Ultimate synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From a project request to separate work contexts
Ultimate is a project-specialized coding application with a browser interface, a standing agent, coordinator branches, bounded subagents and extensible host tools. This account covers its shipped application, prompts, maintenance paths and project-memory loop at Git commit e2b9fa61962f6aecd394dea1c670688fac3ec432, with an applicability cutoff of 2026-11-01. SRC-1 is the frozen HazAT/ultimate repository. The evidence is implementation and doctrine/design, not a successful execution: no target, tests or model calls were run, and no production trace or causal experiment was supplied. The whole-system functional boundary includes the callers that persist and consume state; it does not mean every extension or external dependency was exhaustively inspected (boundary and source register).
The application encloses an external Pi Durable harness. Ultimate owns browser input admission, extension loading, project configuration, active-chapter SQLite storage ownership and process replacement. Pi Durable supplies conversation/task machinery; a model chooses answers and tool calls, and Node-backed executors can affect the host machine. Browser submission returns an asynchronous receipt, not a finished answer. A resumed conversation and persisted history provide continuity, without establishing exactly-once external effects or the correctness of the external scheduler (RT-RTE-browser-input, RT-OBJ-agent-doc).
The representative feature workflow moves from standing Ultimate to a coordinator branch, then optionally to implementation workers and reviewers. It was selected because shipped instructions explicitly prescribe this progression. The coordinator assesses the request, obtains a go under the normal decision policy, retains architecture and final judgment, and delegates bounded work with scope, acceptance commands and an evidence-based return contract. Small fixes deliberately omit workers and review. Hands-off goals can substitute Ultimate's decision for a fresh human choice; assessment and approval instructions are model-interpreted policy, not universal software gates (WF-CLM-feature-protocol, EPI-RTE-goal-authority).
Coordinator branches and named subagents are different context boundaries. A root branch with memory starts from a memory brief rather than the root tally; a branch fork can retain its source context. Named subagents normally receive a self-contained task without parent history, but explicit fork mode inherits selected history. Their final text and model identity return to the parent; detailed test evidence requires separate transcript retrieval. A worktree separates Git work, not filesystem authority. Children and their host effects can outlive the parent's wait or stop (RT-RTE-lineage-branch, RT-RTE-subagent-start, workflow return annotation).
Checking, integration and delivery are separate transitions
Coordinator instructions require reading diffs, citations and the last full test run rather than accepting a worker's DONE claim. Review instructions require an immutable range and acceptance criteria; a reviewer team is asked to work independently and approve only unanimously. These are requested judgments, not an enforced truth oracle or evidence that a review occurred (WF-CLM-feature-protocol).
Worker landing first fast-forwards the coordinator's checkout under Git/state guards; this intermediate operation does not run the final configured check suite. Top-level submission enters a separate lander that prepares a merged candidate, runs configured check groups, permits one rerun of each failed command, rejects twice-failed checks or changed integration state, and serializes the target fast-forward. A failed submission returns repair responsibility to the coordinator. Software submission does not require a stored reviewer APPROVED verdict. Configured checks license conclusions only within their tested domain, and direct shell changes remain outside this gate (WF-RTE-child-land, RT-RTE-change-admission, workflow recovery annotation).
A local landing, optional remote push and user-facing report are distinct outcomes. A push failure need not undo a valid local landing. The lander routes an outcome card to the branch and its starter; Ultimate is instructed to notify the user. This establishes delivery machinery, not actual user receipt or successful feature behavior. Supervised core replacement has its own type-check and storage-owner handover sequence; live extension loading instead evaluates host code before validating registration shape. Neither mechanism is universal semantic validation or rollback (RT-RTE-change-admission, RT-BAP-extension-code, workflow account).
Landed work can change what future work receives
The strongest supported adaptive contribution is a wired route from work history to durable guidance for later invocations. An eligible landing schedules a summary-model task using the landing note, submitter abstract and current full root memory. It proposes add/update/delete operations, defaulting to no change. Software normalizes and structurally validates the batch, can request one repair, and commits admitted text for future branch, lookup and recall consumers. This is an automatic trace-fed durable update; it does not establish faithful extraction, a newly warranted fact or improved future capacity (MEM-RTE-landing-distil, MEM-BAP-memory-guidance).
Root memory is a slug-keyed map of natural-language decisions, facts and lessons with symbolic kind, timestamp and optional source metadata. Ordinary batch admission caps it at 150 items, with bounded line/body lengths. The explicit remember executor checks the standing Ultimate caller, but this is not a write monopoly: internal extension APIs, automatic landing tasks and chapter seed assignment are other writers. Normalization can clip necessary conditions; source syntax checks do not inspect the referenced evidence. Automatic attribution points to the landing, not necessarily each original premise. Structural admission and retention therefore grant operational availability without establishing factual acceptance (MEM-OBJ-memory, MEM-OBJ-memory-addresses, MEM-RTE-memory-edit, EPI-RTE-memory-normalization, EPI-RTE-memory-checks, EPI-RTE-memory-retention).
Read-back is selective. New root branches receive index lines, dates and ten recent landings, not every body or source. An explicit exact-slug lookup returns the fuller item and stored source breadcrumb. Standing Ultimate additionally has message-triggered recall: with a classifier available, known context usage at least 20%, nonempty memory and a new eligible user/branch input, an external classifier sees the incoming message prefix capped at 4000 characters and all memory lines. Code takes at most three known slugs above a 0.25 probability threshold, excluding recent recalls. A non-abort failure or three-second timeout leaves work without recall. Delivered line/body text omits source and update date; probabilities estimate applicability, not truth. Guidance asks the model to act when appropriate but does not compel compliance (MEM-RTE-memory-read, MEM-RTE-recall-message, EPI-RTE-recall-selection, EPI-RTE-recall-injection).
Other continuity mechanisms preserve different projections. Abstracts compress selected branch text; idle catch-up pushes changed work and Git state within an approximately 6000-token budget, prioritizing path overlap without detecting all semantic conflicts. Local ownerless compaction summarizes an edited view with transient entries excluded, while owned-subagent default compaction remains external. Side-chat and optional go-back summaries can return selected work without rolling back files or other effects. Requested history/search and Store loads remain separate retrieval paths. Codemode keeps intermediate results outside model context unless emitted; its nested executors bypass tool hooks, so hook-only safeguards do not automatically cover that route (MEM-RTE-summary-tally, MEM-RTE-tally-catchup, memory compaction annotation, MEM-RTE-rewind-context, RT-RTE-codemode-tools, MEM-OBJ-store).
Review and chapter change afford further adaptation
Memory review and dream instructions ask agents to compare retained content and own-operation records against later evidence, return reasoned proposals and reject unsupported reports. Normally the user chooses after Ultimate weighs in; hands-off goals allow computational substitution. Generic tools afford adoption into memory, code and instructions, but no performed semantic review or evidence-consuming acceptance instance was supplied. A user-controlled chapter transition then deterministically carries selected state into a fresh store: same-slug memories compete by timestamp, open work and eligible automations are retained, and prepared handoff text becomes waking input. This is continuity, not contradiction adjudication. Direct seed assignment bypasses ordinary batch validation, so the ordinary per-root cap is not a universal storage invariant (MEM-RTE-review-memory, EPI-RTE-memory-review, MEM-RTE-chapter-seed).
These contributions support separate, not ladder-like, assessments at the project-specialized application boundary:
| Property | Supported contribution and unresolved conclusion |
|---|---|
| Theory-builder condition 1: localized content | Review/dream guidance is inspectable as a whole procedural unit. Mapping its prescriptive rules to a methodological theory is limited; no retained explanatory theory instance is demonstrated. |
| Condition 2: consumption | Runtime supplies instructions, but no decision trace establishes dependence on what this procedure says. Consumption is uninspected, not absent. |
| Condition 3: content-directed criticism | The procedure requests criticism of named memory content against evidence and changed reliance. Enacted criticism is unobserved. |
| Condition 4: iteration | Chosen changes are intended to shape later work; actual criticism-fed iteration is unobserved. |
| Learning from criticism | Automatic history-derived guidance is wired. Criticism improving capacity for future action is not established; no relevant outcome comparison was supplied. |
| Reflection | Own failures, tools and methods are represented as review targets, affording feedback. Both exercised causal directions remain unestablished. |
| Autonomous theory building | Normal review includes human proposal decisions. Goal-mode substitution affords greater computational control but does not establish every internal theory-building operation. |
| Self-improvement | For fewer evidenced failures and less repeated work across later sessions/chapters, standing routes can install behavior-shaping changes. Exercised membership is not established: no objective-bearing evidence actually shaping a specific update, its installation/live consumer and later causally dependent operation were supplied. Successful improvement is a further unestablished result. |
These distinctions follow the theory assessment in EPI-RTE-memory-review, the afforded procedure in MEM-RTE-review-memory, and the live instruction/code channels in RT-BAP-prompt-guidance and RT-BAP-extension-code. Automatic trace-fed summaries and memory writes remain positive architectural findings even where exercised learning, reflection or self-improvement is unproven; the profile verification preserves that distinction.
For repeated project work with selected configured checks, the supported design is durable coordination plus multiple bounded continuity mechanisms. It is not a demonstrated end-to-end reliability result or a uniformly warranted knowledge base. Traced automatic memory paths do not independently accept factual claims; imported web excerpts retain URL labels but have opaque extraction and bounded output. No candidate evidence establishes checked entailment or a particular non-entailed conjecture. Additional traces linking source evidence, candidate, criticism, adoption and later behavior would change the exercised-property assessment. Independently varied interventions and outcome measures would be needed for component-level benefit attribution. Evidence-consuming memory acceptance, source-preserving delivery, or enforced reviewer prerequisites would change the corresponding warrant and admission assessment, rather than merely adding more retained text (EPI-RTE-web-acquisition, EPI-RTE-landing-candidates, EPI-RTE-memory-checks, WF-CLM-feature-protocol).
Limitations
| limitation | affected SRC IDs and record citations | inspected boundary | conclusion prevented | evidence that would resolve it |
|---|---|---|---|---|
| Implementation is not exercised operation or benefit | SRC-1; RT-RTE-browser-input, MEM-RTE-landing-distil, WF-CLM-feature-protocol | Frozen application source and instructions; no supplied production run or causal experiment | Reliable completion, actual compliance, faithful summaries, realized savings and causal improvement | Inspectable end-to-end traces and scoped outcome comparisons; independent component variation for component attribution |
| External machinery and model state remain opaque | SRC-1; RT-OBJ-agent-doc, RT-CMP-provider-models, MEM-CMP-jev, RT-RTE-context-return | Ultimate integration calls, local producer hooks and configuration, not full upstream harness/provider implementations | External crash consistency, default checkpoint selection, immutable remote weights or complete provider-learning classification | Frozen dependency implementations/contracts and bounded provider or recovery evidence |
| Repair/adaptation inventory is partial | SRC-1; RT-RTE-automation-run, epistemic coverage, report-verification limits | Verifier inspected extensions/automations/supervisor.ts: failed-script delivery asks a watcher to investigate, fix and resume; counters can escalate/disable |
Complete repair-evaluator or behavior-changing script-update comparison; performed repair or improved capacity | Accepted route records tracing supervisor evidence, proposed repair, admission and later consumer, plus execution evidence for operation/benefit |
| Cross-extension context coverage is incomplete | SRC-1; RT-RTE-context-return, memory coverage, report-verification limits | Verifier inspected extensions/voice-live/server.ts: standing Voice can reappend the newest successful brief missing after a head marker, transiently and in the current request |
Complete selector/reload comparison or faithful Voice-content conclusion | Accepted Voice producer/selector/delivery records and scoped content/operation evidence |
| Actual content, arbitrary payloads and archive consumers are incompletely classified | SRC-1; MEM-OBJ-memory, MEM-OBJ-store, MEM-OBJ-call-summaries, MEM-RTE-chapter-seed | Known subsequent local routes; no deployed corpus; arbitrary Store lineage/effects, mixed call-summary derivations and full remote/archive lifecycle unresolved | Complete memory inventory, initialization provenance, cross-chapter retrieval completeness or all-source classification | Provenanced deployed examples and accepted producer-to-consumer records for unresolved branches, including backup/restore and remote reads |
| Consumer projections omit evidence and reasons | SRC-1; MEM-RTE-memory-read, MEM-RTE-recall-message, MEM-RTE-operational-notes, EPI-RTE-memory-normalization | Brief lacks body/source; recall lacks source/date; goal note lacks full decision reasons; automatic text may be clipped | Full provenance or criticism retention at each consumer, semantic preservation, universal withdrawal of obsolete guidance | Consumer payloads retaining checked support/reasons and explicit update/withdrawal propagation, with fidelity evidence |
| Seed capacity consequences are state-dependent | SRC-1; MEM-RTE-chapter-seed, MEM-RTE-memory-edit | Cross-root timestamp selection and direct destination assignment, outside ordinary batch cap enforcement | Universal 150-item storage invariant or an observed over-cap seed and its next-edit behavior | Concrete eligible source/destination state and transition evidence, or a cross-root validation mechanism |
| Review, theory-building and self-improvement remain partly procedural | SRC-1; MEM-RTE-review-memory, EPI-RTE-memory-review, EPI-CLM-dream-learning | Localized review/dream instructions plus generic adoption channels; no actual criticism-to-later-behavior chain | Exercised theory-building conditions 2–4, criticism-induced capacity improvement, exercised reflection, autonomous theory building or exercised self-improvement | Linked criticism, blame, decision, retained update and later causally dependent operation; separate outcome evidence for successful improvement |
| Reported recall metrics are not an inspected evaluation | SRC-1; MEM-CLM-recall-tuning, EPI-RTE-recall-selection | README reports tuning on 55 labelled messages from three chapters; underlying dataset and runs not supplied | Independently observed precision/recall, broad relevance quality, factual correctness or causal benefit | Frozen labelled dataset, scoring procedure and run outputs; separate downstream outcome comparison |
No unresolved report conflict was retained by reconciliation or report verification. Profile verification declares no additional limit. This synthesis made no frozen-source reads and adds no evidence beyond the supplied members.