v0.54.0

Changes

• Releases can now refresh the built-in model catalog's per-token prices from the Artificial Analysis aggregate (https://artificialanalysis.ai/), which republishes each vendor's own list prices. Base input and output prices feed model routing's cost term and its cost-budget filter, and were previously hand-entered and left to drift. Prompt-cache rates, which no catalog entry carried, make the cost figures in car's usage display exact instead of approximations bounded by the base input rate. Only curated model joins are written, and the step is skipped when no API key is configured. • Both catalog-writing release steps now replace the file atomically, so an interrupted write can no longer leave a truncated builtin_catalog.json for the release commit to pick up.

• Document historical draft proposals for Advisor live wiring and offer exchange; these proposals do not represent implemented runtime behavior.

• Enforce the agent builder's advertised end-to-end deadline, report live phase/model/attempt/scenario/elapsed progress through coder.get, and end timeouts with actionable budget_exhausted state. • Mark agent_scenarios_pass passed when an agent build succeeds instead of leaving the builder contract pending. • Stop an agent build's scenario turns at their next turn boundary on coder.cancel, and show the model serving the current repair or scenario instead of the spec-generation model.

• Stop offering the Parslee platform tools (parslee_capabilities, parslee_m365_generate_document) to a generated agent's build unless a Parslee account is signed in, so a signed-out local build is no longer derailed by car auth login guidance flowing back through its scenario checks; signing in and rebuilding offers them again.

• Agent builder：choosing another starting point now keeps the answers you wrote instead of silently erasing them; a "Continue your draft" action returns to them, and a stopped or abandoned build parks its answers for a retry instead of losing them. • "Continue your draft" and the replace confirmation appear only for answers you typed yourself — an untouched template is not a draft. • Picking a template while a draft with answers is waiting asks for confirmation before replacing them; the Home research starter no longer auto-picks its template over a waiting draft. • Leaving setup ("Back to Agents" or cancel) still clears everything.

• car agent new no longer mints over-long agent IDs：ids derived from long descriptions are now capped at 64 characters (well under the 128-character peer-name limit that agent registration enforces) and suffixed with a 64-bit digest of the full name, so a long description always yields an ID that every CAR surface accepts and two different long descriptions are kept apart with collision-resistant odds (car#1492).

• The daemon now keeps token-bearing per-agent manifest.toml files owner-only on Unix and repairs permissive files and directories from older releases before loading them.

• car agent new now accepts long descriptions without creating overlong project directory names, while keeping the complete description as the agent builder's input. • Clarified that a managed project's committed agent.json is build output and the daemon runs the separately registered declagents.json copy.

• Supervised agents can now receive read-only daemon credentials. Add method_allowlist to an agents.upsert record or external-process manifest.toml to restrict that agent token to named non-host daemon methods; excluded calls are rejected before dispatch, while registrations without the field keep their existing access.

• Kept car agent new successful during CLI/daemon version skew by falling back to the CLI's declarative registry path when an older daemon omits registry_path; car agent where now surfaces declarative lookup failures, emits daemon-canonical IDs in JSON, and car info marks corrupt declarative state unavailable instead of reporting zero agents.

• car agent new now reports the exact declarative registry file alongside the new agent id, with agent_id and registry_path available from --json. • Added read-only car agent where <id> human and JSON output for both in-daemon declarative agents and lifecycle-supervised agents, including their registry or manifest and log locations. • car info now summarizes declarative and supervised agent counts and the state directories that hold them; car inspect shows a supervised agent's manifest path.

• The daemon now keeps token-bearing agents.json files owner-only on Unix and repairs permissive files from older releases before loading them.

• New crate car-topology — pick a multi-agent communication topology from a short list, in one pass, scored on measured tokens rather than edge count. Applies arXiv:2609.02264 ("Codebook Agent"). CAR decides how to coordinate agents in two places and neither learns：Coordinator spends an LLM round trip on a hand-written prompt, and car-multi runs whatever the caller named — while every AgentOutput already carries the TokenAccounting that says what the run cost. The crate fits a codebook of reward-surviving topologies, a reward-weighted prior mapping a query embedding to a distribution over them, and a proxy regressed on measured utility and per-task normalized token cost; selection is one predictor pass, at most M decodes, and one batched scoring call, with no search and no message passing. Deterministic and dependency-light (serde only)：no model calls, no network, no unseeded randomness, and a fitted selector round-trips through serde. The fitting and selection core touches no filesystem; journal (below) is the one module that does. • The three measurements the design rests on ship as first-class diagnostics, runnable against your own execution records. design_space_collapse (does the reward-surviving set stay small as codebook capacity grows?), edge_count_token_correlation (is |E| correlated with measured tokens negatively, so that minimizing it raises the bill?), and team_homogeneity / scorer_advice (do the agents share a profile, in which case a message-passing scorer over profile nodes pools identical features and scores every candidate the same?). These are workload properties, not theorems — the paper's numbers are the paper's, measured on their benchmarks under their implementation, and nothing here has been evaluated against them. The third check matters most for CAR, whose teams are usually N clones of one AgentSpec: car_multi::topology::scorer_advice answers it straight off the specs with no encoder involved. • car_multi::topology bridges finished coordination runs into records. execution_record folds a run's outputs into one ExecutionRecord, returning None rather than a zero-cost record when the runner reported no token accounting — a fabricated zero would drag the cost head toward whatever merely went unmetered. select_shape returns a CoordinationShape a car-multi caller can execute. • Records persist, and a persisted record can never be read in the wrong embedding space. car_topology::journal is an append-only JSONL log at <car-home root>/topology/records.jsonl following the car-eventlog idiom (one record per line, torn-line tolerant on load, missing file reads as empty, an exclusive advisory lock for one writer per path); car_multi::topology::record_run is the one-call write path from a finished coordination run. Every line names the encoder that produced its query embedding, because the prior is a kernel over cosine similarity between query directions — fold two encoders into one record set, or query a fitted selector with a third, and nothing fails, slows down, or looks wrong; it just answers from arbitrary similarities. So the tag is enforced at the boundary: RecordJournal::load_records refuses a mixed journal unless you name the encoder, TopologySelector::fit carries the label through, and select_with(query, embedder) refuses a foreign query. An unlabeled selector passes the check — "nobody recorded it" is not "any encoder fits". • The three diagnostics fold into one report you can actually read. car_topology::diagnose returns a serde-serializable DiagnosticsReport over a record set, with render() for a terminal and headline() for a log line. Two design points are deliberate. The homogeneity claim needs the team's agent profiles, which execution records do not carry, so it renders as NOT MEASURED unless you call diagnose_with_profiles — naming the gap beats silently omitting the claim most likely to hold for CAR. And every report carries a Sufficiency verdict against the paper's own collection protocol (300 records over 50 tasks), so a thin journal reads as provisional rather than as a finding; a set with no variation in edge count or cost is reported as Insufficient rather than given a verdict. • The utility label has a principled source, and the ergonomic path cannot fabricate it. car_multi::topology::run_utility derives a coordination run's utility from the agents' structured car_ir::AgentOutcomes, with the mapping dictated by the type rather than chosen：Success to 1.0, PartialSuccess to 0.5, Failure/GiveUp/Timeout to 0.0, and Done to none, because car-ir documents that status as "neutral -- may or may not have succeeded" and scoring it either way would invent the number the crate exists to measure. UtilityEvidence::Grounded additionally refuses a success backed only by the agent's own assessment, mirroring the receipts-decide-completion rule the flagship assistant already applies -- and yields none, never 0.0, since "nobody checked" is not "it failed". record_run_from_outcomes derives both measured numbers and writes nothing when either is mis