Agent Task Economics v1

Status: integrated observation and deterministic export executed locally for exact subject 474c481bf3c3561c144e077f0000460f61af55f2, with both format reports authenticated by the focused bundle. No productivity, latency, model-token or completion claim.

Audience: agent evaluators, workflow integrators, and compiler contributors.

The integrated graph-operational Git workflow records a bounded semaprax.agent-task-economics.v1 observation while exercising the workflow's twelve requirement criteria. This is separate from Agent Context Economics v1, which measures one-file context selection. The task observation describes exact compiler protocol and review traffic for one authored workflow; it is not a model or human study.

The recorder wraps the already configured in-process v5 session used by the regression. Every semantic request still passes through the ordinary bounded VNextSession::handle_frame dispatcher. The wrapper measures the exact response-bearing request and response bytes and computes their digests before parsing the response as JSON for the test. It does not add a protocol method, grant, side channel or runtime authority, and it cannot change a request or response.

Current implementation regression evidence is HOSTED GREEN under the v0.4.0 baseline. The exact-subject local observations above and below remain historical, not new release measurements.

Recorded observations

Each row follows invocation order and records its session, method, associated requirement criteria, request and response byte counts, repository lexical units, SHA-256 digests and success/error outcome. Criterion associations are logical rather than chronological: one signature operation is associated with both signature change and caller migration without becoming two protocol calls. The report aggregates distinct semantic protocol calls and an exact method histogram. This fixed workflow has zero notifications; the recorder makes no claim to capture response-less frames. Request/response counts are compiler protocol traffic, never model tool calls.

The workflow additionally records:

• bounded review material sizes for compact source-review, function-summary, and impact-summary projections;
• one disjoint sibling-candidate reconciliation and the deliberately scripted conflict or authority-control rejections;
• zero stale recoveries, because this scenario does not perform live-source refresh or semantic rebase;
• explicit validation, recovery replay, semantic-delta verification, interpreter-test and target-admission observations;
• twelve criteria rows only after the owning test assertions establish the corresponding snapshot, selection, migration, invariant, evidence, conflict and publication facts.

Scripted SPX-G235, SPX-G286 and SPX-G287 responses exercise control paths. They are not observed agent mistakes. The accepted sibling merge is conflict reconciliation, not stale recovery. Native-C11 and Core-Wasm rows count compiler emission or structural admission; native and Wasm execution remain zero.

All byte counts use exact UTF-8 lengths. Lexical units reuse semaprax.lexical-token.v1 and retain model_tokens: false. Ratios or savings are not inferred from either measure. The pre-publication review traffic is source-deterministic. Complete candidate and semantic-delta reports are replayed inside the verifier rather than transferred through the agent-facing protocol. Commit, commit-report and source-commit-status routes bind the canonical temporary manifest and bare-repository identities, so their exact digests are invocation evidence and carry host_route_bound: true. Portable recovery traffic remains separately labeled. The regression checks schema, limits, relationships, selected compact-route counts and counter self-consistency. It enforces aggregate ceilings without freezing invocation-specific byte, lexical or digest values as exact goldens.

Explicitly unobserved fields

The report keeps model identity, tokenizer identity, model input/output tokens, model or external agent tool calls, wall and CPU time, peak memory, monetary cost, and human review duration null with status not_observed. Validation invocation counts are not validation cost or elapsed time. Review-material bytes are not review effort. A deterministic scripted success is not an agent success rate.

A comparative productivity claim requires a separately captured immutable observation bundle for both semantic and source-first workflows. It must bind exact prompts and contexts, model and tokenizer versions, ordered external tool traffic, corpus/compiler/source revisions, cold/warm state, correctness rubric, validation timing method and human-review protocol. None of those external observations is synthesized from compiler protocol bytes.

Deterministic focused export

When and only when the test process receives an absolute SEMAPRAX_GRAPH_WORKFLOW_EVIDENCE_DIR, each successful format-specific twelve-step scenario writes its already validated compact JSON observation to one fixed name:

• agent-task-economics-sha1.json
• agent-task-economics-sha256.json

The ordinary test invocation sets no directory and writes no report. Export does not change the report schema, protocol traffic, Git provider, candidate, or assertions. The focused execution-evidence runner supplies the directory, authenticates both outputs, and places their digests in a separate exact-commit envelope. A successful test without both required files is not a successful evidence bundle.

The two reports are per-invocation facts. In particular, temporary canonical manifest and repository identities make selected publication-route hashes host-bound. Their byte counts and digests may differ across otherwise valid runs; the outer evidence envelope authenticates the observed files without turning them into cross-host goldens.

Compact workflow cost guard

The current regression rejects any twelve-step workflow that exceeds 22 response-bearing protocol calls, 9 KiB of serialized requests, 48 KiB of serialized responses, or 12,288 semaprax.lexical-token.v1 response units. It also rejects more than 16 KiB or 4,096 lexical units across the three explicit review materials. candidate/query, candidate/impact, and candidate/semantic-delta are excluded from the agent-facing method histogram; the workflow uses candidate/source-review, candidate/function-summary, and candidate/impact-summary while retaining complete authority-free candidate and semantic-delta replay as internal verification. Each compact review route is transferred exactly once. The expected conflict does not re-download immutable review data; the later required recovery replay proves the same candidate digest, exact source-review bytes, and declaration facts instead.

These ceilings are regression budgets for one fixed scripted workflow, not measurements of a model context window. The lexical unit still carries model_tokens: false; the guard makes no tokenizer, usefulness, latency, success-rate, or monetary-cost claim. The authenticated historical bundle below predates this compact-route guard and remains evidence only for its exact recorded subject.

Bounds and evidence

The fixed test supplies two finite sessions and their response-bearing frames. Each request remains under the protocol's 64 KiB request limit, the session retains its existing 1 MiB response limit, and the regression rejects a final report over 256 KiB in addition to the aggregate workflow ceilings above. The compact sorted-key JSON retains digests and counters rather than becoming a second source or candidate archive. The report has source_authority: false, execution_authority: false, and no publication authority; the separately attached Git host remains the sole authority for the scripted commit step.

The integrated SHA-1 and SHA-256 regressions passed in the exact local bundle for subject 474c481bf3c3561c144e077f0000460f61af55f2; the execution envelope authenticates both reports. The ignored managed workflow is not_selected; generated clients, MCP, native/Wasm runtime, hosted CI and the complete programme are separate dimensions and were not selected by this runner. No benchmark value is claimed, and the graph-operational programme remains Partial.