# CLI and agent authoring In the TUI, publishing is labelled **Create pipeline** for a new definition and **Update pipeline** for a new immutable version. The CLI `submit` command and `:submit` shortcut remain compatible. Templates offer description/graph panes: Tab expands the focused pane, `n/p` selects a template, and `hjkl` pans the focused graph. Create opens one confirmation with an editable, prefilled template name; Enter creates the pipeline without a separate naming or editor step. Duplicate names are rejected rather than updating an existing pipeline. `pipelines help` lists commands; `pipelines help edit` (or `edit --help`) shows command-specific flags and examples. Installed commands work from any directory. The installed command is a compiled Go binary. Optional autostart uses the OS user service manager directly. Legacy Python-managed installations require an explicit `PIPELINES_PREFIX`; no configuration is inferred from the executable's location. For source development, use `go run ./cmd/pipelines help`; start the native server separately. Set `PIPELINES_URL` or pass `--url` to select another server. ## Dashboard and status `pipelines --url ws://HOST:PORT/connect ping --json` returns `pong` with the running server's compiled `build.version`, full Git `build.commit`, and `build.dirty`. Uncommitted builds have a `-dirty` suffix. Missing metadata is explicit (`commit: "unknown"`, `dirty: null`); older servers omit `build`. This identifies the server binary, not the CLI or files mounted into the container. Plain `ping` also prints the server version/commit and, beside the local client build, the client/server relationship. `pipelines version --json` (or `pipelines --version`) identifies the local client without connecting to a server. Include both identities when reporting a problem; they can differ. `ping`, `connection`, `status` and every server command now check that relationship. The server stamps its `build` and integer `protocol` onto each response (the unsolicited welcome envelope too). The client compares this build against the server: a same-protocol build mismatch prints exactly one `warning:` on stderr (stdout JSON stays parseable) and a server speaking a newer protocol fails the command non-zero before it acts on the frame. If a response carries a top-level field this client does not model — for example an `artifacts` field from a newer server — the client prints a one-time `server sent unrecognized top-level field(s) ...` notice instead of silently appearing to have no data. Rebuild or reinstall the CLI so it matches the server; the installer also warns when a different `pipelines` resolves earlier on PATH than the freshly installed binary. Home's `r` opens **all** runs, not only failures or the newest 50. Inside a pipeline, `r` and `:runs` show only that pipeline's runs across its versions. In any run list, press `/` for fuzzy search, type to filter, and Enter to return to the results. Search includes pipeline name, version, status, outcome, stage and diagnostic; `#123` searches run-ID prefixes. Esc while searching clears the filter without closing the list. The filter survives inspection/back and live list refreshes. Home's `/` instead opens the omni popup, which searches pipelines, versions, runs, steps/visits, artifact names, and the text of stage inputs, outputs and evidence. Each row carries a short type badge (`pipeline`, `run`, `step`, `artifact` or `match`); a text match also shows a snippet with the matched part bracketed so it is legible without colour. A bare status word (`queued`, `running`, `succeeded`, `failed`, `canceled`, `interrupted`) or a date (`today`, `yesterday`, `this week`, or `YYYY-MM-DD` in local time) selects runs. Enter opens the matched pipeline, run, stage visit, artifact or field. The same query is available non-interactively as `pipelines search QUERY [--type KIND] [--limit N] [--json]`; `--type` is the only kind-narrowing mechanism. Wide terminals show runs and pipelines side by side, with stacked panels on narrow terminals. Home run cards prioritize reported results, status, completed visits and duration; active cards also show the current stage. Start/finish times remain in run details. The usage summary is aggregated over all stored history, including retry attempts: total/active runs, outcomes, success rate, average completed duration, pipeline usage and a zero-filled 14-day daily activity chart (UTC). Success rate is succeeded / (succeeded + failed); active, canceled and interrupted runs are excluded. The average includes all runs with a finish timestamp. It also totals the reported model usage across history: `tokens_total`, an in/out split only when every token-reporting stage reported one, and the summed cost, plus counts of stages that did or did not report a cost. A stage whose provider reported no cost is counted as cost-unknown and never contributes a fabricated or zero cost, and no cost/token count is inferred from output length or transcript text. The totals are the sum of the per-run totals `inspect` shows. `pipelines stats --json` exposes these aggregates. Use `:status` on any TUI page to open server and ntfy information. Scroll with arrows or `j/k`, PgUp/PgDn; `r` refreshes and Esc closes without changing the page. `pipelines status [--json]` returns the same server data, including hostname, start/uptime, the connected server endpoint/port, active interface IPs, ntfy publishing health, public phone URL/topic/port and candidate URLs. It does not start services or send a notification. The ntfy health probe has a two-second deadline; it checks the publishing server, not phone reachability. IPs belong to the **server computer or container**, not necessarily the client. Candidate IP/port combinations are not reachability checks: loopback bindings, firewalls, container port mapping and TLS can prevent access. ntfy candidates are only derived when a configured endpoint identifies this computer; remote ntfy hosts are not misrepresented as local. Publishing URLs and credentials are never returned. Missing public phone configuration is labeled explicitly. `pipelines service status` remains the separate OS autostart status command. ## Search `pipelines search QUERY [--type KIND] [--limit N] [--json]` runs one read-only query across the whole system: pipeline names and their versions, run identity, status and dates, stage and visit identifiers, artifact names and types, and the text of stage inputs, outputs, evidence and transcripts. It never reads artifact contents, never makes a per-result follow-up request, and asks for no confirmation. - Free text matches case-insensitively. A bare status word (`queued`, `running`, `succeeded`, `failed`, `canceled`, `interrupted`) selects those runs, and a lone date expression (`today`, `yesterday`, `this week` — Monday 00:00 local — or a literal `YYYY-MM-DD`) selects the runs created in that window; any other text, including a status word inside a sentence, stays free text. - `--type` is the only narrowing mechanism and takes one kind: `pipeline`, `version`, `run`, `step`, `field` or `artifact`. An unknown kind is an error, never a silently ignored filter. - Ranking: an exact name match (score 3) beats a name substring (2) beats a body-text match (1); equal tiers break by recency, then a stable kind/target order. Every result carries `score`, `matched_field` and `recency` so a client can re-rank. - Bounds: the result cap defaults to 50 and clamps to 200 (`0` = default); every kind scans at most 2000 rows, and field-text scanning stops at 2000 steps or 1 MiB decoded. `truncated` plus `scanned_steps`, `scanned_bytes`, `cap` and `scan_bound_bytes` report when a bound applied. - Targets are stable and navigable: `pipeline:`, `pipeline::v`, `run:`, `run::`, `run:::`, or an artifact URL served by the artifact endpoint. ```sh pipelines search "release notes" --json pipelines search failed --json pipelines search 2026-10-16 --type run --json pipelines search "acceptance check" --type field --limit 10 ``` ## Draft-first harness workflow ```sh pipelines new # the configured harness asks what to build and names it pipelines edit # fuzzy pipeline selection pipelines edit "bug fix" # ask the harness what to change pipelines edit "bug fix" "Add a review step" pipelines new "release notes" "Summarize changes and notify my phone" pipelines new "Add a feature" --template task ``` `new` opens the configured harness (Pi or OpenCode; host by default, or Docker-linked when a checkout is linked) immediately when its name/request is omitted. The skill asks only for missing decisions and records the chosen name in draft metadata. `edit` accepts an exact name or ID, with names taking precedence over numeric IDs. Both retain private drafts and launch the harness once with a short `@pipelines` reference and context-file path. There is no preliminary harness API call, temporary service, session creation, or instruction preload. The prompt's CLI prefix is Docker-aware: a Docker-linked profile embeds the in-container `pipelines` executable and, for this Compose service's own loopback endpoint, scoped transport aliases. Loopback is recognized only for parsed IP literals and the exact `localhost` name (case-insensitive), so a remote DNS name such as `127.remote.example` stays remote with no Docker probe. The single bounded `docker compose port` probe must return a complete `host:port` binding that covers the requested loopback address and port (wildcard or matching family); malformed or non-matching bindings fail clearly, and a loopback `wss` endpoint is rejected as having no supported local TLS mapping rather than downgraded to plaintext. nvim remains a client-side editor. `./install` installs/updates the global skill and format reference. Plain `go install` users run `pipelines skill install` separately. Shared login/config and user-modified skills are preserved; harness permissions remain in effect. The skill encourages editing only the draft, running server validation, summarizing the changes, and leaving submission to the user. This is an instruction, not a security sandbox: an agent with CLI/shell access can submit when explicitly asked. Every submission, including agent-issued `submit`/`rpc`, still uses server validation. ## Drafts in the TUI The server runs one active run per resolved workspace, including overlapping parent/descendant directories (and symlink aliases). Separate sibling worktrees remain independent. A new request key in a different workspace is admitted immediately and executes concurrently; a second run in a workspace an active run is already changing is refused, naming the conflicting run and advising you to wait or use a different workspace (for example another checkout of the repository). Reusing the original request key reattaches idempotently. `cancel ID` / `cancel_run` records `cancel_requested: true` before signaling the executor. The reply can still have `status: running` (or `queued`) and no `finished_at`: this means **stopping**, not stopped. The workspace stays reserved until execution returns and owned-session cleanup is confirmed; late success cannot override cancellation. A parked human run has no executor and cancels immediately. Repeating Stop while stopping is safe. If cleanup cannot be confirmed, the run stays nonterminal with `cleanup_unconfirmed: true`, retains partial step evidence, and reserves its workspace. The web client labels this explicitly and hides duplicate Stop. Successful startup recovery interrupts and confirms recorded sessions before settling the run; without a recoverable session pointer, uncertainty and the reservation survive restart rather than inventing stop confirmation. Recovery does not resume a canceled run. A database-sidecar advisory server lock is held through runner shutdown, so a second server on that database cannot recover or interrupt the first one's live runs. Keep the sidecar file in place; its existence is not ownership (the kernel lock is). This is a local-filesystem lock, not a distributed lease; hard-link database aliases and filesystems without reliable advisory locking are unsupported. Use `:templates` for four simple, manual-only starters: task, review, research, and plan/implement/review. **Submit template** (`s` or Enter) asks for a name without opening an editor. Enter saves a retained draft, then a separate confirmation publishes it as a new pipeline. Esc before naming creates nothing; Esc at publication keeps the draft. **Customize** (`e`) opens the editor picker instead. Both paths inherit the configured default model and preserve existing pipelines/history. The research starter requires configured web search support. Starter stages use goal-appropriate tools: the task worker and the three-stage implementer edit and run checks, the task, planner and review stages may search/fetch official sources only when external knowledge is genuinely needed, the three-stage implementer has no web tools and follows the planner's research, the standalone review stage never edits or runs commands and says so explicitly, and the three-stage reviewer independently inspects the real changes and receives the planner-owned acceptance checks rather than the implementer's summary, approving only when the applicable acceptance checks were independently observed passing and no unresolved substantive issue remains. The task and three-stage templates end in a committer stage that stages the runner's captured changed paths and commits locally without ever pushing; a genuine operator gate is a `type=human` decision stage, never a model stage pretending to approve. Editing a bundled template source changes no installed or published definition; a starter becomes an immutable version only after explicit authoring and publication. Home's `n` / `:new` and **Customize** (`e`) on a template first ask how to author the new pipeline: **nvim** opens a retained Markdown draft file, or a configured harness (Pi/OpenCode) opens the collaborative session with `@pipelines`. Tab/Shift+Tab or arrows select; typing fuzzy-searches the choices; Enter launches and Esc cancels without creating a draft. The choice overrides `VISUAL`/`EDITOR`. New drafts preserve the create-only publication guard and still require explicit submission unless you already published them from the harness. The configured harness can ask for a missing name; for an unnamed nvim draft, `pipelines submit --name NAME` names and publishes it. CLI `pipelines new` retains its existing direct-harness behavior; `--no-open` creates the file without launching an editor. Press `d` on the dashboard or use `:drafts` to browse the same retained drafts as `pipelines drafts`. Use `n/p` to select, `j/k` to scroll source, `e`/Enter to reopen the existing draft, `v` for server validation, and `s` for submission confirmation. `r` refreshes the list. Drafts can be validated, edited, and submitted to whichever server you are connected to. Same-server submissions preserve create-only/version guards. A different destination imports the draft as a new pipeline, ignoring the origin's version number and refusing an existing destination name; the confirmation shows the destination, and opening, validating or submitting a draft whose recorded server differs from the configured one prints a notice naming both URLs so it is never silently retargeted. Retained source and origin metadata stay unchanged. For an edit draft, CLI `submit --expected-version N` can explicitly approve an update against version N on the destination (or `--force` can explicitly skip that version check). For unnamed drafts, submission asks you to type a pipeline name directly in the TUI. Enter saves the name in retained metadata and opens publication confirmation; it does not publish yet. Esc keeps the draft without saving unfinished name input. Draft selection/editor completion never publishes it. On editor exit, the connected TUI checks the authoritative server before asking to submit. If the draft source is already published (including a newer version after the draft's base that was subsequently superseded), it skips confirmation and reloads the latest pipeline/page and Home list. Local-only changes still need explicit submission. A failed check retains the draft and says publication could not be verified; it never assumes publication from an editor's exit code. Original draft conflict guards and retained draft files remain unchanged. New/edit drafts open the configured default harness; `--harness opencode|pi|nvim` selects another configured harness or nvim. TUI pickers show only configured harnesses, default first, plus nvim. See [harnesses and runtime control](harnesses-and-control.md) for fuzzy model setup, separate defaults, shared memory and runner-owned limits. For agents or scripts that should not open a terminal: ```sh pipelines new "release notes" --no-open pipelines edit "bug fix" --json pipelines drafts --json pipelines validate /path/to/pipeline-123.md --json pipelines submit /path/to/pipeline-123.md --json ``` `--json` never opens an editor or picker; provide a name explicitly. It returns draft metadata including `path`, `name`, `server`, `create_only`, and `base_version`. `--no-open` prints the path and review commands. `$PIPELINES_DRAFT_DIR` overrides the shared cache directory. CLI and TUI drafts use private Markdown plus metadata sidecars; `submit FILE` can publish either. External files require `--name NAME`. New drafts use create-only submissions and cannot overwrite existing names. Edits carry a latest-version guard checked atomically by the server. On conflict, inspect/export the latest version and reconcile the draft. `--force` explicitly bypasses the version guard only, **never** syntax, graph, model, tool, schedule, or notification validation. Original versions/sources remain immutable. Submitting activates included schedules, so review them before publishing. ## CLI/TUI capability map | TUI action | Agent-capable CLI | | --- | --- | | Browse/search pipelines | `list --json`, `show NAME --json` | | Search everything | `search QUERY --json` (see [Search](#search)) | | Original source and history | `export NAME --version N`, `versions NAME --json` | | Browse templates/source | `templates --json`, `template KEY --source` | | Create or edit a definition | `new`, `edit`, `drafts` | | Validate and publish | `validate FILE`, `submit FILE` | | Start a run | `run NAME PROMPT [--title TITLE] [--model P/M] [--stage-model S=P/M] --yes --json` | | Run history/inspection | `runs --json`, `inspect ID --json` | | Rename a run | `title ID "TITLE" --json` | | Decide a waiting human stage | `decide ID [STAGE] --option KEY [--feedback TEXT] --json` | | Cancel | `cancel ID --json` | | Connection and phone settings | `connection --json`, `ping --json` | | Server/ntfy status popup | `status --json` | | Usage statistics/charts | `stats --json` | | Run-history retention | `retention report --keep N --json`, `retention apply --keep N --yes` | | Legacy operational diagnostics | managed `logs`, `doctor` | The TUI does not spawn CLI processes for each request. Both front ends share `internal/client`'s command transport and `internal/draft`'s persisted drafts; the server owns all validation, persistence, execution, and schedules. `rpc` exposes the full protocol as an escape hatch, so no server command is TUI-only: ```sh printf '%s' '{"type":"get_pipeline","pipeline_id":1,"version":1}' | pipelines rpc - ``` Existing protocol commands: `ping`, `list_pipelines`, `get_pipeline`, `list_templates`, `validate`, `submit`, `run_pipeline`, `list_runs`, `get_run`, `cancel_run`, `retry_run`, `decide_run`, `set_run_title`, `get_run_evidence`, `get_status`, `search`. `search` takes a `search` request object (`query`, with an optional `kind` and `limit`) and returns a `search_results` response whose ranked `results` each carry a stable `target` the client can open plus the typed payload and matched field. `list_runs` also returns all-history `usage` aggregates. See `internal/protocol/message.go` for JSON envelopes. Mutating RPCs have real effects; they are not previews. TUI-only animation, scrolling, and transient client logs are presentation, not separate server capabilities. ## Chat sessions Structured tools use `chat_answer` with `chat_answer: {session_id, form_id, answers, action?}`. Optional `action` is `accept`, `decline` or `cancel` only when advertised by the pending form's `actions`; ordinary question replies can omit it. Answers retain JSON types. OpenCode cancellation calls the existing form DELETE API. Codex MCP elicitation returns the corresponding action/content; other Codex questions accept answers only. Permission frames may provide `always_label: "Allow for session"` to avoid misrepresenting a session-scoped grant as a permanent policy change. Chat streams may send `chat_event.kind="resync"` to request an authoritative history refresh without inventing text deltas or ending a queued prompt. Native step markers may include `step.turn_id`; a terminal marker for an older turn must not clear a newer turn's Stop control or release its queue slot. Native context compaction is reported as `chat_event.kind="compaction"`, with `compaction: {status: "running" | "completed" | "failed", id?: string}`. Optional `native_turn_id` scopes the activity to its real harness turn; ignore activity for an older turn when a newer native turn is known. This is session activity and may have no `message_id`: clients must not invent an assistant message or tool call for it. Show **Compacting context…** while running; completion clears that phase but does not itself end the model turn. A new subscriber receives any retained active compaction. Clear stale activity before reconnecting, then preserve fresh replayed activity while history loads. This reports native observations only; it does not request or trigger compaction. The brokered OpenCode chat exposes the operator's own sessions, separate from the sessions pipeline runs create. The protocol types are: - `chat_sessions` lists the operator's own sessions only, ordered most-recently-used first. Each carries its `directory`, its `tag` and its recency. A session with no recorded tag is never listed, and no pipeline stage session is ever listed: stage sessions are still recorded in `run_steps`/`run_session_owners` for the run that owns them. The list is fail-closed: with no registry and no store it returns an empty list, never the endpoint's whole session list. A session is never marked `busy`: any number of clients may watch and write one. - `chat_subscribe` attaches a connection to one session's shared live stream (body `session_id`) and replays the session's still-pending approval requests, so every client that watches a session sees every frame any client's turn produces — the other devices' prompts and the turn's streaming output — and can answer a shared approval. Attaching is idempotent per connection; the subscription is released when the connection ends. - `chat_prompt` may come from any client. A prompt whose session already has a turn in flight is accepted into that session's bounded queue (currently 8) and acknowledged with `chat_queued: true`; the agent never receives two interleaved turns. The prompt is echoed to the other watching clients as a `chat_event` of kind `user`, carried with the sender's `client` label (a device name), while the sender's own transcript keeps its local copy. - `chat_reply` is exactly-once: the first answer to a pending request wins, is recorded with the answering client's label and is pushed back to every watching client as the same `chat_permission` frame with `resolved: true`, `decision` and `resolved_by`, so each card clears and shows the outcome. A later answer is refused as already answered (naming the client that won), never silently ignored and never delivered to the agent twice. - `chat_create` takes a `title` (default "New chat") and a `directory` (default `~`, the operator's home; `~`, an omitted or empty value and a `~/...` value all resolve to the home directory). A relative or missing directory is rejected before any OpenCode call. The created session is recorded with the stored tag `operator` and returned as the response's `chat_session`. An optional `ephemeral: true` creates the session for automated verification: it is created, bootstrapped and usable exactly like any other chat and can still be closed and reopened, but its registry row is hidden, so `chat_sessions` never lists it and a check that drives the real UI cannot leave a stray "New chat" in the operator's list. The flag only removes a chat from the shared list, so it cannot reveal or alter another session. - `chat_directories` returns the operator's home first, then the directories of previous operator sessions most-recently-used first, deduplicated and bounded (currently 8). It is a dedicated request so the create dialog can load the choices without first listing sessions; `chat_sessions` stays a pure session list. - `chat_adopt` takes a `session_id` that already exists on the configured endpoint and records it in the operator session registry, using the endpoint's own directory and title (never a guess or the server's working directory). The response echoes the adopted session as `chat_session` and, like `chat_create`, its registry row in `chat_sessions`, so `chat_sessions` then lists it and `chat_prompt` can continue it. A nil payload or an invalid/blank id is refused before any endpoint call; an id the endpoint does not report is refused naming the id and records nothing; an endpoint record with an empty directory is refused rather than substituting the home or working directory. The CLI equivalent is `pipelines chat-adopt ` (`--json` for the full response; otherwise a short `Adopted · ` line). ### Choosing the endpoint By default chat brokers from the single `opencode serve` instance the server spawns for itself, unchanged. To broker from the operator's own OpenCode service instead, set the opt-in `-opencode-endpoint` server flag or the `PIPELINES_OPENCODE_ENDPOINT` environment variable. Precedence is explicit flag, then explicit environment, then the endpoint recorded on a previous run; with none set the spawned default is used. The reference must be an http(s) URL with a host and no credentials, query, fragment or path. It is validated once at startup: the backend is detected from the service's own `/api/info` (V2) or `/global/health` (native) contract, and a streamable `text/event-stream` event path is required. A malformed reference, an unreachable host, a reachable non-OpenCode service, or a healthy service with no SSE each disable chat with an error naming the endpoint — never a silent fallback to another instance. An endpoint that authenticates accepts an optional password through the private file `PIPELINES_OPENCODE_ENDPOINT_PASSWORD_FILE` (a regular 0600 file; never an inline secret), and `PIPELINES_OPENCODE_ENDPOINT_USERNAME` overrides the detected default user. The validated choice is recorded so a restart with no flag or environment keeps brokering from the same place; an unreachable recorded endpoint disables chat with a clear message and no fallback. The endpoint choice is preserved by `pipelines service enable` autostart (the variables, with the password file as a reference). The stored tag is exactly `operator`; the display label `Operator` is derived from it, so there is no second tag vocabulary. Chat uses the server's one OpenCode transport. A session's stream is fanned out to every client watching it, prompts are serialised per session (queued behind the in-flight turn rather than rejected), and approvals are shared and exactly-once; the retained-state bounds, the resync-on-reconnect path, the exclusion of a run's own session and per-connection write serialisation are unchanged. The operator-session registry, filter, directory defaulting and the list bound are covered by the `pipeline`, `opencode` and `server` unit tests. The live end-to-end check is opt-in (`PIPELINES_TEST_OPENCODE_CHAT_LIVE=1`); it runs against a scratch store and never touches the deployed server. Android rendering of the tag is out of scope for the server change; the wire carries `tag: "operator"`. ## Scripted runs Every run records a short, single-line `title` saying what it was for. `pipelines run --title "..."` sets it explicitly; otherwise the server generates it from the prompt with the configured default harness. It is never empty: when generation is unavailable or times out, the server derives the prompt's first line, and `title_source` (`model`, `fallback`, or `explicit` for an operator-set title) plus `title_model` record how it was produced so a diagnosis can tell them apart. `pipelines runs` shows the title beside the pipeline name and version, and `pipelines inspect` returns the title, name and version, so two runs of one pipeline are distinguishable. Rename after the fact with `pipelines title ""`; the run's input, steps and execution are unchanged. The stored prompt is unchanged, and listing/renaming never calls a model. ### Model overrides at launch A pipeline version pins one model per stage. Without republishing, a run can choose different models at launch: ```sh pipelines run "bug fix" "Fix the layout" --json --yes \ --model openai/gpt-5 \ --stage-model review=openai/o4-mini ``` - `--model PROVIDER/MODEL` overrides the model for **every** model stage of that run. - `--stage-model STAGE=PROVIDER/MODEL` overrides **one** stage and is repeatable, so any number of stages can be overridden. A repeated key for the same stage is last-wins. - Precedence: a per-stage override beats the whole-run override, which beats the model pinned in the pipeline version. Overrides apply to that run only. The pipeline definition and its immutable version are never modified, `show <name> --version N` still returns the pinned source and graph, and the run records the `model` and `stage_models` it was launched with so it stays reproducible; `retry` reuses them. At launch every value must be a `provider/model` pair of the configured form and every stage key must exist in the pinned version — an unknown key fails naming the valid keys and no run is created. A per-stage key that names a non-model stage is refused because that stage never runs a model; `--model` on a pipeline with no model stage is accepted and inert. The effective model reaches the stage and its repair/formatter path from the same resolved definition, and each step's evidence records the provider/model it actually ran with (`inspect <id> --json` exposes per-step `provider`/`model`, and the run's graph shows the effective pair). Each model step also records the usage the harness reported, carried from the same accounting the stage already used for its budget: `usage.tokens_total`, the `tokens_in`/`tokens_out` split when the harness reported it (`split_known`), `cost_usd` when reported, and the explicit `tokens_known`/`cost_known`/ `split_known` markers. The run detail totals those stages as `usage`, so a mixed-model run shows each stage's model, tokens and cost and the run total. A stage whose provider reported no cost shows its tokens with `cost_known: false` (never a fabricated or zero cost), and a harness that reports only a total (Pi) shows `tokens_total` with the split marked unavailable rather than presenting the total as input tokens. The CLI is the deliverable; a TUI picker for choosing overrides is a documented follow-up, not half-built. On either the runs list or run inspector, press `P` (or use `:pipeline`) to open the selected run's exact pipeline version. Navigation does not cancel the run. Initial input includes `pwd`, the resolved run directory (normally the directory you invoked Pipelines from), alongside the task and any custom input fields. It is recorded by the server and shown in the run's Input section. Press `r` or use `:retry` on a failed run to restart it from the first stage. CLI equivalent: `pipelines retry <run-id>` (optional `--yes`, `--json`, `--wait`, `--wait-timeout`, and `--request-key`). The new attempt retains the initial input, directory, pinned pipeline version, and edit scope, and records `retry_of`. Original history is unchanged. Legacy runs gain `pwd` from their recorded workspace on retry. Retry replays the entire workflow; earlier edits or notifications are not undone and may happen again. A missing original directory blocks retry rather than silently using your current directory. Retry is hidden and rejected while a run is executing or agent cleanup remains active, even if a terminal status was already persisted. Run RPCs report `executing` alongside status so clients can distinguish cleanup from completion. A pipeline may contain a `type=human` decision stage (see the authoring skill's `references/format.md`). Its step becomes `waiting`, the run reports that it is waiting on a person, and no later stage runs. `runs`/`inspect --json` show the `waiting` status, the stage and its recorded instructions and options. Answer it with `pipelines decide <run-id> [stage] --option KEY [--feedback TEXT] [--json]` or, in the run inspector, press `c` to open the option picker, choose with `↑/↓`, type optional feedback with `Tab`, and confirm with Enter. The choice and feedback are recorded as the step's evidence and the run resumes along `edges[KEY]`. An unknown option is refused with the valid list, a decision to a stage that is not waiting is refused, and a repeated decision for an already-decided visit does not double-apply. The wait is persisted in the run rows, so it survives a server restart and `decide` still works afterwards; `cancel` and `retry` keep working while waiting and after a decision, and a human stage is never killed by a stage timeout or `max_visits`. The run inspector's upper half shows the entire graph from the run's immutable pinned version using the shared renderer and viewport. Executed stages and taken routes retain their colors; unexecuted stages, unchosen routes and unreached terminals are dimmed. Queued runs show the full graph before the first visit. Repeated visits remain individually inspectable in the stage sidebar, while the graph highlights every audited route through a loop. On small terminals details take precedence. Older servers without a graph fall back to a matching loaded definition or the recorded execution path, never a different pipeline/version. Press `g` to focus the graph, arrows to pan, and `j/k` or PgUp/PgDn to scroll. Press `g` or Tab to return to details. When not focused, the graph follows the latest visit or selected stage. Existing polling refreshes the graph once per second, and running durations show elapsed time rather than “in progress.” The server can also push run changes. A client sends `subscribe` scoped to a workspace (and optionally to explicit run ids); the server first returns the current state of the matching runs, then a compact `run_event` delta on each transition — which run changed, its status, outcome, current stage, completed-step count and timestamps, and, when a single step changed, that step's sequence, stage, status, outcome, target and diagnostic. It is a delta, never the whole run, and rapid changes are coalesced to at most a few frames per second per run. `unsubscribe` or disconnecting stops the frames. This lets a client drop the once-per-second poll: the Android app consumes these pushes instead, while the TUI still polls in this release. Run-list summaries and subscribe snapshots additionally carry compact `graph_evidence: {complete: true, visits: [...]}`. Each visit contains only `sequence`, `stage`, `status`, `outcome` and `target`: the last visit of each stage and each distinct taken route, ordered by sequence. Repeated loops do not grow the summary indefinitely, and private step inputs, outputs, transcripts and diagnostics are excluded from this projection. Live `run_event.graph_visits` preserves changed visits across coalescing (unlike the single latest `step`). Clients merge these by sequence and retain last-stage/distinct-route evidence. Missing evidence from an older peer means unknown, not that every stage shares the overall run result. These fields are additive; older peers ignore them. In the run inspector, `n`/`p` select the next/previous recorded execution visit, including repeated visits in loops, from any pane. Selection stops at the last visit or overview; it does not wrap. The selected stage has a full-row highlight; the focused stage/details pane has a highlighted title and pointer. In details, the section containing the top visible row is highlighted and named in the pane title. `Tab` cycles stages → details → graph (Shift+Tab reverses); `g` switches graph focus. Each graph stage title shows live progress from its latest execution visit. New run pages open with the graph focused and using roughly three quarters of the available height. Tab expands the focused pane and contracts the others; the stage list also widens when focused, and shrinks to a sidebar for details. Pending stages are dimmed. The graph is centered horizontally/vertically when it fits; the box-and-arrow diagram is retained at every viewport size, never replaced with a stage list to fit. Oversized diagrams remain pannable with `h/j/k/l` or arrows while the graph is focused; `c` recenters the selected/current stage. `P` opens the pinned pipeline (lowercase `p` remains an alias only on the runs list). The pipeline page also gives most height to its focused graph or stage-details pane; Tab switches focus and resizes the two panes. Focused titles use the same background highlight as the run page. Graph boxes are sized to their contents rather than the longest label anywhere in the graph. Fork columns share only their own required width, route tracks are adjacent, and branch spacing follows stacked boxes rather than tall trunk/merge nodes. All stages, outcome labels, ports, and terminal arrows remain visible in the full pannable diagram. Incoming route IDs are shown on the left and outgoing outcome labels on the right of shared rows, without separate input/output sections or divider rows. This reduces height while allowing enough width to keep both labels readable. Failure-bound error exits remain at the bottom. Pipeline and run graph panes both center horizontally and vertically when the diagram fits, and clamp scroll/pan offsets on resize or graph changes. Press `a` to select the active/latest retained configured-harness agent session and open its tool evidence. Session ID is available during execution; tool transcripts appear when persisted (currently after generation). This does not attach a live harness client or stream its messages. Supporting live attachment consistently across the shared and privately owned harness backends is deliberately deferred. ## Stage turn files Each model-stage visit receives a file-based turn under the run's artifact directory. `pipelines inspect <run-id> --json` and the run inspector list these files as that step's artifacts: ``` <artifacts>/<run>/stages/01-implementer/input.md <artifacts>/<run>/stages/01-implementer/memory.md <artifacts>/<run>/stages/01-implementer/output.md <artifacts>/<run>/stages/01-implementer/private/ <artifacts>/<run>/public/ ``` `input.md` is the run prompt on the first visit and the previous visit's output body (without its outcome section) afterwards. `memory.md` is the accumulated markdown of completed turns plus the files each turn added to the shared `public/` directory. A text-only stage (an explicit empty `tools` block) receives both files' contents in its prompt instead, because it has no file-reading tool. The model returns the complete `output.md` markdown; only its `# outcome` section is interpreted by the server. Everything else is free-form and preserved under the step's `markdown` output. `private/` belongs to its own visit, while `public/` is shared by every stage in the run. Artifacts are served from `/artifacts/<run>/<step>/<name>`; shared files appear with a `public/` prefix. ### Run notification modes and receipts `--notify auto|on|off` records `notify_mode` on the run (also returned in summaries and `run_event` frames). Auto retains stage policies and run-level waiting/failure defaults. On also alerts terminal success without requiring a stage hook. Off suppresses stage, run-level system and browser tray alerts. `inspect --json` / `get_run` includes top-level `notifications` receipts with `sequence`, `event`, `channel`, `status`, `attempts`, and optional `diagnostic` and `message_id`. Status is pending, sending, sent, failed, or covered (already published by the stage). Run-level system alerts persist and retry at most three times, recovering pending work after restart. A publisher acknowledgement is not proof of reception; a crash before saving it can duplicate a message. See [`notes/audit-notifications.md`](notes/audit-notifications.md) for limits. ## Investigating runs with a configured harness On the runs list or inspector, press `D` or use `:debug`; CLI equivalent: `pipelines debug <run-id>`. This opens a fresh conversation in the configured harness (Pi or OpenCode), not the implementing session. It supplies exact commands for `inspect <run-id> --json` (including stage history, exact outputs, diagnostics, notifications, session IDs and bounded transcript previews), `show <pipeline-id> --version N --json`, and `runs --pipeline <pipeline-id> --json`. The commands reuse the same Docker-aware CLI prefix as authoring, so a Docker-linked profile runs them in-container against the original logical endpoint. All retrieval remains available through the CLI and can be repeated while execution is active. No new model API or live attachment integration is required. The injected server-side session evidence is the latest bounded, credential-redacted evidence for up to the 3 most recent session-bearing steps; when a preview is bounded or omitted, inspect the retained session itself rather than assuming a complete tool history. `debug --json` returns the context and prompt without launching the harness; `--no-open` prints the prompt. `--harness opencode|pi|nvim` selects a configured harness or nvim; `--opencode` selects the host executable. Debugging works for active and terminal runs and does not automatically retry, cancel, or resume them. If the recorded directory is unavailable, the harness launches from your current directory and retrieves historical evidence through the CLI. ## Editor choice and draft cleanup On the pipeline detail page, `e` / `:edit` opens an editor picker: **nvim** or the configured harness (Pi/OpenCode, default first). Type to fuzzy-search; Tab/Shift+Tab (or arrows) choose; Enter opens; Esc cancels without creating a draft. Explicit choice overrides VISUAL/EDITOR. Both choices edit a retained copy with the normal version guard, never the published definition. Submission remains separate and explicit. In the draft browser, `x` asks to delete the selected draft; `y` confirms and Esc or `n` keeps it. The CLI equivalent, `pipelines draft-delete <draft.md>`, is an explicit deletion command and needs no server connection; `--json` reports the result. It asks for confirmation; `--yes` explicitly approves scripted deletion. Only managed drafts inside the configured draft cache are accepted. Source and metadata symlinks/directories and external Markdown files are rejected. Published definitions and run history are not deleted. Cleanup can also remove orphaned metadata when a draft source was already removed. ```sh cd ~/repos/project pipelines run "bug fix" "Fix the layout" --json pipelines run "bug fix" "Fix the layout" --json --wait --wait-timeout 10m pipelines run "bug fix" "Fix the layout" --json --request-key my-unique-task ``` Runs default to the client's current directory and latest accepted version. The server resolves that directory without a repository allowlist. `--version` chooses an immutable version in JSON mode. Starting a run enables its stage-declared tools without a second approval. Editing supports all file types and new files within the run directory; no file list is required. Shell commands use the repository's actual tooling. `--write-paths` is an optional narrower scope, not a prerequisite. `--yes` skips action confirmation for explicitly approved automation; `--allow-tools` remains a compatibility flag and is not required. `--wait` polls until terminal status, returns nonzero on failure/cancellation, and does not cancel a run when waiting times out. `--wait-timeout` bounds only the client wait after admission and is independent of the per-request `--timeout`; a reached wait limit leaves the admitted run active and never cancels or replays it. When it is reached, inspect read-only with `pipelines inspect <run-id> --json`, or keep waiting by rerunning `--wait` and reusing the original `--request-key` rather than starting a duplicate run. Reuse `--request-key` to prevent duplicate work. A supplied pipeline/task opens an action-confirmation popup before starting with a live inspector (unless `--yes` was supplied). When a task is omitted, Enter inserts a newline and Ctrl+S opens the confirmation for the composed task. There is no file-list question or typed RUN confirmation. Shell runs with your user's authority: the run directory and tool policy are not an OS sandbox. Run, retry, cancellation, CLI publication, and draft deletion require confirmation. CLI prompts go to stderr and default to No; use `--yes` for non-interactive approval. The TUI's run/retry/cancel popup accepts `y`; Enter does not implicitly approve, and `n`/Esc cancels without sending the action. Retry restarts with new sessions and original input/pwd, never the failed conversation; side effects may repeat. Publication retains its explicit submission confirmation and version conflict guards. Quit with `:quit`, not `q`; Ctrl+C remains an interrupt. JSON results go to stdout, errors to stderr (JSON error envelopes when requested), and failures exit nonzero. Export/template `--source` deliberately prints raw Markdown. `--timeout` bounds each server request; `--wait-timeout` bounds the optional client wait independently and, when reached, leaves admitted work active. Neither changes the stage execution timeout. ### Stage limits Model-stage `configs` support these optional settings: ```text retries=2 timeout=1h budget_usd=1.50 budget_tokens=50000 ``` Retries default to **2** (three total attempts); timeout defaults to **one hour per attempt**. Omitted budgets are unlimited; configured budgets must be positive. Each budget covers all attempts, format-only repairs, and repeated visits to that stage in one run. A new run starts new accounting. Tokens include input, output, reasoning, and cache reads/writes. USD amounts use OpenCode's reported cost. Missing or invalid required accounting stops the stage rather than counting zero. Usage is polled during execution and read again afterward; an in-flight request may exceed a limit before usage arrives. These are not provider-side hard caps. Automatic retries cover retryable transport failures and per-attempt timeouts. An explicit read-only stage (a tools block listing neither `edit` nor `shell`) may also retry a post-admission deadline or transport failure and re-emit a malformed final result up to `retries`; a stage whose tools include `edit`/`shell` or that has no tools block never replays admitted tool work or invalid results. Budget failures always stop. Cleanup must be confirmed before another attempt. Existing explicit `retries=0` settings remain unchanged. The old 30-minute run deadline no longer truncates stage timeouts; the 30-step and five-visits-per-stage limits still apply. Native ntfy stages also accept `timeout` (default `1h`), but never automatically retry potentially delivered notifications and do not accept model budgets. Pinned definitions contain the configured limits. Run transcripts retain limits, attempt counts and reported budget usage, also shown in the TUI's Evidence pane. The same reported usage is recorded on the run as typed per-step evidence and totalled for the run, so `inspect <id> --json` exposes tokens in/out, the total, the cost when reported and the model each stage ran with. A stage revisited within a run shares one budget, so each step records only that visit's increment of it and the run total is not the cumulative budget counted again. Missing or unknown cost or split is reported as unavailable (never zero or inferred), and a historical or asked-for run recorded before this evidence existed simply omits `usage`. Exact stage outputs are authoritative; a transcript over the 32 KiB retained bound is stored as a valid, credential-redacted preview envelope that preserves available stage limits and failure/format diagnostics and names the retained session instead of slicing to malformed JSON. ### Result-first views Home shows retained result previews and completed visits alongside each run and pipeline. Pipeline pages show the latest result for the **selected version**, not a result from a different published definition. Run opens focused on results, with output before input and the graph below. Failed or ongoing runs preserve partial output and completed stage reports without labeling them final success. These previews quote recorded output; they do not independently verify model claims. Tab and Shift+Tab toggle **Details ↔ Graph** only. The visit selector lives inside Details; `n`/`p` still navigate recorded visits, including repeated stages, and `0` returns to the overview. `j`/`k` scroll the report; the focused graph uses `hjkl` for panning and `c` to center. Secondary progress/input/evidence/execution sections are collapsed initially. Small inactive graph previews collapse instead of showing severed boxes. See [UI/UX review](ui-ux-review.md) for screen-by-screen feedback. ## Run retention Nothing prunes run history automatically. `pipelines retention report` and `pipelines retention apply` are the only commands that remove run history or run artifacts, they are local and operator-controlled, and they never run as a side effect of any other command, server startup, `newRunner`, the scheduler, `list_runs`, `stats`, `runs` or `inspect`. Pipeline definitions and their immutable versions are never touched: retention is about run history and the artifacts runs produced. ```sh # Read-only: show exactly what a policy would remove. pipelines retention report --db ~/.local/share/pipelines/data/pipelines.db --keep 200 --json # Explicit removal of that same set (artifacts go with their run). pipelines retention apply --db ~/.local/share/pipelines/data/pipelines.db --keep 200 --yes --json ``` Both commands accept `--older-than DURATION`, `--keep N`, `--db FILE`, `--artifacts DIR` and `--json`; `apply` also accepts `--yes`. At least one policy flag is required; an unknown subcommand, a negative count and a non-positive age are refused. `--db` must name an existing database and is never created. With only `--db`, the artifacts root defaults to `artifacts/` beside the database, matching `cmd/pipelines-server/main.go`; `--artifacts` overrides it. The commands are dispatched before server URL resolution, so they work without a running server (for example `--url not-a-server`) and never require, restart or redeploy the deployed server. ### Policy - `--older-than` selects runs created **strictly before** `now - DURATION`. The age is a rolling UTC duration, not a calendar-day boundary. It accepts Go durations plus a count of days or weeks (`720h`, `30d`, `12w`). - `--keep N` retains the newest `N` runs by run id and selects ranks `N+1` and beyond. `--keep 0` selects every eligible run; a count larger than the history selects none. - Both together are the conservative **intersection**: a run is selected only when it is both older than the age bound and outside the newest `N`. ### Protected runs A policy-selected run is nonetheless never removed when it: - is `queued` or `running`; - is `waiting` on a person (a `waiting` run or a `waiting` step); - has any step with `agent_active=1` (cleanup still active); or - is referenced by a run that will remain through retry provenance (`run_retries.original_run_id`), propagated transitively along a retry chain. If both an original and its retry are selected, no retained link remains and both may be removed. Protected runs are reported with their reason and their artifacts are never counted as reclaimable. Free-text run ids in prompts are not treated as dependencies; retry provenance is the only structured cross-run dependency the store records. If a selected run becomes protected between the report and the delete — a new retry reference, a waiting/active state or new agent activity — `apply` aborts with the run ids and reasons, deletes no rows or artifacts, and un-stages anything already renamed. A removal never half-completes. ### What apply removes and reclaims `apply` stages only the would-remove run directories into one `artifacts/.retention-staging-<timestamp>-<pid>/` directory (a same-filesystem rename), deletes the rows for those runs in one immediate write transaction (`step_notifications`, `step_file_changes`, `step_control`, `step_artifacts`, `step_verdicts`, `step_session_exports`, `step_usage`, `run_session_owners`, `run_steps`, `run_retries`, then `pipeline_runs`, batched at 500 ids), commits, and removes the staging directory. A symlink at `<root>/<id>` is removed as a link without touching its target; a regular file at `<root>/<id>` is refused. A symlinked artifacts root is followed consistently: the report sizes it and apply stages through the same directory, so the artifacts really go. All regular files under a removed run directory are counted, including files beyond the server's 200-file/8-MiB/64-MiB inventory caps, which bound the recorded inventory, not disk ownership. The database is then compacted with one `VACUUM` and one `PRAGMA wal_checkpoint(TRUNCATE)`, retried at most three short times on `SQLITE_BUSY`. `database_bytes_reclaimed` and `space_reclaimed_bytes` are measured, not estimated. If `VACUUM` fails the command returns an error, sets `vacuum_error` and claims no reclaimed database pages while still reporting the artifact bytes it did remove; a later `apply` with nothing left to remove can retry the `VACUUM`. If staging cleanup fails the report still gives the removed run count and totals, names the staging path and the remaining bytes/files, and does not count the leftovers as reclaimed. ### Measured example (this machine) Reported before, applied, reported after, against the live database while its server kept running (server PID 99492 unchanged throughout; the server was also creating new runs concurrently, which is why the artifact `du` moved independently): ```text report --keep 200: total 207 · would remove 7 · protected 0 · retained 200 artifacts 418 files · 279668627 bytes (266.7 MiB) database 27897856 bytes · free 0 · reclaimable 279668627 apply --keep 200: removed 7 runs · 418 files · 266.7 MiB vacuumed · database reclaimed 1.5 MiB · now 25.1 MiB · free 0 total space reclaimed 268.2 MiB report --keep 200: total 200 · would remove 0 · retained 200 artifacts 0 files · 0 bytes database 26353664 bytes · free 0 pipelines 19 · pipeline_versions 45 (unchanged) ``` Independent observation: `du -sk` of the artifacts root fell from 2815316 to 2572456 KiB (the running server added artifacts during the same window), the database file from 27897856 to 26353664 bytes, the WAL was truncated to zero and the logs stayed at 18404 KiB. A retained run's artifact still served (HTTP 200, `/artifacts/100/1/00_run94_artifacts.png`) while a removed run's URL returned 404 (`/artifacts/83/1/...`). The measured numbers above come from that before/apply/after output, not an estimate. ## Verification scope CLI integration tests exercise draft/validate/submit/history/conflict handling and scripted run/inspection/cancellation through a real local WebSocket server (the run test uses a controlled executor, not a model). The isolated installed CLI also exercised the actual OpenCode API's session creation and skill activation, while capturing interactive launch arguments to avoid a model turn. Full model-driven authoring and Linux execution are **not tested**. ## Per-run report contract `pipelines run NAME PROMPT --output-format html` requires a server-written `public/report.html` before run success; `--output-format markdown` requires `public/report.md`. Omit the option for historical behavior. The `run_pipeline` wire request and returned run record carry `output_format` (`html` or `markdown`), and retries inherit it. Request-key replays must retain the same format. Reports use nonempty terminal text output and are bounded to 2 MiB; missing, empty, unsafe or unverifiable files fail the run. This checks structure, not factual correctness. See [audit notes](notes/audit-deliverables.md). `pipelines repair-session-export RUN_ID SEQUENCE [--json]` asks the runtime to read back that settled visit's exact recorded native session/backend. It never executes a stage or changes a definition. It writes a versioned repair artifact and appends evidence without replacing original files or outcomes. The protocol command is `repair_session_export` with `run_id` and `export_sequence`; the reply is a run detail containing each visit's `session_export_repairs`. CLI JSON emits the repair record with explicit completeness `status`, and incomplete repair exits nonzero. Recovery still depends on upstream session availability. Incomplete native stage-session exports also recover automatically after the run settles: up to three durable runtime attempts, with backoff and restart-safe leases. `inspect RUN_ID --json` includes each visit's `session_export_recovery` status/attempt count/diagnostic and preserved `session_export_repairs`. Exhausted automatic recovery remains a warning; manual versioned repair is still available.