@evolvingmachines/evolve 0.0.67 → 0.0.68

package/spec/openapi.yaml

@@ -101,9 +101,9 @@ info:
       directory and runs it as a new local Harbor job
       (their src/harbor/analyze/analyzer.py `run_analyze`). Here analysis is
       a NATIVE capability of the platform — the worker runs the same
-      analyzer (claude-code over the trial's Harbor-shape tree plus its
-      task, rubric-driven, `analysis.json` validated criterion-for-
-      criterion) server-side, and the result is stored and served on the
+      analyzer (the chosen agent, claude-code by default, over the trial's
+      Harbor-shape tree plus its task, rubric-driven, `analysis.json`
+      validated criterion-for-criterion) server-side, and the result is stored and served on the
       trial body (`Trial.analysis`) instead of written into a local
       directory. Two named deviations from Harbor's shape: (1) server-side
       execution — no wrapper job appears anywhere on this surface; the
@@ -111,8 +111,8 @@ info:
       `JobCreate.analyze` runs the analyzer per trial as each trial settles,
       which Harbor has no equivalent of (their analyze is always a manual
       follow-up command; `POST /api/jobs/{jobId}/analyze` is that manual
-      verb here). Vocabulary is Harbor's: `model_name`, rubric criteria
-      `{name, description, guidance}` (their cli/analyze.py `--model` /
+      verb here). Vocabulary is Harbor's: `agent`, `model_name`, rubric criteria
+      `{name, description, guidance}` (their cli/analyze.py `--agent` / `--model` /
       `--rubric`, cli/quality_checker/models.py RubricCriterion), result
       `{summary, checks{outcome, explanation}, estimated_cost_usd}` (their
       analyze/models.py AnalyzeResult). Selection and width are Harbor's
@@ -132,7 +132,7 @@ info:
       own record: the caller uploads the task directory (or a directory of
       task directories) as one archive (`POST /api/checks`, multipart —
       `archive` + `config`), the worker runs the SAME checker per task
-      (claude-code over the task's byte-sacred tree, the platform's default
+      (the chosen agent, claude-code by default, over the task's byte-sacred tree, the platform's default
       check prompt and rubric rendered through Harbor's own tokens,
       `check-result.json` validated criterion for criterion) server-side,
       and the report is served as the `Check`
@@ -140,12 +140,14 @@ info:
       deviations: server-side execution (no wrapper job on this surface);
       a stored record instead of a local report file; a typed `failure
       {phase, message}` per task where Harbor's QualityCheckResult carries
-      `error: str` (the TrialAnalysis dialect); the model default
-      `openrouter/deepseek/deepseek-v4.1-flash`, the analyzer's, where Harbor's check defaults to
-      `claude-sonnet-4-6` (one roster, one default for both rubric agents);
+      `error: str` (the TrialAnalysis dialect); an omitted model resolving
+      as the analyzer's — the platform's pick,
+      `openrouter/deepseek/deepseek-v4.1-flash`, on every agent whose roster
+      carries it, else the agent's own default — where Harbor's check
+      defaults to `claude-sonnet-4-6` (one rule for both rubric agents);
       and the platform's `reasoning_effort` and `sandbox_provider` beside
       Harbor's knobs, exactly as on `AnalyzeConfigInput`. Vocabulary is
-      Harbor's: `model_name`, `rubric`, `prompt` (their cli/analyze.py:88-
+      Harbor's: `agent`, `model_name`, `rubric`, `prompt` (their cli/analyze.py:88-
       103), `n_concurrent` (:117-119), `include_task_names` /
       `exclude_task_names` / `n_tasks` (:121-131; fnmatch globs and the cap
       of checker.py:132-138), result `{task_name, checks{outcome,
@@ -1459,9 +1461,11 @@ paths:
                 │   │                    at Harbor's own tee name for the
                 │   │                    harness (claude-code.txt, codex.txt,
                 │   │                    gemini-cli.txt, qwen-code.txt,
-                │   │                    kimi-code.txt, opencode.txt;
-                │   │                    droid.txt, and stdout.log for a
-                │   │                    harness Harbor has no name for)
+                │   │                    kimi-code.txt, opencode.txt, pi.txt,
+                │   │                    antigravity-stream.jsonl; droid.txt,
+                │   │                    prime-agent.txt, dsh.txt, zcode.txt, and
+                │   │                    stdout.log for a harness Harbor has
+                │   │                    no name for)
                 │   ├── stderr.log       the harness process's stderr stream,
                 │   │                    when stored
                 │   ├── trace-parsed.jsonl  the parsed event trace, one
@@ -1569,7 +1573,8 @@ paths:
         `CancelledError`) and any still-QUEUED trials of a cancelled source.
         The source job must be terminal. An UPLOADED job is refused
         (`job_uploaded`, 409): it is a record of a run that happened
-        elsewhere, not something the platform can re-run.
+        elsewhere, not something the platform can re-run. A selection whose
+        arm is a retired agent is refused (`agent_retired`, 409).
 
         Supports `Idempotency-Key` with a fingerprint namespaced to this
         route, so a create key can never replay a resume.
@@ -1643,7 +1648,8 @@ paths:
         caller-SELECTED trials of the source, linked back via `source_jobs`
         (`action: "retry"`). The source job is never mutated. An UPLOADED
         source is refused (`job_uploaded`, 409): a record of a run that
-        happened elsewhere cannot be re-run here.
+        happened elsewhere cannot be re-run here. A selection whose arm is a
+        retired agent is refused (`agent_retired`, 409).
 
         The body selects the trials — `trial_ids` XOR `failed_only`:
         - Omitted body (or `{}`): every trial of the job. The source must be
@@ -1801,9 +1807,9 @@ paths:
       summary: Analyze a job's trial traces (rubric-driven)
       description: |
         Runs Harbor-style trace analysis over a terminal job's trials,
-        server-side: for each trial the analyzer agent (claude-code, in its
-        own sandbox, under the network Harbor's analyze wrapper task gives
-        it — open by default) reads the trial's Harbor-shape tree plus its
+        server-side: for each trial the analyzer agent (`agent`, claude-code
+        by default, in its own sandbox, under the network Harbor's analyze
+        wrapper task gives it — open by default) reads the trial's Harbor-shape tree plus its
         original task and rules every rubric criterion, storing
         `{summary, checks, estimated_cost_usd}` on the trial
         (`Trial.analysis`) and the aggregate on the job
@@ -1823,7 +1829,7 @@ paths:
         `max_concurrent_analyses`).
 
         An analyzer run that completes without a valid `analysis.json` is
-        automatically re-run at most once (fresh sandbox, same model and
+        automatically re-run at most once (fresh sandbox, same agent, model and
         rubric); the analysis object's `attempts` records how far the
         ladder went. A run cut by its budget is not that class: it settles
         `failed` with phase `timeout` at once and is never re-run — a
@@ -3274,7 +3280,8 @@ paths:
         Harbor's hosted `trial retry <id>`) address the trial, not the job.
         The source JOB may still be running: a settled trial's facts are
         final, so "run it again" is well-defined the moment it settles. A
-        trial of an UPLOADED job is refused (`job_uploaded`, 409).
+        trial of an UPLOADED job is refused (`job_uploaded`, 409), and a
+        trial whose arm is a retired agent by `agent_retired` (409).
 
         Supports `Idempotency-Key` — the identical fingerprint the job-level
         door computes for this one-trial selection, because it IS the
@@ -4999,17 +5006,30 @@ paths:
         body, as the TEMPLATE a caller's own `prompt` replaces — with its
         `{trial_path}`, `{task_section}` and `{criteria_guidance}` slots
         unrendered; the output section Harbor appends is not part of it),
-        `reasoning_effort` (the default model's per-model effort) and
+        `reasoning_effort` (the resolved model's per-model effort) and
         `sandbox_provider` (the fleet default). Read them to see what a
         bare `evolve analyze` runs under, or to start an edited rubric or
-        prompt from the platform's own. `model_name` and `rubric` sent
-        back resolve as if nothing was sent; `prompt`, `reasoning_effort`
-        and `sandbox_provider` sent back are stored as named, so the
-        analysis echoes the prompt text and keeps that effort and provider
-        even when the fleet default later changes. Harbor has no such verb:
+        prompt from the platform's own. With `agent`, the same five knobs
+        as `analyze: {agent}` resolves them — that agent's default model
+        and the effort it takes (`evolve analyze --show-defaults -a`).
+        `model_name` and `rubric` sent back with the same `agent` resolve
+        as if nothing was sent; `prompt`, `reasoning_effort` and
+        `sandbox_provider` sent back are stored as named, so the analysis
+        echoes the prompt text and keeps that effort and provider even
+        when the fleet default later changes. Harbor has no such verb:
         its defaults are files on disk (analyze/prompts/analyze.txt and its
         default analyze rubric), which a hosted caller cannot open — the
         hosted reading of the same files.
+      parameters:
+        - name: agent
+          in: query
+          description: >
+            Resolve the defaults for this agent instead of the default one
+            (`AnalyzeConfigInput.agent`'s vocabulary, `GET /api/meta`
+            `analyze.agents[].name`); an unknown name is refused
+            `invalid_input` naming `agent`.
+          schema:
+            type: string
       responses:
         '200':
           description: The resolved defaults.
@@ -5017,6 +5037,8 @@ paths:
             application/json:
               schema:
                 $ref: '#/components/schemas/AnalyzeDefaults'
+        '400':
+          $ref: '#/components/responses/BadRequest'
         '401':
           $ref: '#/components/responses/Unauthorized'
         '429':
@@ -5692,8 +5714,8 @@ paths:
 
             analyze-<trial>__<7>/
             ├── config.json              TrialConfig, partial: the wrapper
-            │                            task's name, the agent (claude-code,
-            │                            the model), and under `x_evolve`
+            │                            task's name, the agent (its Harbor
+            │                            name, the model), and under `x_evolve`
             │                            the frozen policy — rubric, prompt,
             │                            reasoning_effort — plus the analyzed
             │                            trial/job/task and the provider
@@ -5721,13 +5743,15 @@ paths:
             │                            validator (an infrastructure phase)
             ├── agent/
             │   ├── claude-code.txt      the analyzer's stdout stream at
-            │   │                        Harbor's own tee name for claude-code
+            │   │                        Harbor's own tee name for its agent
+            │   │                        (claude-code.txt, codex.txt, …)
             │   ├── stderr.log           its stderr, when stored
             │   ├── trace-parsed.jsonl   the run's parsed event trace
             │   ├── .claude/ …           the captured home at its real names,
             │   ├── agent-home.json      the capture record beside it
-            │   └── sessions/ …          Harbor's own copy of the .claude/
-            │                            subtree (claude_code.py:482)
+            │   └── sessions/ …          Harbor's own copies of the home, as
+            │                            on a trial (claude: the .claude/
+            │                            subtree, claude_code.py:482)
             ├── verifier/                ONLY when the validator ruled —
             │   │                        completed, or failed `invalid_result`
             │   ├── test-stdout.txt      the validator's printed reasons
@@ -5804,8 +5828,8 @@ paths:
         gzipped tarball (`archive`), or a published dataset is named
         instead (`dataset`, the hosted form; `CheckRequest.dataset` states
         the deviation and its reason), and for each task directory it holds
-        the checker agent (claude-code, Harbor's own check agent, in its own
-        sandbox, under the network Harbor's check wrapper task gives it —
+        the checker agent (`agent`, claude-code by default — Harbor's own
+        check agent — in its own sandbox, under the network Harbor's check wrapper task gives it —
         open by default) reads EVERY file of the task and rules every rubric
         criterion `pass`, `fail`, `not_applicable` or `unknown` with a
         rationale and the evidence behind it, writing `check-result.json`
@@ -5978,18 +6002,31 @@ paths:
         body, as the TEMPLATE a caller's own `prompt` replaces — with its
         `{task_path}`, `{file_tree}` and `{criteria_guidance}` slots
         unrendered; the output section Harbor appends is not part of it),
-        `reasoning_effort` (the default model's per-model effort) and
+        `reasoning_effort` (the resolved model's per-model effort) and
         `sandbox_provider` (the fleet default). Read them to see what a
         bare `evolve check` runs under, or to start an edited rubric or
-        prompt from the platform's own. `model_name` and `rubric` sent
-        back resolve as if nothing was sent; `prompt`, `reasoning_effort`
-        and `sandbox_provider` sent back are stored as named, so the check
+        prompt from the platform's own. With `agent`, the same five knobs
+        as a check with that `agent` resolves them — its default model and
+        the effort it takes (`evolve check --show-defaults -a`).
+        `model_name` and `rubric` sent back with the same `agent` resolve
+        as if nothing was sent; `prompt`, `reasoning_effort` and
+        `sandbox_provider` sent back are stored as named, so the check
         echoes the prompt text and keeps that effort and provider even
         when the fleet default later changes. Harbor
         has no such verb: its defaults are files on disk
         (cli/quality_checker/default-rubric.toml, analyze/prompts/
         check.txt), which a hosted caller cannot open — the hosted reading
         of the same files.
+      parameters:
+        - name: agent
+          in: query
+          description: >
+            Resolve the defaults for this agent instead of the default one
+            (`CheckConfigInput.agent`'s vocabulary, `GET /api/meta`
+            `analyze.agents[].name`); an unknown name is refused
+            `invalid_input` naming `agent`.
+          schema:
+            type: string
       responses:
         '200':
           description: The resolved defaults.
@@ -5997,6 +6034,8 @@ paths:
             application/json:
               schema:
                 $ref: '#/components/schemas/CheckDefaults'
+        '400':
+          $ref: '#/components/responses/BadRequest'
         '401':
           $ref: '#/components/responses/Unauthorized'
         '429':
@@ -10856,6 +10895,10 @@ components:
         - agent_config_unsupported
         - agent_config_key_refused
         - agent_preset_unsupported
+        # A new run naming a retired built-in agent: job create, analyze or
+        # check (400), or resume/retry of a job whose arm it is (409). Details
+        # name each one and its replacement (`retired_agents` in GET /api/meta).
+        - agent_retired
         - provider_unsupported
         - job_not_found
         - job_not_terminal
@@ -11663,7 +11706,10 @@ components:
       properties:
         name:
           type: string
-          description: Agent name — a built-in or one registered under /api/agents.
+          description: >
+            Agent name — a built-in or one registered under /api/agents. A
+            built-in listed under `GET /api/meta` `retired_agents` is refused
+            (`agent_retired`, details naming its replacement).
         model_name:
           type: string
           description: >
@@ -12359,7 +12405,7 @@ components:
       type: object
       description: |
         Trace-analysis configuration — Harbor's `harbor analyze` vocabulary
-        (their cli/analyze.py: `--model`, `--rubric`, `--prompt`). PRESENCE of this
+        (their cli/analyze.py: `--agent`, `--model`, `--rubric`, `--prompt`). PRESENCE of this
         object is the switch: on `JobCreate.analyze` it arms the embedded
         trigger (each trial is analyzed server-side right after it settles;
         CANCELLED trials are skipped — withdrawn work is not billed an
@@ -12371,13 +12417,13 @@ components:
         `GET /api/analyses/defaults`; Harbor's own analyze-rubric.toml
         still runs when passed as `rubric`).
 
-        The analyzer always runs the claude-code harness (Harbor's default
-        analyze agent) in its own sandbox, under the network Harbor's
-        analyze wrapper task gives it (open by default; the model door is
-        unchanged) — on the provider `sandbox_provider` names, or the
-        platform's analysis default when it names none; its spend is capped
-        per analysis and metered as its own line, never blended into the
-        trial's own bill.
+        The analyzer runs the agent `agent` names (claude, Harbor's own
+        default claude-code, when it names none) in its own sandbox, under
+        the network Harbor's analyze wrapper task gives it (open by default;
+        the model door is unchanged) — on the provider `sandbox_provider`
+        names, or the platform's analysis default when it names none; its
+        spend is capped per analysis and metered as its own line, never
+        blended into the trial's own bill.
 
         Which trials, and how wide, are Harbor's own analyze options with
         their exact names: `n_concurrent` (`-n/--n-concurrent`), `passing`
@@ -12393,9 +12439,7 @@ components:
         takes it; on `JobCreate.analyze` the trials do not exist yet.
 
         Harbor's other analyze options, and why they are not here:
-        `-a/--agent` — the analyzer harness is claude-code, fixed (the
-        instruction templates, the roster check and the sandbox policy are
-        written against it); `--job-name` and `-o/--jobs-dir` — a hosted
+        `--job-name` and `-o/--jobs-dir` — a hosted
         analysis attaches to its source job and is served on that job's
         trials, so there is no wrapper job to name and no directory to
         write; `-k/--n-attempts` (k analyses per trial) — deferred: each
@@ -12410,30 +12454,52 @@ components:
         (task quality checks, their cli/analyze.py:126-137) and are not
         adopted here.
       properties:
+        agent:
+          type: string
+          default: claude
+          description: >
+            The agent the analyzer runs on — Harbor's `-a/--agent` (their
+            cli/analyze.py:258, default claude-code), named in the
+            `agents[].name` vocabulary (claude, not claude-code) so one set
+            of names covers arms and reviewers: any built-in agent
+            `analyze.agents` lists on `GET /api/meta`. `model_name` must be
+            on this agent's roster and `reasoning_effort` must be one it
+            accepts; omitted, `model_name` is this agent's
+            `analyze.agents[].default_model`. A registered (bring-your-own) agent
+            is not accepted. A built-in listed under `GET /api/meta`
+            `retired_agents` is refused `agent_retired` naming
+            `analyze.agent`, details naming its replacement; analyzing the
+            trials a retired agent ran, with a live agent, stays allowed.
+            Anything else is refused `invalid_input` naming
+            `analyze.agent`. Stored as given and frozen into each analysis
+            (`TrialAnalysis.agent`).
         model_name:
           type: string
-          default: openrouter/deepseek/deepseek-v4.1-flash
-          description: >
-            Model the analyzer agent runs — Harbor's `--model`. The default
-            is openrouter/deepseek/deepseek-v4.1-flash on this platform's
-            claude roster (DeepSeek V4.1 Flash served through OpenRouter, at
-            its default effort high — the owner's ruling 2026-09-10: far
-            more parallel capacity through OpenRouter's provider pool than
-            one pinned Fireworks host) — a recorded deviation from Harbor's default
-            analyze model (their cli/analyze.py `claude-haiku-4-5`): analysis
-            is input-dominated, and this is the roster's intelligence-per-
-            input-dollar pick; `glm-5.3-flash` (at max, the effort its
-            published scores use) and `haiku` stay on the roster as
-            alternatives, `glm-5.3` to escalate, and the same model on its
-            Fireworks route, `fireworks/deepseek-v4.1-flash`, is a further
-            option (the OpenRouter id stays the default). The value
-            speaks the same vocabulary as `agents[].model_name`: either
-            advertised spelling is accepted and stored AS GIVEN (the
-            default is the roster alias), the wire id is resolved only when
-            the analyzer runs, and every stored analysis serves the
-            spelling it was created under. Must be on
-            the claude roster (`GET /api/meta`, `agents[].models`, also
-            `analyze.models`); anything else is refused at accept
+          description: >
+            Model the analyzer agent runs — Harbor's `--model`. Omitted, the
+            agent's default: the platform's pick,
+            openrouter/deepseek/deepseek-v4.1-flash (`GET /api/meta`
+            `analyze.default_model`), on every agent whose roster carries it,
+            else that agent's own default (`GET /api/meta`
+            `agents[].default_model`, the model the SDK runs that agent on
+            when none is named). The platform's pick is DeepSeek V4.1 Flash
+            served through OpenRouter, at its default effort high (the
+            owner's ruling 2026-09-10: far more parallel capacity through
+            OpenRouter's provider pool than one pinned Fireworks host) — a
+            recorded deviation from Harbor's default analyze model (their
+            cli/analyze.py `claude-haiku-4-5`): analysis is input-dominated,
+            and this is the roster's intelligence-per-input-dollar pick;
+            `glm-5.3-flash` (at max, the effort its published scores use)
+            and `haiku` stay on the claude roster as alternatives, `glm-5.3`
+            to escalate, and the same model on its Fireworks route,
+            `fireworks/deepseek-v4.1-flash`, is a further option (the
+            OpenRouter id stays the pick). The value speaks the same
+            vocabulary as `agents[].model_name`: either advertised spelling
+            is accepted and stored AS GIVEN (the default is the roster
+            alias), the wire id is resolved only when the analyzer runs, and
+            every stored analysis serves the spelling it was created under.
+            A named model must be on the agent's roster (`GET /api/meta`,
+            `analyze.agents[].models`); anything else is refused at accept
             (`invalid_input`, roster in the message).
         rubric:
           $ref: '#/components/schemas/Rubric'
@@ -12472,20 +12538,22 @@ components:
           type: string
           description: >
             Reasoning effort the analyzer runs at — the platform's
-            `agents[].reasoning_effort` vocabulary applied to the analyzer,
-            which IS the claude harness: the accepted values are
-            `analyze.reasoning_efforts` on `GET /api/meta` (the claude
-            agent's `effort_support` is `level`, so the whole list), and an
-            unknown value is refused `invalid_input` with the same sentence
-            an arm's is (`analyze.reasoning_effort must be one of: ...`).
+            `agents[].reasoning_effort` vocabulary applied to the analyzer's
+            agent, held to that agent's own rule exactly as an arm on it:
+            the accepted values are `analyze.agents[].reasoning_efforts` on
+            `GET /api/meta` (claude's `effort_support` is `level`, so the
+            whole list; an agent that takes none, gemini, accepts none), and
+            a refused value answers `invalid_input` with the same sentence
+            an arm's does (`analyze.reasoning_effort must be one of: ...`).
             Omitted, the PER-MODEL default applies —
-            `analyze.models[].default_reasoning_effort`: `high` on
-            openrouter/deepseek/deepseek-v4.1-flash (the default model;
+            `analyze.agents[].models[].default_reasoning_effort`: `high` on
+            openrouter/deepseek/deepseek-v4.1-flash (the platform's pick;
             DeepSeek's own documented default, the owner's ruling
             2026-09-10), `max` on glm-5.3-flash (the platform's ruling
-            2026-09-08, the effort its published scores use), the claude
-            harness default (`agents[].default_effort`) on every
-            other roster model. Whatever resolves is passed to the analyzer
+            2026-09-08, the effort its published scores use), on any agent
+            that can honor them; the agent's own default
+            (`agents[].default_effort`) on every other roster model; none
+            on an agent that takes none. Whatever resolves is passed to the analyzer
             EXPLICITLY on every call — never left to a provider's default —
             and recorded as `TrialAnalysis.reasoning_effort`. A hosted
             extension, recorded as such: Harbor's `harbor analyze` has no
@@ -12602,11 +12670,14 @@ components:
         always states the policy it executes (same law as RetryConfig).
         Echoed on the job body when the job was created with `analyze`;
         each analysis additionally carries the exact policy IT ran under
-        (`Trial.analysis.model_name` / `.rubric` / `.prompt` /
+        (`Trial.analysis.agent` / `.model_name` / `.rubric` / `.prompt` /
         `.reasoning_effort`), which a later manual re-analysis may have
         changed.
-      required: [model_name, rubric, prompt, reasoning_effort, sandbox_provider, n_concurrent, passing, failing, n_trials, trial_ids]
+      required: [agent, model_name, rubric, prompt, reasoning_effort, sandbox_provider, n_concurrent, passing, failing, n_trials, trial_ids]
       properties:
+        agent:
+          type: string
+          description: The agent this policy's analyses run on — as named, or the default (claude).
         model_name:
           type: string
         rubric:
@@ -12618,14 +12689,15 @@ components:
             null = the built-in prompt (the platform's default analyze body).
         reasoning_effort:
           x-wave: 3
-          type: string
+          type: [string, 'null']
           description: >
             The effort this policy's analyses run at. Named at create it is
             served as stored; when the create named none, this echoes the
-            per-model default of the day for `model_name` — the value the
-            next enqueue under this policy stamps (the same nuance as
-            `sandbox_provider` below, for the same reason: the default is a
-            platform table, resolved where an analysis is enqueued).
+            per-model default of the day for `model_name` on `agent` — the
+            value the next enqueue under this policy stamps (the same nuance
+            as `sandbox_provider` below, for the same reason: the default is
+            a platform table, resolved where an analysis is enqueued). Null
+            when the agent takes no effort.
         sandbox_provider:
           allOf:
             - $ref: '#/components/schemas/SandboxProvider'
@@ -13964,6 +14036,7 @@ components:
         - name
         - status
         - source
+        - agent
         - model_name
         - reasoning_effort
         - sandbox_provider
@@ -13992,11 +14065,14 @@ components:
         source:
           $ref: '#/components/schemas/CheckSource'
           description: What was checked; the row's DATASETS cell (the dataset form's `name@version`; an archive shows as an archive with its sha256).
+        agent:
+          type: string
+          description: The checker's agent, as Check.agent.
         model_name:
           type: string
           description: The checker's model; the row's AGENTS cell, shown like an arm with its effort under it.
         reasoning_effort:
-          type: string
+          type: [string, 'null']
           description: The effort every task's checker ran at, as Check.reasoning_effort.
         sandbox_provider:
           allOf:
@@ -14769,7 +14845,11 @@ components:
             that failed validation, its reasons in the message — never
             re-run: a timeout is deterministic, a second attempt cannot
             finish in less time than the first; the message names the
-            budget, the seconds used and the exit code), or an infrastructure
+            budget, the seconds used and the exit code), `agent_retired`
+            (an embedded analysis whose job's stored `analyze.agent` was
+            retired after the job was created: recorded when the trial
+            settles, nothing booted or spent, never re-run; the message
+            names the replacement), or an infrastructure
             stage of the analyzer run (`mint_key`, `boot`,
             `harness_install`, `agent`, `artifact_read`, `lease_expired`,
             ...) — infrastructure stages never auto re-run.
@@ -14783,7 +14863,7 @@ components:
         AnalyzeResult (their analyze/models.py: `summary`, `checks` keyed
         by criterion, `estimated_cost_usd`; `trial_name` is the enclosing
         trial) with each check carrying the platform's evidence list and
-        fourth outcome (AnalysisCheck); the rest is provenance — which model, rubric and prompt THIS analysis ran
+        fourth outcome (AnalysisCheck); the rest is provenance — which agent, model, rubric and prompt THIS analysis ran
         under, its lifecycle status, and its typed failure when it failed.
 
         `estimated_cost_usd` is the analyzer agent's OWN metered spend (its
@@ -14797,6 +14877,7 @@ components:
         - job_id
         - task_name
         - status
+        - agent
         - model_name
         - reasoning_effort
         - rubric
@@ -14841,6 +14922,12 @@ components:
             `not_applicable` are their exact words) and the ladder follows
             its object's dialect; the UI capitalizes on render, and nothing
             ever compares this ladder's words against another ladder's.
+        agent:
+          type: string
+          description: >
+            The agent THIS analysis ran on (`AnalyzeConfigInput.agent`, or
+            the default claude); every analysis from before the choice
+            existed ran on claude and says so.
         model_name:
           type: string
         reasoning_effort:
@@ -14850,8 +14937,8 @@ components:
             The reasoning effort THIS analysis ran at — passed to the
             analyzer explicitly, so it is what the model was asked for
             (`AnalyzeConfigInput.reasoning_effort`, or the per-model
-            default). Null only on analyses recorded before the effort was
-            stamped; never a guessed value.
+            default). Null when its agent takes no effort, or on analyses
+            recorded before the effort was stamped; never a guessed value.
         rubric:
           $ref: '#/components/schemas/Rubric'
         prompt:
@@ -15006,13 +15093,14 @@ components:
       type: object
       description: |
         Task-check configuration — Harbor's `harbor check` vocabulary
-        (their cli/analyze.py:84-148 check_command). The three rubric-agent
+        (their cli/analyze.py:84-148 check_command). The four rubric-agent
         knobs are the analyze door's, with the SAME rules and refusals
         (`AnalyzeConfigInput` states them; the param prefix here is
-        `check.`): `model_name` (`-m/--model`; Harbor's check default is
-        `claude-sonnet-4-6`, :101-103 — this platform's is the analyzer's
-        `openrouter/deepseek/deepseek-v4.1-flash`, one roster and one default for both rubric
-        agents, a recorded deviation), `rubric` (`-r/--rubric`; the default
+        `check.`): `agent` (`-a/--agent`, :100; default claude, Harbor's
+        claude-code), `model_name` (`-m/--model`; Harbor's check default is
+        `claude-sonnet-4-6`, :101-103 — this platform's is the analyzer's,
+        the agent's default per `AnalyzeConfigInput.model_name`, one rule for
+        both rubric agents, a recorded deviation), `rubric` (`-r/--rubric`; the default
         is the platform's check rubric, served by `GET /api/checks/defaults`;
         Harbor's own default-rubric.toml still runs when passed),
         `prompt` (`-p/--prompt`, the file's TEXT: it replaces the built-in
@@ -15035,10 +15123,9 @@ components:
         matched against the directory NAME.
 
         Harbor's other check options, and why they are not here: the same
-        list as the analyze door's — `-a/--agent` (the checker is
-        claude-code, fixed), `--job-name` and `-o/--jobs-dir` (a hosted
-        check is its own record, no wrapper job to name and no directory
-        to write), `-k/--n-attempts` (the one hosted auto re-run replaces
+        list as the analyze door's — `-o/--jobs-dir` (a hosted check is its
+        own record, no directory to write; `--job-name` is `name`),
+        `-k/--n-attempts` (the one hosted auto re-run replaces
         it; see `TaskCheck.attempts`), `-c/--config`, `--ak`, `--ae`,
         `--ek` (local-runner knobs with no hosted meaning), `-q` (a client
         rendering option, not a wire field).
@@ -15051,9 +15138,18 @@ components:
             The caller's name for the check — Harbor's `--job-name` on
             `harbor check` (cli/analyze.py:138). Omitted, the accept
             timestamp in Harbor's own shape, `YYYY-MM-DD__HH-MM-SS` (UTC).
+        agent:
+          type: string
+          default: claude
+          description: >
+            The agent the checker runs on — the analyze door's `agent`, same
+            rule: a retired built-in is refused `agent_retired` naming
+            `check.agent`, anything else off the list `invalid_input`.
         model_name:
           type: string
-          default: openrouter/deepseek/deepseek-v4.1-flash
+          description: >
+            Model the checker agent runs — the analyze door's `model_name`,
+            same rule: omitted, the agent's default; named, on the agent's roster.
         rubric:
           $ref: '#/components/schemas/Rubric'
         prompt:
@@ -15213,6 +15309,7 @@ components:
         - name
         - status
         - source
+        - agent
         - model_name
         - reasoning_effort
         - rubric
@@ -15239,11 +15336,14 @@ components:
           enum: [queued, running, completed]
         source:
           $ref: '#/components/schemas/CheckSource'
+        agent:
+          type: string
+          description: The agent every task's checker ran on — named at create, or the default (claude); every check from before the choice existed ran on claude.
         model_name:
           type: string
         reasoning_effort:
-          type: string
-          description: The effort every task's checker ran at — named at create, or the model's default, resolved at accept and stamped.
+          type: [string, 'null']
+          description: The effort every task's checker ran at — named at create, or the model's default on the agent, resolved at accept and stamped; null when the agent takes none.
         rubric:
           $ref: '#/components/schemas/Rubric'
         prompt:
@@ -15293,14 +15393,18 @@ components:
       type: object
       description: >
         The policy an empty `AnalyzeConfigInput` resolves to (GET
-        /api/analyses/defaults) — each key the same value `AnalyzeConfig`
-        echoes for a job created with `analyze: {}`, except `prompt`, which
+        /api/analyses/defaults), or one naming only `agent` when the door's
+        `?agent=` names it — each key the same value `AnalyzeConfig`
+        echoes for a job created with that config, except `prompt`, which
         `AnalyzeConfig` serves as null (the built-in body) and this serves as
         that body's template text.
-      required: [model_name, rubric, prompt, reasoning_effort, sandbox_provider]
+      required: [agent, model_name, rubric, prompt, reasoning_effort, sandbox_provider]
       properties:
+        agent:
+          type: string
         model_name:
           type: string
+          description: The model an omitted `model_name` takes on `agent` (`GET /api/meta` `analyze.agents[].default_model`).
         rubric:
           $ref: '#/components/schemas/Rubric'
         prompt:
@@ -15310,8 +15414,8 @@ components:
             `AnalyzeConfigInput.prompt` to run the default body explicitly,
             or edit it from here.
         reasoning_effort:
-          type: string
-          description: The effort the default model runs at when the config names none.
+          type: [string, 'null']
+          description: The effort `model_name` runs at on `agent` when the config names none; null if that agent takes none.
         sandbox_provider:
           allOf:
             - $ref: '#/components/schemas/SandboxProvider'
@@ -15320,14 +15424,18 @@ components:
       type: object
       description: >
         The policy an empty `CheckConfigInput` resolves to (GET
-        /api/checks/defaults) — each key the same value `Check` echoes for
-        a check created with no config, except `prompt`, which `Check`
+        /api/checks/defaults), or one naming only `agent` when the door's
+        `?agent=` names it — each key the same value `Check` echoes for
+        a check created with that config, except `prompt`, which `Check`
         serves as null (the built-in body) and this serves as that body's
         template text.
-      required: [model_name, rubric, prompt, reasoning_effort, sandbox_provider]
+      required: [agent, model_name, rubric, prompt, reasoning_effort, sandbox_provider]
       properties:
+        agent:
+          type: string
         model_name:
           type: string
+          description: The model an omitted `model_name` takes on `agent` (`GET /api/meta` `analyze.agents[].default_model`).
         rubric:
           $ref: '#/components/schemas/Rubric'
         prompt:
@@ -15337,8 +15445,8 @@ components:
             `CheckConfigInput.prompt` to run the default body explicitly,
             or edit it from here.
         reasoning_effort:
-          type: string
-          description: The effort the default model runs at when the config names none.
+          type: [string, 'null']
+          description: The effort `model_name` runs at on `agent` when the config names none; null if that agent takes none.
         sandbox_provider:
           allOf:
             - $ref: '#/components/schemas/SandboxProvider'
@@ -16141,7 +16249,10 @@ components:
             script exited non-zero, or installed no executable): the agent's
             fault, never retried automatically; the trial's stored stderr
             stream holds the install's output. Remove the registration, add
-            it again, and resume. `budget_exhausted` — the
+            it again, and resume. `agent_retired` — the arm's built-in agent
+            is retired (`GET /api/meta` `retired_agents`): refused before any
+            key or sandbox, never retried automatically; `failureDetail`
+            names the replacement. `budget_exhausted` — the
             one phase whose status is `BUDGET`, not `INFRASTRUCTURE_ERROR`
             — is a ceiling ABOVE the run's own per-trial cap refusing it:
             `failureDetail` carries the subject right after the
@@ -18203,6 +18314,18 @@ components:
         latest_version:
           type: [string, 'null']
 
+    RetiredAgent:
+      type: object
+      description: >
+        One retired built-in agent and the agent to use instead — the same
+        row the `agent_retired` refusal carries in `details.retired_agents`.
+      required: [agent, replaced_by]
+      properties:
+        agent:
+          type: string
+        replaced_by:
+          type: string
+
     AgentModelOption:
       type: object
       description: One model alias an agent offers, for a picker's option list.
@@ -18332,41 +18455,71 @@ components:
           type: integer
         agents:
           type: array
-          description: Built-in agents and their declared capabilities.
+          description: >
+            Built-in agents a new job may name, and their declared
+            capabilities. A retired agent is not listed here; it appears in
+            `retired_agents`.
           items:
             $ref: '#/components/schemas/AgentCapability'
+        retired_agents:
+          type: array
+          description: >
+            Built-in agents that are retired: a new job, resume, retry,
+            analysis or check naming one is refused (`agent_retired`), while
+            the jobs, trials and traces it produced stay readable under its
+            name.
+          items:
+            $ref: '#/components/schemas/RetiredAgent'
         analyze:
           x-wave: 3
           type: object
           description: >
-            The trace analyzer's roster and defaults: the model an omitted
-            `AnalyzeConfigInput.model_name` takes, the efforts
-            `AnalyzeConfigInput.reasoning_effort` accepts (the claude
-            harness's — the analyzer IS that harness), and every roster
-            model with the effort an omitted `reasoning_effort` takes for
-            it. Published because the effort changes what an analysis costs
-            and how it reads, and a client comparing two waves must know
-            what "omitted" meant in each.
-          required: [default_model, reasoning_efforts, models]
+            The rubric agents' roster and defaults (the analyzer's and the
+            checker's alike): the agents `AnalyzeConfigInput.agent` and
+            `CheckConfigInput.agent` accept and the one an omitted agent
+            takes (`default_agent`); the platform's model pick
+            (`default_model`); and per agent the model an omitted
+            `model_name` takes (`agents[].default_model`: the platform's
+            pick on every agent whose roster carries it, else the agent's
+            own), the efforts its `reasoning_effort` accepts and every
+            roster model with the effort an omitted `reasoning_effort`
+            takes for it. Published because the agent and the effort change
+            what an analysis costs and how it reads, and a client comparing
+            two waves must know what "omitted" meant in each.
+          required: [default_agent, default_model, agents]
           properties:
+            default_agent:
+              type: string
             default_model:
               type: string
-            reasoning_efforts:
-              type: array
-              items:
-                type: string
-            models:
+            agents:
               type: array
               items:
                 type: object
-                required: [alias, model_id, default_reasoning_effort]
+                required: [name, default_model, reasoning_efforts, models]
                 properties:
-                  alias:
+                  name:
                     type: string
-                  model_id:
-                    type: string
-                  default_reasoning_effort:
+                  default_model:
                     type: string
+                    description: The model an omitted `model_name` takes on this agent; one of its `models[].alias`.
+                  reasoning_efforts:
+                    type: array
+                    description: Empty for an agent that takes no effort.
+                    items:
+                      type: string
+                  models:
+                    type: array
+                    items:
+                      type: object
+                      required: [alias, model_id, default_reasoning_effort]
+                      properties:
+                        alias:
+                          type: string
+                        model_id:
+                          type: string
+                        default_reasoning_effort:
+                          type: [string, 'null']
         agent_registration:
           type: object
           description: Rules a bring-your-own agent registration must satisfy.