Apache Maka synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From request to durable execution
Apache Maka is an enclosing agent runtime and workspace, not only a model client or memory library. This account covers the shipped system at commit 1751c399239b0f6b717e98e2a5e2b875c3dd11ce, with analysis cutoff 2026-10-15. SRC-1 is the frozen Apache Maka repository. The boundary and Source register include clients, runtime, persistence, extension mechanisms, evaluators and optional assistant services. The evidence is inspected implementation and attributed design, not observed runs or causal experiments. Whole-system scope does not mean exhaustive inspection of every included path. This synthesis uses the accepted records; no additional source files were read.
Maka declares that its purpose is finishing tasks at measured cost and that its event log is the authority for execution, prompts and recovery. The inspected paths support bounded parts of that architecture, not its comparative benchmark claims: RT-CLM-task-runtime.
In the traced CLI route, the user supplies text, workspace and model target. The client connects to a Runtime Host, prepares a prompt, requires a Run identity and observes a Turn. Session, Turn and Run have distinct identities. Execution and durable state belong to the Host, not the observing CLI: RT-RTE-cli-turn. The Host resolves a configured model connection and composes instructions, advertised skills, eligible memory and tools. The serving model's exact weights are not certified by the configured model name: RT-CMP-ai-model.
The model proposes the next text or tool action; code decides dispatch eligibility. The main turn loop resolves calls against the active tool map, settles attempts through durable boundaries and reconstructs continuation from durable events. Code Mode nested calls re-enter settlement. A terminal answer is therefore a delivered runtime outcome, not an independent proof of task fulfillment: RT-RTE-tool-settlement.
This boundary is not a universal sandbox. Built-in managed filesystem and the inspected Bash branches fail closed when their required execution facilities are missing. Bypass, injected workspaces and client capabilities have different effect envelopes; client preparation can be admitted under Bypass or ask. Arbitrary plugin I/O is not constrained merely by receiving boundary metadata. Plugin executors are alternate backends, explicitly without Maka sandbox-boundary requests, and plugin LLM adapters can generate outside the main model/tool loop. Local dispatch guarantees cannot be projected through those seams: RT-BAP-tool-boundary, RT-RTE-executor-plugin.
Delegation and keeping context bounded
The representative multi-step workflow is a foreground request that delegates a bounded task, retrieves selected child output and continues to a client-visible answer. The foreground model chooses whether and what to delegate. Host code validates the profile, active parent and durable lineage, then creates a separately identified child with explicit task and tool ceiling. Children do not implicitly inherit local memory or personalization; required background must be supplied in the task. Executor-backed children use a separate execution path with no local tool list. Parent cancellation is rechecked before initial child admission: RT-RTE-child-graph.
Only selected results cross back; bounded output reads can request more. This keeps intermediate child history out of the parent window without deleting it. Durable, independently addressable child sessions differ from disposable call frames. For worktree children, terminal/quiescence checks can admit capture and publication of workspace.patch. Publication and returned artifact IDs do not apply the patch to the parent or verify its quality. The selected workflow has admission and structural result checks, not a mandatory independent task-quality verifier before the foreground answer: WF-RTE-patch-publication, workflow account.
Within a model context, Maka distinguishes the full ledger from effective history. Code preferentially removes oversized tool bodies and can fold history into a retained checkpoint. Later requests receive a text summary plus replay tail, or opaque provider state through the provider boundary. Identity, coverage and compatibility checks govern reuse; they do not check summary meaning. The encrypted alternative cannot be classified from its readable wrapper. Omitted tool evidence can be requested through a session-bound event address and reloaded from ledger evidence, subject to integrity and size checks. An address does not make the body resident or automatically dereferenced: RT-RTE-context-projection, MEM-OBJ-text-checkpoint, MEM-OBJ-provider-checkpoint, MEM-RTE-archive-read.
Retention, acceptance and later reliance
Maka has several distinct memory admission planes. Local Markdown memory supports proposals, approval, rejection, lifecycle changes and whole-document replacement. Foreground composition automatically selects active, policy-eligible, session-compatible bodies under a 12,000-character cap. Selection is not task-relevance search. The composer labels the text untrusted context; that wording requests model behavior, not executable permission enforcement. Producer-written user-approval labels do not establish per-content human review across every caller, and the queue is not the sole admission route: MEM-RTE-local-memory, MEM-OBJ-bundle-metadata.
The stronger evidence-sensitive mechanism is trace-derived fact extraction. Eligible conversational evidence feeds proposal and canonicalization, then software/model decisions select durable MemoryItems without requiring the Markdown approval queue. Deterministic citation checks establish lexical occurrence, not entailment. A separate model reconstructs assertions from user quotations under a criterion of full evidence support; code validates result identities, reattaches citations, repeats admission checks and rechecks policy before commit. This is a wired, fallible evidence-consuming acceptance decision for later global/workspace reuse as a supported user assertion. It is not independent verification of what the user said, and no concrete output establishes warranted novel derivation or accepted ampliative production: MEM-RTE-fact-extraction, EPI-RTE-memory-evidence, EPI-RTE-memory-canonicalize, EPI-RTE-memory-disposition.
Recall is a separate, requested consumer. It selects facts by key prefixes and transcripts by literal matching/ranking, with bounded passage delivery. Stored fact provenance does not all reach that consumer: the Host returns content, kind and time, without stored citation links or canonicalization rationale. Retrieval therefore does not add acceptance or deliver complete warrant. Archived sessions remain searchable, unlike archived facts omitted by the default fact lookup: MEM-RTE-recall-history, EPI-RTE-memory-commit.
Optional assistant indexes retain agent-authored documents, original references and exact coverage/progress state. Body writes and range completion are separate admissions; stale ranges or revisions can be rejected, but completion is not certified semantic completeness. Directories and links are navigation, not body delivery. Matters retain continuing objective state and independently reviewed settlement proposals. Rejection returns feedback and keeps the activation active; intermediate checkpoints do not share every final-settlement gate. Later activations are directed to read current state files, not shown to consume every historical review narrative. These are distinct from independent extraction files, which do not automatically enter an index, and daily-review archives, whose later client query does not establish automatic method revision: MEM-RTE-index-maintenance, MEM-RTE-matter-state, MEM-RTE-history-extract, MEM-OBJ-review-archive.
Continuation, evaluation and change
Persistent work is not limited to a foreground message. Scheduled fires and Graph wakes have separate execution identities and admission checks; recovery checks durable authority rather than blindly replaying displayed text. Admission or fire settlement is not final task success: RT-RTE-durable-wake, RT-RTE-child-graph. An additional ordinary-session heartbeat alternative remains outside the detailed report inventory, preventing a complete account of proactive modes: report verification limits.
Goal evaluation has real operational force. A separate tool-free model call judges recent conversation, and code consumes its flags to stop, wait or prepare another turn. The input is the last six nonempty user/assistant messages with long-message excerpts, not independent tool re-execution. Evaluator failure permits continuation without being counted as genuine progress evidence. A reason-derived pending prompt can steer a later turn, but an achieved label is not independently verified goal truth: EPI-RTE-goal-judge, EPI-RTE-goal-selection. The distinct Eval route imports external reward/status for reporting; completed verification can coexist with score zero and does not automatically revise the runtime: EPI-RTE-eval-verification.
The strongest supported learning-related contribution is architectural: automatic fact extraction, text compaction, history-fed indexes, Matter state and retained goal steering connect experience to durable material with later consumers. This establishes trace-fed update wiring, not improved capacity for future action. No before/after capacity comparison or component intervention is supplied. Task-state criticism and evidence-responsive continuation are positive mechanisms; they do not establish that criticism of a consumed formulated theory improved future capacity. The accepted runtime does not assess theory-builder conditions 1–4 separately, so this synthesis does not infer localization, consumption, content-directed criticism or iteration of a theory from memory and judging alone. Nor does it infer a reflective or autonomous theory builder. Computational continuation and autonomous triggers exist in the named routes; complete system autonomy and reflection remain unestablished, rather than evidenced absent. These independent limits do not negate the supported retained-context mechanisms: MEM-RTE-fact-extraction, RT-RTE-context-projection, MEM-RTE-matter-state, EPI-RTE-goal-selection.
For a long task that exceeds the context window, the architecture provides isolated child work, reduced effective history and selective evidence reload. For continuing personalized work, it provides bounded memory delivery and evidence-gated fact reuse. Neither scenario has observed success here. Plugin installation can change capabilities and prompts through committed desired authority followed by runtime convergence; installability is not evaluated self-revision. At this whole-system boundary and across later invocations, no exercised self-improvement pathway is established: retained outputs and live consumers are partly wired, but subsequent operation causally depending on an improvement-directed organizational change is not shown. Successful improvement is a further unanswered question. Traces linking an evidence-responsive update to later live consumption would change the exercised-path assessment; matched outcome comparisons would be needed to assess benefit. Semantic output checks, independent task evidence, deployment recovery/isolation traces and completion of the named alternative-route inspections would separately strengthen the corresponding conclusions: RT-RTE-plugin-admission, EPI-RTE-goal-selection, boundary evidence limits.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Implementation is not exercised behavior or benefit | SRC-1; RT-RTE-cli-turn, RT-CLM-task-runtime, EPI-RTE-goal-selection | Frozen source and accepted reports; no supplied runs or experiments | Task success, reliable compliance/recovery, measured superiority, improved capacity or exercised self-improvement | Linked executions showing update, actual consumer and subsequent dependence; objective outcome comparisons with attribution appropriate to the intervention |
| External models, executors and effect envelopes remain opaque | SRC-1; RT-CMP-ai-model, RT-RTE-executor-plugin, RT-BAP-tool-boundary | Host composition and selected local/client guards, not serving weights or remote internals | Exact model fixity, universal ledger completeness, uniform sandboxing or deployed isolation | Weight/version identity and adapter contracts, relevant implementation and deployed effect/isolation evidence |
| Approval labels do not authenticate all callers | SRC-1; MEM-OBJ-bundle-metadata, MEM-RTE-local-memory | Mutation handlers, replacement and prompt projection | Universal per-content human review, or immediate erasure from already composed context | End-to-end caller authorization and per-content decisions; prompt refresh/withdrawal traces |
| Semantic fidelity and provider checkpoint contents unresolved | SRC-1; RT-RTE-context-projection, MEM-OBJ-provider-checkpoint, EPI-RTE-memory-canonicalize, MEM-RTE-archive-read | Structural coverage, citation admission, text/native checkpoint seams and event-address reload | Lossless summaries, universal entailment, hidden payload form/reasons, all legacy reload guarantees | Concrete input/output semantic checks, provider payload contract and inspection of remaining overflow/tail and legacy reload paths |
| Fact warrant and withdrawal are not complete at the consumer | SRC-1; MEM-OBJ-item-index, MEM-RTE-recall-history, EPI-RTE-memory-commit | Storage controls and bounded Recall projection | Fully cited later evidence, retrieval completeness, authorized end-to-end correction or contradiction withdrawal | Citation-preserving consumer evidence and traced edit/withdrawal callers, lifecycle and eligibility filters |
| Import, source revocation and index coverage have distinct unresolved semantics | SRC-1; MEM-RTE-session-import, MEM-OBJ-index-cursors, MEM-RTE-index-maintenance | Staged import and selected index bodies/range checks | Native-trace equivalence, semantic completeness, complete freshness/deletion or remote permission propagation | Adapter conversion and role-filter tracing, source deletion/revocation and maintenance recovery inspection, semantic coverage evidence |
| Retained products do not establish all later adoption or delivery | SRC-1; MEM-RTE-matter-state, MEM-OBJ-matter-reviews, MEM-RTE-history-extract, MEM-OBJ-review-archive, WF-RTE-patch-publication | Matter settlement/state-read workflow, generated products and patch publication | Every future review-narrative read, exactly-once notification, extraction adoption, automatic archive-driven revision, parent patch application or semantic approval | Consumer/adoption traces and full cancellation, delivery, artifact and parent integration paths |
| Goal and benchmark judges have bounded evidence and force | SRC-1; EPI-RTE-goal-judge, EPI-RTE-goal-selection, EPI-RTE-eval-verification | Recent text projection and imported reward classification | Independent task verification, demonstrated prevention of premature completion, oracle validity or causal benchmark advantage | Independent outcome evidence, complete recovery/provider wiring, inspected task verifiers and controlled comparisons |
| Assistant heartbeat inventory is incomplete | SRC-1; report verification limits, MEM-RTE-index-maintenance, MEM-RTE-matter-state, RT-RTE-durable-wake | Verifier found ordinary-chat resume/follow-up with history/idle wait, distinct from index workers and Matter checkpoints | Complete proactive-trigger, context-reuse or delivery comparison; universal Matter settlement or index-worker isolation | Accepted route inventory tracing ordinary-session heartbeat state, consumers and delivery alongside the other assistant modes |
| Comparison units are coarser than individual mechanisms | SRC-1; profile verification limits, MEM-RTE-index-maintenance, MEM-OBJ-index-cursors, MEM-RTE-matter-state, MEM-OBJ-matter-reviews, MEM-OBJ-text-checkpoint | Grouped index, Matter and checkpoint authority/admission units | Assigning every grouped value to every part, counting units as independent mechanisms, fine-grained route comparison | Split classification units by retained part, actual consumer and admission mechanism; current existence findings remain supported |
| Remaining included paths prevent whole-system enumeration | SRC-1; MEM-BAP-local-guidance, MEM-RTE-index-maintenance, RT-RTE-child-graph, RT-RTE-plugin-admission | Selected paths, not all Goal/plan/Todo/WorkHub, Feishu, changed guidance/plugin state, Graph edits, native/client/MCP/shell or package-trust internals | Complete memory/curation/admission/learning inventory, all-client equivalence, universal security or recovery; no system-complete theory/reflection/autonomy negative | Accepted producer–admission–maintenance–consumer records for the named alternatives and evidence specific to each assessed property |