Future AGI Files

1.48.0 (2026-10-09)

⚠️ Contains database migrations. This release adds four migrations in tracer (0111 through 0114) for Error Feed grouping: three add columns and one is a data migration. Apply them with manage.py migrate when rolling out the backend image.

• 0111_grouping_budget_wait, 0112_grouping_failure_recovery and 0113_error_group_redirect add new columns to the grouping work, grouping attempt and error group tables.
• 0114_backfill_group_redirects reads every stored grouping decision, points each merged or split error group at the group that replaced it, and soft-deletes groups already marked retired. It works in batches of 500 outside a single transaction, so it takes longer the more grouping decisions are stored. Rolling back does not undo it.

Features

• error-feed: give the investigator the provider log's warnings and errors (0d288af)
• error-feed: let an unrecovered finding reach the Feed without a failed requirement (cc3a693)
• error-feed: let an unrecovered finding reach the Feed without a failed requirement (TH-8420) (c0c7a3c)
• error-feed: serve compact conversation evidence to the investigator (7ec0b20)
• error-feed: serve compact conversation evidence to the investigator (3d2150b)
• gateway: tag provider models not served on chat completions (9ec4535)
• gateway: tag the provider models served on /v1/responses, not chat completions (06c1e8e)
• simulate: environment API to re-grade evals on a finished run and edit one eval (0b400b0)
• simulate: environment API to re-grade evals on a finished run and edit one eval [TH-8352] (495e393)
• simulate: show added eval columns and re-grade evals on a finished run [TH-8121] (c501f49)

Bug Fixes

• agent-playground: keep Agent node discoverable in orgs with no published agent (TH-4549) (3189122)
• agentcc: filter gateway analytics by gateway key id, not a UUID (75b44dc)
• agentcc: filter gateway analytics by gateway key id, not a UUID (9915911)
• agentcc: stop doubling /v1 in the Go gateway's Anthropic inference paths too (4b52c66)
• agentcc: stop doubling /v1 when the base URL already has it (e391443)
• agentcc: stop model discovery doubling a /v1 the base URL already has (b28d2bb)
• alerts: drop a scoring alert's stored choice on update instead of rejecting it (be2d37a)
• alerts: handle an unresolved eval list and a label-less preview (6824914)
• alerts: preview only the eval the picker has settled on (9411679)
• alerts: scoring-eval alerts measure the score, not a label (51c5738)
• ci: stop CI installs and tests from registering fake signups (9077af0)
• ci: stop CI installs and tests from registering fake signups (a38a574)
• contracts: refresh MCP manifest after grouping recovery merge (0292f18)
• contracts: regenerate grouping failure API contracts (0310aa2)
• contracts: regenerate grouping failure API contracts (669791a)
• dashboards: anchor stacked line charts at zero (36a211f)
• dashboards: auto-scale the widget y-axis and give each side one scale [TH-7680] (0a56d8f)
• dashboards: drop dense-series markers and stop stacked tooltips reading 0 (cc8f26c)
• dashboards: keep explicit bounds for negative bars and a round step for one typed bound (4d2aacf)
• dashboards: measure the series extent in one pass (ffd3800)
• dashboards: stop minute-granularity widgets freezing the browser [TH-7757] (c47cab9)
• e2e: match trace pagination parameters in POST bodies (caa64e5)
• e2e: match trace pagination parameters in POST bodies (27ed557)
• error-feed: accept harmless protected example refreshes (ee96692)
• error-feed: accept harmless protected example refreshes (b5aff94)
• error-feed: accept harmless protected example refreshes (79da083)
• error-feed: accept harmless protected refreshes (TH-8367) (7eb7db6)
• error-feed: accept legacy merges and cover redirects (7a33d17)
• error-feed: declare the real shape of the conversation evidence response (c43c463)
• error-feed: keep a completed report when a later attempt fails (31c41df)
• error-feed: leave a job that gets its report while the requeue command runs (725834b)
• error-feed: mask the call participants' own numbers in the conversation record (8f570c2)
• error-feed: preserve group IDs and redirect retired issues (TH-8431) (a6c5e41)
• error-feed: preserve merge identity and redirect retired groups (8a24124)
• error-feed: publish sampled merges with complete membership (TH-8368) (2726c81)
• error-feed: publish sampled merges with complete membership (TH-8368) (44688e2)
• error-feed: read compact evidence at the attempt's cutoff and keep every transcript event (85e6309)
• error-feed: recover budget-paused grouping and reopen recurring issues (0c74f7b)
• error-feed: recover failed grouping work and reconcile dead leases (654bec3)
• error-feed: recover failed grouping work and reconcile dead leases (6119101)
• error-feed: recover failed grouping work and reconcile dead leases (TH-8428) (75c2f79)
• error-feed: retry failed trace investigations and bound the evidence read (c249afe)
• error-feed: retry failed trace investigations and bound the evidence read (9ed9dd5)
• error-feed: use a completed report that lands late and keep jobs that hold one (a2a5236)
• gateway: stop tagging hand-typed models as Responses-API-only (3e51273)
• mcp: refresh manifest hash after rebasing grouping contracts (6783a6a)
• mcp: refresh manifest hash after rebasing grouping contracts (42c8280)
• observe: keep same-level agent graph nodes on one level (TH-4321) (9e44fc3)
• observe: label each Users row with its project (TH-5037) (a3e7dd1)
• observe: report a failed stereo split instead of an endless loader (f3b12cb)
• observe: reset the play button when a recording finishes (c3266e0)
• observe: span-list-only readable neutral status text (TH-4088) (1d78900)
• observe: stop the pager inventing a page count that keeps growing (848a3f3)
• observe: tell the user when a voice recording cannot be played [TH-7758] (df4cdb9)
• prompt-workbench: keep image/audio metadata on save (TH-7926) (c173698)
• restore hosted simulation provider configuration (797573a)
• restore hosted simulation provider configuration (e86d717)
• settings: keep the date disclosure open after a tap or a click (40b54ab)
• settings: render dates as neutral month-name in the browser timezone (777221f)
• settings: use the local month-name date on Workspaces and API keys (00dc142)
• share: only offer share links and access states the server confirmed (TH-4565) [agent] (ed9314a)
• share: report share-link failures once, in the dialog (36ae317)
• share: split the share dialog and never offer an unconfirmed state (1ff03b8)
• simulate: cap the Columns menu height so a long eval list scrolls (427d18b)
• simulate: cap the Columns menu width and cut long eval names with an ellipsis (c5826aa)
• simulate: honour page_size on the agent-version call executions (3becabc)
• simulate: keep the evals in the Columns menu while calls load (ec1ffa3)
• simulate: keep the user's own Expand all when the affected-calls chip is dismissed (1adf147)
• simulate: leave soft-deleted calls out of the runs list counts (0f23dd9)
• simulate: leave soft-deleted calls out of the runs list counts [TH-8392] (de6940d)
• simulate: list each evaluation in the run table's Columns menu (0b69fe4)
• simulate: offer each evaluation in the run table's Columns menu (7e9005c)
• simulate: prefetch the next call-log page from the current page's read (947b8bb)
• simulate: refresh call details when an eval is removed from a run (722110e)
• simulate: report the visible call total in the runs list calls field [TH-8392] (b4f229b)
• simulate: seal hosted harness usage on the backend, not the runner (7dceb02)
• simulate: send the clicked status chip as the run calls filter (3e5a0cf)
• simulate: show the affected calls when Debug failures hands them to the table (0ff0724)
• simulate: tighten the gap between a Columns menu checkbox and its label (aace95c)
• simulate: use the common error envelope for the new re-grade and eval-edit errors (db98f36)
• simulate: use the common error envelope for the new re-grade and eval-edit errors (136da2f)

Reverts

• error-feed: take the wider grouping gate out of this PR (5b52b24)