Claude Workstream Kit report verification

Type: agentic-system-analyses/types/agentic-system-verification.md

Verification

The pinned reports support a bounded account of file-backed work continuity and model-mediated maintenance. I checked the declarations, source anchors, status choices, annotations, containment and reconciliation against SRC-1, not merely the analysts' check reports. The frozen checkout HEAD matches the boundary. No target code, test, example or provider was executed. The supplied structural check says none; all four correction-answer inputs are empty, so there are no correction arguments to adjudicate. The boundary remains supported: shipped installer, prompts, hooks, script clients and evaluator are inside the target, while host model execution and native goal/compaction semantics are not attributed to the kit.

Runtime and workflow

I traced settings to both hooks, the complete SessionStart and Stop implementations, work/create/capture/review/extract/close/handoff instructions, agent definitions, installer deployment and seed writes, upgrade instructions, and the complete three-session evaluator. These support RT-RTE-session-reload, RT-RTE-stop-gloss, RT-RTE-install-upgrade, RT-RTE-work-goal, RT-RTE-agent-packet, RT-RTE-state-lifecycle and RT-RTE-acceptance-driver. SRC-1, .claude/settings.json, .claude/hooks/session-start.sh, .claude/hooks/bare-id-guard.py, install.sh, .claude/skills/workstream-work/SKILL.md, tests/va3-acceptance.sh.

The consequential distinctions survive inspection: automatic delivery prints the pointer and signals, not the full workstream; user gates are instructions rather than filesystem enforcement; agent tools are not a sandbox; the Stop hook has explicit fail-open and repeated-stop paths. Installer force/non-Git exceptions and missing-jq handling prevent an unconditional overwrite or integration guarantee. Agent definitions and evaluator calls do not pin model versions or expose provider parameter updates, consistent with RT-CMP-claude. Installed policy and retained planning state have different consumers and force, as stated in RT-BAP-skill-guidance and RT-BAP-resume-state.

The workflow account faithfully follows one connected shipped scenario. Session B requests delegation but its artifact assertions do not prove delegation occurred. Session C supplies approval conditionally in advance, unlike the ordinary interactive gate. The HELD predicate bypasses archive checks, not deliverable/clean-tree checks; failure takes precedence over HELD. Local tagging and retained logs are not remote delivery or automatic upstream adoption. These are source-level findings, not observed successful sessions.

Retention, transformations and epistemic force

The pointer and workstream containers RT-OBJ-active and RT-OBJ-workstream are supported by the rule and actual consumers. Their declared parts are valid: MEM-OBJ-active-selector has fixed selection consumers, whereas MEM-OBJ-active-prose supplies interpreted context. MEM-OBJ-workstream-index, MEM-OBJ-workstream-content, EPI-OBJ-workstream-claims, EPI-OBJ-workstream-markers and EPI-OBJ-workstream-guidance select overlapping but nonidentical parts or consumption roles. None substitutes for the valid container. SRC-1, .claude/rules/workstreams-rule.md, .claude/scripts/workstream_state.py, .claude/skills/workstream-review/SKILL.md.

I checked the separate file/metadata products MEM-OBJ-notes, MEM-OBJ-project, MEM-OBJ-handoffs, MEM-OBJ-archive and MEM-OBJ-context against their producers and consumers. Notes disposition, PROJECT's explicit selection list, receiver filing before deletion, tag retrieval and telemetry selection are distinct mechanisms. A history locator is not automatically delivered history; telemetry is not a semantic summary. SRC-1, .state-seed/PROJECT.md, .claude/skills/handoff/SKILL.md, .claude/skills/workstream-close/SKILL.md, .claude/scripts/status-line.sh.

Direct inspection of marker measurement and self-check fixtures supports EPI-OBJ-record-learnings and EPI-RTE-marker-measure. Their entailment is grammar-relative, not a license for destination truth. Capture, review, extraction and status instructions support EPI-RTE-capture-synthesis, EPI-RTE-still-true, EPI-RTE-disposition-check and EPI-RTE-status-statement. Separating their checking, retention, permission and disposition functions is warranted. No actual input/output pair warrants classifying capture synthesis as ampliative or entailed. Requested governed-file adoption supports MEM-BAP-artifact-adoption, not demonstrated improvement. SRC-1, .claude/scripts/workstream-record.py, .claude/skills/workstream-capture/SKILL.md, .claude/skills/workstream-status/SKILL.md.

I independently traced standard finish, records/Decisions reduction, the forwarding shim, Learning sentence selection and loss checks, rotation selection, removed-reference sweep and rotation admission. RT-RTE-rewrite-state, MEM-RTE-learning-condense, EPI-RTE-learning-condense and MEM-RTE-state-rotate preserve their correct path-specific qualifications. The epistemic Linux example follows the inspected predicates without execution: first statement and terminal sentence survive while the separate qualifier need not. Rotation deliberately bypasses the common fingerprint and checks different retained structure. Tag existence does not establish snapshot equality. Thus RT-CLM-rewrite-preservation correctly rejects universal preservation. SRC-1, .claude/scripts/workstream-rewrite.py, .claude/scripts/condense-completed-records.py.

The later-addendum guard occurs in Learning condensation, not shared disposition or rotation selection. The damage suite explicitly expects terminal scoring after ROUTED then EXTENDED. MEM-CLM-marker-reopening is therefore a supported source-contract mismatch, not a contradiction between analyst findings. SRC-1, .claude/scripts/workstream_state.py, .claude/scripts/workstream-rewrite.py, tests/damage-acceptance.sh.

EPI-CLM-grounded-claims and EPI-CLM-gate-validation accurately attribute design rhetoric and reported prior operation. The inspected driver cannot isolate skill text as the cause of safety or substantiate every user-authority constraint. No report upgrades those claims to observed or causal evidence. SRC-1, docs/design.md, tests/va3-acceptance.sh.

The reconciliation correctly supersedes the epistemic Learning-condensation declaration by the memory declaration: both identify one algorithm, not sequential transformations. It preserves the epistemic deduction and both original IDs. Other Part of relations have valid containment; side files, measurements and authority paths are not mistaken for duplicate containers. Integration issues about recovery, marker reopening, host events and adoption are retained without selecting a stronger status. No unresolved disagreement between the reports requires a conflict paragraph.

Omission questions were selected from the frozen recursive tree and producer/consumer dependencies. Could another shipped writer bypass rewrite guards? Installer writes, model edits and delegated Bash already make that explicit. Could another consumer restore all history or decide semantic validity? Full work/status/history instructions and hook output preserve requested retrieval versus automatic supply, and measures versus judgment. Could another evaluator establish actual semantic acceptance or exercised self-change? The model-calling driver retains only scenario outcomes and inspection opportunities; offline suite searches show fixtures rather than supplied execution evidence. Could source-maintenance doctrine add another improvement-directed route? README versioning and design fitness review do; the partial inventory limit below preserves that distinction without invalidating any emitted negative about observed operation.

Further omission expansion stopped because these consequential challenges have supported dispositions, with one bounded inventory omission retained below. Unread handler branches and other fixtures are not evidenced absence, but the reports do not claim exhaustive detection or a system-complete negative about checking routes. There is no identified unsupported comparison value or material contradiction left without disposition.

Blockers

none

Limits

  • The maintenance-route inventory is partial. MEM-BAP-artifact-adoption, RT-RTE-acceptance-driver and the epistemic bounded conclusion explain governed edits and evaluator intake, but do not separately trace the release-round protocol and artifact-fitness review in SRC-1, README.md and docs/design.md. These prescribe live feedback, fixes or named queuing, renewed rounds after fixes, and review against current standards. They add improvement-directed doctrine, not supplied execution or causal benefit. Retain the positive file-retention, adoption and narrow evaluator findings, but withhold an exhaustive comparison of kit-maintenance or iterative improvement architectures from these records.

    The pass converges in ROUNDS: a fix built in response to one round is untested code until the next, so a round that produces fixes owes another, targeted at the paths those fixes touch, and the release is declared only by a round that produces none. --- README.md

  • The source's clear-event contract remains unresolved: RT-RTE-session-reload and the memory integration issues retain capture's statement that clear fires no hook versus the SessionStart comment naming clear. SRC-1, .claude/skills/workstream-capture/SKILL.md, .claude/hooks/session-start.sh. File retention and registered reload remain supported; withhold a guarantee of automatic post-clear injection without the external host event contract.