Claude Workstream Kit synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Establishing a durable work boundary
Claude Workstream Kit is a repository-local work-memory system for Claude Code, not an enclosing model runtime. It preserves goals, tasks, decisions, findings and resume scope across sessions. This analysis covers its shipped installer, instructions, hooks, state utilities and evaluators at commit 15705c38a4fffb206b868b3e6bc5ffb4fd8b9d64, with cutoff 2026-10-15. The sole source, SRC-1, supplies inspected implementation, design instructions and attributed reports of earlier operation. No target code, tests or providers were run, and no live consumer trace or causal experiment was supplied. The code-grounded boundary therefore supports executable predicates and prescribed workflows, not successful deployment or reliable model compliance. Claude Code, native goals, compaction, account memory and provider behavior remain external (boundary, Source register).
Installation copies skills, agents, scripts and hook registrations and seeds missing state. It admits instructions and executable capabilities, not just reference content. Drift checks can refuse some overwrites, but force/non-Git exceptions and incomplete settings integration without jq prevent an unconditional deployment guarantee. Creation then asks a foreground model to interview the user, obtain approval of exact state wording, commit it and stop; approval of creation is not permission to begin work (RT-RTE-install-upgrade, RT-RTE-state-lifecycle).
Resuming and executing bounded work
The central checkpoint has two different roles. ACTIVE.md carries symbolic selectors and interpreted Now/Next/Blockers prose. A workstream file carries plans, completion evidence, Decision reasoning, Learnings and questions alongside parser-visible IDs, checkboxes and markers. These are derived work records rather than raw transcript copies. Their prose can guide later planning; their structure can route reads and transformations. Neither role proves the recorded claims true, and the pointer deliberately does not claim whether a native goal is armed (MEM-OBJ-active-selector, MEM-OBJ-active-prose, MEM-OBJ-workstream-index, MEM-OBJ-workstream-content).
SessionStart automatically prints the whole pointer plus selected metadata, roster and conditional warnings. It does not inject the full workstream. Active identity matching selects active detail; broader availability, age and size predicates supply coarse signals. The work skill separately requests detailed state and current artifacts before deriving a bounded critical-path task. Goal arming remains the user's action. A 65,536-byte warning and context telemetry inform budgeting but do not implement truncation or automatic compaction (RT-RTE-session-reload, RT-RTE-work-goal, MEM-OBJ-context).
The foreground model retains sequencing, scope and acceptance. It can send a scout a read question, a worker an objective/file-scope/check/stop packet, and a fresh verifier the same specification. Workers can attempt an in-scope repair twice, then return failure; missing scope or out-of-scope needs return to the orchestrator. Verifier PASS/FAIL is evidence for foreground admission, not an enforced veto. Selected packets and compact returns limit foreground context exposure, while the external host owns actual context isolation. Read-only agent policies and declared tools do not establish an OS sandbox. Shared-visible or irreversible actions require a user stop unless specifically authorized, but that is a model-facing policy, not a universal shell-write filter (RT-RTE-agent-packet, RT-RTE-work-goal, workflow).
Checking, retaining and reducing state
Routine capture can admit new Decisions, Learnings and resume scope without per-item human approval. Substantive restructuring and closure have separate user gates. Review asks whether dated premises remain true against current artifacts, and extraction checks whether a claimed application actually reached the file that governs it. An installed or generated copy is only a provisional destination until its governing source changes. Handoff likewise separates sender destination/visibility approval from receiver verification and action/backlog/decline. Retention, permission and semantic acceptance are distinct operations (RT-RTE-state-lifecycle, EPI-RTE-still-true, EPI-RTE-disposition-check, MEM-OBJ-handoffs).
The strongest implemented checks concern file form. Marker measurements establish counts under a grammar, not that a Learning reached its destination. Standard rewrites can reject heading/checkbox changes, while Learning condensation retains selected statement, target and disposition sentences. Those predicates do not preserve every qualification: the epistemic report deduces a conforming case in which a separate Linux-only qualifier disappears. This is a source-based deduction, not an observed loss. Rotation intentionally removes completed structure and uses different checks. A later EXTENDED protects an entry from Learning condensation, but shared scoring and rotation can still treat its earlier terminal marker as terminal. Direct edits bypass the script guards (EPI-RTE-marker-measure, MEM-RTE-learning-condense, report verification, MEM-RTE-state-rotate, MEM-CLM-marker-reopening).
Extraction precedes trimming so that durable substance can reach governed artifacts before leaving live state. Closure additionally requires per-criterion evidence and user approval, then prescribes committing the done record and ledger, tagging it, deleting the directory, conditionally resetting ACTIVE.md and committing again. History is recoverable only if the required commits and faithful snapshots exist; rotation checks tag existence, not equality to current bytes. Removed detail requires requested Git retrieval rather than automatic reload. Publication of tags or cross-machine state is a separate decision (RT-RTE-state-lifecycle, MEM-OBJ-archive, MEM-RTE-state-rotate).
What the representative scenario establishes
One shipped driver connects three model sessions: create a greeting-script workstream, implement two tasks with requested worker/verifier packets, then conditionally approve closure and archive. It checks the greeting, smoke test, held user gate, evidence-shaped notes, tag snapshot, ledger, deletion, pointer and clean tree. A recorded hold yields HELD rather than archive success; failures take precedence. Delivery in this scenario is local committed work and resolvable history, not merely a final answer (RT-RTE-acceptance-driver, workflow).
For that narrow scenario the driver provides a concrete outcome oracle. It does not independently prove every requested delegation or every citation's substance. Its calls bypass interactive permissions, its prompts supply approvals in advance, and Session B substitutes for rather than exercises native goal arming. Saved logs and fixture history afford maintainer investigation but do not automatically update kit source. Accordingly, the design's grounding and text-only gate-safety claims have protocol and fixture support, not observed all-constraint compliance or causal attribution to skill text (EPI-CLM-grounded-claims, EPI-CLM-gate-validation, RT-RTE-acceptance-driver).
Experience-driven change and assessment boundaries
The strongest supported contribution toward learning is an explicit experience-to-durable-guidance pathway: capture derives checkpoints from session findings, and extraction can change governed standing instructions for later readers. This is prescribed capture and afforded adoption, not just storage of unused raw logs. Automatic model admission and explicit human gates coexist. The profile verification preserves positive trace-fed update findings at those bases without equating them with improved capacity or parameter learning (EPI-RTE-capture-synthesis, MEM-BAP-artifact-adoption, RT-BAP-skill-guidance, profile verification).
The unresolved question is whether criticism of a consumed account changes future capability. The records do not identify an actual theory instance establishing separately localized content, consumption, content-directed criticism and iteration—the four theory-builder conditions. They also do not establish improved future capacity from such criticism. Bounded computational work is afforded, but human decisions remain internal to the prescribed lifecycle; autonomous theory building is not demonstrated. Work-state self-description and revision alone do not establish reflection of the kit's own organization. These are independent evidential limits, not a combined negative grade (EPI-RTE-capture-synthesis, EPI-RTE-still-true, RT-RTE-state-lifecycle).
For the kit's own self-change, governed guidance adoption is a possible route, but no supplied episode connects objective-relevant evidence to a kit-source update, its live instruction consumer and subsequent behavior causally depending on that update. Exercised self-improvement and successful improvement therefore remain unestablished. The report verifier also found additional release-round and artifact-fitness doctrine outside the separately traced maintenance inventory; this prevents an exhaustive architectural verdict, not the positive adoption finding. A retained before/after episode with criticism, governing-source changes, later consumption and objective-relative outcome evidence would materially change this assessment. Independent variation would be needed to attribute an effect specifically to a component (MEM-BAP-artifact-adoption, RT-RTE-acceptance-driver, report-verification limits).
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Maintenance-route inventory is partial; release-round feedback/fix/retest and current-standard artifact-fitness review add improvement-directed doctrine. | SRC-1; MEM-BAP-artifact-adoption, RT-RTE-acceptance-driver, epistemic conclusion, report-verification limits | Accepted records plus verifier's inspection of README.md and docs/design.md at the frozen commit | Exhaustive comparison of kit-maintenance or iterative improvement architectures; no inference of absent improvement-directed design | Separate records tracing those protocols, their admission and subsequent consumers; execution evidence for exercised effects |
| Clear-event contract is unresolved: capture says no hook, while SessionStart commentary includes clear. | SRC-1; RT-RTE-session-reload, memory integration issues, report-verification limits | Registered hook and capture instructions; external host sequencing excluded | Guaranteed automatic post-clear injection | Versioned host event contract and trace showing delivery after clear |
| No supplied live consumer execution or causal comparison. | SRC-1; boundary, RT-RTE-work-goal, RT-RTE-agent-packet, RT-RTE-acceptance-driver | Source code, instructions and attributed earlier operation only | Successful deployment, compliant gates/delegation, reliable retention and safety/effectiveness attribution | Inspectable deployment/session artifacts; appropriate interventions and comparisons for causal claims |
| Native memory, goal evaluation, compaction, actual tool grants and model identity/update behavior remain external or uninspected. | SRC-1; RT-CMP-claude, RT-RTE-work-goal, RT-RTE-session-reload | Kit integrations, not host/provider implementation or consumer configuration | Complete host-memory account, exact reconstructed context, pinned model behavior or universal permission enforcement | Host/provider contracts, deployed configuration and relevant context/control traces |
| Preservation checks are path-specific and syntactic; history fidelity depends on caller commits/tags. EXTENDED has different consumers. | SRC-1; RT-CLM-rewrite-preservation, MEM-RTE-learning-condense, MEM-RTE-state-rotate, MEM-CLM-marker-reopening, report verification | Inspected standard condensation, Learning and rotation paths; direct edits bypass them | Universal semantic preservation, reopened-entry retention under rotation or unconditional full-text recovery | Scope-sensitive semantic checks, consistent reopening rules and verified pre-rewrite snapshot fidelity |
| Initial notes/PROJECT provenance and admission, destination-specific effects, completion-note shapes and mixed trace origins remain unresolved locally. | SRC-1; MEM-OBJ-notes, MEM-OBJ-project, RT-RTE-rewrite-state, MEM-BAP-artifact-adoption, profile verification | Accepted retained-memory inventory; some branches only partly inspected or consumer-dependent | Complete provenance, agency, curation, authority or trace-source enumeration | Branch-specific records and concrete consumer initialization, input and admission evidence |
| Candidate synthesis, semantic acceptance, governed adoption and improved capacity lack a connected instance. | SRC-1; EPI-RTE-capture-synthesis, EPI-RTE-still-true, EPI-RTE-disposition-check, MEM-BAP-artifact-adoption | Prescribed capture/review/extract pathways and afforded later guidance use | Accepted ampliative knowledge, exercised theory-building, criticism-caused learning or exercised self-improvement | Linked inputs/output, evidence-consuming acceptance, criticism/revision records, governing update, later consumer and objective-relative outcomes |
| Remaining CLI handlers and offline suites are not comprehensively audited. | SRC-1; epistemic coverage, memory coverage | Selected parser/rewrite paths and source searches; no test execution | Exhaustive detection claims or a system-complete negative about other checking routes | Targeted inspection of unresolved handlers and suites with records for consequential additional routes |