Tracecraft reconciliation

Type: agentic-system-analyses/types/agentic-system-reconciliation-report.md

Reconciliation

This reconciliation connects the four supplied reports within the frozen boundary and Source register. It uses their records and retained passages, not additional source reads. No correction answers or refusal were supplied. Supersessions below identify referents; they do not replace findings, raise statuses or correct the declaring reports.

Duplicate supersessions

Amendment: EPI-OBJ-handoff is superseded by MEM-OBJ-handoff. Both identify steps/<sid>/handoff.json, produced by completion, with note, next-action, symbolic routing/state fields and optional Git-derived paths. Their retained passages identify the same write in SRC-1 sdk/tracecraft/cli/steps.py. The epistemic inventory, completion-disposition row and Git-evidence row therefore concern the memory-declared handoff, not an additional artifact. This leaves the epistemic findings about content, warrant and consumers unchanged.

Amendment: WF-OBJ-handoff is superseded by MEM-OBJ-handoff. The source-native key, schema, producer, field set and overwrite horizon coincide; both records retain the handoff write from SRC-1 sdk/tracecraft/cli/steps.py. Workflow references to construction, delivery and recovery concern that same document. Its prose, routing fields and optional changed-files list remain distinguished by their uses; supersession does not turn them into verified task results.

Amendment: EPI-OBJ-session-parts is superseded by MEM-OBJ-session-parts. Both identify the numbered UUID-suffixed JSONL part objects written by mirroring, with Hermes row serialization as an alternative input conversion. The supporting part-key and serialization passages are in SRC-1 sdk/tracecraft/cli/session.py and sdk/tracecraft/harness/hermes.py. The epistemic import, scrub, retention and compact rows concern these same retained payloads. Memory's additional distinctions about checkpoint inputs and decoded display do not establish another payload identity.

Amendment: EPI-CLM-history-loss is superseded by MEM-CLM-history-immutability. Both quote the same immutable-snapshot/no-version-lost assertion in SRC-1 sdk/tracecraft/cli/memory.py and compare it with snapshot-before-live ordering and direct mutation. The epistemic retention and claim-comparison rows address that same source claim. Neither report establishes enforced immutability or universal no-loss; the amendment selects no stronger claim or status.

Distinct records and independent convergence

The handoff declarations above are not duplicates of RT-OBJ-step-state. That record identifies claim/status documents consumed by fixed polling rules, whereas handoff prose and optional paths reside in a separate object. Nor do the handoff objects duplicate their producer RT-RTE-step-control. The object/route distinction preserves both the workflow's completion receipt and the memory report's unresolved handoff content-reader. The reports converge that a successful waiter can precede the handoff write and does not deliver its text. Evidence is the records and annotations on that route in SRC-1 sdk/tracecraft/cli/steps.py and sdk/tracecraft/protocol.py.

Likewise, the live value and history snapshot, message content and inbox cursor, artifact file, session payload/meta/local cursor, and agent registration declared in the memory member are material outputs or state parts of runtime routes, not replacement declarations of those routes. Registration is remote descriptive state, not RT-OBJ-config, which holds local access and attribution configuration. Session metadata and local ingestion position are not transcript payload. No further object identity with runtime records is asserted.

EPI-RTE-git-files is a subroute of RT-RTE-step-control, not a duplicate of the whole coordination route or of the handoff object. EPI-CLM-git-truth is the associated source warrant claim, not another declaration of the Git mechanism. Memory and workflow independently preserve the narrower license: tracked diff paths relative to HEAD when the subprocess succeeds; failure and no tracked change can both yield an empty list. The source passages retained by the epistemic records are SRC-1 sdk/tracecraft/cli/steps.py. No report treats that list as task correctness evidence.

EPI-RTE-regex-scrub is a transformation inside RT-RTE-session-mirror, not another declaration of mirroring or of RT-CLM-secret-redaction. The latter owns the source claim. Runtime's bounded pattern coverage, memory's transformed-trace account and epistemic uncertainty about semantic preservation are compatible findings about different consequences of the same scrub. The epistemic example is a deduction from source, not an observed invocation. Their retained evidence is SRC-1 sdk/tracecraft/redact.py and sdk/tracecraft/cli/session.py.

EPI-RTE-read-verification is the specific doctrine-only checking request, not the broader instruction-delivery path RT-BAP-agent-prompt, nor the implemented storage reads. The workflow annotation and memory annotation independently distinguish requested read-back from substantive verification and observed compliance. All retain the checklist in SRC-1 examples/hermes-agent/SKILL.md. Its doctrine-only architectural status is compatible with wired read commands whose operation is uninspected; these are not competing statuses for one route.

MEM-RTE-evaluator-readback and EPI-RTE-bucket-evaluation overlap but are not coextensive records. Memory groups sampled SDK tests, benchmark preflight/trial read-back, cleanup and optional trial JSONL emission. Epistemic isolates benchmark result checks, with a sampled memory assertion as additional evidence. Their shared invariant and distinct-message calculations in SRC-1 benchmarks/contention_bench.py and benchmarks/messaging_bench.py support independent convergence, not identity of their complete endpoints and coverage. No whole-record supersession is warranted. Neither duplicates ordinary mail exchange, task coordination or RT-RTE-package-release: evaluator checks, product operations and publication have different consumers and effects. Both evaluator records retain wired implementation and uninspected operation, and neither establishes a numerical result, causal benefit or automatic policy adoption.

Integration-issue dispositions

The memory integration issues produce the following connections rather than disagreements requiring a finding to be selected:

• Handoff consumer: MEM-OBJ-handoff, RT-RTE-step-control and the workflow delivery account agree that direct addressability is not a shipped handoff-fetch-and-execute consumer. The source docstring's next-agent promise remains a source-intent gap, not a conflict among these reports. Next-agent receipt and use remain unestablished.
• Compaction provenance and replay: MEM-OBJ-session-meta adds loss of per-target cursor/UUID/redaction detail to RT-RTE-session-compact. Runtime already limits byte preservation to the normal path and recognizes duplicate display after interruption; the epistemic compact row also retains metadata loss. Payload preservation and metadata replacement are different findings, so no correction or supersession of the compact route follows. Detailed lineage preservation and interrupted read equivalence remain unavailable conclusions. Supporting evidence is SRC-1 sdk/tracecraft/cli/session.py in those records.
• Cursor scope: MEM-OBJ-mirror-state adds that local cursor filenames depend only on session ID and the loader does not validate stored source/harness identity. RT-RTE-session-mirror does not assert universal destination isolation. The added conditional collision risk is compatible with runtime's adapter discovery account; actual collisions remain unobserved. Evidence is SRC-1 sdk/tracecraft/cli/session.py retained by the memory record.
• History assertion: the claim supersession above resolves duplicate identity. RT-RTE-memory-write already retains nontransactional ordering and direct mutability. The memory and epistemic qualifications address stronger source wording, not an incompatible runtime guarantee.
• Registration versus liveness: MEM-OBJ-agent-registration adds the lack of heartbeat refresh in the inspected claim/completion/mirror paths. Runtime describes the five-minute label as age rather than measured process liveness. These findings coexist; direct-writer capacity establishes neither maintenance in operation nor a live scheduler. The memory record retains SRC-1 initialization and consumer evidence.
• Alternative trace provenance: MEM-OBJ-session-parts adds OpenClaw checkpoint/topic inputs and distinguishes host derivation from Tracecraft's ingestion. RT-RTE-session-mirror and the epistemic import account make no claim to complete host reasoning or checkpoint lineage. Missing host provenance remains a local limit, not a disagreement about the implemented import.
• Evaluator scope and record identity: evaluator overlap is retained without equating complete records, as explained above. The memory objects' part relations do not require superseding valid runtime containers. The independently declared handoff, transcript-part and history-claim duplicates across specialist reports do require the explicit amendments above; their lack of a runtime counterpart does not make their specialist referents distinct.

No unresolved cross-report disagreement was identified in the supplied members. Their common execution limit remains unchanged: inspected code and shipped instructions support bounded architecture, not observed agent uptake, successful deployment or demonstrated benefit. The boundary's exclusions and the analysts' uninspected evaluator outputs prevent those conclusions without converting missing evidence into absence. This reconciliation supplies no verification judgment or public synthesis.