OpenResearch synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
OpenResearch is a local research workspace that coordinates installed coding agents, experiment branches, asynchronous compute and retained evidence. This account covers the repository's shipped software at commit 2c45ff3183b494aac4001bd44eefbe57a3c249ae, with an applicability cutoff of 2026-10-20. The whole-system boundary includes the desktop/UI callers, standalone CLI, host, adapters, instructions and state consumers. Its evidence is SRC-1 implementation and doctrine/design, not a newly executed workflow or causal experiment. External harness internals, model services and compute-provider implementations are outside the inspected implementation boundary. Included but unassessed paths remain coverage limits, not exclusions or evidence of absent behavior (boundary, Source register).
The shipped assignment spans ideation, literature review, hypothesis formulation, experiment execution and artifacts. The discriminating architecture is the division of responsibility: OpenResearch owns session admission, persistence and process supervision; the nested coding agent owns semantic decisions; the project's program supplies measurements; the foreground model is instructed to judge what those measurements support. Tools enabling that assignment do not establish successful science (RT-CLM-research-assistance, EPI-CLM-results-evidence).
From a project request to an executable experiment
A UI/API message reaches ChatHost, which starts or durably queues the turn and dispatches the selected harness. In the inspected Claude path, a private Git worktree, rendered project playbook and skill folders prepare the session. Native context is resumed where available; missing native context can instead receive a lossy ORX transcript snapshot. The model then chooses tools and text. On return, normalized events and assistant output are persisted and presented to the UI. Startup reconciles unfinished turns and resumes queues; these are implemented continuity mechanisms, not proof of exactly-once external effects (RT-RTE-chat-turn).
The playbook supplies natural-language directions and changing orientation facts, including experiment/run counts, rather than full research evidence. For Claude, rewriting its file each turn does not establish that a resident child rereads it: appended system guidance enters at spawn/respawn. Optional model names do not pin immutable weights, and provider parameter changes remain uninspected (RT-OBJ-playbook, RT-BAP-playbook-injection, RT-CMP-claude-model, RT-CMP-codex-model).
Permissions are route-specific. Claude bridge requests receive token/busy-turn checks and mode-dependent allow, deny or human review. Codex Plan uses a different policy mapping, including workspace-write. A separate shipped operator-shell endpoint runs Bash without consulting model permission mode. This is a reachable human mutation path, not a demonstrated model exploit. Neither Plan nor a private worktree is a system-wide read-only or OS-sandbox guarantee; bridge enforcement also depends on native runtime cooperation (RT-RTE-permission-review, RT-RTE-user-shell).
For literature work, discovery and paper commands import attributed external material. Ranking and source identifiers help locate candidates, not certify their claims. The default alphaXiv report is not original text suitable for quotation; instructions request reading original evidence. OpenAlex token-position reconstruction preserves reading order only under its stated input assumptions. Neither import nor this reshaping establishes new accepted scientific knowledge (EPI-RTE-literature-import, EPI-RTE-abstract-reshape).
For execution, the agent/user proposes research code and a run command. orx exp run selects the experiment branch's committed tip, creates a content-addressed source archive, validates launch settings and reserves a run. The default concurrent-run guard can be deliberately relaxed with --force; source selection still applies. The inspected local controller extracts the snapshot and launches the command with configured environment, while detached supervision maintains status and mirrors logs. Standalone dispatch also reaches this submission path. The snapshot binds recorded code, not data, dependencies, secrets or services; arbitrary shell commands do not inherit this compute contract (RT-OBJ-source-snapshot, RT-RTE-experiment-run).
From completion to an evidence-bearing answer
A representative source-supported workflow is a locally configured request to run and summarize an experiment: inspect baseline, create hypothesis-relevant child branches, commit, submit, await completion, inspect evidence and return a cited explanation. It is a scenario reconstructed from connected shipped responsibilities, not an observed session. Compute delegation is to a controller/process, not another reasoning agent (workflow account).
Submission returns a run identity, not a measured answer. A prior explicit wake subscription can later cause the host to claim an idle session and start a hidden turn. That turn receives run/status text and directions to inspect evidence, not full measurements. Waiting likewise supplies a completion signal; instructions require rereading run states to reconcile other completions (RT-RTE-run-wakeup, workflow account).
The log command returns a full-file locator and a short tail: at most 500 characters derived from a 2048-byte suffix. The next evidence step is an explicit targeted file read. Preview omission cannot establish that a metric is absent from the log. The instructed reporting check asks for variant/effective configuration, final metrics and summary, long-run trajectory and supporting citation lines. This is semantic guidance to the model, not a software gate preventing unsupported reports. No recorded evidence-consuming acceptance event was established (MEM-RTE-log-retrieval, EPI-RTE-results-check, WF-RTE-result-handoff).
The instructed next-step policy distinguishes repair, refill, promote and stop. An unintended failure normally leaves the question unanswered, but an OOM may answer a memory hypothesis. Answered nodes should remain frozen and changes should branch; repeated unanswered repairs require user escalation. These are requested model judgments, not enforced scientific counters or an expected-answer oracle. Carrying a winning branch forward can support local optimization without establishing an explanation of why it won or transfer beyond the comparison (EPI-RTE-winner-promotion, workflow account).
Reports become project artifacts immediately on filesystem write. Descriptions likewise retain caller text without scientific approval. Tree moves change comparison organization without rebasing branch code; archive flags hide or restore nodes rather than refuting their claims. The chart helper does enforce numerical/schema conditions before its own new write, but fabricated finite values can pass and arbitrary writers bypass it. Thus structural validation, retention and scientific acceptance have distinct domains; none establishes post-acceptance knowledge integration here (MEM-RTE-artifact-access, EPI-RTE-description-store, EPI-RTE-tree-organize, EPI-RTE-chart-render).
What persists and can change later work
OpenResearch's strongest supported trace-fed contribution is retained continuation context. Fallback compaction automatically turns recorded work and prior bootstrap material into a summary, persists it, clears the native session ID and supplies that seed to a later ordinary turn without native context. Persisted selected-path ACP/legacy side-chat seeds provide another continuation path. These implemented trace-fed writes can shape later work; they do not demonstrate improved capacity for future action (MEM-RTE-chat-compaction, MEM-RTE-branch-context, MEM-BAP-bootstrap-injection, profile verification).
Three stores must remain distinct: ORX transcript rows, native session payloads and bootstrap prose. Fallback compaction consumes an all-branch, newest-32-KiB flattened snapshot plus earlier seed; full tool bodies have already been omitted. Side-chat preparation instead selects a fixed parent active path. Native compaction preserves a native session through provider-specific mechanisms whose contents and budgets remain opaque. Keeping transcripts does not guarantee their full return to a model window, and requesting preservation of reasons does not guarantee that reasons survive (MEM-OBJ-chat-messages, MEM-OBJ-native-session, MEM-OBJ-bootstrap-context, MEM-RTE-chat-compaction).
Ordinary composition supplies retained goals, gated bootstrap and selected recent shell exchanges; slash-selected skills enter as instructions, while shell output and selected excerpts are labeled untrusted data. prepared_input bypasses ordinary composition. Side chats have separate conversational state but share a worktree; documented independent chats instead have no automatic result return. Neither establishes a universal native-subagent isolation contract (MEM-RTE-turn-context, MEM-RTE-branch-context, MEM-CLM-independent-chat).
Reusable skill upload admits structurally valid content and distributes managed copies for later use; it does not certify quality or guarantee immediate resident-harness reload. Verification also found persisted model screening of native catalog entries and machine-wide compute/CUSTOM.md guidance outside the detailed analyst inventories. Uploads bypass that catalog screening; custom compute guidance has implemented revision-checked replacement but only instructed later retrieval. These omissions prevent a complete instruction/adaptation inventory, not the supported summary and upload findings (RT-RTE-skill-upload, report verification limits).
Theory-building properties remain separate. In the assessed experiment-policy route, condition 1, localized theory content, is afforded by whole descriptions/code; condition 2, content-dependent consumption, is claimed; condition 3, content-directed criticism assigning blame and changing reliance, is uninspected; condition 4, iteration shaped by the result, is claimed. Winner selection alone supplies no formulated criticism. Whether criticism improved future capacity is unestablished. Reflection on OpenResearch's own operation and autonomous computational performance of every internal theory-building role are also unestablished, not inferred absent. At this boundary, no exercised self-improvement pathway is established: research-code optimization concerns an external product, summary delivery does not prove later causal dependence or improvement-directed self-change, and skill/release update paths alone do not close those links. These limits coexist with implemented retained-context updates (EPI-RTE-winner-promotion, MEM-RTE-chat-compaction, RT-RTE-skill-upload, RT-RTE-release-update).
For the local experiment scenario, the supported result is a connected architecture for committed-code execution, asynchronous continuation and evidence-oriented reporting instructions. Exercised reliability, scientific acceptance and improved research capacity remain unresolved. Traces linking source, measurements, actual checks, dispositions and subsequent decisions would change the operation assessment. Controlled comparisons would be needed for benefit or component attribution. An enforced claim-admission mechanism would change the reporting guarantee; traced evidence-responsive updates to OpenResearch's own organization, followed by causally dependent operation, would change the self-improvement assessment.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Implementation and instructions are not exercised research or benefit. | SRC-1; RT-RTE-experiment-run, EPI-RTE-results-check, WF-RTE-result-handoff | Frozen shipped local workflow; no execution/causal layer assessed | Successful end-to-end research, actual acceptance, theory criticism, reliable compliance or causal improvement | Linked execution/check/acceptance/next-decision traces; controlled comparisons for benefit and attribution |
| Harness and backend contracts remain partly external or unassessed. | SRC-1; RT-RTE-permission-review, RT-RTE-user-shell, RT-RTE-experiment-run, RT-CMP-claude-model, RT-CMP-codex-model | Sampled Claude/Codex controls and local controller | Universal permission, tenant-isolation, backend-parity, immutable-model or full reproducibility guarantees; operator shell is not a proved model bypass | Complete native enforcement and backend traces/contracts; model identity and dependency/data provenance |
| Native context and summary fidelity remain unresolved. | SRC-1; MEM-OBJ-native-session, MEM-RTE-chat-compaction, MEM-OBJ-bootstrap-context | Native resume/compact bridges and known fallback chain | Complete payload/form/provenance, preserved rationale, native budgets or erasure; full restoration from transcript display | Native checkpoint/selector inspection and before/after context evidence with later consumer traces |
| Additional initialization, provenance and lifecycle paths are untraced. | SRC-1; MEM-OBJ-bootstrap-context, MEM-OBJ-project-artifacts, MEM-OBJ-run-log, RT-RTE-skill-upload | Demo/legacy seeds, arbitrary artifacts, discovered skills, remote logs, literature-cache and platform alternatives in the accepted inventory | Complete memory lineage, trace-source, storage, recovery or forgetting inventory; demo-route authentication | Producer/admission/maintenance/later-consumer records for each unresolved alternative |
| Context separation is not a complete child lifecycle contract. | SRC-1; MEM-RTE-branch-context, MEM-CLM-independent-chat | Side-chat selection and independent-chat doctrine; native subagents uninspected | Universal child permission envelope, filesystem isolation, inherited context or automatic subordinate return | Spawn, child input/state, permission, cancellation and return-path implementation and traces |
| Literature and rendering checks have limited warrant. | SRC-1; EPI-RTE-literature-import, EPI-RTE-abstract-reshape, EPI-RTE-chart-render | Client acquisition/reshaping and chart helper, not external services or all renderers | Source truth, complete corpus coverage, universally faithful extraction or scientifically valid plotted data | Original-source comparisons, service/evaluator provenance and independent measurement checks |
| Native catalog screening is omitted from the detailed instruction inventory. | SRC-1; RT-RTE-skill-upload, MEM-RTE-turn-context; verification limit | Verifier inspected persisted import decisions in src/local/user_skills.rs and catalog refresh in src/commands/up.rs; uploads bypass screening |
Every catalog admission being structural/human selection; complete instruction-selector inventory; screening proving quality or improvement | Records tracing screening inputs, persisted decisions, refresh and actual downstream loading, with outcome evidence for quality |
| Machine-wide compute guidance is omitted from the detailed retained-guidance/workflow inventory. | SRC-1; MEM-RTE-turn-context, WF-RTE-result-handoff; verification limit | Verifier inspected src/commands/compute.rs storage/revision-checked replacement and agent-skills/orx-compute/SKILL.md retrieval instruction for compute/CUSTOM.md |
Complete cross-project guidance, trace-origin or behavior-adaptation comparisons; exercised trace-fed update or self-improvement from this mechanism alone | Authoring/provenance and later retrieval/consumption records, followed by evidence of actual dependent behavior |
No source files were read for this synthesis. It uses the handed accepted members and their retained passages. The report verification declared the two inventory limits above; profile verification declared no additional limits. No unresolved cross-report conflict was supplied.