,IAAI,EAAE,YAAY,EAAE,KAAK,EAAE,UAAU,GAAG,IAAI,CAWvF;AAED,wCAAwC;AAExC,wBAAgB,aAAa,CAAC,IAAI,EAAE,YAAY,GAAG,IAAI,CAMtD;AAED,0CAA0C;AAE1C,wBAAgB,mBAAmB,CAAC,IAAI,EAAE,YAAY,EAAE,OAAO,EAAE,OAAO,GAAG,MAAM,CAMhF;AAED,8EAA8E;AAE9E,wBAAgB,yBAAyB,CAAC,IAAI,EAAE,YAAY,GAAG,gBAAgB,GAAG,SAAS,CAS1F;AAGD,wBAAgB,sBAAsB,CAAC,IAAI,EAAE,YAAY,EAAE,MAAM,EAAE,YAAY,EAAE,WAAW,EAAE,YAAY,GAAG,IAAI,CAchH;AAED,kDAAkD;AAElD,wBAAsB,mBAAmB,CAAC,IAAI,EAAE,YAAY,EAAE,KAAK,EAAE,UAAU,GAAG,OAAO,CAAC,IAAI,CAAC,CAuE9F;AAED;;;;GAIG;AAEH,wBAAgB,SAAS,CAAC,IAAI,EAAE,YAAY,EAAE,QAAQ,EAAE,yBAAyB,GAAG,MAAM,IAAI,CAU7F;AAED;;;;GAIG;AAEH,wBAAgB,oBAAoB,CAAC,IAAI,EAAE,YAAY,GAAG,IAAI,CAK7D;AAED;;;GAGG;AAEH,wBAAgB,iBAAiB,CAAC,IAAI,EAAE,YAAY,GAAG,IAAI,CAG1D;AAED;;;GAGG;AAEH,wBAAgB,OAAO,CAAC,IAAI,EAAE,YAAY,GAAG,IAAI,CAQhD;AAMD,uBAAuB;AAEvB,eAAO,MAAM,yBAAyB;;;;;;;;;;;;;;;;;;CAkBrC,CAAC","sourcesContent":["import { disposeSessionAsyncJobManager } from \"./async/session-manager.js\";\nimport type { AgentEvent, AgentMessage } from \"@earendil-works/pi-agent-core\";\nimport type { AssistantMessage, Message, TextContent } from \"@earendil-works/pi-ai/compat\";\nimport { cleanupSessionResources } from \"@earendil-works/pi-ai/compat\";\nimport { formatCopilotProviderError } from \"./copilot-errors.ts\";\nimport type { AgentSessionInternalSurface as AgentSession } from \"./agent-session-methods.ts\";\nimport { customMessageExcludesContext, isSingleGenericAbortTextContent, replacementAbortContent, type AgentSessionEvent, type AgentSessionEventListener } from \"./agent-session-types.ts\";\nimport type { MessageEndEvent, MessageStartEvent, MessageUpdateEvent, ToolExecutionEndEvent, ToolExecutionStartEvent, ToolExecutionUpdateEvent, TurnEndEvent, TurnStartEvent } from \"./extensions/index.ts\";\n\nexport function _emit(this: AgentSession, event: AgentSessionEvent): void {\n\tfor (const l of this._eventListeners) {\n\t\tl(event);\n\t}\n}\n\n\nexport function _emitQueueUpdate(this: AgentSession): void {\n\tthis._emit({\n\t\ttype: \"queue_update\",\n\t\tsteering: [...this._steeringMessages],\n\t\tfollowUp: [...this._followUpMessages],\n\t});\n}\n\n/** Internal handler for agent events - shared by subscribe and reconnect */\n\nexport function _handleAgentEvent(this: AgentSession, event: AgentEvent): void {\n\t// Create retry promise synchronously before queueing async processing.\n\t// Agent.emit() calls this handler synchronously, and prompt() calls waitForRetry()\n\t// as soon as agent.prompt() resolves. If _retryPromise is created only inside\n\t// _processAgentEvent, slow earlier queued events can delay agent_end processing\n\t// and waitForRetry() can miss the in-flight retry.\n\tthis._createRetryPromiseForAgentEnd(event);\n\n\tthis._agentEventQueue = this._agentEventQueue.then(\n\t\t() => this._processAgentEvent(event),\n\t\t() => this._processAgentEvent(event),\n\t);\n\n\t// Keep queue alive if an event handler fails\n\tthis._agentEventQueue.catch(() => {});\n}\n\n\nexport function _createRetryPromiseForAgentEnd(this: AgentSession, event: AgentEvent): void {\n\tif (event.type !== \"agent_end\" || this._retryPromise) {\n\t\treturn;\n\t}\n\n\tconst settings = this.settingsManager.getRetrySettings();\n\tif (!settings.enabled && this._fallbackModels.length === 0) {\n\t\treturn;\n\t}\n\n\tconst lastAssistant = this._findLastAssistantInMessages(event.messages);\n\tif (!lastAssistant) {\n\t\treturn;\n\t}\n\n\tconst shouldRetry = this._isRetryableError(lastAssistant) || this._isEmptyCompletion(lastAssistant) || this._isSafetyRefusal(lastAssistant);\n\tif (!shouldRetry) {\n\t\treturn;\n\t}\n\n\tthis._retryPromise = new Promise((resolve) => {\n\t\tthis._retryResolve = resolve;\n\t});\n}\n\n\nexport function _findLastAssistantInMessages(this: AgentSession, messages: AgentMessage[]): AssistantMessage | undefined {\n\tfor (let i = messages.length - 1; i >= 0; i--) {\n\t\tconst message = messages[i];\n\t\tif (message.role === \"assistant\") {\n\t\t\treturn message as AssistantMessage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\n\nexport async function _processAgentEvent(this: AgentSession, event: AgentEvent): Promise<void> {\n\t// When a user message starts, check if it's from either queue and remove it BEFORE emitting\n\t// This ensures the UI sees the updated queue state\n\tif (event.type === \"message_start\" && event.message.role === \"user\") {\n\t\tthis._overflowRecoveryAttempted = false;\n\t\tthis._fallbackAttemptedKeys.clear();\n\t\tconst messageText = this._getUserMessageText(event.message);\n\t\tif (messageText) {\n\t\t\t// Check steering queue first\n\t\t\tconst steeringIndex = this._steeringMessages.indexOf(messageText);\n\t\t\tif (steeringIndex !== -1) {\n\t\t\t\tthis._steeringMessages.splice(steeringIndex, 1);\n\t\t\t\tthis._emitQueueUpdate();\n\t\t\t} else {\n\t\t\t\t// Check follow-up queue\n\t\t\t\tconst followUpIndex = this._followUpMessages.indexOf(messageText);\n\t\t\t\tif (followUpIndex !== -1) {\n\t\t\t\t\tthis._followUpMessages.splice(followUpIndex, 1);\n\t\t\t\t\tthis._emitQueueUpdate();\n\t\t\t\t}\n\t\t\t}\n\t\t}\n\t}\n\n\tthis._applyInterruptAbortMessage(event);\n\tthis._applyProviderErrorGuidance(event);\n\n\t// Emit to extensions first\n\tawait this._emitExtensionEvent(event);\n\n\t// Notify all listeners\n\tthis._emit(event);\n\n\t// Handle session persistence\n\tif (event.type === \"message_end\") {\n\t\t// Check if this is a custom message from extensions\n\t\tif (event.message.role === \"custom\") {\n\t\t\t// Persist as CustomMessageEntry\n\t\t\tthis.sessionManager.appendCustomMessageEntry(\n\t\t\t\tevent.message.customType,\n\t\t\t\tevent.message.content,\n\t\t\t\tevent.message.display,\n\t\t\t\tevent.message.details,\n\t\t\t\tcustomMessageExcludesContext(event.message),\n\t\t\t);\n\t\t} else if (\n\t\t\tevent.message.role === \"user\" ||\n\t\t\tevent.message.role === \"assistant\" ||\n\t\t\tevent.message.role === \"toolResult\"\n\t\t) {\n\t\t\tif (event.message.role === \"assistant\") {\n\t\t\t\tthis._normalizePersistedGeminiToolArgs(event.message);\n\t\t\t}\n\t\t\t// Regular LLM message - persist as SessionMessageEntry\n\t\t\tthis.sessionManager.appendMessage(event.message);\n\t\t}\n\t\t// Other message types (bashExecution, branchSummary) are persisted elsewhere\n\n\t\t// Track assistant message for auto-compaction (checked on agent_end)\n\t\tif (event.message.role === \"assistant\") {\n\t\t\tthis._lastAssistantMessage = event.message;\n\n\t\t\tconst assistantMsg = event.message as AssistantMessage;\n\t\t\t// Treat degenerate empty completions (no content, zero output tokens) and\n\t\t\t// intercepted canned safety refusals as failures alongside stopReason ===\n\t\t\t// \"error\". Otherwise such a turn that stops with reason \"stop\"/\"length\"\n\t\t\t// would reset the retry counter on every attempt, causing unbounded\n\t\t\t// retries instead of honoring maxRetries.\n\t\t\tconst assistantFailed = assistantMsg.stopReason === \"error\" || this._isEmptyCompletion(assistantMsg) || this._isSafetyRefusal(assistantMsg);\n\t\t\tif (!assistantFailed) {\n\t\t\t\tthis._fallbackAttemptedKeys.clear();\n\t\t\t\tthis._overflowRecoveryAttempted = false;\n\t\t\t}\n\n\t\t\t// Reset retry counter immediately on successful assistant response\n\t\t\t// This prevents accumulation across multiple LLM calls within a turn\n\t\t\tif (!assistantFailed && this._retryAttempt > 0) {\n\t\t\t\tthis._emit({\n\t\t\t\t\ttype: \"auto_retry_end\",\n\t\t\t\t\tsuccess: true,\n\t\t\t\t\tattempt: this._retryAttempt,\n\t\t\t\t});\n\t\t\t\tthis._retryAttempt = 0;\n\t\t\t}\n\t\t}\n\t}\n\n\t// Check auto-retry and auto-compaction after agent completes\n\tif (event.type === \"agent_end\" && this._lastAssistantMessage) {\n\t\tconst msg = this._lastAssistantMessage;\n\t\tthis._lastAssistantMessage = undefined;\n\n\t\t// Check for retryable errors first (overloaded, rate limit, server errors,\n\t\t// transient provider finish_reason errors, degenerate empty completions,\n\t\t// or intercepted canned safety refusals)\n\t\tconst retryableError = this._isRetryableError(msg);\n\t\tconst emptyCompletion = !retryableError && this._isEmptyCompletion(msg);\n\t\tconst safetyRefusal = !retryableError && !emptyCompletion && this._isSafetyRefusal(msg);\n\t\tif (retryableError || emptyCompletion || safetyRefusal) {\n\t\t\tif (emptyCompletion && !msg.errorMessage) {\n\t\t\t\t// Surface a clear reason in the retry banner; empty completions carry no\n\t\t\t\t// provider error message of their own.\n\t\t\t\tmsg.errorMessage = \"Provider returned an empty completion\";\n\t\t\t} else if (safetyRefusal && !msg.errorMessage) {\n\t\t\t\tmsg.errorMessage = \"Provider returned a canned safety refusal\";\n\t\t\t}\n\t\t\tconst didRetry = await this._handleRetryableError(msg);\n\t\t\tif (didRetry) return; // Retry was initiated, don't proceed to compaction\n\t\t}\n\n\t\tthis._resolveRetry();\n\t\tawait this._checkCompaction(msg);\n\t}\n}\n\n\nexport function _applyInterruptAbortMessage(this: AgentSession, event: AgentEvent): void {\n\tconst abortMessage = this._activeInterruptAbortMessage;\n\tif (!abortMessage) return;\n\n\tif (event.type === \"tool_execution_end\" && event.isError && isSingleGenericAbortTextContent(event.result.content)) {\n\t\tevent.result.content = replacementAbortContent(abortMessage);\n\t\treturn;\n\t}\n\n\tif (event.type !== \"message_start\" && event.type !== \"message_end\") return;\n\n\tif (event.message.role === \"toolResult\" && event.message.isError && isSingleGenericAbortTextContent(event.message.content)) {\n\t\tevent.message.content = replacementAbortContent(abortMessage);\n\t\treturn;\n\t}\n\n\tif (event.message.role === \"assistant\") {\n\t\tconst assistantMessage = event.message as AssistantMessage;\n\t\tif (assistantMessage.stopReason === \"aborted\") {\n\t\t\tassistantMessage.errorMessage = abortMessage;\n\t\t}\n\t}\n}\n\n\nexport function _applyProviderErrorGuidance(this: AgentSession, event: AgentEvent): void {\n\tif (event.type !== \"message_start\" && event.type !== \"message_update\" && event.type !== \"message_end\") return;\n\tif (event.message.role !== \"assistant\") return;\n\n\tconst assistantMessage = event.message as AssistantMessage;\n\tif (assistantMessage.stopReason !== \"error\" || !assistantMessage.errorMessage) return;\n\n\tassistantMessage.errorMessage = formatCopilotProviderError(\n\t\tassistantMessage.provider,\n\t\tassistantMessage.errorMessage,\n\t);\n}\n\n/** Resolve the pending retry promise */\n\nexport function _resolveRetry(this: AgentSession): void {\n\tif (this._retryResolve) {\n\t\tthis._retryResolve();\n\t\tthis._retryResolve = undefined;\n\t\tthis._retryPromise = undefined;\n\t}\n}\n\n/** Extract text content from a message */\n\nexport function _getUserMessageText(this: AgentSession, message: Message): string {\n\tif (message.role !== \"user\") return \"\";\n\tconst content = message.content;\n\tif (typeof content === \"string\") return content;\n\tconst textBlocks = content.filter((c) => c.type === \"text\");\n\treturn textBlocks.map((c) => (c as TextContent).text).join(\"\");\n}\n\n/** Find the last assistant message in agent state (including aborted ones) */\n\nexport function _findLastAssistantMessage(this: AgentSession): AssistantMessage | undefined {\n\tconst messages = this.agent.state.messages;\n\tfor (let i = messages.length - 1; i >= 0; i--) {\n\t\tconst msg = messages[i];\n\t\tif (msg.role === \"assistant\") {\n\t\t\treturn msg as AssistantMessage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\n\nexport function _replaceMessageInPlace(this: AgentSession, target: AgentMessage, replacement: AgentMessage): void {\n\t// Agent-core stores the finalized message object in its state before emitting message_end.\n\t// SessionManager persistence happens later in _processAgentEvent() with event.message.\n\t// Mutating this object in place keeps agent state, later turn/agent events, listeners,\n\t// and the eventual SessionManager.appendMessage(event.message) persistence in sync.\n\tif (target === replacement) {\n\t\treturn;\n\t}\n\n\tconst targetRecord = target as unknown as Record<string, unknown>;\n\tfor (const key of Object.keys(targetRecord)) {\n\t\tdelete targetRecord[key];\n\t}\n\tObject.assign(targetRecord, replacement);\n}\n\n/** Emit extension events based on agent events */\n\nexport async function _emitExtensionEvent(this: AgentSession, event: AgentEvent): Promise<void> {\n\tif (event.type === \"agent_start\") {\n\t\tthis._turnIndex = 0;\n\t\tawait this._extensionRunner.emit({ type: \"agent_start\" });\n\t} else if (event.type === \"agent_end\") {\n\t\tawait this._extensionRunner.emit({ type: \"agent_end\", messages: event.messages });\n\t} else if (event.type === \"turn_start\") {\n\t\tconst extensionEvent: TurnStartEvent = {\n\t\t\ttype: \"turn_start\",\n\t\t\tturnIndex: this._turnIndex,\n\t\t\ttimestamp: Date.now(),\n\t\t};\n\t\tawait this._extensionRunner.emit(extensionEvent);\n\t} else if (event.type === \"turn_end\") {\n\t\tconst extensionEvent: TurnEndEvent = {\n\t\t\ttype: \"turn_end\",\n\t\t\tturnIndex: this._turnIndex,\n\t\t\tmessage: event.message,\n\t\t\ttoolResults: event.toolResults,\n\t\t};\n\t\tawait this._extensionRunner.emit(extensionEvent);\n\t\tthis._turnIndex++;\n\t} else if (event.type === \"message_start\") {\n\t\tconst extensionEvent: MessageStartEvent = {\n\t\t\ttype: \"message_start\",\n\t\t\tmessage: event.message,\n\t\t};\n\t\tawait this._extensionRunner.emit(extensionEvent);\n\t} else if (event.type === \"message_update\") {\n\t\tconst extensionEvent: MessageUpdateEvent = {\n\t\t\ttype: \"message_update\",\n\t\t\tmessage: event.message,\n\t\t\tassistantMessageEvent: event.assistantMessageEvent,\n\t\t};\n\t\tawait this._extensionRunner.emit(extensionEvent);\n\t} else if (event.type === \"message_end\") {\n\t\tconst extensionEvent: MessageEndEvent = {\n\t\t\ttype: \"message_end\",\n\t\t\tmessage: event.message,\n\t\t};\n\t\tconst replacement = await this._extensionRunner.emitMessageEnd(extensionEvent);\n\t\tif (replacement) {\n\t\t\tthis._replaceMessageInPlace(event.message, replacement);\n\t\t}\n\t} else if (event.type === \"tool_execution_start\") {\n\t\tconst extensionEvent: ToolExecutionStartEvent = {\n\t\t\ttype: \"tool_execution_start\",\n\t\t\ttoolCallId: event.toolCallId,\n\t\t\ttoolName: event.toolName,\n\t\t\targs: event.args,\n\t\t};\n\t\tawait this._extensionRunner.emit(extensionEvent);\n\t} else if (event.type === \"tool_execution_update\") {\n\t\tconst extensionEvent: ToolExecutionUpdateEvent = {\n\t\t\ttype: \"tool_execution_update\",\n\t\t\ttoolCallId: event.toolCallId,\n\t\t\ttoolName: event.toolName,\n\t\t\targs: event.args,\n\t\t\tpartialResult: event.partialResult,\n\t\t};\n\t\tawait this._extensionRunner.emit(extensionEvent);\n\t} else if (event.type === \"tool_execution_end\") {\n\t\tconst extensionEvent: ToolExecutionEndEvent = {\n\t\t\ttype: \"tool_execution_end\",\n\t\t\ttoolCallId: event.toolCallId,\n\t\t\ttoolName: event.toolName,\n\t\t\tresult: event.result,\n\t\t\tisError: event.isError,\n\t\t};\n\t\tawait this._extensionRunner.emit(extensionEvent);\n\t}\n}\n\n/**\n * Subscribe to agent events.\n * Session persistence is handled internally (saves messages on message_end).\n * Multiple listeners can be added. Returns unsubscribe function for this listener.\n */\n\nexport function subscribe(this: AgentSession, listener: AgentSessionEventListener): () => void {\n\tthis._eventListeners.push(listener);\n\n\t// Return unsubscribe function for this specific listener\n\treturn () => {\n\t\tconst index = this._eventListeners.indexOf(listener);\n\t\tif (index !== -1) {\n\t\t\tthis._eventListeners.splice(index, 1);\n\t\t}\n\t};\n}\n\n/**\n * Temporarily disconnect from agent events.\n * User listeners are preserved and will receive events again after resubscribe().\n * Used internally during operations that need to pause event processing.\n */\n\nexport function _disconnectFromAgent(this: AgentSession): void {\n\tif (this._unsubscribeAgent) {\n\t\tthis._unsubscribeAgent();\n\t\tthis._unsubscribeAgent = undefined;\n\t}\n}\n\n/**\n * Reconnect to agent events after _disconnectFromAgent().\n * Preserves all existing listeners.\n */\n\nexport function _reconnectToAgent(this: AgentSession): void {\n\tif (this._unsubscribeAgent) return; // Already connected\n\tthis._unsubscribeAgent = this.agent.subscribe(this._handleAgentEvent);\n}\n\n/**\n * Remove all listeners and disconnect from agent.\n * Call this when completely done with the session.\n */\n\nexport function dispose(this: AgentSession): void {\n\tthis._extensionRunner.invalidate(\n\t\t\"This extension ctx is stale after session replacement or reload. Do not use a captured pi or command ctx after ctx.newSession(), ctx.fork(), ctx.switchSession(), or ctx.reload(). For newSession, fork, and switchSession, move post-replacement work into withSession and use the ctx passed to withSession. For reload, do not use the old ctx after await ctx.reload().\",\n\t);\n\tdisposeSessionAsyncJobManager(this._asyncJobManager, this._asyncJobManagerSessionId);\n\tthis._disconnectFromAgent();\n\tthis._eventListeners = [];\n\tcleanupSessionResources(this.sessionId);\n}\n\n// =========================================================================\n// Read-only State Access\n// =========================================================================\n\n/** Full agent state */\n\nexport const agentSessionEventsMethods = {\n\t_emit,\n\t_emitQueueUpdate,\n\t_handleAgentEvent,\n\t_createRetryPromiseForAgentEnd,\n\t_findLastAssistantInMessages,\n\t_processAgentEvent,\n\t_applyInterruptAbortMessage,\n\t_applyProviderErrorGuidance,\n\t_resolveRetry,\n\t_getUserMessageText,\n\t_findLastAssistantMessage,\n\t_replaceMessageInPlace,\n\t_emitExtensionEvent,\n\tsubscribe,\n\t_disconnectFromAgent,\n\t_reconnectToAgent,\n\tdispose,\n};\n"]}
+{"version":3,"file":"agent-session-events.d.ts","sourceRoot":"","sources":["../../src/core/agent-session-events.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,UAAU,EAAE,YAAY,EAAE,MAAM,+BAA+B,CAAC;AAC9E,OAAO,KAAK,EAAE,gBAAgB,EAAE,OAAO,EAAe,MAAM,8BAA8B,CAAC;AAG3F,OAAO,KAAK,EAAE,2BAA2B,IAAI,YAAY,EAAE,MAAM,4BAA4B,CAAC;AAC9F,OAAO,EAA0F,KAAK,iBAAiB,EAAE,KAAK,yBAAyB,EAAE,MAAM,0BAA0B,CAAC;AAI1L,wBAAgB,KAAK,CAAC,IAAI,EAAE,YAAY,EAAE,KAAK,EAAE,iBAAiB,GAAG,IAAI,CAIxE;AAGD,wBAAgB,gBAAgB,CAAC,IAAI,EAAE,YAAY,GAAG,IAAI,CAMzD;AAED,4EAA4E;AAE5E,wBAAgB,iBAAiB,CAAC,IAAI,EAAE,YAAY,EAAE,KAAK,EAAE,UAAU,GAAG,IAAI,CAe7E;AAGD,wBAAgB,8BAA8B,CAAC,IAAI,EAAE,YAAY,EAAE,KAAK,EAAE,UAAU,GAAG,IAAI,CAuB1F;AAGD,wBAAgB,4BAA4B,CAAC,IAAI,EAAE,YAAY,EAAE,QAAQ,EAAE,YAAY,EAAE,GAAG,gBAAgB,GAAG,SAAS,CAQvH;AAGD,wBAAsB,kBAAkB,CAAC,IAAI,EAAE,YAAY,EAAE,KAAK,EAAE,UAAU,GAAG,OAAO,CAAC,IAAI,CAAC,CAyH7F;AAGD,wBAAgB,2BAA2B,CAAC,IAAI,EAAE,YAAY,EAAE,KAAK,EAAE,UAAU,GAAG,IAAI,CAsBvF;AAGD,wBAAgB,2BAA2B,CAAC,IAAI,EAAE,YAAY,EAAE,KAAK,EAAE,UAAU,GAAG,IAAI,CAWvF;AAED,wCAAwC;AAExC,wBAAgB,aAAa,CAAC,IAAI,EAAE,YAAY,GAAG,IAAI,CAMtD;AAED,0CAA0C;AAE1C,wBAAgB,mBAAmB,CAAC,IAAI,EAAE,YAAY,EAAE,OAAO,EAAE,OAAO,GAAG,MAAM,CAMhF;AAED,8EAA8E;AAE9E,wBAAgB,yBAAyB,CAAC,IAAI,EAAE,YAAY,GAAG,gBAAgB,GAAG,SAAS,CAS1F;AAGD,wBAAgB,sBAAsB,CAAC,IAAI,EAAE,YAAY,EAAE,MAAM,EAAE,YAAY,EAAE,WAAW,EAAE,YAAY,GAAG,IAAI,CAchH;AAED,kDAAkD;AAElD,wBAAsB,mBAAmB,CAAC,IAAI,EAAE,YAAY,EAAE,KAAK,EAAE,UAAU,GAAG,OAAO,CAAC,IAAI,CAAC,CAuE9F;AAED;;;;GAIG;AAEH,wBAAgB,SAAS,CAAC,IAAI,EAAE,YAAY,EAAE,QAAQ,EAAE,yBAAyB,GAAG,MAAM,IAAI,CAU7F;AAED;;;;GAIG;AAEH,wBAAgB,oBAAoB,CAAC,IAAI,EAAE,YAAY,GAAG,IAAI,CAK7D;AAED;;;GAGG;AAEH,wBAAgB,iBAAiB,CAAC,IAAI,EAAE,YAAY,GAAG,IAAI,CAG1D;AAED;;;GAGG;AAEH,wBAAgB,OAAO,CAAC,IAAI,EAAE,YAAY,GAAG,IAAI,CAQhD;AAMD,uBAAuB;AAEvB,eAAO,MAAM,yBAAyB;;;;;;;;;;;;;;;;;;CAkBrC,CAAC","sourcesContent":["import { disposeSessionAsyncJobManager } from \"./async/session-manager.js\";\nimport type { AgentEvent, AgentMessage } from \"@earendil-works/pi-agent-core\";\nimport type { AssistantMessage, Message, TextContent } from \"@earendil-works/pi-ai/compat\";\nimport { cleanupSessionResources } from \"@earendil-works/pi-ai/compat\";\nimport { formatCopilotProviderError } from \"./copilot-errors.ts\";\nimport type { AgentSessionInternalSurface as AgentSession } from \"./agent-session-methods.ts\";\nimport { customMessageExcludesContext, isSingleGenericAbortTextContent, replacementAbortContent, type AgentSessionEvent, type AgentSessionEventListener } from \"./agent-session-types.ts\";\nimport type { MessageEndEvent, MessageStartEvent, MessageUpdateEvent, ToolExecutionEndEvent, ToolExecutionStartEvent, ToolExecutionUpdateEvent, TurnEndEvent, TurnStartEvent } from \"./extensions/index.ts\";\nimport { normalizeMessageContent } from \"./messages.ts\";\n\nexport function _emit(this: AgentSession, event: AgentSessionEvent): void {\n\tfor (const l of this._eventListeners) {\n\t\tl(event);\n\t}\n}\n\n\nexport function _emitQueueUpdate(this: AgentSession): void {\n\tthis._emit({\n\t\ttype: \"queue_update\",\n\t\tsteering: [...this._steeringMessages],\n\t\tfollowUp: [...this._followUpMessages],\n\t});\n}\n\n/** Internal handler for agent events - shared by subscribe and reconnect */\n\nexport function _handleAgentEvent(this: AgentSession, event: AgentEvent): void {\n\t// Create retry promise synchronously before queueing async processing.\n\t// Agent.emit() calls this handler synchronously, and prompt() calls waitForRetry()\n\t// as soon as agent.prompt() resolves. If _retryPromise is created only inside\n\t// _processAgentEvent, slow earlier queued events can delay agent_end processing\n\t// and waitForRetry() can miss the in-flight retry.\n\tthis._createRetryPromiseForAgentEnd(event);\n\n\tthis._agentEventQueue = this._agentEventQueue.then(\n\t\t() => this._processAgentEvent(event),\n\t\t() => this._processAgentEvent(event),\n\t);\n\n\t// Keep queue alive if an event handler fails\n\tthis._agentEventQueue.catch(() => {});\n}\n\n\nexport function _createRetryPromiseForAgentEnd(this: AgentSession, event: AgentEvent): void {\n\tif (event.type !== \"agent_end\" || this._retryPromise) {\n\t\treturn;\n\t}\n\n\tconst settings = this.settingsManager.getRetrySettings();\n\tif (!settings.enabled && this._fallbackModels.length === 0) {\n\t\treturn;\n\t}\n\n\tconst lastAssistant = this._findLastAssistantInMessages(event.messages);\n\tif (!lastAssistant) {\n\t\treturn;\n\t}\n\n\tconst shouldRetry = this._isRetryableError(lastAssistant) || this._isEmptyCompletion(lastAssistant) || this._isSafetyRefusal(lastAssistant);\n\tif (!shouldRetry) {\n\t\treturn;\n\t}\n\n\tthis._retryPromise = new Promise((resolve) => {\n\t\tthis._retryResolve = resolve;\n\t});\n}\n\n\nexport function _findLastAssistantInMessages(this: AgentSession, messages: AgentMessage[]): AssistantMessage | undefined {\n\tfor (let i = messages.length - 1; i >= 0; i--) {\n\t\tconst message = messages[i];\n\t\tif (message.role === \"assistant\") {\n\t\t\treturn message as AssistantMessage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\n\nexport async function _processAgentEvent(this: AgentSession, event: AgentEvent): Promise<void> {\n\t// When a user message starts, check if it's from either queue and remove it BEFORE emitting\n\t// This ensures the UI sees the updated queue state\n\tif (event.type === \"message_start\" && event.message.role === \"user\") {\n\t\tthis._overflowRecoveryAttempted = false;\n\t\tthis._fallbackAttemptedKeys.clear();\n\t\tconst messageText = this._getUserMessageText(event.message);\n\t\tif (messageText) {\n\t\t\t// Check steering queue first\n\t\t\tconst steeringIndex = this._steeringMessages.indexOf(messageText);\n\t\t\tif (steeringIndex !== -1) {\n\t\t\t\tthis._steeringMessages.splice(steeringIndex, 1);\n\t\t\t\tthis._emitQueueUpdate();\n\t\t\t} else {\n\t\t\t\t// Check follow-up queue\n\t\t\t\tconst followUpIndex = this._followUpMessages.indexOf(messageText);\n\t\t\t\tif (followUpIndex !== -1) {\n\t\t\t\t\tthis._followUpMessages.splice(followUpIndex, 1);\n\t\t\t\t\tthis._emitQueueUpdate();\n\t\t\t\t}\n\t\t\t}\n\t\t}\n\t}\n\n\tthis._applyInterruptAbortMessage(event);\n\tthis._applyProviderErrorGuidance(event);\n\n\t// Emit to extensions first\n\tawait this._emitExtensionEvent(event);\n\n\t// Notify all listeners\n\tthis._emit(event);\n\n\t// Handle session persistence\n\tif (event.type === \"message_end\") {\n\t\t// Check if this is a custom message from extensions\n\t\tif (event.message.role === \"custom\") {\n\t\t\t// Persist as CustomMessageEntry\n\t\t\tthis.sessionManager.appendCustomMessageEntry(\n\t\t\t\tevent.message.customType,\n\t\t\t\tevent.message.content,\n\t\t\t\tevent.message.display,\n\t\t\t\tevent.message.details,\n\t\t\t\tcustomMessageExcludesContext(event.message),\n\t\t\t);\n\t\t} else if (\n\t\t\tevent.message.role === \"user\" ||\n\t\t\tevent.message.role === \"assistant\" ||\n\t\t\tevent.message.role === \"toolResult\"\n\t\t) {\n\t\t\tif (event.message.role === \"assistant\") {\n\t\t\t\tthis._normalizePersistedGeminiToolArgs(event.message);\n\t\t\t}\n\t\t\t// Regular LLM message - persist as SessionMessageEntry\n\t\t\tthis.sessionManager.appendMessage(event.message);\n\t\t}\n\t\t// Other message types (bashExecution, branchSummary) are persisted elsewhere\n\n\t\t// Track assistant message for auto-compaction (checked on agent_end)\n\t\tif (event.message.role === \"assistant\") {\n\t\t\tthis._lastAssistantMessage = event.message;\n\n\t\t\tconst assistantMsg = event.message as AssistantMessage;\n\t\t\t// Treat degenerate empty completions (no content, zero output tokens) and\n\t\t\t// intercepted canned safety refusals as failures alongside stopReason ===\n\t\t\t// \"error\". Otherwise such a turn that stops with reason \"stop\"/\"length\"\n\t\t\t// would reset the retry counter on every attempt, causing unbounded\n\t\t\t// retries instead of honoring maxRetries.\n\t\t\tconst assistantFailed = assistantMsg.stopReason === \"error\" || this._isEmptyCompletion(assistantMsg) || this._isSafetyRefusal(assistantMsg);\n\t\t\tif (!assistantFailed) {\n\t\t\t\tthis._fallbackAttemptedKeys.clear();\n\t\t\t\tthis._overflowRecoveryAttempted = false;\n\t\t\t\tthis._outputBudgetErrorContinuationAttempts = 0;\n\t\t\t}\n\n\t\t\t// A non-truncated assistant response means the length-continuation loop\n\t\t\t// made progress (or the turn completed cleanly), so reset the bounded\n\t\t\t// output-cap continuation counter.\n\t\t\tif (assistantMsg.stopReason !== \"length\") {\n\t\t\t\tthis._lengthContinuationAttempts = 0;\n\t\t\t}\n\n\t\t\t// Reset retry counter immediately on successful assistant response\n\t\t\t// This prevents accumulation across multiple LLM calls within a turn\n\t\t\tif (!assistantFailed && this._retryAttempt > 0) {\n\t\t\t\tthis._emit({\n\t\t\t\t\ttype: \"auto_retry_end\",\n\t\t\t\t\tsuccess: true,\n\t\t\t\t\tattempt: this._retryAttempt,\n\t\t\t\t});\n\t\t\t\tthis._retryAttempt = 0;\n\t\t\t}\n\t\t}\n\t}\n\n\t// Check auto-retry and auto-compaction after agent completes\n\tif (event.type === \"agent_end\" && this._lastAssistantMessage) {\n\t\tconst msg = this._lastAssistantMessage;\n\t\tthis._lastAssistantMessage = undefined;\n\n\t\t// Check for retryable errors first (overloaded, rate limit, server errors,\n\t\t// transient provider finish_reason errors, degenerate empty completions,\n\t\t// or intercepted canned safety refusals)\n\t\tconst retryableError = this._isRetryableError(msg);\n\t\tconst emptyCompletion = !retryableError && this._isEmptyCompletion(msg);\n\t\tconst safetyRefusal = !retryableError && !emptyCompletion && this._isSafetyRefusal(msg);\n\t\tif (retryableError || emptyCompletion || safetyRefusal) {\n\t\t\tif (emptyCompletion && !msg.errorMessage) {\n\t\t\t\t// Surface a clear reason in the retry banner; empty completions carry no\n\t\t\t\t// provider error message of their own.\n\t\t\t\tmsg.errorMessage = \"Provider returned an empty completion\";\n\t\t\t} else if (safetyRefusal && !msg.errorMessage) {\n\t\t\t\tmsg.errorMessage = \"Provider returned a canned safety refusal\";\n\t\t\t}\n\t\t\tconst didRetry = await this._handleRetryableError(msg);\n\t\t\tif (didRetry) return; // Retry was initiated, don't proceed to compaction\n\t\t}\n\n\t\tthis._resolveRetry();\n\t\tawait this._checkCompaction(msg);\n\t}\n}\n\n\nexport function _applyInterruptAbortMessage(this: AgentSession, event: AgentEvent): void {\n\tconst abortMessage = this._activeInterruptAbortMessage;\n\tif (!abortMessage) return;\n\n\tif (event.type === \"tool_execution_end\" && event.isError && isSingleGenericAbortTextContent(event.result.content)) {\n\t\tevent.result.content = replacementAbortContent(abortMessage);\n\t\treturn;\n\t}\n\n\tif (event.type !== \"message_start\" && event.type !== \"message_end\") return;\n\n\tif (event.message.role === \"toolResult\" && event.message.isError && isSingleGenericAbortTextContent(event.message.content)) {\n\t\tevent.message.content = replacementAbortContent(abortMessage);\n\t\treturn;\n\t}\n\n\tif (event.message.role === \"assistant\") {\n\t\tconst assistantMessage = event.message as AssistantMessage;\n\t\tif (assistantMessage.stopReason === \"aborted\") {\n\t\t\tassistantMessage.errorMessage = abortMessage;\n\t\t}\n\t}\n}\n\n\nexport function _applyProviderErrorGuidance(this: AgentSession, event: AgentEvent): void {\n\tif (event.type !== \"message_start\" && event.type !== \"message_update\" && event.type !== \"message_end\") return;\n\tif (event.message.role !== \"assistant\") return;\n\n\tconst assistantMessage = event.message as AssistantMessage;\n\tif (assistantMessage.stopReason !== \"error\" || !assistantMessage.errorMessage) return;\n\n\tassistantMessage.errorMessage = formatCopilotProviderError(\n\t\tassistantMessage.provider,\n\t\tassistantMessage.errorMessage,\n\t);\n}\n\n/** Resolve the pending retry promise */\n\nexport function _resolveRetry(this: AgentSession): void {\n\tif (this._retryResolve) {\n\t\tthis._retryResolve();\n\t\tthis._retryResolve = undefined;\n\t\tthis._retryPromise = undefined;\n\t}\n}\n\n/** Extract text content from a message */\n\nexport function _getUserMessageText(this: AgentSession, message: Message): string {\n\tif (message.role !== \"user\") return \"\";\n\tconst content = message.content;\n\tif (typeof content === \"string\") return content;\n\tconst textBlocks = content.filter((c) => c.type === \"text\");\n\treturn textBlocks.map((c) => (c as TextContent).text).join(\"\");\n}\n\n/** Find the last assistant message in agent state (including aborted ones) */\n\nexport function _findLastAssistantMessage(this: AgentSession): AssistantMessage | undefined {\n\tconst messages = this.agent.state.messages;\n\tfor (let i = messages.length - 1; i >= 0; i--) {\n\t\tconst msg = messages[i];\n\t\tif (msg.role === \"assistant\") {\n\t\t\treturn msg as AssistantMessage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\n\nexport function _replaceMessageInPlace(this: AgentSession, target: AgentMessage, replacement: AgentMessage): void {\n\t// Agent-core stores the finalized message object in its state before emitting message_end.\n\t// SessionManager persistence happens later in _processAgentEvent() with event.message.\n\t// Mutating this object in place keeps agent state, later turn/agent events, listeners,\n\t// and the eventual SessionManager.appendMessage(event.message) persistence in sync.\n\tif (target === replacement) {\n\t\treturn;\n\t}\n\n\tconst targetRecord = target as unknown as Record<string, unknown>;\n\tfor (const key of Object.keys(targetRecord)) {\n\t\tdelete targetRecord[key];\n\t}\n\tObject.assign(targetRecord, replacement);\n}\n\n/** Emit extension events based on agent events */\n\nexport async function _emitExtensionEvent(this: AgentSession, event: AgentEvent): Promise<void> {\n\tif (event.type === \"agent_start\") {\n\t\tthis._turnIndex = 0;\n\t\tawait this._extensionRunner.emit({ type: \"agent_start\" });\n\t} else if (event.type === \"agent_end\") {\n\t\tawait this._extensionRunner.emit({ type: \"agent_end\", messages: event.messages });\n\t} else if (event.type === \"turn_start\") {\n\t\tconst extensionEvent: TurnStartEvent = {\n\t\t\ttype: \"turn_start\",\n\t\t\tturnIndex: this._turnIndex,\n\t\t\ttimestamp: Date.now(),\n\t\t};\n\t\tawait this._extensionRunner.emit(extensionEvent);\n\t} else if (event.type === \"turn_end\") {\n\t\tconst extensionEvent: TurnEndEvent = {\n\t\t\ttype: \"turn_end\",\n\t\t\tturnIndex: this._turnIndex,\n\t\t\tmessage: event.message,\n\t\t\ttoolResults: event.toolResults,\n\t\t};\n\t\tawait this._extensionRunner.emit(extensionEvent);\n\t\tthis._turnIndex++;\n\t} else if (event.type === \"message_start\") {\n\t\tconst extensionEvent: MessageStartEvent = {\n\t\t\ttype: \"message_start\",\n\t\t\tmessage: event.message,\n\t\t};\n\t\tawait this._extensionRunner.emit(extensionEvent);\n\t} else if (event.type === \"message_update\") {\n\t\tconst extensionEvent: MessageUpdateEvent = {\n\t\t\ttype: \"message_update\",\n\t\t\tmessage: event.message,\n\t\t\tassistantMessageEvent: event.assistantMessageEvent,\n\t\t};\n\t\tawait this._extensionRunner.emit(extensionEvent);\n\t} else if (event.type === \"message_end\") {\n\t\tconst extensionEvent: MessageEndEvent = {\n\t\t\ttype: \"message_end\",\n\t\t\tmessage: event.message,\n\t\t};\n\t\tconst replacement = await this._extensionRunner.emitMessageEnd(extensionEvent);\n\t\tif (replacement) {\n\t\t\tthis._replaceMessageInPlace(event.message, normalizeMessageContent(replacement));\n\t\t}\n\t} else if (event.type === \"tool_execution_start\") {\n\t\tconst extensionEvent: ToolExecutionStartEvent = {\n\t\t\ttype: \"tool_execution_start\",\n\t\t\ttoolCallId: event.toolCallId,\n\t\t\ttoolName: event.toolName,\n\t\t\targs: event.args,\n\t\t};\n\t\tawait this._extensionRunner.emit(extensionEvent);\n\t} else if (event.type === \"tool_execution_update\") {\n\t\tconst extensionEvent: ToolExecutionUpdateEvent = {\n\t\t\ttype: \"tool_execution_update\",\n\t\t\ttoolCallId: event.toolCallId,\n\t\t\ttoolName: event.toolName,\n\t\t\targs: event.args,\n\t\t\tpartialResult: event.partialResult,\n\t\t};\n\t\tawait this._extensionRunner.emit(extensionEvent);\n\t} else if (event.type === \"tool_execution_end\") {\n\t\tconst extensionEvent: ToolExecutionEndEvent = {\n\t\t\ttype: \"tool_execution_end\",\n\t\t\ttoolCallId: event.toolCallId,\n\t\t\ttoolName: event.toolName,\n\t\t\tresult: event.result,\n\t\t\tisError: event.isError,\n\t\t};\n\t\tawait this._extensionRunner.emit(extensionEvent);\n\t}\n}\n\n/**\n * Subscribe to agent events.\n * Session persistence is handled internally (saves messages on message_end).\n * Multiple listeners can be added. Returns unsubscribe function for this listener.\n */\n\nexport function subscribe(this: AgentSession, listener: AgentSessionEventListener): () => void {\n\tthis._eventListeners.push(listener);\n\n\t// Return unsubscribe function for this specific listener\n\treturn () => {\n\t\tconst index = this._eventListeners.indexOf(listener);\n\t\tif (index !== -1) {\n\t\t\tthis._eventListeners.splice(index, 1);\n\t\t}\n\t};\n}\n\n/**\n * Temporarily disconnect from agent events.\n * User listeners are preserved and will receive events again after resubscribe().\n * Used internally during operations that need to pause event processing.\n */\n\nexport function _disconnectFromAgent(this: AgentSession): void {\n\tif (this._unsubscribeAgent) {\n\t\tthis._unsubscribeAgent();\n\t\tthis._unsubscribeAgent = undefined;\n\t}\n}\n\n/**\n * Reconnect to agent events after _disconnectFromAgent().\n * Preserves all existing listeners.\n */\n\nexport function _reconnectToAgent(this: AgentSession): void {\n\tif (this._unsubscribeAgent) return; // Already connected\n\tthis._unsubscribeAgent = this.agent.subscribe(this._handleAgentEvent);\n}\n\n/**\n * Remove all listeners and disconnect from agent.\n * Call this when completely done with the session.\n */\n\nexport function dispose(this: AgentSession): void {\n\tthis._extensionRunner.invalidate(\n\t\t\"This extension ctx is stale after session replacement or reload. Do not use a captured pi or command ctx after ctx.newSession(), ctx.fork(), ctx.switchSession(), or ctx.reload(). For newSession, fork, and switchSession, move post-replacement work into withSession and use the ctx passed to withSession. For reload, do not use the old ctx after await ctx.reload().\",\n\t);\n\tdisposeSessionAsyncJobManager(this._asyncJobManager, this._asyncJobManagerSessionId);\n\tthis._disconnectFromAgent();\n\tthis._eventListeners = [];\n\tcleanupSessionResources(this.sessionId);\n}\n\n// =========================================================================\n// Read-only State Access\n// =========================================================================\n\n/** Full agent state */\n\nexport const agentSessionEventsMethods = {\n\t_emit,\n\t_emitQueueUpdate,\n\t_handleAgentEvent,\n\t_createRetryPromiseForAgentEnd,\n\t_findLastAssistantInMessages,\n\t_processAgentEvent,\n\t_applyInterruptAbortMessage,\n\t_applyProviderErrorGuidance,\n\t_resolveRetry,\n\t_getUserMessageText,\n\t_findLastAssistantMessage,\n\t_replaceMessageInPlace,\n\t_emitExtensionEvent,\n\tsubscribe,\n\t_disconnectFromAgent,\n\t_reconnectToAgent,\n\tdispose,\n};\n"]}

package/dist/core/agent-session-events.js

@@ -2,6 +2,7 @@ import { disposeSessionAsyncJobManager } from "./async/session-manager.js";
 import { cleanupSessionResources } from "@earendil-works/pi-ai/compat";
 import { formatCopilotProviderError } from "./copilot-errors.js";
 import { customMessageExcludesContext, isSingleGenericAbortTextContent, replacementAbortContent } from "./agent-session-types.js";
+import { normalizeMessageContent } from "./messages.js";
 export function _emit(event) {
     for (const l of this._eventListeners) {
         l(event);
@@ -115,6 +116,13 @@ export async function _processAgentEvent(event) {
             if (!assistantFailed) {
                 this._fallbackAttemptedKeys.clear();
                 this._overflowRecoveryAttempted = false;
+                this._outputBudgetErrorContinuationAttempts = 0;
+            }
+            // A non-truncated assistant response means the length-continuation loop
+            // made progress (or the turn completed cleanly), so reset the bounded
+            // output-cap continuation counter.
+            if (assistantMsg.stopReason !== "length") {
+                this._lengthContinuationAttempts = 0;
             }
             // Reset retry counter immediately on successful assistant response
             // This prevents accumulation across multiple LLM calls within a turn
@@ -278,7 +286,7 @@ export async function _emitExtensionEvent(event) {
         };
         const replacement = await this._extensionRunner.emitMessageEnd(extensionEvent);
         if (replacement) {
-            this._replaceMessageInPlace(event.message, replacement);
+            this._replaceMessageInPlace(event.message, normalizeMessageContent(replacement));
         }
     }
     else if (event.type === "tool_execution_start") {