Spacebot synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Spacebot is an enclosing runtime for persistent conversational, delegated and background agent work. It owns message admission, model context, tool grants, worker supervision, retained state and outbound routing. This account covers its shipped functional loop at commit 5c2ee47b397e09c956dc15000ea576ecd28e92bb, with an analysis cutoff of 2026-10-14. The evidence is source implementation and doctrine/design from SRC-1, not execution: no target tests, model calls, deployment traces or causal experiments were supplied or run. Whole-system scope does not mean exhaustive file inspection. Provider internals, external OpenCode and remote services are outside the frozen implementation boundary (boundary, Source register).
From request to delegated effects
An inbound message is resolved to an agent and conversation channel. Code constructs the context, selects model routing and exposes tools; the channel model chooses its next action within those grants. Effects can include replies, task changes, filesystem work, subprocesses and remote calls. Branches receive copied conversation context and return conclusions. Built-in workers receive their own preambles, tools, task context and clean or forked history. Required skills supply full procedure bodies; suggested skills supply hints. Neither delivery path proves obedience (RT-RTE-channel-turn, RT-RTE-delegated-work, RT-CLM-required-skills).
The representative workflow is task-bound, noninteractive built-in worker execution. It exposes three different admission decisions. First, spawning checks authoritative task status, live attempts and required-skill availability, then claims an attempt. Ready or InProgress can proceed; PendingApproval, Backlog and terminal tasks cannot. Second, dispatch attempts to commit a terminal outcome before emitting its normal completion event. The receiving channel rereads that durable result and checks versions, kind and owner before queuing it. Third, the result retriggers the channel model, which decides what to relay through permitted outbound paths. These checks establish execution permission and lifecycle consistency, not satisfaction of the user's requirements (RT-RTE-delegated-work, WF-RTE-completion-admission).
Recovery is local rather than universal. A timeout can preserve available partial output; terminal commits have bounded retries; some retrigger failures retain pending results. Exhausted commits can suppress notification, other queue failures can drop pending results, and a model can skip relay. Completion, queue acceptance, outbound routing and actual user receipt are different milestones. No workflow execution or delivery receipt establishes their reliability here (workflow account).
Tool topology is not global containment. Direct channels and autonomy Act can expose execution tools outside workers. Task approval applies to task-bound execution, not every unbound, reference-only or direct action. OpenCode is started without Spacebot's sandbox wrapper, and its permission events receive automatic allowance rather than human veto. Configured sandboxing can fall back to unsandboxed subprocess execution. These are implemented scope limits, not observed exploits (RT-CLM-role-grants, RT-RTE-opencode-work).
What survives and reaches later work
Spacebot retains distinct surfaces: typed claim text and selectors in per-agent SQLite, association graphs, dense retrieval indexes, temporal event narratives, durable conversation text, chronicle checkpoints, filesystem procedures and identity text, and instance-wide wiki versions. These parts do not have interchangeable authority. Prose informs models; fixed code interprets selectors and lifecycle state (MEM-OBJ-memory, MEM-OBJ-memory-metadata, MEM-OBJ-retrieval-index, MEM-OBJ-wiki-page, MEM-BAP-memory-selectors).
Automatic context is a limited view. Global memory rendering selects first lines under fixed type, importance, recency and budget rules. Participant context separately matches human keys; temporal context supplies budgeted narratives and event tails. Explicit recall uses hybrid or metadata selection and returns deeper content to eligible consumers. In standard channel operation a branch mediates raw memory recall; direct and other configured processes can recall themselves. Requested recall and unsolicited context supply are separate routes (MEM-RTE-context-supply, MEM-RTE-memory-recall).
Context reduction also changes what remains recoverable. Rolling compaction leaves a session-live summary and recent tail. Chronicles retain durable summaries or explicit unsummarized coverage markers, then trim selected prefixes. Branches and workers keep private context out of the parent and return selective results. Durable conversation expansion reconstructs text, not original tool-call structures. A stored checkpoint or identifier therefore does not preserve the entire former model window. Failure markers and emergency cuts must not be counted as successful semantic summaries (RT-RTE-context-control, MEM-OBJ-conversation-log, MEM-RTE-chronicle-retention).
Retention checks do not supply a truth oracle
Conversation persistence and file ingestion ask models to select useful statements. Save handlers validate structure and pipeline outcomes. The terminal contract compares reported IDs with successful saves and governs process completion; it does not check truth or extraction completeness. Model-produced statements may preserve, distort or extend their inputs; no actual input/output comparison establishes which occurred (MEM-RTE-memory-admission, MEM-RTE-file-ingestion, EPI-RTE-completion-contract, EPI-RTE-memory-extraction).
Maintenance changes later reliance without necessarily judging propositions. Write-time consolidation advice arrives after saving; explicit batches replace survivors and withdraw absorbed rows. SQL commitment precedes warn-only vector upkeep. Similarity associations and contradiction weights alter ranking, not truth adjudication. Forgetting excludes ordinary retrieval but does not retract copied prompts or derived procedures. Wiki versioning preserves edits; its verification instruction is not an enforced evidence gate (MEM-RTE-memory-curation, EPI-RTE-memory-selection, EPI-RTE-wiki-mutation).
Two source deductions show why deterministic processing alone cannot guarantee preserved meaning. First-line rendering can omit a necessary condition on a later line. Substring-based merging or byte clipping can suppress distinct content. These are code-supported counterexamples to general guarantees, not observed incidents. Accordingly, public claims of staleness-free context and write-time duplicate merging support narrower conclusions: a current store view and an advisory consolidation path, not current truth, complete content or unconditional merge-on-save (EPI-RTE-memory-render, EPI-RTE-memory-maintenance, EPI-CLM-memory-context, EPI-CLM-write-consolidation).
Retained adaptation, reflection and autonomous work
The strongest supported improvement-related contribution is a wired, trace-fed procedure-change path. Eligible persistence branches can inspect conversation and worker records, decide whether a reusable lesson exists, write or patch skill files, reload them and supply procedures to later processes. The objective is reusable procedure improvement that avoids repeated procedural errors. Model abstention and structural/origin checks can reject changes, but there is no enforced procedural-correctness oracle. Agent-origin pin/read protections apply to that mutation protocol; User-origin, installation, API and generic file/shell paths have different controls (RT-RTE-skill-change, RT-BAP-skill-instruction).
Conversation extraction, temporal synthesis and continuation summarization likewise wire automatic trace-fed retained outputs to later consumers. This is a supported adaptation capacity, including non-parametric changes, not demonstrated improved future capacity. Retaining summaries or procedures does not by itself show that criticism of a consumed theory improved future action (MEM-RTE-temporal-synthesis, MEM-RTE-chronicle-retention, RT-RTE-skill-change).
The procedure route's independent assessments are:
- Theory-builder condition 1, localized content: afforded architecturally through whole procedure bodies and replacement; particular formulated theory content is uninspected.
- Condition 2, content-dependent consumption: operation uninspected. Instruction supply is wired, but dependent decisions are not observed.
- Condition 3, content-directed criticism: claimed in reflection guidance. Formulated criticism with blame placement is not demonstrated.
- Condition 4, iteration: operation uninspected. Reload permits another round, but no retained criticism is shown shaping it.
- Reflection: the architecture wires records of its own operation into a writer that can change later instructions. Exercise and quality of both causal directions remain unobserved.
- Self-improvement: improvement-directed self-change capacity is supported at the skill surface across sessions. No supplied run closes the four links from objective-bearing evidence to a particular update, accepted self-change, live consumption and causally dependent behavior. Exercised or successful self-improvement is not established.
These conclusions follow the separate assessments in RT-RTE-skill-change, not an inference from one property to another. Autonomous operation has its own wired route: a supervisor admits resident-channel epochs under the lesser of agent level and instance ceiling, pause and scheduling/wake conditions, then supplies briefings and level-dependent tools. This supports autonomous work capacity, not observed useful work, reliable criticism or an autonomous theory builder (RT-RTE-autonomy-epoch).
For persistent conversational assistance, the supported architectural contribution is continuity through selective context and delegated work. For task delivery, permission and durable-result consistency are stronger than semantic verification or receipt guarantees. For procedure improvement, source wiring supports a candidate self-change mechanism, while the public claim that use makes the system better remains untested. The shipped behavioral fixtures evaluate first-action routing with a four-of-five threshold, not final task correctness or the effect of a changed procedure (EPI-CLM-better-use, EPI-RTE-behavioral-fixtures).
A retained before/after procedure, its source evidence and criticism, subsequent prompt consumption and dependent task outcomes would materially change the learning assessment. An intervention varying that procedure against an appropriate comparison would support attribution at that level. Failure/restart traces and delivery receipts would change the workflow assessment. Implemented semantic acceptance, derivative withdrawal or deployment-wide containment would alter different conclusions and would require their own route evidence. No single missing evaluator or trace decides all these properties.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No observed operation or causal comparison | SRC-1; RT-RTE-channel-turn, RT-RTE-skill-change, EPI-RTE-behavioral-fixtures | Frozen implementation, prompts and evaluator definitions; no execution evidence | Production reliability, exercised theory criticism, useful learning, beneficial self-improvement or component effect | Provenance-linked runs and artifacts; scoped outcome evidence and controlled comparisons for attribution |
| Human filtering and cached participant refresh, carried from report verification | SRC-1; MEM-RTE-memory-recall, MEM-OBJ-human-anchor; report limits | Recall schema/parser and session anchor cache | Complete human-filter coverage or guaranteed immediate refresh after writes/org edits; general recall and participant rendering remain supported | Matching parser support and traced cache invalidation/refresh across relevant mutation paths |
| Saved-attachment read-back is omitted from declared inventory, carried from report verification | SRC-1; MEM-OBJ-conversation-log, MEM-RTE-chronicle-retention, RT-RTE-channel-turn; report limits | Verification inspected conditional attachment registration and bounded text-file delivery versus image/path metadata; acquisition and all payload consumers remain incomplete | Complete multimodal recovery or complete read-back inventory comparisons; text hydration still does not reconstruct original tool structures | Accepted acquisition, storage and later-consumer records for text/image/binary alternatives, plus recovery traces |
| Installation agency is conservatively grouped, carried from profile verification | SRC-1; RT-RTE-capability-admission; profile limits | Model-tool installation arguments lead to software download, persistence and reload without per-content human approval; the profile does not separately classify that automatic branch | Mechanism-complete comparisons of installation admission against administrative and generic filesystem writes; the automatic existence finding remains valid | Separate classification units grounded in records for model installation and other caller-controlled paths |
| Recovery, indexing and withdrawal guarantees remain local | SRC-1; MEM-RTE-memory-curation, MEM-RTE-file-ingestion, MEM-RTE-wiki-editing, WF-RTE-completion-admission | SQL/index sequencing, chunk processing, archive/read differences and selected built-in worker completion path | Cross-store atomicity, complete derivative retraction, universal effect rollback or guaranteed delivery | Coordinated index/retraction mechanisms, complete failure-path records and crash/delivery evidence |
| Available detail and external internals remain unresolved | SRC-1; boundary, memory inventory, epistemic coverage, RT-CMP-spacebot-model, MEM-CMP-embedding-model | Selected source bodies; incomplete org propagation, auxiliary task/goal/wake revisions, restored-worker alternatives, API mutations and profile-card behavioral consumption; external implementations excluded | Complete memory/authority enumeration, system-complete negative about acceptance, immutable model identity, security certification or dependency-internal conclusions | Additional accepted source-grounded records for included routes; separately frozen dependency evidence where needed; deployment-specific configuration and traces |