tascan_create_cycle
Start an unattended build-review-decide cycle (protocol v0.2). Queues T1 CODE: (or SHELL:) with your build_brief on the AI Inbox and T2 REVIEW: with your review_brief, born blocked on T1. T1 (2026-09-24 throughput): kind=review_only mints T1 as CAPTURE: instead — the executor stores the bundle from the repo at HEAD without a model call and the review runs on that; dry_run=true runs the dispatch PREFLIGHT only (nothing queued) and prints every problem at once (codes title_too_long, brief_rule, brief_names_unbundled_path, artifact_over_cap, migration_as_context (tascan repo only), idempotency_replay, kind_invalid — bad kind, bad revisable_by, review_only + task_type SHELL/RESEARCH, review_only + max_questions 0), then what the API could NOT check (unchecked[]: artifact_missing_at_head always — only the executor sees the repo, and a CAPTURE naming a missing path fails with "[artifact_missing_at_head]" in its error; artifact_over_cap for any path without a byte count — pass artifact_bytes; idempotency_replay only if its lookup failed) and warnings[] (a build_brief naming a context file outside the bundle); a real create that the API refuses prints the same problems[] list (when the transport hands the tool only the first problem's text, the tool re-runs the preflight and prints the whole list). The local executor builds, stores the exact bytes of artifact_paths as a bundle (build_ref = sha256 over the manifest), the independent reviewer reviews THAT bundle, an approve verdict mints a Decision task for the human authority (one SMS), Approve mints an Integrate task for the deploy id. Revise verdicts spawn revisions (cap max_revisions, default 3); reject, human Reject, scope violations or exhausted revisions PARK the cycle (a Parked task with Resume with notes / Close). REQUIRES agent:dispatch:code. Duplicate protection: the same idempotency_key, or (keyless) the same briefs + paths on a live root, within 24 h is refused by the preflight as idempotency_replay (400, nothing queued, the existing root_id in the problem) and printed as DUPLICATE with that root_id; only a replay the preflight could not see (a race) comes back from the RPC as 200 created=false, printed the same way. Optional reviews[] attaches a multi-lens review panel (design item 14a) in place of the single OpenAI review — one review task mints per lens and every blocking lens must approve before the Decision task mints. Track with tascan_get_cycle_report (root_id). Nothing spawns a cycle on its own.
Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| kind | No | Cycle kind (default build). review_only = no model builds anything: T1 is "CAPTURE: <title>", a local-executor task that stores artifact_paths from the repo at HEAD as the bundle (same build_ref, size cap and scope check as a CODE build) and completes in seconds; T2 REVIEW then reviews that bundle. Use it to get an independent review of documents or of code already committed by hand. Stored as coord.kind_of_cycle on the root. | |
| repo | No | Which codebase on the executor the build runs in — an alias from the executor's allowlist (tascan-agent/repos.json), e.g. "tascan" (default), "merchskipper", "inkskipper", "rangerlizzy", "cardvault", "take1", "fitcheck", "safegrid", "eternity", "maniacs". artifact_paths are relative to that repo. An unknown alias is refused by the executor (repo_unknown) and the cycle parks. | |
| title | Yes | Short human title (1-200 chars). T1 becomes "CODE: <title>" ("CAPTURE: <title>" when kind=review_only), T2 "REVIEW: <title>". | |
| dry_run | No | true = PREFLIGHT only: the API validates the whole dispatch and returns every problem at once (code, path, detail) plus the unchecked[] list (what only the executor / the RPC can decide) and warnings[] — nothing is queued, no idempotency key is consumed. Recommended before every real create. Default false. | |
| reviews | No | Optional multi-lens review panel (design item 14a) instead of today's single OpenAI review_brief lens — 1-8 entries, each: {lens: slug matching ^[a-z][a-z0-9-]{0,39}$ unique per array (e.g. "code-correctness", "security", "ui-ux"), brief: 1-8000 chars of cycle-specific context for that lens, provider: "openai"|"anthropic"|"gemini", model?: 1-100 chars, blocking?: boolean (default true — at least one entry must be blocking or the whole array is refused), max_tool_calls?: integer (openai only)}. One REVIEW: task mints per lens, all blocked on T1; the Decision task mints only once every blocking lens has approved the same build. Omit for today's single OpenAI code review. | |
| task_type | No | T1 prefix (default CODE). CAPTURE is never caller-settable: kind=review_only makes T1 "CAPTURE:" itself. With kind=review_only the value is sent as given — CODE is accepted (T1 is still CAPTURE), SHELL / RESEARCH are refused by the API as kind_invalid. | |
| project_id | Yes | Working project (UUID). Its Decisions and Agent Questions lists are created on the first cycle (coord_ensure_lists). Requires the org human authority to be configured (coord_set_authority) — otherwise 403. | |
| build_brief | Yes | The executor prompt (1-40000 chars). Executed verbatim by the local Claude Code runner as a CODE:/SHELL: task — write it as a complete instruction, name the files, forbid nothing the runner already forbids (deploy/commit/push are refused by its context). | |
| preview_url | No | Optional https preview link shown on the Decision task. | |
| review_brief | Yes | The reviewer prompt (1-40000 chars). The reviewer reads the stored bundle (tascan_get_build / tascan_get_build_file), the task text and the filtered trail; it returns approve, revise, reject or needs_input as strict JSON bound to the build_ref. Still required even when `reviews[]` is given (kept as T2's legacy description text). | |
| revisable_by | No | review_only cycles only — who fixes a revise verdict. dispatcher (default): the cycle parks on revise as today and you re-dispatch. executor: a revise spawns a revision CAPTURE task (up to max_revisions) that is born blocked on a question task "Revise the documents, then answer to release" addressed to you — fix the files in the repo, answer that question (tascan_post_message kind=answer), and the capture re-runs on the fixed HEAD. Build cycles always revise through the executor regardless of this field. | |
| max_questions | No | Questions a runner may ask per task before the attempt fails (default 3). | |
| max_revisions | No | Revision cap (default 3): at most max_revisions + 1 builds and reviews. | |
| artifact_bytes | No | Optional {path: bytes} for artifact_paths — the byte length of each file in your checkout. The API cannot read the repo, so artifact_over_cap (262144-byte reviewable cap per file) is only checked for paths you give a byte count; every other path is reported under unchecked (artifact_over_cap, unchecked_at_api) — never guessed from an earlier stored bundle. Same on dry_run and the real create. | |
| artifact_paths | Yes | Repo-relative paths the build binds (1-64). Exactly these files are stored as the bundle and hashed into build_ref; files the executor touches OUTSIDE them fail the scope check and park the cycle. No .., not absolute, no drive letter, no backslash. | |
| idempotency_key | No | Optional caller key (≤ 200 chars). The same key within 24 h is refused as idempotency_replay (nothing queued) and printed as DUPLICATE with the existing root_id — safe to retry after a lost answer. | |
| integrate_title | No | Optional title template for the Integrate task (default "Integrate: <title>"). | |
| checkpoint_title | No | Optional title template for the Decision task (default "Decision: <title>"). | |
| max_cost_micro_usd | No | Per-cycle spend cap summed over every attempt, in micro-USD (default 5000000 = USD 5). A claim that could overrun it is refused (budget_exhausted). | |
| integrate_description | No | Optional description template for the Integrate task. | |
| checkpoint_description | No | Optional description template for the Decision task; the build_ref, preview and findings summary are appended. |