Six computed Flutter runs stopped at the export handoff gate because the rendered session.html cap was 2 MiB while their complete sessions (well inside the 8 MiB source read bound) rendered between 2.25 and 4.28 MiB once html escaping and the readable projection duplicated native text. Raise SessionExportFileLimit to 8 MiB, the server's existing per-artifact file bound, so a complete export is still listed and served. Normal export verification, versioned repair and the handoff gate keep using the single shared constant. The source read bound is unchanged, and an over-cap render is still visibly truncated and recorded partial. Add focused regressions for a complete export above the old cap, an over-new-cap file rejected as partial, repair above the old cap leaving original evidence untouched, the gate accepting the larger repair, and the constant staying inside the artifact file bound. Update the artifact, handoff-gate and operations docs and record the diagnosis in docs/notes/session-export-render-cap.md.
31 KiB
System, storage and run evidence
Pipelines turns a reusable process into sequential stages with explicit handoffs, outcome routing and retained evidence. A Go CLI/TUI connects to an execution server; the server owns execution and SQLite history, OpenCode owns model/tool sessions, and optional ntfy delivery supplies notifications. It is technology-neutral, not a parallel/distributed task engine or a guarantee that model reports are correct.
Lifecycle and location
- Drafts are editable Markdown plus JSON metadata. Publishing validates and creates immutable versions; drafts remain retained. Keep context create-only/base-version guards. Five optional generic starters are task, three-stage, review, research and debug; a fresh database starts empty. Templates can be created directly with an editable prefilled name or customized first. Editing a bundled repository template changes no installed or published definition; only explicit authoring then submission creates a new immutable version.
- A run pins its exact definition version, initial input and resolved directory. The client invocation directory is the default, but files/tools execute on the server. Docker needs that directory mounted at the same absolute path. A remote client does not upload its repository or make host credentials available remotely.
- Host and Docker OpenCode credentials are separate. Preserve setup-selected model inheritance; do not select a provider just because its ID appears in an example.
- Runtime caps are 30 stage steps and five visits to a stage. Runs in different
resolved workspaces execute concurrently; only 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). Agent-session quiescence is scoped to the run: a run fails only when
one of its own sessions cannot be confirmed idle, so one run's active agent never
fails another's concurrent run. An active session left over by a restart with no
owning active run is an orphan and still refuses new work until interrupted.
Reusing that active run's request key reattaches idempotently.
type=humandecision stages exist and pause the run (seereferences/format.md); there are no parallel branches. Optional worktrees, full backup/restore, automatic database sync and independent verifier stage types are future ideas, not existing capabilities. Do not invent them in definitions. - Services are native or Docker-based. With an existing Docker installation do not start a competing native server. Server APIs are unauthenticated: trusted networks only; never recommend public exposure. Tool scope is not a universal OS sandbox.
- Docker setup resolves the current supported OpenCode V2 release from the single
documented complete
GET .../git/matching-refs/tags/v2response (a bounded complete array with no pagination parameters; never the movinglatesttag or the registry's insertion order) and the current Pi release from npm, resolves the selected tag's immutable GHCRsha256digest, buildsFROMthat digest, and verifies the built image reports exactly those executable versions plus a Node-compatible Pi engine. It fails closed, keeping the last verified Dockerfile fallback pins, when discovery, digest resolution, completeness, build or compatibility cannot be established: never a hardcoded fallback, never a silent stale-pin substitution. A full install skips a rebuild only when the release receipt and the actual built image's release and source identity labels already match; a dirty source always invokes the Docker build and relies on layer caching. A model/login-only reconfiguration compares only the verified release image identity, so a changed source alone does not rebuild../install --opencode|--pireconfigures inside the existing container and refuses an image that does not match; it does not rebuild or restart services (an explicit ordinary./installdoes when the user asks).
Where data lives
SQLite retains pipelines/immutable versions, runs, recorded stage visits, inputs/outputs, status/errors, audit transcripts, session ownership/references, reported per-stage model usage (tokens/cost), retry links, schedules/state, notification receipts and file changes. Not everything is in that database: drafts/settings are client files; complete OpenCode sessions and credentials use OpenCode storage; repository artifacts are files; ntfy has its own storage. A pipeline ZIP is not a full backup of these components. Never merge or sync live SQLite files using generic file synchronization. Do not delete databases, sessions, volumes or credentials to repair a workflow.
Inspect results first
Use the authoritative CLI_PREFIX from the prompt:
CLI_PREFIX runs --json: run history (optional--pipeline NAMEscope).CLI_PREFIX search QUERY [--type KIND] [--limit N] --json: one read-only query across pipelines, versions, runs, steps, field text and artifacts. QUERY is free text, a bare status word (queued/running/succeeded/failed/canceled/interrupted) selecting those runs, or a lone date expression (today,yesterday,this week— Monday 00:00 local — orYYYY-MM-DD) selecting the runs created in that window; anything else stays free text.--typeis the only narrowing mechanism and takes one ofpipeline,version,run,step,field,artifact; an unknown kind is an error. Ranking is exact name (3) > name substring (2) > body text (1), equal tiers by recency, and every hit reportsscore,matched_fieldandrecency. The cap defaults to 50 and clamps to 200; every kind scans at most 2000 rows and field scanning stops at 2000 steps / 1 MiB and reportstruncated/scanned_steps/scanned_bytes/cap/scan_bound_bytes. Targets are stable:pipeline:<id>,pipeline:<id>:v<version>,run:<id>,run:<id>:<sequence>,run:<id>:<sequence>:<field>, or an artifact URL.CLI_PREFIX runs --json: run history (optional--pipeline NAMEscope). Each run carries a bounded, single-linetitlesaying what it was for, set explicitly withCLI_PREFIX run ... --title "..."or generated from its prompt by the shared summariser on the configured default harness. It is never empty: when generation is unavailable or times out the server derives the prompt's first line, andtitle_source(model,fallback, orexplicit) plustitle_modelrecord how it was produced, so a diagnosis can tell them apart. Humanrunsandinspectshow the title beside the pipeline name and version, so two runs of one pipeline are distinguishable;CLI_PREFIX title RUN_ID "..."renames a run after the fact without changing its input or steps. The stored prompt is unchanged, and listing/renaming never calls a model.CLI_PREFIX inspect RUN_ID --json: selected run, visits and retained evidence. Each model step carries the usage its harness reported (usage.tokens_total,tokens_in/tokens_outonly whensplit_known,cost_usdonly whencost_known, plus the model the step actually ran with); the run carries the totalledusageover those stages, so a mixed-model run shows what each stage cost. An unreported cost or split is shown as unavailable, never as zero and never inferred from output length or transcript text; a historical or pre-upgrade run simply omitsusage. Exact stage outputs are authoritative; a transcript over the 32 KiB retained bound is a valid, credential-redacted preview envelope (with preserved stage limits, failure/format diagnostics and the retained session pointer), not the complete tool history.CLI_PREFIX decide RUN_ID [STAGE] --option KEY [--feedback TEXT] [--json]: answer a run that iswaitingon atype=humanstage. The status and the stage's instructions/options are visible inrunsandinspect --json(and the TUI picker); the stage is inferred from the run's current stage unless named. The choice and feedback are recorded as the step's evidence (its output and derived verdict; the feedback is never an error diagnostic) and the run resumes alongedges[KEY]. An unknown option is refused with the valid list; a stage that is not waiting is refused; a repeated decision for an already-decided visit does not double-apply. The wait is persisted, sodecidestill works after a server restart.cancelandretrykeep working while waiting and after a decision; a human stage is never killed by a stage timeout ormax_visits(the shared 30-step run budget still applies).CLI_PREFIX run ... --json --wait --wait-timeout DURATION:--waitpolls until terminal and--wait-timeoutbounds only the client wait after admission, independent of per-request--timeout. A reached wait limit leaves the admitted run active; inspect it read-only withinspect RUN_ID --json, or keep waiting by rerunning--waitand reusing the original--request-key. Never start a duplicate run, and never expect an automatic replay or cancel of admitted work.CLI_PREFIX show NAME --version N --json: exact pinned definition, not whichever version happens to be latest now.CLI_PREFIX status --json: connection details; candidate addresses are not proof they are reachable from another device.CLI_PREFIX ping --json:pong.build.version,build.commit(full Git revision) andbuild.dirtyidentify the compiled server, not the client or its runtime checkout. A dirty build includes-dirty;commit: "unknown"ordirty: nullmeans metadata was unavailable. Older servers omitbuild. Never infer a deployed fix from source/skill contents. Native Go builds use embedded VCS metadata; Docker setup stamps build arguments. For an explicitly requested Docker rebuild,python3 scripts/build_metadata.pyresolves the current releases and stampsdocker compose buildfrom the checkout with the immutable digest; it does not restart/deploy. A plaindocker compose buildwithout that resolved digest cannot build the image. Plainpingalso displays server identity;CLI_PREFIX version --json(alias--version) identifies the client offline. Include both in sanitized bug reports. Two dirty builds at the same commit can differ. Docker setup verifies its started server identity against the build arguments before declaring setup complete.
Pipelines is a single-user first-version preview for supervised work, not a
multi-user service. Recommend a disposable repository for initial trials. Tool
permissions and same-path home mounts do not isolate valuable files. Never collect
raw credential/session databases or publish private inspect output as diagnostics.
In a source checkout, python3 scripts/smoke-first-install.py checks a disposable
source-built test binaries, draft/publication/transfer and restart persistence without
inference, notifications or deployment. It is not a live-provider or phone test.
That is a developer test fixture, not the native installer. End-user installation
runs natively through ./install by default (scripts/setup-native.py, with
--native equivalent); the Docker server + ntfy installation is explicit with
./install --docker.
In the TUI, Home advertises / omni search and : commands. Home / opens one
popup over pipelines, versions, runs, steps/visits, artifact names and stage
input/output/evidence text; each row shows a short type badge and a text match
shows a bracketed snippet, and a bare status word or date selects runs. The same
query is available as pipelines search QUERY [--type KIND], where --type
narrows the kind. Run Tab switches only
Details ↔ Graph. n/p select recorded visits, including repeated stages; d
focuses/toggles file changes, A opens the run's grouped artifacts view (name,
content type, size and URL per entry, PgUp/PgDn scroll so a trailing visit
that recorded none stays reachable, Enter/o to open the selected URL and
y to copy it), a selects the active/latest retained
configured-harness agent session's bounded, redacted tool evidence, D/:debug
opens fresh investigation, and P opens the pinned pipeline version. :quit
exits; q does not. Results are retained reports, not independently confirmed
achievements. Distinguish partial output from full success.
Retained stage file changes
Each newly executed visit compares a bounded, read-only workspace snapshot just
before execution with one after execution/cleanup. SQLite retains unified patches
and metadata in steps[].file_changes; snapshots themselves are transient memory.
The stage's File changes section highlights added/removed lines and hunk headers.
- The baseline is stage-start content, not HEAD/index. Pre-existing edits and earlier stages' changes are not attributed to this visit. Added/deleted/modified files, modes and binary-change metadata are recorded; renames are delete/add.
- Git scope is tracked and nonignored untracked files within the run directory; outside Git, regular workspace files. Symlinks/special files, Git internals, node_modules and common environment/key/credential/database paths are excluded. This is path-based protection, not secret-content scanning. Patches may still contain sensitive text; never publish them without approval.
- Limits: 10,000 entries, 1 MiB/file, 32 MiB/snapshot, five-second snapshot context, 5,000 combined text lines, 64 KiB/patch, 128 KiB serialized file evidence/visit. Missing/skipped/truncated captures must be described as partial/unavailable.
- Failed/canceled stages retain changes when execution returns; unconfirmed agent cleanup is partial. Crashes may have no after-capture. Historical visits say not recorded—never infer no edits from absent evidence.
- These are observed workspace changes, not exclusive authorship. Concurrent edits, changes outside scope, ignored new files and changes reverted before capture can be absent or mixed in. Capture failure does not change the declared run verdict.
Do not add model output fields, new stages or extra model calls merely to obtain file diffs: current capture is server-owned. Do not promise full files/rollback.
Retained stage artifacts
Every model visit gets a file-based turn under the data directory
(<data>/artifacts/<run>/stages/<seq>-<stage>/) with input.md, memory.md,
output.md and private/; public/ is run-level and shared by every stage.
The root is configurable and defaults next to the database. The writable
private/ and shared public/ directories are stated in the agent prompt for
every harness, and PIPELINES_ARTIFACTS_DIR points at the turn directory for a
locally spawned Pi process. OpenCode V2 tools run inside the shared serve
process, so there the prompt, not the environment, is the reliable carrier.
Turn-local files belong in private/; reports, logs and screenshots for later
stages belong in public/. The server owns input.md, memory.md and
output.md and never asks the model to write them directly.
After the visit the server makes one bounded inventory of the turn and public
directories and records each file's name, size, content type (extension, sniff
fallback) and a relative url; these appear under each step's artifacts in
inspect RUN_ID --json and the TUI's A artifacts view, which groups them by
stage and visit, marks images and shows the recorded count,
capture-pending/not-recorded wording and partial-capture notes. Public entries
carry a public/ prefix. input.md, memory.md and output.md are the primary
artifacts of every model visit.
GET /artifacts/<run>/<step>/<name> serves the bytes (correct Content-Type and
Content-Length, HEAD, and Range/206/416 via http.ServeContent), refusing
traversal, absolute paths, symlink escapes and unknown runs/steps. A pre-turn
<run>/<step>/ path stays readable for runs recorded before the file protocol
landed.
- Limits: 200 files, 8 MiB/file, 64 MiB total per visit; over-cap files are skipped and counted; symlinks and non-regular files are excluded.
- Artifacts are read-only evidence, not a gate: absent, unreadable, symlinked or oversized directories record nothing and never change the run verdict. The database keeps the recorded list, not file copies.
- Do not invent artifact files or claim a stage produced one without the recorded evidence; the server observes the directory after the stage.
Session export
- When an OpenCode-harness model stage finishes, the server exports its real
session read-only (v1
opencode export <id>, v2opencode session export <id>) and writes exactly one file,session.html, into the same turn artifacts directory; it lists and serves like any other artifact. - The
session.mdexport is retired: nothing writes it any more and it is never presented. A run recorded before the HTML-only change still listssession.mdamong its stored artifacts (dated evidence kept on disk, not deleted); every listing — app, TUI, web spike — hides it rather than showing it beside the html. - The source read is capped at 8 MiB and the rendered file at 8 MiB, matching the
ordinary per-artifact file bound; a cut source or a cut rendered file carries a
visible truncation banner at both ends and
session_export.truncated: true, and a truncated export is never called whole. - The step records
session_export(session_id,backend,source, the html file name,html_bytes,messages,truncated,diagnostic) ininspect RUN_ID --json. A record written before the change still carriedmarkdownandmarkdown_bytes; readers ignore those fields. - A failed export, or a Pi stage with no native export path, records a diagnostic
(Pi:
source: "unavailable") and never fails the run or changes the stage result. An existingsession.htmlis never overwritten.
Operator-only retention
Run history and its artifacts are never pruned automatically. The operator-only
pipelines retention report (strictly read-only) and pipelines retention apply
(removal after explicit approval) are the only commands that remove them; no
server startup, scheduler, list_runs, stats, runs or inspect path calls
retention, and there is no background goroutine, timer or startup hook. A policy
selects by rolling UTC age (--older-than), newest-N count (--keep), or their
intersection. Runs that are queued/running, waiting on a person, running a cleanup
agent, or referenced by a retained run's retry provenance are protected and
reported with a reason; their artifacts are kept. Removing a run removes its
<root>/<run> artifact directory and reclaims database space with one VACUUM
and one WAL checkpoint, reporting measured bytes. Pipeline definitions and their
immutable versions are never touched. The commands are local, require an existing
--db, and need no running server.
Brokered OpenCode chat
The protocol declares the brokered chat_* surface (chat_sessions,
chat_create, chat_directories, chat_prompt, chat_messages,
chat_interrupt, chat_permission, chat_reply). The server dispatches all of
them, and pushes chat_event frames derived from OpenCode's global SSE firehose
(/event on the native serve, /api/event on the V2 API), with retained
per-session/per-message state so a reconnecting client re-syncs instead of
replaying deltas.
At startup cmd/pipelines-server calls opencode.OpenChat, which detects the
installed backend and opens exactly one owned opencode serve endpoint for it:
the native wire for v1, the V2 API for the installed V2 CLI. Chat is enabled only
when that connection carries a URL-bearing, SSE-capable event endpoint; otherwise
chat stays disabled and every chat_* request is answered with a protocol error
naming the missing configuration rather than hanging. Stage generation still
reaches V2 through the opencode api CLI, so the owned serve process is chat's
single transport and starts no process for stages. The adapter skips firehose
events it does not surface before decoding their data, so an unrelated event
whose delta has a different shape (session.instructions.updated carries an
object) never fails the stream.
The operator can instead broker chat from their own OpenCode service with the
opt-in -opencode-endpoint flag or PIPELINES_OPENCODE_ENDPOINT, with precedence
flag > environment > the endpoint recorded on a previous run and the spawned
instance as the default. opencode.OpenEndpoint validates the reference once
(the established http(s)/host/no-credentials/no-query/no-path idiom), detects the
backend from the service's own /api/info (V2) or /global/health (native)
contract and requires a streamable text/event-stream event path; it returns a
connection with no owned process and never calls startServe. A malformed
reference, an unreachable host, a reachable non-OpenCode service or a
health-ok-but-no-SSE service each disable chat with an error naming the endpoint
and never fall back to the spawned instance. An authenticating endpoint accepts an
optional password only through the private file
PIPELINES_OPENCODE_ENDPOINT_PASSWORD_FILE (regular, 0600, single-owner), and
PIPELINES_OPENCODE_ENDPOINT_USERNAME overrides the detected default user. The
validated choice is recorded in the additive server_settings table so a restart
with no flag/environment reuses it; an unreachable recorded endpoint disables chat
with a clear message and no fallback. pipelines service enable preserves the
endpoint and the password-file reference in autostart.
chat_sessions lists the operator's own chat sessions, not the sessions pipeline
runs create. Operator sessions are recorded in the dedicated operator_sessions
table (session_id primary key, plus directory, title, tag, created_at,
last_used_at), separate from run_steps/run_session_owners, so an operator
chat is a first-class thing and a pipeline stage session — recorded as run
ownership, never in the operator registry — is never listed. The handler reads the
registry once (bounded to 128 rows, most-recently-used first), calls OpenCode's
session list once, and keeps only registry rows the endpoint still reports, so a
session with no recorded tag is never listed and a deleted upstream session does
not linger. It fills each session's directory, tag and created_at/
updated_at from the registry and marks busy when another client currently
holds the session (an in-server claim, released when the stream ends or the client
disconnects); a second client's chat_prompt for a held session is rejected
before any OpenCode call. It is fail-closed: with no registry configured it
returns an empty list, never the endpoint's whole session list. The OpenCode
session list is read unscoped by directory (the V2 GET /api/session otherwise
defaults to the caller's directory), because an operator session may live in any
directory.
chat_create creates one operator session and records it in the same registry
with the stored tag operator (the display label Operator is derived from it,
so there is no second tag vocabulary). It takes a title (default "New chat") and
a directory (default ~): an omitted or empty value, ~, and any ~/... value
all resolve to os.UserHomeDir, never the server's working directory; a relative
path or a missing directory is rejected before any OpenCode call, and a session
created but not recordable is reported as an error naming its id. Create posts the
verified V2 shape POST /api/session {"title":...,"location":{"directory":...}}.
chat_directories returns the operator's home first, then the directories of
previous operator sessions most-recently-used first, deduplicated and bounded (8);
it is a dedicated request so the create dialog can load the choices without first
listing sessions and chat_sessions stays a pure, fail-closed session list. A
successful chat_prompt touches the session's last_used_at in one UPDATE, so it
moves to the front of the list; a non-operator session id touches no row.
chat_adopt records a session that already exists on the configured endpoint in
the same operator registry, using the endpoint's own directory and title
verbatim (never a guess, the server's working directory or the home directory).
It reads the endpoint once through the shared Chat.Session (which reuses
Sessions/decodeChatSessions), records through the existing
RecordOperatorSession with the operator tag, and echoes the adopted session
plus its registry row like chat_create, so chat_sessions lists it and
chat_prompt continues it. A nil payload, a blank or foreign-shaped id, a nil
broker or a nil registry 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 a directory. The CLI
equivalent is pipelines chat-adopt <session-id> (--json prints the full
response, otherwise a short Adopted <id> · <directory> line).
While a chat_prompt subscription is open the server also polls
GET /api/permission/request on a fixed small interval (first poll immediately) and
pushes one chat_permission frame per newly pending approval request, deduped by
request id so a repeated poll never re-pushes the same request. The app answers with
chat_reply, which the server forwards as
POST /api/session/{id}/permission/{requestID}/reply with
{"decision":"once"|"always"|"reject"}. A missing session/request id or an invalid
decision is rejected before any OpenCode call, and the poller stops when the stream
ends or the client disconnects.
The permission list and reply are implemented against the V2 HTTP API only: their
field names match the service's Permission.Request/Permission.Reply schemas
(verified against the installed service's /openapi.json). The native v1 serve
translation rejects them explicitly instead of guessing an endpoint, so on v1 the
permission poller stops quietly rather than reporting the documented limitation as
a per-prompt error.
Verification evidence (2026-10-17)
Verified against the installed OpenCode 2.0.20 on 2026-10-17:
opencode --versionreports2.0.20;opencode --helpadvertisesapiandserve.GET /openapi.jsonon an ownedopencode servelistsGET /api/event(SSE) and the session/prompt/interrupt/permission endpoints.- A live turn streamed token-level
session.text.deltaandsession.reasoning.deltaevents (each withassistantMessageIDandordinal), plussession.step.*andsession.usage.updated; the V2 API server authenticates the fixedopencodebasic-auth user. - The opt-in live end-to-end test (
PIPELINES_TEST_OPENCODE_CHAT_LIVE=1 go test ./internal/server -run TestChatLiveEndToEnd -v) opened an owned V2 endpoint, created an operator session throughchat_createin a chosen directory, listed exactly that session throughchat_sessions(with its directory and tagoperator), recorded a genuine pipeline stage session inrun_steps/run_session_ownersand showed it never appeared in the list, and streamed a completed turn as[step:started,text,step:streamed,step:ended](first delta"pong"). The operator-session registry, the filter, directory defaulting and the recents bound are also covered by thepipeline,opencodeandserverunit tests, which run without the live gate. - Not verified live: a genuine pending-approval round trip and any backend other than the installed 2.0.20. The permission wire shapes are unit-tested over an httptest fixture, not against a live pending request.
- Operator-supplied endpoint selection and
chat_adopt(added 2026-10-17):opencode.OpenEndpointand the flag/env/recorded precedence, the URL-shape validation failures (scheme, empty host, userinfo, query, fragment, non-empty path), the unreachable/non-OpenCode/no-SSE/auth-mismatch failures, the no-owned-process property (conn.process == nil) and the password-file rule are unit-tested over httptest fixtures ininternal/opencode/endpoint_test.go.chat_adopt(endpoint directory/title recorded with theoperatortag, echo, missing id refused naming the id and recording nothing, nil payload/blank id/nil broker/nil registry refused before any endpoint call, empty endpoint directory refused) is unit-tested over the server fake broker ininternal/server/chat_test.go; the recorded-endpoint round-trip across a store Close/Open is ininternal/pipeline/settings_test.go; the protocol round-trip and the CLI dispatch (pipelines chat-adopt) are covered ininternal/protocol/chat_test.goandinternal/cli/chat_adopt_test.go. The live scratch-server check against the operator's real endpoint (adopt, list, prompt) was not run in this change; only the unit fixtures above were executed.
Deployment state: not deployed. This is reviewed and committed locally only; deploying it (and restarting the server) is the operator's step. Do not describe brokered chat as already deployed.
Debug and retry
Debug opens a new conversation with retained failure context, exact definition and session evidence; it does not resume implementation or grant permission to edit. Inspect first, identify evidence-supported causes, and say what remains unknown. Server-side session evidence is the latest bounded, credential-redacted evidence for up to the 3 most recent session-bearing steps; OpenCode V2 selects the latest assistant message with the same bounded readback as the final result, so a small final result is visible behind large prior tool content. Inspect the retained session itself when a preview is bounded or omitted; never assume a complete tool history from a preview. Retry is allowed for a failed run after cleanup stops: a fresh run/session starts at entry with the original input/directory/pinned version/edit scope. Previous side effects remain and can happen again. Never describe retry or cancellation as rollback, or resume/replay admitted tool work just to recover formatting.
Retained model-stage failures carry a fixed sanitized category. FINAL RESULT VALIDATION names the exact duplicate-field/syntax detail and byte offset for a
strictly invalid final result; HARNESS TRANSPORT covers request, active-session
poll and final-readback failures. A validation failure never re-prompts the
admitted tool-using stage: completed edits and tool side effects are retained and
not undone, so a full retry repeats work. An explicit read-only stage is the one
exception: a malformed result there mutates nothing, so it may be re-emitted up
to retries with its format_repairs evidence retained, and a post-admission
deadline or transport failure may be retried. A failure to reach the harness
before any prompt or tool call is admitted (connecting, or creating the session)
replays nothing, so it is retried within retries for any stage, mutating or not.
A stage whose tools include
edit/shell, or that has no tools block (the legacy read/glob/grep/edit/shell
bundle), is never replayed. Neither category exposes provider/model bodies or
credentials. The isolated formatter's deadline is proportional to the stage
timeout and clamped so it never extends the stage's own deadline.