Tracecraft synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Establish a shared coordination space
Tracecraft is a durable coordination and session-retention system for externally hosted agents. It owns neither the model nor the agent's next-action loop. Its shipped CLI and Python storage interface preserve shared values, messages, task ownership, handoffs, files and host transcripts in a project-prefixed bucket. The claimed work is continuity and coordination across hosts and non-overlapping agent sessions, not substantive verification of agent reasoning.
This account covers the shipped product at Git commit 97c6b7447f8e696f0116357d9f86bb0ece25e5b1, with an applicability cutoff of 2026-10-14. SRC-1 is the frozen arrmlet/tracecraft repository. The evidence is inspected implementation and source-native doctrine, not an executed deployment: no target code, tests, examples or services were run, and no real agent trace was supplied. CLI drivers, storage wrappers, all four session adapters, shipped agent instructions and sampled evaluators are included. External model hosts, service implementations and deployed credential grants are excluded. Available benchmark outputs were not independently assessed. See the boundary and evidence and Source register.
A caller initializes backend, bucket, project and agent configuration. Local configuration takes precedence over global configuration, and a process environment variable can replace the agent ID. That ID supplies attribution and mailbox selection; it is not an authenticated worker identity. Bucket credentials supply the actual service principal, while project prefixes organize keys rather than enforce access isolation. Configuration is written before bucket access is checked, so failed initialization can leave local configuration behind. Subsequent commands obtain an S3-compatible or HuggingFace storage wrapper through RT-CMP-get-store and RT-RTE-init-config.
Claim work, exchange progress and declare completion
The representative workflow is prescribed by the shipped Claude Code project prompt and Hermes skill: an already-running agent checks peers and inbox, chooses and claims a step, works outside Tracecraft, shares progress and leaves completion context. It is not a workflow observed to succeed. Task creation, decomposition, agent launch and substantive execution remain external; Tracecraft does not turn the procedure into an enclosing scheduler. The host must load and follow RT-BAP-agent-prompt. The workflow account traces this ordered procedure rather than joining unrelated capabilities.
Claiming a normalized step ID requests a conditional claim write and then writes in-progress status. S3 arbitration depends on the service honoring conditional PUT; HuggingFace uses an exists-then-write check that permits a race. A claim is cooperative ownership, not a lease or security boundary. Completion compares the stored owner with the caller's agent string, but --force bypasses that comparison, and direct store writes bypass CLI checks. Neither path verifies whether the work was actually done. See RT-RTE-step-control.
The progress channels have different retained parts and consumers:
- Shared memory stores an arbitrary caller string as the live value and attempts a separate version snapshot before replacing it. Exact-key get returns the value, normally without writer metadata; history selects the newest requested number of snapshots and displays those oldest first. Snapshots preserve submitted values, not reasons, approval or proof of successful live replacement. Direct writes and deletes defeat enforced immutability. See MEM-OBJ-memory-live, MEM-OBJ-memory-history and MEM-CLM-history-immutability.
- Messages append caller text with attribution and time. Requested inbox reads combine direct mail and broadcasts, suppressing the caller's own broadcasts. Incremental reads use a durable cursor advanced after display. Its five-minute grace window and 1000-key seen cap bound late-arrival and repeat handling; full inbox is a separate cursor-ignoring route. This is selection bookkeeping, not semantic triage or acknowledgment that an agent understood the message. See RT-RTE-mail-exchange and MEM-OBJ-inbox-cursor.
- Artifacts transfer arbitrary bytes under a step or shared prefix. Download delivers the entire file to disk; the SDK does not parse, redact or approve it. The shipped skill affords agent use, but its meaning and downstream effects remain unresolved. See RT-RTE-artifact-share and MEM-OBJ-artifact-file.
Completion has a concrete operational consequence: wait-for returns only when all selected step states equal complete, fails on blocked state, and continues polling on needs-review until completion or timeout. These are enforced dependency gates within the waiter, not judgments of work quality. Needs-review does not launch a reviewer or require a distinct reviewer before clearance. Recovery requires an external agent or operator to resolve the issue and write a new state. See RT-BAP-step-polling and the workflow failure path.
The separate handoff document preserves optional note, next action, recipient and Git-derived paths. It records delegation intent without dispatching the named agent. Inspected status and wait commands do not display this content, and status is written before the handoff, allowing a waiter to return before context exists. The architecture documentation declares a handoff-read command, but that command is not registered in the shipped CLI. Thus declared consumer intent exists without an established executable handoff-content route. See MEM-OBJ-handoff, RT-RTE-step-control and the report-verification limits. The optional Git field measures tracked diff paths relative to HEAD when its subprocess succeeds; failure and no tracked changes both yield an empty list. It is not a certificate of all work or content correctness: EPI-RTE-git-files.
Retain sessions and inspect them later
Session mirroring is a separate loop started by a caller, cron job or host command. A registry selects Claude Code, Codex, OpenClaw or Hermes. File adapters read byte deltas; Hermes reads SQLite message rows and serializes them into JSONL. Claude Code discovery is cwd-scoped, whereas the other adapters ignore cwd. Their all-session mode is therefore not project isolation. OpenClaw checkpoint files can be ingested, but their names do not establish how a host derived them. See RT-CMP-harness-registry, RT-RTE-session-mirror and MEM-OBJ-session-parts.
Each flush selects a delta using local cursor state, optionally redacts recognizable credential patterns, uploads a uniquely named part, overwrites cumulative metadata and then advances the local cursor. One-shot returns; follow repeats against its selected session; follow-all rediscovers sessions. These are code-controlled ingestion policies, not model reasoning. Regex replacement is bounded pattern removal, not a disclosure guarantee or a truth check. Disabled redaction and artifact/direct-store uploads take different paths, and independently processed chunks can split a match. Removing details also leaves semantic preservation indeterminate for arbitrary payloads. See EPI-RTE-regex-scrub and RT-CLM-secret-redaction.
Later list/show/status commands reload retained metadata. Session list ranks entries by recency or size before applying its limit. Positive show-tail downloads every matching part, concatenates bytes and displays the last requested number of decoded lines. A line limit bounds display, not download memory or a model's token window. This is requested inspection, not automatic context injection, restoration of the original host conversation or host sub-agent isolation. See MEM-OBJ-session-meta, MEM-OBJ-session-parts and RT-RTE-session-mirror.
Normal-path compact concatenates bytes and reduces object count; it does not extract lessons or summarize meaning. Replacing the manifest loses detailed per-flush cursor, UUID and redaction provenance. Interrupted deletion can leave merged and original files that show subsequently reads together. Likewise, a mirror upload followed by metadata or cursor failure can replay bytes. Local checkpoints are keyed only by session ID and do not revalidate stored source identity, so independent ingestion also depends on stable source/destination and suitable IDs. Stop removes local state and stamps end metadata, but neither deletes retained transcripts nor terminates an already-running follow process. See RT-RTE-session-compact, MEM-OBJ-session-meta, MEM-OBJ-mirror-state and RT-CLM-mirror-idempotence.
What the checks establish, and what would change the assessment
The strongest supported improvement-relevant contribution is a wired storage-evaluation plane. Sampled tests and benchmarks compare durable reads with expected values, distinct message bodies or exactly one consistent claim winner. Optional trial JSONL retains measurements and check results. These narrow evaluators could expose coordination defects, but their inspected source does not establish passing results, numerical performance or later adoption of a correction. The Hermes skill additionally requests write/read verification without supplying an observed checked result. None of these checks licenses message truth or coding quality. See MEM-RTE-evaluator-readback, EPI-RTE-bucket-evaluation and EPI-RTE-read-verification.
The resulting epistemic distinction is between retained assertions and warranted conclusions. Tracecraft imports caller and host content with unresolved source warrant; structural admission and later availability do not endorse it. Symbolic ownership, status and cursors have fixed operational force, while progress text supplies only afforded reference/context to external agents. All established dynamic deliveries are requested reads. Missing handoff wiring and uninspected durable evaluator consumers prevent claiming a completely pull-only inventory. See the epistemic bounded conclusion and profile verification.
Retained traces and future-readable progress are partial continuity mechanisms, not demonstrated improvement in capacity for future action. No accepted record establishes a qualifying automatic trace-to-behavior-shaping-artifact update with a later guidance consumer; opaque caller artifacts and host checkpoints prevent a system-wide negative. The runtime does not supply a formulated-theory assessment, so theory-builder conditions 1–4 are each unestablished here, rather than inferred from storage, polling or evaluator scores. Improvement caused by criticism of a consumed theory is likewise unestablished. Reflection is not established by the retained account, and autonomous theory building is not established by automatic ingestion or fixed checks: next-action selection and substantive recovery remain external. These are independent findings, not a ladder of capability. See RT-RTE-session-mirror, MEM-RTE-evaluator-readback and the epistemic bounded conclusion.
Exercised self-improvement is also unestablished at this shipped-product boundary. Evaluation output has no inspected evidence-responsive consumer that revises the SDK's own behavior-determining organization and exercises that revision later. Source/package release is a separate wired change path, not proof of self-improvement; CI is separate from the release job, whose source does not explicitly require test-suite success. See RT-RTE-package-release.
For a scenario requiring agents to exchange durable progress and explicit dependency states across hosts, the inspected mechanisms support that architectural capacity, conditional on backend contracts and agents making the reads. For a scenario requiring verified work, automatic handoff execution, faithful host-context reconstruction or exact-once recovery, this evidence supports no such guarantee. A registered handoff reader with delivery traces would alter the recipient assessment. Backend-specific contention and interruption evidence would sharpen arbitration and replay conclusions. Linked traces of host prompt loading, retained-content consumption and subsequent action would establish uptake. A retained criticism-to-update-to-later-operation chain, with independently specified objectives and appropriate comparisons, would be needed to establish learning or self-improvement rather than continuity alone.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Declared handoff-reader intent is incompletely inventoried; the documented command is not shipped | SRC-1; MEM-OBJ-handoff, RT-RTE-step-control, workflow delivery account, report-verification limits | Verification inspected docs/s3-architecture.md and CLI registrations; documentation illustrates an older-key tracecraft handoff GET |
Neither “no read procedure was ever declared” nor usability of that documented command; recipient content delivery remains unwired in the inspected CLI | A registered reader matching current keys/schema and evidence of recipient retrieval and use |
| Protection receipt says “0 lost” without measuring losses; preservation-claim inventory is not exhaustive | SRC-1; RT-RTE-session-mirror, MEM-OBJ-session-meta, RT-CLM-mirror-idempotence, report-verification limits | Verification traced receipt construction in sdk/tracecraft/cli/session.py and metadata totals in sdk/tracecraft/protocol.py |
A measured zero-loss result or an exhaustive list of preservation claims | Source-to-retained completeness comparisons across failure cases, and a receipt tied to those measurements |
| Implementation and instructions were inspected, not exercised | SRC-1; RT-BAP-agent-prompt, EPI-RTE-read-verification, RT-RTE-step-control | Shipped CLI, example prompts and selected workflow; external hosts and actual deployments excluded | Successful multi-agent execution, reliable prompt compliance or demonstrated benefit | Linked execution evidence showing host loading, calls, returned material and downstream actions |
| Service contracts and cooperative identities bound task guarantees | SRC-1; RT-RTE-init-config, RT-RTE-step-control, RT-CMP-get-store | S3/HF wrappers, config override, force and direct-store paths; service internals and deployed grants excluded | Backend-independent atomic arbitration, authenticated ownership, project access isolation or durable-service guarantees | Deployment-specific authorization and conditional-write evidence, plus contention/recovery results |
| Multiobject writes and bounded cursors do not establish exact-once delivery or immutable history | SRC-1; RT-RTE-mail-exchange, MEM-OBJ-inbox-cursor, MEM-CLM-history-immutability, RT-CLM-mirror-idempotence, MEM-OBJ-mirror-state | CLI write ordering, full/incremental mail, direct mutation and session-ID-only local checkpoints | Unconditional no-loss history, exact-once replay, all-late-message delivery or independent cursor namespaces | Failure/concurrency traces and mechanisms enforcing transaction, retention and source/destination identity contracts |
| Imported payloads and host checkpoints retain unresolved content provenance | SRC-1; MEM-OBJ-memory-live, MEM-OBJ-artifact-file, MEM-OBJ-session-parts, EPI-RTE-regex-scrub | Arbitrary caller values/files and adapter ingestion; no concrete payload instances or external-host derivation evidence | Complete form/lineage classification, source truth, full reasoning preservation, universal secret removal or semantic equivalence after redaction | Concrete payloads with provenance, host derivation records and transformation/completeness checks |
| Storage compaction loses detailed provenance and can alter interrupted display | SRC-1; RT-RTE-session-compact, MEM-OBJ-session-meta | Concatenation, manifest replacement and show's actual-file listing | Detailed lineage preservation or read equivalence during interrupted/concurrent maintenance | Preserved per-flush provenance and interruption/concurrency evidence covering later reads |
| Evaluation outputs and downstream adoption are incompletely assessed | SRC-1; MEM-RTE-evaluator-readback, EPI-RTE-bucket-evaluation, RT-RTE-package-release | Sampled evaluator source and optional JSONL emitter; frozen numerical results and later report consumers uninspected; CI/release source only | Independent numerical results, exhaustive later-consumer coverage, qualifying trace learning, criticism-driven capacity improvement or exercised self-improvement | Assessed trial outputs/design, a recorded adoption consumer and retained update-to-later-operation evidence with appropriate outcome comparisons |
No unresolved cross-report conflict was supplied. Profile verification declares no additional limits. Source-read log: none; this synthesis uses the supplied members only.