5.7 KiB
Explicit run reports and session export completeness
Launch with pipelines run NAME PROMPT --output-format html (or markdown),
or choose Report format in the web run dialog. The wire field is
output_format: "html" | "markdown"; omission preserves prior behavior. It is
stored with the run, compared for request-key replay and inherited by retries.
This never changes a pipeline definition and is not inferred from prompt text.
Before success, the server takes the first nonblank string in the terminal
output's markdown, report, result, summary, message fields, renders a
report deterministically and records the verified artifact on the terminal
visit. Fixed filenames are public/report.html and public/report.md.
Both source and rendered output are capped at 2 MiB. Existing targets and
symlinked directories are refused; reports are not overwritten. Missing or
empty textual output, invalid targets and failed verification fail the run.
HTML has inline styling, no executable raw HTML, and a restrictive content
security policy preventing external resource loading. This is a structural
file contract, not factual validation or proof that every task was fulfilled.
Session export is owned by the runtime, not the model stage. No stage prompt,
definition or execution contract changes. Ownership is recorded during execution;
export evidence also snapshots the original session ID/backend/directory before
the run finishes. Evidence records status as complete, empty, failed,
partial, or unavailable, plus attempt count and capture time. Zero-message
readbacks are retried within the runtime's 10-second export budget and then
recorded as empty with a warning, without producing success-looking HTML.
Known temporary transport/store failures and interrupted JSON with verified
identity can retry (at most 3 attempts). Wrong-session responses, unsupported
envelopes and permanent CLI failures do not retry. A 200 ms delay does not prove
that settling writes caused an invalid readback. Expected array/data envelopes
are normalized, always retaining exact session identity validation.
pipelines repair-session-export RUN_ID SEQUENCE --json invokes the
repair_session_export protocol command (run_id, export_sequence). It is
read-only to the native session: no generation, resumption, interruption or
stage execution. It requires a settled run/session and exact original ownership
metadata, never guesses the backend, and appends a repair attempt to the run.
Successful readback writes public/session-repair-SEQUENCE-RANDOM.html and
records its bounded inventory entry. Original exports/evidence and run outcomes
are preserved. Failed/empty/partial attempts remain inspectable warnings; CLI
JSON returns the repair record itself and exits nonzero for incomplete evidence.
There are at most 20 manual repair attempts per visit. Repairs cannot guarantee
recovery if the upstream session or original directory is gone. The web pane
explains original failure and lists repair attempts/links separately.
Durable automatic recovery
Incomplete native exports now enqueue a session_export_jobs row in the same
transaction as FinishStep. The snapshot records the exact original session ID,
backend, directory and harness reference. Runtime startup also reconciles missing
jobs from persisted ownership/steps, including interrupted initial capture; it
never guesses missing ownership or edits original evidence. Reconciliation is
bounded to 256 newly discovered jobs per sweep (startup, then every 30 seconds).
A runtime-owned worker claims jobs only for settled runs/steps, with no active
agent in that run or using that session, and with cleanup_unconfirmed=0 on
current main. The runner's in-memory cleanup gate also defers execution without
spending a retry. Each automatic attempt uses the existing versioned, read-only
repair path and verifies ownership against the durable snapshot before readback.
No stage prompt, pipeline definition, stage execution or run outcome changes.
The first attempt is due after 5 seconds; further attempts wait 20 and 60 seconds.
There are at most 3 automatic claims, each with the existing 10-second readback
budget. A 45-second persisted lease recovers interrupted workers after restart;
attempt counts/backoff are not reset, including process kills. Saved complete
repair evidence can acknowledge a job after a crash without exporting again.
Expired final-attempt leases become exhausted, not an infinite retry loop.
Wrong-session readbacks never qualify as complete evidence. Failed/empty/partial
repairs and exhausted-job diagnostics remain visible in run inspection and the
web visit pane (session_export_recovery). Originals remain unchanged.
Worker shutdown follows runner context cancellation and current main's
Store.Done() (including cancellation of in-flight readback). The small startup
helper detects that lifecycle interface without modifying this branch's older
store/server-lock implementation. Its compatibility fallback exits on a closed
database; current main's lifecycle/cleanup implementations must be retained when
integrating. Upstream loss, missing original ownership/directory, inventory
limits or sessions exceeding export caps can still leave exhausted evidence;
manual versioned repair remains available. Browser captures for recovery status
are not captured under the no-live-runs audit constraint.
No production service, live run or historical data was modified for this audit. Unit/fixture tests cover reports, failure gates, retry inheritance, idempotency, migration, transient/permanent export readback, empty-export diagnostics and versioned repairs preserving originals. Browser screenshots are not captured; new web states remain partial pending a permitted isolated browser review.