Changelog

All notable changes to Oddish are documented in this file.

The format is based on Keep a Changelog.

[2026-08-07]

Changed

The verdict now says accept or reject instead of is_good: true/false. Stored payloads keep is_good too, so old rows, the dashboard queries, and the Slack alert still work. The badge shows "Accepted" or "Rejected".
The verdict judge used to bury its hard rules inside exceptions, and it accepted a task whose own audit had found a must_fix leak — on tests the untouched base model already passed (0.96 against a 0.25 threshold). The prompt (verdict_prompt.txt) is rewritten as two steps: first look for evidence that rejects the task by itself (a leak, weak tests, a failed baseline), and only then weigh the trials' opinions, which need agreement.
The task overview panel used to list only the current experiment's trials, but the verdict is computed over every trial of the task — so the panel could show a verdict whose deciding trial it refused to list. It now shows every trial of the version. Trials from other experiments carry a dashed "elsewhere" chip and open in a new tab. Long subtypes also stopped pushing the "View trial" button out of its row.
The verdict badge used to hide its rerun button once a verdict existed, and the button that did exist re-classified every trial from scratch. Tasks with a verdict now show "Rerun verdict" (qa/backfill with force: false), which keeps the stored trial analyses and redoes only the verdict. The full re-classify stays on qa/retry.
Submitting new trials used to delete the task's verdict immediately, and the task had no verdict until QA finished the new trials. The old verdict now stays until the new QA run replaces it.

Removed

The cc_chat dashboard chat feature is gone end to end: the /chat-sessions backend router and orchestrator, the chat drawer/button UI and its /api/chat-sessions proxies in the frontend, the ChatSession / ChatSessionEvent / ChatTurn models, and the chat tables themselves (dropped by backend migration dropchat001; api_keys.is_internal stays — internal key minting also serves probe credentials and the sandbox analyzer). The chat-only settings ODDISH_CC_CHAT_DAYTONA_SNAPSHOT and ODDISH_PUBLIC_API_BASE_URL are removed with it. ⚠️ Deployments that set only ODDISH_CC_CHAT_DAYTONA_SNAPSHOT must now set ODDISH_AGENT_DAYTONA_SNAPSHOT (same snapshot name) or analyzer sandboxes fall back to installing claude-code + harbor at provision time.
The shared sandbox infrastructure the chat feature grew — Daytona client, provisioner, Claude Code runtime, stream renderer — survives because the hosted analyzer runs on it; it moved from backend/api/services/cc_chat/ to backend/api/services/sandbox/, and the analyzer cohort modules (analyzer_block_runner, analyzer_parse, analyzer_prompt) moved to backend/api/services/blocks/analyzer/.

Fixed

Quota cancellation, retry, and append reconciliation no longer hide a preserved accepted verdict by leaving its payload paired with a missing status. Verdict lifecycle changes now use one state-transition module: replacement QA retains the published payload while queued/running, cancellation or a no-op restores it to SUCCESS, and only terminal QA failure discards it. A database constraint repairs and prevents invalid payload/status pairs.
Worker heartbeats used to stop as soon as the agent finished, but the worker still had to upload and save the results. When that took over 15 minutes, the cleanup sweep marked the trial "Worker heartbeat stalled for over 15 minutes", threw away the finished result, and re-ran the whole trial. The heartbeat now runs until the results are saved and settled.

[2026-08-05]

Added

opencode trials can now run on closed-internet tasks. Stock opencode self-installs (nvm/Node/opencode-ai) during agent SETUP, which runs under the ENVIRONMENT baseline network policy — the agent-phase allowlist (extra_allowed_hosts, runtime-host merges) only applies around agent.run(), so no agent-phase declaration can save a self-installing agent: the trial died at DNS during setup (curl: (6) Could not resolve host: raw.githubusercontent.com) before the model was ever reached. The fix mirrors the existing claude-code installer arm in run_harbor_trial_async: -a opencode now merges OPENCODE_INSTALL_HOSTS plus the model transport host (via outbound_hosts_for_model, which resolves openrouter/tencent/hy3 → openrouter.ai) into env_config.extra_allowed_hosts, which harbor folds into the environment baseline so the allowlist spans install and run. On legacy closed tasks ([environment] allow_internet=false → no-network baseline for every phase, e.g. the GDM SWE-Marathon samples) this is the only channel that works at all; on modern swe-marathon-shaped tasks (public setup → restricted agent) harbor ignores baseline extras on the public baseline and the agent phase keeps its model-host-only allowlist, so no install hosts leak into agent run there. _build_agent_config still routes -a opencode through the OddishOpenCode wrapper. Note: required_outbound_domains — the hook two earlier revisions of this change relied on, and which several wrapper docstrings describe as "Harbor builds the Modal egress allowlist from this hook" — has no consumer in oddish or harbor; it is kept declarative-only for interface parity (both failed approaches were validated end-to-end on the PR preview backend before landing on this one).

[2026-07-27]

Changed

Enforced quotas now cancel every quota-counted nonterminal trial as soon as live or settled spend reaches the payer's rolling 24-hour cap or the organization's monthly cap. User caps stop that payer's trials; org caps stop all trials in the org, including queued and retrying work, and remote workers are terminated after the cancellation transaction commits.

[2026-07-20]

Changed

The task-level QA worker job now leases concurrency from the analysis model's queue key (get_qa_queue_key() returns normalize_queue_key(analysis_model), currently anthropic/claude-sonnet-5) instead of the verdict model's. The bulk of a QA job's LLM work is the per-trial classification pass on the analysis model; keying the lease off the verdict model capped QA throughput at the verdict bucket's default (48) while the analysis bucket sat idle. ANALYZER jobs share the QA queue key and move with it (#802).
Raise the baked anthropic/claude-sonnet-5 queue-key concurrency override in the Modal deploy from 128 to 256, giving the relocated QA jobs and the analysis model's trials more headroom; operators can still override the whole JSON via the env var / oddish-prod secret (#802).

[2026-07-18]

Changed

The shared analysis model (ODDISH_ANALYSIS_MODEL — trajectory graph, trajectory summary, trial classifier, probe analysis) now defaults to Claude Sonnet 5 as the plain Anthropic-style id claude-sonnet-5, replacing the Bedrock inference-profile id global.anthropic.claude-sonnet-4-6. Plain Claude ids route analysis calls to the direct Anthropic API, and the analysis queue key changes accordingly to anthropic/claude-sonnet-5 (#794).
Bake a per-model ODDISH_MODEL_CONCURRENCY_OVERRIDES default into the Modal deploy that raises the anthropic/claude-sonnet-5 queue-key concurrency lease to 128 (up from the 48 default), giving the relocated analysis model the same headroom its predecessor queue key had; operators can still override the whole JSON via the env var / oddish-prod secret (#795).
Bake a per-model ODDISH_MODEL_CONCURRENCY_OVERRIDES default into the Modal deploy that raises the global.anthropic.claude-sonnet-4-6 queue-key concurrency lease to 128 (up from the 48 default) — the queue key every Sonnet 4.6 trial id spelling normalizes to; operators can still override the whole JSON via the env var / oddish-prod secret (#796).

Fixed

Dashboard queue stats no longer fold the trajectory-analysis and verdict pipeline counts into the analysis/verdict model's queue bucket. They now live under reserved analysis / verdict queue keys, so trials awaiting or undergoing classification can no longer masquerade as that model's queued/running trial workers (an incident showed 4k+ phantom "running" rows under one model's queue while the model's real trials were misrouted into the "analyses" pipeline). The reserved buckets report the QA job bucket's concurrency instead of a meaningless per-model default.
A QA job that dies or is cancelled mid-classification no longer strands trials in a non-terminal analysis_status. The stale-heartbeat reap now resets the dead job's task trials inline (RETRYING → QUEUED, exhausted → FAILED), the append-supersede cancel requeues in-flight rows, and a new _reset_orphaned_trial_analysis cleanup phase heals any remaining orphans: never-classifiable rows (superseded / skipped / gate-skipped / bulk-imported trials, soft-deleted tasks, or terminal tasks with no active QA job) are finalized FAILED, while rows a future QA attempt will re-classify are moved back to QUEUED. Previously these accumulated forever as phantom in-flight analyses. Orphan-finalized rows carry an Analysis orphaned: sentinel prefix on analysis_error, and resurrecting a task by appending trials reopens them so the fresh QA pass classifies them instead of inheriting a permanent verdict gap. Every reset selects its trial rows FOR UPDATE SKIP LOCKED so the sweep can never deadlock against the trials-then-task lock order the cancel path takes.

[2026-07-16]

Fixed

A grok trial killed by an xAI rate limit (You've hit your team's API rate limit) is no longer thrown away mid-run: the resume loop that already rescues idle-timeout deaths now also resumes rate-limited ones, sleeping first with a doubling backoff (60s, 120s, 240s) before each replay. The case was previously excluded on purpose, since an immediate grok -c re-hits the same wall — the throttle is on the account, not on one replica — but xAI's limits are refilling token buckets, so a resume that waits often lands, and a limit that never clears just fails as it did before. Idle timeouts still resume with no delay. Observed on a trial that spent 19 minutes and 437k tokens, announced its next step, and died to the limit; the truncated trajectory was then graded as a model failure rather than an infra one (#758).

[2026-07-15]

Added

Analyzers: a new cross-experiment trajectory-analysis feature that gathers finished trials from one or more experiments and synthesizes four evidence-backed narrative sections (bad failures, good failures, universal capabilities, headroom) via a Haiku agent-team map/reduce pipeline, with inline [trajectory](...) deep links, new analyzers REST endpoints, an oddish analyzer create CLI, and a dashboard Analyzers tab with list/create/detail pages (#706). Analyzer pipeline logs are now prefixed with the driving job's kind (e.g. [ANALYZER]) for easier attribution in mixed worker logs (#710), and the reduce-stage prompt that produced an analyzer's sections is now persisted on the row for debugging/reproducibility, though not exposed via the API (#711).
Task pages can now promote any stored task version to be the default: a new PUT /tasks/{task_id}/versions/{version}/default endpoint updates the task's current-version pointer (and legacy storage mirrors), and the task page gained a "Make default" action with an optimistic update and inline error handling (#713).
Owner-directed expense alerts (expensive experiment/trial) and a new failed-experiment alert (fires when a finished experiment has no active trials and at least half — configurable — of its current trials are FAILED) can now be DMed to the owner on Slack via SLACK_ALERT_BOT_TOKEN, matched to their Slack account by account email, alongside the existing webhook and email channels (#703).
Oddish task, experiment, and public-share links posted in a configured Slack workspace now unfurl with outcome glyphs, run details, and a compact task-by-agent result matrix for smaller experiments, via a new signed POST /webhooks/slack/events endpoint bound to one workspace/org (#700).
Trial drawer gained an adaptive Verifier Results card: test-based tasks report Common Test Report Format (CTRF) passed/failed/skipped/pending counts, benchmark-style tasks show scalar metrics, and other tasks fall back to the reward score; historical trials without a persisted summary lazily discover and parse their verifier/ctrf.json artifact (#699).

Changed

Markdown-rendered hyperlinks (analyzer reports, probe summaries, and other markdown content) now render in a theme-aware blue instead of a hard-to-recognize brown/off-white, so they read as clickable links (#712).
The trial Live tab now shares its step, tool-call, and observation rendering with the Trajectory tab, grouping streamed events into collapsible per-turn steps where the newest step auto-expands as the previous one collapses; Claude live-tail events now carry a turn_id/block_index/text_mode so streamed text deltas merge into one step instead of duplicating (#675).

Fixed

Analyzer reports no longer wait behind the QA backlog they share a queue with: ANALYZER and QA jobs both land on the QA queue key, and the claim orders by priority DESC, running_count ASC, created_at ASC — but every enqueue site left priority at 0, so the first two keys tied and claims fell through to pure FIFO, stranding an analyzer behind whatever QA burst a sweep had just produced (one report waited ~59 minutes to start). Analyzer jobs now enqueue at priority=1, so a draining worker picks them up ahead of that backlog (#744).
Non-Meta mini-swe-agent trials (e.g. Claude, GPT models) no longer crash deterministically on their first model call with ModuleNotFoundError: orjson — the litellm[proxy] reinstall that previously only applied to Meta-model trials now applies to all mini-swe-agent trials via a shared OddishMiniSweAgent base class (#714).
Experiment cost rollups now attribute spend only to a trial's home experiment (trials.experiment_id), so collection/rollup views that render trials gathered from other experiments no longer double-count that spend on both the collection and the trial's owning experiment (#702).
Pricing lookups now walk a specificity-ordered, case-insensitive chain of model-id candidates (exact id, path suffixes, provider vocabulary aliases, spelling variants, then generic provider prefixes) instead of one-off guesses, resolving previously-unpriced production model ids while preserving provider-specific rates; token-bearing trials that still settle to an unpriced NULL cost now emit a structured trial_cost_unpriced warning to logs and Logfire (#660).

[2026-07-07]

Added

oddish preflight <path> checks local tasks for integrity problems before they cost a trial: solution/tests baked into the agent image, repo fetches or .git directories that expose branch history, unjustified open internet, patch-file solutions, and brittle source-scanning anti-cheat. --json emits findings for CI.
oddish run now runs preflight before upload. --force submits anyway and still prints the findings. oddish upload is gated the same way (same --force override) — closing a two-step bypass where uploading a leaky task directly, then running it by ID, skipped preflight entirely.

Changed

oddish run aborts when any resolved task fails preflight. Previously a broken task in a multi-task run was reported and silently skipped while the rest proceeded. Use --force for the old behaviour.
API key creation is now self-service for every organization, gated on the caller's role in their current org instead of membership in the hardcoded Abundant org. can_create_api_keys no longer checks an org-slug/Clerk-org allowlist — any admin or member (Clerk-JWT auth only) may create keys for their own org, admins minting full/tasks/read and members minting tasks/read. API-key auth still cannot mint keys, and listing/revoking all org keys stays admin-only. Removed API_KEY_CREATOR_ORG_SLUGS / API_KEY_CREATOR_CLERK_ORG_IDS and refreshed the stale @abundant.ai/Abundant-org wording in the settings UI, endpoint errors, and docs (#617).

Fixed

Admin cost dashboard "Cost by user" rows now link to the per-user drilldown whenever the row resolves to a real oddish user (billed user or submitting credential), even when some or all of its trials are unbilled — previously a row was only clickable when every trial was billed, so real users with pre-billing spend (e.g. created before quota billing stamping shipped) or offboarded/unlinked spend were shown as non-clickable "unbilled". A new CostUserBreakdown.has_unbilled_spend field drives an "unbilled" chip: on a linkable (registered-user) row the tooltip explains the drilldown counts billed spend only, so its total may be less than the row total; GitHub-handle-only / Unattributed rows that are not registered users stay non-clickable with the existing "not a registered user" wording.

[2026-07-06]

Added

Org-wide aggregate calendar-month (UTC) spend cap, layered on top of the per-user rolling-24h cap. Admission now sums every payer's settled org spend (including unattributed NULL-billed spend) plus the org's in-flight reservation and blocks when it reaches the effective org limit (override row org_quotas ?? ODDISH_DEFAULT_ORG_MONTHLY_QUOTA_USD ?? none); over-cap submissions get HTTP 402 under enforce and log reason=org_over_budget under shadow. Advisory-lock order is org → payer → row locks (ENFORCE-only, org lock only when a cap is configured). Admins set/clear the cap via PUT /quotas/org and see month-to-date org usage on GET /quotas; any member reads the org budget snapshot plus an adaptive daily-goal via the new GET /quotas/org. Ships inert (no default, no override rows). oddish core reads org_quotas via raw text() SQL to preserve the oddish→backend package boundary; per-user rolling-24h behavior is unchanged.

Changed

Per-user quotas now use a rolling 24-hour window instead of a UTC-midnight reset. quota_window_start() replaces start_of_today_utc() in admission and quota usage reads, while ODDISH_DEFAULT_DAILY_QUOTA_USD keeps its existing name and value.

[2026-07-02]

Added

Trial-collection experiments: oddish experiment create --name "..." <trial_id...> and POST /experiments/collections (TASKS scope) gather existing trials into a new read-only is_collection experiment via create_trial_collection_core, without moving trials out of their home experiment. Membership is additive through a new experiment_trials join table (plus task_experiments for parent tasks); read paths — dashboard aggregates, task listing/effective-version resolution (including older task versions), export, and public/share views — treat membership as home experiment_id OR a gathered row via shared experiment_membership helpers, with probes still excluded from every public surface. New runs and sweep-appends targeting a collection are rejected. (First landed in #536, reverted, relanded in #552, reverted again after a prod migration deadlock, and relanded unchanged in #556 with a deadlock-safe exp_trials_join_001 migration — each DDL step now runs in its own autocommit_block, FKs are added NOT VALID then VALIDATE CONSTRAINT, and a lock_timeout guards against lock-order conflicts with hot-table DML.) (#536, #552, #556)
Task browser overhaul: browse_tasks_core / GET /tasks/browse accept a large set of new filters — task metadata (status, priority, verdict, has-link, run-analysis/probe, created-date presets/custom range, experiments), trial-level EXISTS filters (agent, model, agent·model pair, provider, environment, trial status, origin, analysis classification, has-error, has-trajectory, attempts, tokens/steps/reward ranges), on-the-fly aggregate filters/sort (avg score, total tokens, trial counts, pass/partial/fail/harness buckets, run time), agent/model "Compare A vs B" and "top performer" comparisons, and an or_groups DNF combinator for "match any of" condition groups. A new GET /tasks/browse/facets endpoint supplies sidebar option values. The frontend tasks page moves from client-side SWR to URL-driven, server-rendered filtering with a sticky filter sidebar (draft-then-Apply for heavy fields), Suspense/skeleton loading, link-based pagination, and whole-URL (v2) saved filters. (#540)
Migration Head Guard CI workflow: a new .github/workflows/migration-head-guard.yml runs alembic heads on the PR's merge ref for both the oddish and backend Alembic trees and fails if either would have more than one head after merging into main, catching stale down_revision forks (the kind that caused a prior experiments.is_collection prod incident) before they reach main. (#549)

Changed

Harbor artifact extraction is now centralized: a new oddish.core.harbor_artifacts module holds shared trajectory-metrics, timing, token, reward, and error extraction from Harbor TrialResults; both the live worker outcome path (workers/harbor/outcome.py) and CLI/server trial import (cli/api.py) reuse it instead of duplicating logic. Sauron trial-directory discovery now uses Harbor's JobScanner (with legacy fallback preserved), zip-import task-name inference prefers Harbor config/result model parsing over nested legacy JSON discovery, and several edge cases were fixed along the way (worker reward propagation, multi-trial token precedence, invalid trajectory cost handling, non-numeric verifier reward handling). (#557)
Task version content hashing (compute_task_content_hash) now hashes only execution-relevant task contents: it uses Harbor's Packager.collect_files publishable-file selection (respecting default ignores like __pycache__) and semantically parses task.toml via TaskConfig, so descriptive [metadata]/[task] edits no longer create a new Oddish task version — only changes to runtime fields (verifier, agent, environment, steps, etc.) or other files do. (#546)
Dispatch planning is now shared across hosts: a new build_dispatch_plan / DispatchPlan in oddish.dispatch.cycle centralizes queue discovery, counts, held-slot accounting, concurrency limits, and spawn-plan calculation. Modal's poll_queue (backend/worker/functions.py) and the standalone self-hosted run_polling_worker() (now routed through run_dispatch_loop with a new InProcessDispatcher) both consume the shared planner, while assigned single-queue Docker/Kubernetes workers are unchanged. The shared dispatch loop also now catches and retries on transient failures instead of propagating them. (#545, #547)
Backend router decomposition: hosted task-submission identity resolution, GitHub attribution, experiment-owner stamping, and auto-publish logic are extracted out of the large backend/api/routers/tasks.py into a new task_submission.py module; Claude Code's Anthropic-compatible environment setup in workers/harbor/agent_config.py is now shared across the OpenRouter, Fireworks, z.ai, MiniMax, and Moonshot routes instead of being duplicated per provider. (#554)
Trial environment is now surfaced directly in trial responses: TrialResponse.environment (including the compact experiment-page response) exposes trials.environment, added to the compact eager-load load_only set in list_tasks_core; the trial detail drawer's sandbox badge and rerun command now prefer this field over worker-job metadata, so the badge renders instantly instead of waiting on jobs data (worker-job metadata remains the fallback for legacy rows). (#531)
Locked harbor git dependency (rishidesai/harbor@main) advances from aeadaf4b to 2ae61e86, with HARBOR_DEFAULT_SHA and both uv.lock files updated in lockstep. (#543)
Probe detail panel: the trial ID on the probe run detail view now renders as a labeled "Trial: " line (matching the existing "Preset:" styling) instead of small unlabeled muted text; the panel also no longer shows the red trial-error block when error_message contains "exit 137" (SIGKILL/OOM-style termination). (#532, #534)

Fixed

Off-Modal dispatch no longer over-spawns workers: run_dispatch_cycle now sizes its spawn plan against max(running, held) per queue_key, using a new count_held_queue_slots helper, instead of worker_jobs RUNNING counts alone. An event-triggered cycle re-firing before newly-spawned workers show as RUNNING previously over-planned, causing wasted Docker/Kubernetes container churn as the extras lost the queue_slots acquire race and exited; Modal's production poll_queue is unaffected since it doesn't pass held counts. (#539)
An empty Alembic merge migration (merge_mca01_wjtoken_heads) unified two Alembic heads left by concurrently-merged migrations, which had blocked alembic upgrade head with "Multiple head revisions" and stalled the production Supabase DB migration deploy. (#538)
Grok Build trials no longer fail deterministically on large instructions: OddishGrokBuild.run() now uploads the rendered task instruction into the sandbox as a file (environment.upload_file, chmod 0644) and reads it back via grok -p "$(cat ...)" instead of inlining it into the exec argv (embedded up to three times across CLI fallbacks); large instructions previously produced exec commands exceeding Modal's 65536-byte ARG_MAX limit and failed at agent start. (#535)

Removed

The hosted per-user probe auto-opt-in default is removed: the users.run_probe_default column and its submission-time hook are dropped, so run_probe is now only ever true when explicitly requested on a submission — no per-user server-side default can silently enable probe trials. (#553)

Security

The unauthenticated public-experiments list endpoint (list_public_experiments) no longer queries or returns any experiment rows — it always returns an empty list — so share tokens can no longer be discovered by enumeration; direct /public/experiments/{public_token} lookups for a token a caller already has continue to work unchanged. (#558)

[2026-07-01]

Added

Grok Build trials are now converted into full ATIF trajectories: streamed grok-build.json events are parsed into reasoning/tool-call/tool-result/message steps with final metrics (tokens, cost) via a new agent wrapper, and the trajectory endpoint falls back to synthesizing ATIF from agent/grok-build.json for older trials that predate the conversion (#530)

Fixed

Completed Grok Build trials are now marked as having a fetchable trajectory even when the DB's has_trajectory flag is stale, so the dashboard no longer skips trajectory loading for them (#530)

[2026-06-30]

Added

oddish-query CLI gains five probe-only commands: solution cat, solution fetch, verifier source, harbor src, and verify run; probe-only assets are now staged to a root-owned hidden directory (/opt/oddish-probe, via ODDISH_PROBE_STAGE_DIR) off the agent's browsable tree so the agent can't passively stumble on verifier/solution content; every command output is wrapped in a PROBE-ONLY boundary banner (carried in a note field for verify run's JSON) so the boundary travels with the data through subagents; both local and cloud runners updated for identical container layout (#504)
CI guard (oddish/scripts/load_only_guard.py) that statically diffs columns read on the compact /tasks response-builder path against columns declared in load_only(...) sets in list_tasks_core, failing the build on any gap; prevents the MissingGreenlet class of 500 from shipping silently; triggered on PRs touching oddish/src/oddish/** (#495)

Changed

Probe details now open in a sliding ResizableDrawer panel from both the experiment trials matrix and the task probe-history table instead of navigating to a full page; a new shared ProbeDetailPanel component handles self-fetching, SWR polling, on-demand artifact loading, prev/next navigation across a task's probes, and agent-process keyword filtering; the standalone /tasks/[id]/probe/[trial_id] URL continues to render a full page via ProbeDetailPanel in contentOnly mode so deep links are preserved (#505, #508)
Task Analysis card on the task page now shows the latest run of each distinct probe type as a separate labeled section (probe-type header with agent · model sub-line) instead of collapsing all probe runs for a version to a single newest trial; polling continues while any per-type latest run is still in-flight (#512)
Reverted experiment grid to use GET /tasks?include_trials=true instead of the short-lived slim-tasks endpoint; removes the dedicated GET /experiments/{id}/slim-tasks backend route, GET /trials/{trial_id} single-trial detail route, and associated Next.js proxies; also reverts the Usage page Cost tab, CostingPanel SSR component, cost CSV export, and the series_top_n admin costs parameter (#507)
Bake a per-model ODDISH_MODEL_CONCURRENCY_OVERRIDES default into the Modal deploy that raises the xai/redacted-model queue-key concurrency lease to 128 (up from the 48 default), giving Harbor's grok-build arm more throughput; operators can still override the whole JSON via the env var / oddish-prod secret

Security

Probe trials are no longer returned by any public unauthenticated endpoint; get_public_task strips is_probe trials before returning the task, list_public_experiment_tasks excludes probes when scoping each task's trials (and now filters unconditionally regardless of experiment-id resolution), and list_public_task_trials always passes probe=False with the public probe query parameter removed — probe data never reaches the browser regardless of UI guards (#513)

[2026-06-29]

Added

Probe instructions now include a REFERENCE SOLUTION section when the golden/oracle solution is staged, telling the probe it may copy or adapt it into /app as a baseline before pursuing the operator directive; omitted when no solution/ directory was staged (#500)
Probe instructions always include a SUBAGENTS section encouraging the probe to fan out parallel Task-tool subagents for independent investigation threads, noting the one-level nesting limit so all parallel work is dispatched directly (#500)
Probe run page now shows a Preset line (between the run header and Summary section) displaying harbor_config.probe_name, falling back to the agent name for older or preset-less runs (#502)

Changed

Probe launch form "Instructions" field now populated from the selected skill's SKILL.md body (frontmatter stripped) instead of the legacy operator_prompt; skills without a SKILL.md file fall back to operator_prompt (#501)
Probe launch form fields renamed: "Extra instructions" → "Instructions", "Result focus (optional)" → "Output JSON / Result Focus" (#501)
extractSkillMdBody extracted from skills-client.tsx into a new shared frontend/src/lib/skill-md.ts module, reused by both the skills editor and the probe form (#501)
CLAUDE_CODE_SUBAGENT_MODEL is now pinned to the main agent's normalized model id for probe claude-code trials on both cloud (agent_config.py) and local (local_runner.py) paths, ensuring Task-tool subagents have an explicit model on the direct Anthropic API path where Harbor's run() does not set it; a pre-set value is never overridden (#500)

Fixed

Probe summaries using a result_focus JSON Schema with oneOf (e.g. from Pydantic v2 discriminated unions) no longer fail with BadRequestError: Schema type 'oneOf' is not supported; normalize_findings_schema now rewrites oneOf → anyOf at analysis time since the two are equivalent for constrained generation, transparently unblocking already-saved skills without operator action (#499)

[2026-06-28]

Added

Batch probe-analysis backfill Modal script (backend/scripts/backfill_analysis.py) that accepts comma-separated task names, finds eligible probe trials with S3 artifacts whose analysis isn't SUCCESS, resets their analysis state, and re-runs the analyzer; dry-run by default, --execute to write; reports per-name match counts and trials skipped for having no S3 artifacts (#493)
LLM-powered result_focus repair (core/result_focus_repair.py): malformed-but-JSON-intended result_focus values (trailing commas, single quotes, code fences) are coerced into valid JSON via a cheap Haiku pass before driving probe analysis or being stored on a skill; falls back gracefully on any failure, leaving the original value unchanged (#492)

Changed

Experiment owner and PR link are now stamped set-once on the experiment itself (new experiments.owner / experiments.link columns, backfilled from each experiment's earliest linked task) from the creating run's submitter; re-runs of shared tasks no longer overwrite the original experiment's provenance; dashboard and experiment detail view prefer experiment-level fields with task-derived fallback for un-backfilled rows; taskPrUrl now prefers link over github_meta.pr_url (#358)
Probe directive fields (operator prompt, evaluation metric, result focus) removed from the skill upload/edit form as they duplicate configuration handled elsewhere; the form now only shows upload folder, name, description, SKILL.md body, and additional files (#497)

Fixed

Probe summaries using a structured-output result_focus schema no longer crash with TypeError: unexpected keyword argument 'output_config'; the field is now forwarded via the Anthropic SDK's extra_body escape hatch, compatible with the pinned anthropic==0.76.0 (#493)

[2026-06-27]

Fixed

Probe result cells in the experiment trials table now navigate to the probe's dedicated run page (/tasks/{id}/probe/{trial_id}) instead of opening the trial drawer; non-probe cells are unchanged (#490)

[2026-06-26]

Added

Admin cost breakdown dashboard tab (/admin → Costs) with per-window totals (24h/7d/30d/all-time), cost-over-time chart stacked by model/user/agent dimensions, and ranked tables by user, model, and experiment; GET /api/admin/costs backend endpoint aggregates globally using native cost_usd when present and per-model token estimates otherwise; cost split between native and estimated spend is surfaced per entry (#452)
Run Probe tab in the QA tab bar (/qa/run) with a full-page task search that replaces the old "+ New probe run" dialog on the Probe Runs page; typing filters tasks, clicking one navigates to /tasks/{id}/probe to configure and launch; default landing on /qa still shows Probe Runs (#478)

Changed

Probe presets and skills unified into a single Skills feature: SkillModel gains optional operator_prompt, result_focus, and evaluation_metric directive columns; probe form now selects a Skill (not a preset); skills mount only when explicitly selected at launch via skill_ids rather than auto-mounting into every probe; probe_presets table, router, and schemas fully removed; /qa/presets redirects to /qa/skills; existing presets migrated into skills (ids preserved); 13 built-in directive and bundle seed skills seeded on fresh databases; auto-probe default repointed to the cheat-detector seed skill (#477)
Probe trials now render as a sorted-last "Probe" agent group inside the normal trials grid on both the task detail page and the experiment matrix, scoped to the task's effective version; the separate Probe tab is removed from the experiment view; backend batch-loads effective-version probe trials and merges them into each task's trials while keeping aggregate counts (total/completed/reward) probe-free (#471)
Probe launch buttons (task detail header, experiment trials table icon, and experiment "New probe") now navigate directly to /tasks/{id}/probe instead of opening an inline modal; the "Submit a probe run" CTA interstitial on the probe page is removed so the form renders immediately (#474)
Preview database schema bootstrap reverted to model-based creation (Base.metadata.create_all + alembic stamp head) instead of running the full Alembic migration chain on rebuild; migration fingerprint removed from schema trust marker (#469)

Fixed

Production incident: every GET /tasks (experiments page) 500-ing and all worker jobs failing with InvalidRequestError: One or more mappers failed to initialize — OrganizationModel.api_keys and UserModel.api_keys lost their join condition after #466 dropped DB-level FKs; fixed by adding explicit primaryjoin with foreign() annotation and viewonly=True on both relationships (#468)
Trials with verification disabled (verifier.disable: true) that complete with reward=None no longer consume all retry attempts before failing; the worker now terminates them as SUCCESS on the first attempt; the UI shows a new scoreless status (slate "SCORELESS" badge, minus-circle icon) instead of treating them as perpetually pending (#462)

[2026-06-25]

Added

oddish backfill-analysis CLI command to (re)run trial analysis (LLM trajectory classification + task verdict) for an experiment, a task, or a single trial; POST /tasks/{task_id}/qa/backfill backend endpoint on both cloud and local server with force, enable_analysis, and trial_ids options; rerun_task_qa_core refactored to delegate to the new backfill_task_analysis_core primitive (#456)
"Open task page" link button in the experiment trials table for direct navigation from the experiment view to a task's dedicated page; hidden on read-only share view since /tasks/[id] requires authentication (#442)

Changed

Chat button across all scopes (global /tasks header, per-task detail, per-experiment header) switched from outline to solid blue fill (bg-blue-600) for improved visibility (#465)
Preview database schema rebuilt by running alembic upgrade head for both stacks in order instead of Base.metadata.create_all + alembic stamp head; eliminates missing DB objects (e.g. queue_runtime_status, tag_projection_sweep_state, partial unique indexes) that only exist as raw DDL in migrations and are invisible to the ORM graph; production DROP SCHEMA guard re-introduced (#460)
Preview schema trust marker now folds in a migration fingerprint (SHA256 of both stacks' Alembic head revisions) in addition to the model-graph fingerprint, so a cached schema is also invalidated when a migration is added without any ORM model change (#460, #440)
Core package layout reorganized into focused subpackages: oddish.core.tags (service, projection, enqueue, filter_ast, naming, permissions, policies, profanity, ownership_transfer, saved_filters), oddish.core.sharing (public, helpers, documents), oddish.core.ingest (trial_imports, zip_imports, extraction), oddish.core.probe (auto_probe, presets); workers reorganized into oddish.workers.agents (claude_code, codex) and oddish.workers.harbor (runner, ephemeral, agent_config, outcome, storage, patches, modal_debug) (#438)

Fixed

api_keys cross-stack foreign keys dropped from org_id and created_by_user_id columns; new migration apk01dropfk drops constraints IF EXISTS so production converges; fixes NoReferencedTableError that prevented the oddish Alembic chain from bootstrapping independently (e.g. in the schema-parity CI job) (#466)
Probe and offline trials now get network_mode = "public" (and allowed_hosts cleared) instead of the legacy allow_internet = true flag in enable_local_internet; fixes silent no-op on Harbor tasks that set network_mode/allowed_hosts explicitly, which caused claude-code installs to SYN-timeout (~127s, curl exit 28) and the oddish-query CLI to lose egress; applies to both local Docker and cloud Modal probe paths (#464)
Chat session provisioning (POST /chat-sessions) no longer fails with "could not start chat" when the harbor pip install errors; harbor install is now best-effort (logs a warning and continues) since chat reads trial data through the oddish-query CLI, not the harbor package; claude-code install remains fatal (#458)
uq_worker_jobs_tag_project_active partial unique index declared on WorkerJobModel.__table_args__; prevents silent omission from model-built schemas where ON CONFLICT … DO NOTHING index inference for TAG_PROJECT job coalescing had nothing to infer against; no new migration (index already exists in the DB via aa00ta01core) (#454)
Preview bootstrap _rebuild_schema now calls _assert_preview_branch before DROP SCHEMA, refusing to proceed when ODDISH_DATABASE_URL resolves to production (matched via SUPABASE_PROJECT_REF or PREVIEW_SAMPLE_SOURCE_DB_URL) (#450)
Cloud auth migrations a1b2c3d4e5f6 and r4s5t6u7v8w9 made replay-safe against schemas built from the current model graph: supabase_user_id index creation guarded on column existence; userrole enum rebuild guarded on owner still being an enum value; fixes Supabase DB Migrations workflow failing with column "supabase_user_id" does not exist (#443)
Daytona dependency floor raised from >=0.165.0 to >=0.185.0; fixes Daytona trials failing with a misleading MissingExtraError caused by Harbor importing GpuType from the daytona SDK, which was only added in version 0.185.0 (#439)

[2026-06-24]

Added

Configurable per-run Harbor source via --harbor <spec> flag (or ODDISH_HARBOR env / oddish.toml [harbor] manifest): resolves the spec to a concrete commit SHA at submit time, stamps trials.harbor_sha and worker_jobs.harbor_variant_id, and executes the run on that exact Harbor version; blessed variants use digest-pinned worker images while arbitrary refs run in an ephemeral out-of-process engine (uv run --no-project --with harbor@<sha>); dispatcher routes on (queue_key, harbor_variant_id) so variants are isolated but share per-queue-key provider caps; Harbor commit shown in the trial detail drawer (#413)
Trajectory viewer keyword search bar: filters steps by message text, reasoning, tool call names/arguments, and observations; step count shows "N of M steps" while a filter is active; clicking a hidden step in the timing bar clears the filter to reveal it (#425)
trials.total_steps column (nullable integer) persists the total agent trajectory step count; extracted from trajectory.json at trial completion on cloud workers, the local runner, and CLI imports; surfaced in trial API responses and model usage aggregation (#396)

Changed

PR preview pipeline: backend stop step folded into prepare-database; Vercel and backend deploys decoupled via a deterministic Modal URL formula, enabling parallel execution; preview_api_url single-sourced as a workflow job output; stop_previous_preview_backend.sh removed (#430)
Preview database bootstrap now rebuilds the public schema from the combined oddish+backend model graph (Base.metadata.create_all) when the schema is untrusted (no oddish-preview:schema-built-from-base namespace comment); trusted branches continue to run alembic upgrade head only; re-seeds when a rebuild drops data (#429, #430)
Preview database schema is now upgraded to head on any backend deploy, not only when migration files change, so code-only PRs on reused branches cannot run against a stale schema; seeding remains gated to new branches and explicit migration runs (#424, #430)
Preview seed sample sizes drastically reduced (random experiments 4000→8, trials per experiment 100→50, skills/documents/presets 200→10) to cut seed time; code-only pushes no longer trigger re-seeding of an already-populated branch; prepare-preview-database workflow job capped at 12 minutes (#419)

Fixed

Experiment page 500-ing with MissingGreenlet on the compact trials path: harbor_sha was not included in the load_only set in list_tasks_core, causing a deferred-column lazy-load outside the async greenlet; added alongside its sibling harbor_config; CLAUDE.md updated to document the trap for future columns (#433)
Task cancellation database errors (e.g. deadlocks) now return HTTP 503 with a clear user message instead of propagating as an opaque 500 (Internal Server Error); full diagnostic detail (sqlstate, failing SQL, traceback) logged server-side; the Next.js cancel proxy now guards JSON.parse so a plain-text error body is forwarded correctly instead of surfacing a misleading parse exception (#421)
New preview database branches no longer replay the full Alembic migration history against an already-at-head schema: the bootstrap script reads each stack's revision from the parent database and stamps the branch to that revision before running upgrade head, so only this PR's new migrations run (#417)

[2026-06-23]

Added

Per-run container-registry credentials: any way a run is triggered (the oddish run CLI's --registry-login flag / ODDISH_DOCKERHUB_* env, the experiments-repo and harbor-forge CI workflows, a direct POST /tasks/sweep registry_auth field, or a trial retry request) can now supply its own Docker login so the trial sandbox's inner Docker-in-Docker daemon authenticates compose image pulls — fixing multi-service tasks failing at setup with Docker Hub toomanyrequests. The credential is per-user (never a shared Modal/oddish secret), Fernet-encrypted as it crosses the queue on worker_jobs.payload (never written to trials.harbor_config), passed to docker login before compose build/up via the Harbor DinD shim, then logged out on teardown. The encrypted ciphertext stays on the transient worker_jobs row while automatic retries remain possible and is scrubbed from the payload once the row reaches a terminal state.
GET /experiments/{id}/trials read endpoint returns all non-superseded trials for an experiment, gated by READ scope with org-scoped access control; trial rows now include an is_probe flag (#414)
Probe agent receives short-lived read-only API credentials and the oddish-query CLI on launch, enabling it to pull trial trajectories and logs on demand; a credentials-mint failure now fails the trial with a stored error rather than silently proceeding without access (#414)

Changed

oddish-query CLI ported from Python to a dependency-free Node.js script so it works in all claude-code environments where Node is guaranteed but python3 is not; old Python CLI removed (#414)
APIKeyModel, APIKeyScope, and key helpers relocated from backend/models.py into oddish.db.models and oddish.core.api_keys so workers can mint credentials; backend/models.py re-exports both for backwards compatibility (#414)
cc_chat scope CLAUDE.md templates rewritten around the oddish-query CLI (experiments trials, tasks trials, trials logs); per-scope artifact file mounting (experiment_files.py, task_files.py, file_loader.py) removed in favour of query-on-demand access (#414)
Probe trials are now forced to allow_internet=true to give the oddish-query CLI egress to the Oddish API (#414)
Experiment probe-launch button in the trials table is now always visible instead of appearing only on row hover; group/task-row hover marker removed (#400)
backend/uv.lock synced to record the missing jsonschema dependency edges under the oddish editable package, fixing dirty lockfile state on every uv invocation and uv sync --frozen compatibility in CI (#415)

Fixed

Codex trials on openai/* models (e.g. openai/gpt-5.5) no longer time out with zero tokens: AzureCompatibleCodex now implements required_outbound_domains to declare the configured Azure OpenAI endpoint host and per-trial OPENAI_BASE_URL alongside the OpenAI defaults, so Harbor's Modal egress firewall allowlists the Azure host and requests reach the model (#416)

[2026-06-21]

Changed

Automated daily changelog updated with entries for 2026-06-20 changes (#409)

[2026-06-20]

Added

POST /tasks/sweep/batch endpoint on cloud and standalone servers for submitting N task-sweeps in one request with per-item partial-success; each item runs inside its own savepoint so a failure in one item neither aborts the batch nor rolls back siblings; returns HTTP 207 Multi-Status when at least one item fails; CLI prefers the batch path and falls back to per-task only on 404/405 (#406)
Adaptive AIMD in-flight limiter for oddish run task submission: replaces fixed one-at-a-time concurrency with an additive-increase/multiplicative-decrease controller clamped to [4, 16]; backs off on 429/5xx, request timeouts, slow pool checkouts, and EWMA latency overshoot; S3 presigned-PUT step uses a separate smaller bound ([1, 6]) to avoid polluting the API backpressure signal; configurable via --submit-concurrency flag or ODDISH_TASK_UPLOAD_CONCURRENCY env (#404)
Submission idempotency for POST /tasks/sweep: CLI stamps a stable Idempotency-Key (canonical digest of experiment + task_id + sweep spec) on every sweep call; server deduplicates retried submissions and replays the stored response from a new submission_idempotency table (24h TTL, unique-insert-wins savepoint); same key + different body → 409 Conflict; prevents duplicate trials on transient 5xx retries (#399)
Opt-in task-submission timing harness in oddish/tests/perf/ measuring throughput (tasks/min), per-call latency (p50/p95/max per phase), 5xx count, and client-vs-server split; skipped in CI unless ODDISH_PERF and ODDISH_API_URL are both set (#391)

Changed

Batch sweep submission now chunks payloads client-side into groups of at most ODDISH_SWEEP_BATCH_MAX_TASKS (default 10, tunable via env) to stay under Modal's per-request ceiling; first-chunk 404/405 still falls back to per-task; later-chunk failures surface as an error to prevent double-submission after earlier chunks already committed (#407)
Task tarballing in upload_task is now deferred until the server returns a presigned upload URL; dedup hits (content-hash match on /tasks/upload/init) skip archiving entirely, saving CPU on re-runs where the task content hasn't changed (#390)
Sweep re-runs reconcile to exactly k trials per task in the target experiment: unchanged tasks (same version) count existing live trials and only add the shortfall; changed tasks (new version) still get a full k; the --add opt-out flag is removed — reconcile-to-N is now unconditional (#386)

Fixed

/tasks/sweep server-side DB round-trips cut: deduplicated browse-projection recompute in create_task (was running twice — once on incomplete pre-version state, once after trials); per-row session.add loops replaced with a single INSERT … SELECT unnest(…) WITH ORDINALITY statement for both trials and worker_jobs, keeping the statement shape constant under Supavisor transaction pooling (#397)
Upload init and complete calls now retry on 429/500/502/503/504 with capped exponential backoff, full jitter, Retry-After header support, and a token-bucket retry budget (≤10% of requests); transient blips no longer abort a submission entirely (#397)
probe_presets_001 Alembic migration now adds ratio_unit and ratio_verb columns idempotently before the bulk seed insert, fixing alembic upgrade head failures on fresh databases where 000_initial_schema creates probe_presets from current models that no longer carry those columns (#401)

[2026-06-19]

Added

Experiment-level probes: submit a probe whose agent can read artifacts from all trials in an experiment (per-task-balanced, up to 30 trials); "New probe" button added to the experiment Probe tab in both empty and populated states (#372)
Probe launch button in the experiment trials table (hover icon next to version badge) and task detail header (labeled "Launch probe"), both opening the probe submit form in a dialog (#385)
DinD Docker daemon failure diagnostics: when dockerd fails to start in a Harbor DinD worker, its logs (tail 200), process state, memory, and disk are captured from the still-alive VM and folded into exception.txt to aid root-cause analysis (#361)

Changed

All claude-code trials now route to the direct Anthropic API instead of Bedrock (Bedrock credentials were unavailable); trial classifier also switched off Bedrock to the direct Anthropic API (#374, #377)
claude-code model id now matches the transport: when Bedrock env signals are absent, _build_agent_config emits a plain Anthropic API model id (e.g. claude-sonnet-4-6) instead of a Bedrock inference-profile id (global.anthropic.claude-sonnet-4-6), preventing HTTP 400 "Operation not allowed" errors (#359)
Probe result-focus redesigned: result_focus is now a JSON-schema structure with enforced ResultFocusFindings fields; action items lead instead of free-form text; ratio metric removed; full JSON summary accessible via toggle (#370, #378)
Probe result display overhauled: cheating badges, "investigation steps," and "task is gameable" tally removed from the summary panel; summary row and latest-probe card now show action-items count (no action items / N action items · M must-fix) instead of cheat verdicts; reward 0.0 fallback replaced with "awaiting analyzer" (#365, #379)
Probe transcripts no longer clipped before summarization; probe analyzer upgraded to a larger model (#364)
Auto-probe now defaults to the Task Construction Auditor preset (#376)
Subscription auth route (sub/<model> prefix, OAuth token / Codex auth.json path) and the claude-opus-4-8 sub-bucket concurrency bump reverted; related config symbols, tests, and BYOC credential infrastructure removed (#371)
Modal worker container cap raised 768 → 2688; connection-budget comment updated to document the 2882/3000 worst-case client-connection estimate (#387)
nop/oracle queue concurrency default raised 32 → 256 in both core config and the Modal deployment (#380)
Harbor runner split into focused submodules (harbor_agent_config.py, harbor_modal_debug.py, harbor_outcome.py, harbor_storage.py); harbor_runner.py retained as an orchestration facade for backward-compat imports (#383)
Project repositioned as "batch execution and continuous QA for Harbor-compatible RL environments" in the landing page hero copy, README tagline, site metadata, and pyproject.toml keywords (#388)
API key creation permission check extended to recognize the Abundant production Clerk org ID (org_39ufkEqie8rLlVhoK4YMm4IMx0L) in addition to org slugs (#382)

Fixed

Cancel deadlocks with concurrent worker progress: cancel_tasks_runs now locks trial rows first then parent task rows (matching the worker domain-write order), and uses a WITH … FOR UPDATE CTE to lock matching worker_jobs rows before cancelling them (#393, #395)
Unresolved Git LFS pointer files are now detected before archiving and uploading a task directory; oddish run fails fast with the affected relative paths and a git lfs pull remediation hint (#394)
Probe summary crash on token-cap-truncated analyzer JSON (#375)
Probe build error: second result-focus panel now rendered via ResultFocusFindings (#373)
Experiment agent and model icons: agent-harness logos shown in experiment table column headers; model logos used for model rows and pass@k legend; / separator replaces @ in model-scoped experiment labels (#381)

[2026-06-18]

Changed

Official provider logo assets: Gemini, Kimi, and MiniMax now use official SVG assets (google-gemini.svg, kimi-k-only.svg, minimax-vertical.svg) instead of third-party icon library glyphs; a shared ProviderLogoImage component consolidates the rendering (#355)
Harbor pin updated to beabbb7a picking up the agent-tools image update that bakes ripgrep into closed-internet Codex trials so runtime installs are skipped when codex and rg are already prebaked (#357)
sub/claude-opus-4-8 deploy-time concurrency raised 4 → 8 in modal-deploy.yml, roughly halving wall-clock for the claude-code eval arm running against the OAuth-multiplexed Claude Max subscription (#352)
Task and trial ID columns widened: tasks.id VARCHAR(64→128), task_versions.id VARCHAR(128→160), trials.id VARCHAR(128→160), and all FK references widened correspondingly; Alembic migration long_task_ids_001 handles the resize under ACCESS EXCLUSIVE lock with FK rebuild (#354)
Org role model simplified to admin / member; the legacy owner role is removed (existing owners promoted to admin via migration r4s5t6u7v8w9) and API key creation is now gated to admins with an @abundant.ai email in the Abundant org; new GET /api-keys/permissions endpoint reports whether the current user may create keys (#170)
CI preview pipeline now computes component diffs via the GitHub compare API instead of dorny/paths-filter, which ignored its base input on pull_request events and always diffed the whole PR, forcing unnecessary ~15-min Supabase DB seed runs on frontend-only pushes (#346)

Fixed

Experiment-scope chat sessions now mount the jobs artifact tree inside the Daytona sandbox; previously the experiment branch of _resolve_scope_inputs left files empty so the sandbox only received CLAUDE.md and the agent reported "(no trial data available yet)"; new collect_experiment_files mirrors collect_task_version_files and uploads artifacts at jobs/{experiment_id}/{trial_id}/… as the CLAUDE.md template promises (#347)

[2026-06-17]

Added

Fireworks routing: GLM / MiniMax / Kimi (and other open models) can run on the stock claude-code agent via Fireworks' single Anthropic-compatible endpoint. Opt in with an explicit fireworks/ (or fw/) prefix (e.g. fireworks/glm-5.2, fireworks/minimax-m3, fireworks/kimi-k2.7-code), which gets its own fireworks/<id> provider/queue bucket and authenticates with ${FIREWORKS_API_KEY}; bare glm/minimax/kimi ids keep their existing direct-provider routes. Default claude-code settings (no forced thinking/effort)
Global chat scope (global) with an oddish-query read-only CLI injected into the sandbox; the agent can search, inspect, and drill into trial logs across all org tasks; a short-lived internal API key is minted per-session (45-min TTL) for credential isolation; global-scope Chat button added to the tasks page (#332)

Changed

Chat sandbox provisioning now supports an optional pre-baked Daytona snapshot (ODDISH_CC_CHAT_DAYTONA_SNAPSHOT): ClaudeCodeRuntime.install() skips tools already present and installs any missing tools concurrently instead of sequentially, reducing provisioning from ~1 min to a few seconds when a snapshot is configured (#340)
API container concurrency reduced from 8 → 3 and max containers raised 24 → 64 (peak throughput unchanged at 192); experiment-scoped GET /tasks now loads only the requested experiment's non-probe trials in SQL instead of fetching every trial for each task and filtering in Python; GET /tasks limit parameter capped at 2000 (#337)

Fixed

Chat session creation (POST /chat-sessions) returned 500 in prod because the Daytona region only permits ephemeral sandboxes; create_sandbox now passes ephemeral=True (#333, #339)
Chat messages returned session_not_found when the API routed the request to a different autoscaled container than the one that provisioned the session; sandbox handles are now reconnected from the DB-persisted sandbox_id so any container can serve any session (#334, #339)
First chat message showed nothing for ~10 seconds during sandbox provisioning and the composer stayed live, allowing a second send to race in; user bubble is now echoed and composer locked before ensureSession() runs (#335)
Per-task Chat button showed "no trial data available yet" because it used the stub task_probes scope; button now uses the task scope with the full trial tree staged per version; probe trials are marked (probe) in the agent's CLAUDE.md with a "Regular runs vs probe runs" explanation (#336)
GET /chat-sessions (chat history list) took 19–43 s under prod DB load; a composite index on (org_id, scope_kind, scope_id, last_activity desc) now serves the filter and sort directly, and turn counts are batched into one grouped query instead of a correlated subquery per row (#338)
Claude Code ran headless (--print) without a permission flag, so every tool call (Bash, Read, etc.) blocked on an approval gate nothing could answer; --permission-mode bypassPermissions is now passed on both the chat (stream_chat) and probe (run_once) launch paths (#341)
After ~30 min idle, Daytona auto-stops and deletes the ephemeral sandbox; the next message raised session_not_found; send() now transparently self-heals by calling resume() to re-provision the sandbox and restore from the per-turn archive (#342)
Container startup sweep (sweep_orphan_chat_sessions) ran on every API container start and marked all active chats broken — since the API autoscales across many containers with no session affinity, any new container (autoscale-up or deploy rollout) was killing every live chat globally; startup sweep removed from lifespan; recovery is now lazy (reconnect by sandbox_id, self-heal via resume(), Daytona idle auto-stop as backstop) (#343)

[2026-06-16]

Added

Claude Code chat sessions (Phase 1): durable chat_session_events append-only log and chat_turns table (one-running-turn-per-session enforced by a partial unique index) with a full orchestration engine — Daytona sandbox provisioner, Claude Code runtime, idle reaper, and restart sweep that marks orphaned running turns failed while preserving the event log; API routes POST /chat-sessions, GET /chat-sessions/{id}, SSE POST /chat-sessions/{id}/messages, events-replay GET /chat-sessions/{id}/events?since=<seq>, and DELETE /chat-sessions/{id}; sessions survive page refresh and backend container restarts (#306)
Chat task scope (Phase 2a): chat sessions scoped to a task download trial log files from S3 and upload them into the Daytona sandbox as jobs/v{version}/{trial_id}/… (byte-capped at 50 MB); a version-aware CLAUDE.md highlights the current version as the default focus and de-emphasizes past versions (#307)
Task detail page now shows a "Probe runs" card for the selected version: the latest probe run's agent/preset name, run status, cheat/blocked/neutral result, and prioritized action items from the analyzer; auto-polls while the probe is in-flight and links to the full probe result page (#310)
Admin GET /queue-health endpoint and dashboard overview card exposing throughput, per-queue-key capacity fill, and persisted dispatcher/reconciler heartbeats so operators can self-diagnose "queued but not running" without querying psql or Modal logs; backed by a new queue_runtime_status table written at the end of each dispatcher/reconciler cycle (#312)

Changed

Probe submit page now shows a prominent "Submit a probe run" button that expands to reveal the agent picker and form on click (previously the form rendered inline on agent selection); probe history list sorted newest-first; task ID on the probe run detail page rendered as a clickable link to the experiment page (#313)
Modal function CPU/memory resource floors now configurable via env vars (ODDISH_MODAL_API_CPU/MEMORY_MB, ODDISH_MODAL_WORKER_CPU/MEMORY_MB, ODDISH_MODAL_DISPATCHER_CPU/MEMORY_MB, ODDISH_MODAL_RECONCILER_CPU/MEMORY_MB); API defaults to 2 CPU / 4 GiB (was unconstrained fractional-core), reducing latency spikes under concurrent load; WORKER_MAX_CONTAINERS raised 320 → 448 (#312)
Probe submit form converted to shadcn Button, Input, and Select components; unused frontend exports flagged by knip removed; pre-commit hooks (ruff, black, mypy, prettier) pass cleanly across the full repo; dead code removed: TrialClassifier.classify_trials batch method and AUTO_PROBE_INSTRUCTIONS constant (#311)

Fixed

Worker dispatcher no longer starved by a slow or deadlocking reconciliation sweep: reconcile_queue_state now runs as its own dedicated Modal scheduled function (240s interval, 600s timeout) instead of inline inside poll_queue; a SIGKILL mid-sweep previously left orphaned idle in transaction locks that deadlocked the next sweep cycle and spawned zero workers; poll_queue now only discovers queue keys and spawns workers, with MAX_WORKERS_PER_POLL raised 64 → 128 (#309)

[2026-06-15]

Added

Copy button in the task file viewer header copies the raw file content to clipboard with a 2-second check-icon confirmation; resets on file switch and cleans up its timeout on unmount (#299)

Changed

Trajectory analysis is now a single task-level QA worker job instead of one classification job per trial plus a separate verdict job: when every trial of a run_analysis task finishes, one QA job classifies all live trials (unchanged taxonomy/evidence/reasoning, still written to trials.analysis) and then synthesizes the task verdict (tasks.verdict), so a sweep of T tasks × N trials enqueues T jobs instead of T × (N + 1). The whole surface uses one "QA" concept: worker-job kind QA (migration qa01/qa02 adds it and repoints old VERDICT rows; ANALYSIS/VERDICT remain only as legacy enum values for historical/in-flight rows), one run_task_qa_job handler, one set of endpoints (POST /tasks/{id}/qa/retry, POST /tasks/{id}/qa/cancel), one CLI surface (oddish run --retry --qa, oddish cancel --qa), and one dashboard control (Run QA / Cancel QA). The per-trial analysis and separate verdict retry/cancel endpoints, CLI flags (--analysis/--verdict), and UI buttons were removed (#315) (#315)
Probe agent container now receives the full staged task directory via a Harbor AGENT_START hook, with all probe-only material (tests/, solution/, harbor_src/, related_trials/, AGENT_BRIEF.md) staged under /probe-harness/ instead of /app, keeping the real agent's workspace pristine; probe instruction reframes the task spec as a "REAL AGENT BRIEF" with an auto-generated visibility map, eliminating false-positive vulnerability reports for files the real agent cannot access (#300, #301)
Probe analyzer prompt gains a SCOPE section instructing it not to emit recommendations premised on probe-only paths (under /probe-harness/) being agent-reachable, and to preserve the probe agent's own hedges rather than upgrading them to must_fix (#301)
Harbor bumped to 07a576944 picking up MiniMax M3 and Kimi K2.7 long-run hardening: streaming/timeout env vars (API_TIMEOUT_MS=3.6M, idle stream timeout, eager flush, max output tokens), Claude Code pinned to 2.1.167 (fixes MiniMax exit-137 mid-stream stalls), and plan-mode tools (EnterPlanMode, ExitPlanMode, AskUserQuestion) disabled for Kimi variants (fixes K2.7 plan-mode no-op bail) (#303)
QA Probe Runs listing (/qa/runs) now aggregates in SQL using window functions — one row per task — rather than fetching all probe trial rows and folding them in Python; backed by a new partial index on (org_id, task_id, created_at DESC) WHERE is_probe, making load time scale with tasks-per-org rather than total probe trial count (#286)

[2026-06-14]

Added

Advanced free-text grammar for the task browser search box: space-separated terms AND together in any order, "quoted text" matches as a contiguous phrase, a leading - (or uppercase NOT) excludes a term, and uppercase OR makes either side of a group match; a ? icon inside the input opens a syntax cheatsheet tooltip; parse_search_query lives in oddish/core/helpers.py so the dashboard, standalone server, and cloud API all share one grammar; LIKE metacharacters remain escaped as literals (preserving #285 semantics); needles capped at 16 (#295)
"Delete tag for everyone…" action in the tag chip editor: an inline confirm panel shows the tag's current name (refreshed after 409 races) and its direct-assignment count before the destructive click; sends cascade=true to flip ACTIVE assignments; onDeleted drops the chip locally without a redundant unassign call; DELETE passthrough added to the /api/tags/[tag_id] Next.js proxy route (#293)

Changed

Bake per-model ODDISH_MODEL_CONCURRENCY_OVERRIDES defaults into the Modal deploy that raise the global.anthropic.claude-haiku-4-5-20251001-v1:0 (also the analysis-model queue key) and openai/gpt-5.4-mini queue-key concurrency leases to 128 (up from the 48 default); trajectory analysis gets more headroom; operators can still override the whole JSON via the env var / oddish-prod secret (#297)

Fixed

DELETE /tags/{id} backend route now accepts a ?cascade= query parameter so callers can consent to flipping ACTIVE assignments to REMOVED; previously the flag was unreachable from HTTP and tag deletion always failed for any still-assigned tag (#293)

[2026-06-13]

Added

Task page now lists all affiliated experiments as linked chips (dot-separated) instead of just the primary one; public share view only exposes public experiment names to prevent private experiment name leakage, and GET /public/tasks/{id}?include_trials=false no longer 500s for tasks in multiple public experiments (#288)
Tag chips on dashboard experiment rows and tag: / -tag: / OR / NOT filter syntax in the experiments search box, matching the existing /tasks grammar; tag chips hydrated in a single batch query per page with graceful degradation on failure (#291)
Admin "Tag Policy" tab now fully functional: numeric limits, who-can-create and profanity-mode toggles, comma-list editors for reserved prefixes and allow/deny lists, and a 403 → "Admins only" error state (#291)
Probe summary now includes a prioritized "Action items" block with must_fix / should_fix / optional recommendations, color-coded and sorted by severity; an empty probe returns a "No fixes needed — task held up to probing" confirmation; legacy probe rows without the field render nothing (#284)
Probe trials now upload the staged task directory (including related_trials/, harbor_src/, tests/, solution/) into the agent container at start time via a Harbor AGENT_START hook, so the agent can actually access the reward-hack surface the probe instruction references (#282)
Skill create/edit form gains an "Upload folder" button that reads a SKILL.md plus supporting files and auto-fills the form; strips dotfiles and binary files; shows a "skipped N files" notice for anything dropped (#281)
Saved-filter bookmark menu beside the tasks search bar: lists org-shared and private saved filters, applies one as tag: search text, and saves the current query under a name with Private/Org visibility; filters persist stable tag IDs so they survive renames and merges; deletes are optimistic with SWR rollback on failure (#280)

Changed

Dashboard "Avg score" column and experiment page KPI tile now use a task-weighted average (mean over tasks of per-task mean reward) with nop/oracle baselines excluded everywhere; backend computes and returns this as a new avg_score field; a loading spinner is shown on the KPI tile while trial pages are still streaming in; both surfaces explain the calculation on hover (#292)
Tag mutations (assign, unassign, delete, archive, merge, set-visibility) now invalidate the dashboard cache, so tag chip and filter changes appear immediately without waiting for the 30-second TTL (#291)

Fixed

Probe runs no longer pollute the task browser: trial counts, reward stats, experiment chips, and last_run_at (which drives page ordering) now all exclude is_probe trials, consistent with probes having their own tab (#285)
LIKE wildcards in task browser and document search are now escaped as literals: searching _ no longer matches every task, and % and \ behave as plain characters (#285)
Browse page query performance improved: the org-wide trial aggregate now computes only the max(activity) needed for ordering; per-task counters are fetched as a separate targeted query over the visible page only, reducing latency ~40% at prod volume (#285)
GET /tasks no longer intermittently 500s with MissingGreenlet; TrialModel.is_probe added to the compact-trials load_only allowlist so it is loaded eagerly on the async path (#283)
Probe summarizer no longer reports "received no output" for skills: user-turn text blocks (the mechanism claude-code uses to deliver skill bodies) are now captured as injected_context so the summarizer sees the full skill content (#289)

[2026-06-12]

Added

QA tab consolidating probe runs, presets, skills, and documents under a new /qa route group; includes a Probe Runs list (one row per task with probe activity, ordered most-recent-first), a full Presets CRUD management page, and a new org-wide GET /probes backend endpoint; old /skills and /documents routes redirect transparently to their new /qa/* homes (#266)
Probe run summary accessible from the experiment task-file drawer sidebar: a "Latest probe run" entry below the file tree opens a compact card with status, agent/model, headline, metric chips, and cheating verdict; links to the full probe detail page; hidden on public share view; auto-polls while the probe is still running (#265)

Changed

Auto-probe is now opt-in and off by default; maybe_enqueue_auto_probe previously fired unconditionally on every sweep; now gated on a new run_probe: bool field on TaskSubmission, TaskSweepSubmission, and TaskStatusResponse, a run_probe DB column with migration run_probe_001, and a --run-probe CLI flag on oddish run; append-mode submissions flip the flag on first opt-in, mirroring the existing run_analysis pattern (#271)
Org switcher moved from Settings > Workspace into the top nav bar so the active workspace is always visible and switchable; settings page retains a read-only current-workspace card with updated copy; SWR cache is flushed on org change to prevent stale data from the previous workspace leaking through (#274)
Nav "QA" label renamed to "Agents" (route, active-state checks, and QA Verdict/QA Review strings elsewhere are unchanged) (#270)
Probe run detail page and experiment task-drawer probe card now share a single ProbeRunSummary component, bringing the drawer card to full parity with the detail page; the card previously omitted key_actions and tool_insights sections (#269)
Preview environments now seeded with a pseudo-random, deterministic subset of real production data (rows drawn by md5(id || PR_NUMBER)) instead of curated fixtures; reviewers authenticate with their real org credentials in preview; in-flight tasks/trials are normalized to FAILED on import to prevent the preview's safety nets from enqueuing real analysis/verdict jobs; convergence tracked via a private _preview_seed_state table; per-row savepoints prevent a single constraint collision from aborting the run (#264)

Fixed

Probe trials no longer appear in the experiment trial matrix; filtered server-side in the experiment_id branch of list_tasks before version resolution, preventing probe runs from cluttering the main task grid and preventing a probe-only version from skewing the effective version display; public /public/tasks/{task_id}/trials endpoint gains an optional ?probe= filter param (true/false/omit) (#276)
?task= deep links on the experiment page now fall back to matching by task name when no exact task ID is found, fixing hand-written links such as ?task=ghsa-rpfr-x88x-xwcw that previously opened nothing (#275)
Task version badges (v{n}) in the public experiment table are now hidden for unauthenticated viewers; oddish run reproduction command in the trial drawer hidden on public share pages; timing row in the trial drawer now shows the viewer's local timezone abbreviation (e.g. PDT) (#277)
GLM/z.ai provider icon corrected from the ChatGLM mammoth glyph to the ZAI "Z" logo (#277)
Footer "by Abundant AI" link updated from abundantdata.com to abundant.ai (#277)

[2026-06-11]

Added

Auto-enqueue one probe trial per task version on sweep submit using round-robin model rotation; new GET /experiments/{experiment_id}/probes endpoint lists the latest probe trial per task; experiment page fetches and displays the default probe preset (#248)
Dashboard experiment owner filter (Org / Mine / admin member picker) via experiments_author parameter on /dashboard; defaults the Recent Experiments table to the signed-in user's experiments; legacy rows matched via GitHub username and email for accurate attribution (#214)
Self-healing experiments owner backfill runs on every queue-poll tick, converging owner_user_id so the dashboard Mine filter stays on its indexed fast path; adaptive filter uses a pure indexed owner_user_id seek once all experiments in the org are attributed, with an __unattributed__ sentinel for unowned rows (#255)
ShareNav component for public experiment share pages with Oddish logo, "by Abundant AI" branding, and theme toggle; replaces the full app Nav on share pages (#260)
Cursor provider icon support: cursor and cursor-cli agents and the composer-* model family now resolve to the Cursor icon in the queue key icon display (#260)
Claude Fable 5 mapped to Bedrock global cross-region inference profile (global.anthropic.claude-fable-5); note this is a Covered Model requiring AWS account data retention mode set to provider_data_share (#254)

Changed

Pass@k graph and leaderboard default to visible on public experiment share pages; authenticated views are unchanged (#256)
Tasks/Probe tab toggle hidden on public share view — probing requires an authenticated session (#260)
OpenRouter models (Gemini 2.5 Flash, DeepSeek Chat) removed from probe model rotation; only claude-haiku-4-5 remains (#253)

Fixed

Pass@k calculation now uses per-agent per-task attempt count instead of the global maximum, fixing incorrect curves when agents in the same experiment have different trial counts (e.g. oracle with 1 trial shown as 33%/67%/100% instead of its observed flat rate) (#258)
Dashboard Mine filter no longer degenerates into a near-full-table scan; an adaptive indexed owner seek is used once owner_user_id is backfilled; attribution profiles are served from cache instantly with background refresh; frontend keeps existing rows visible during polling revalidation instead of blanking to "Loading…" (#255)
GitHub PR branch link (experiment PR chip) hidden on public share view to avoid exposing internal repository context to anonymous viewers (#261)
skills_001 and documents_001 Alembic migrations made idempotent for data-less preview branches, fixing "relation already exists" errors that caused PR preview deploys to fail at alembic upgrade head (#251)

[2026-06-10]

Added

Skills library and agent doc store: org-scoped skills with CRUD, a frontend Skills page, probe sandbox injection (.claude/skills/ via overlay), and Harbor AgentConfig.skills delivery; a document library with text/markdown/PDF/CSV ingestion, LLM digest generation (summary + tags), and keyword+tag search; oddish-docstore MCP server exposes search, get, and inspect tools; Skills and Documents added to the app nav (#217)
is_probe boolean column on trials with migration and backfill from harbor_config->>'mode'; ?probe=true/false filter on GET /tasks/{id}/trials; probe history table now filters server-side (?probe=true) instead of client-side; inline probe summary generated in the cloud trial handler after each probe trial completes (#231)
Probe skill and MCP tool usage captured as structured signal in probe artifacts: _classify_tool_use tags every transcript tool_use entry as skill, mcp, or builtin; _summarize_tool_usage produces a deterministic tool_usage roll-up (skill slugs and MCP server/tool pairs, ordered by first appearance) surfaced under extract_probe_artifacts (#244)
Probe summary "Tools & skills used" section: when an agent invoked skills or MCP servers, run_probe_analyzer appends per-tool tool_insights entries (name, kind, one-sentence note grounded in the transcript); rendered as a labeled list with Skill/MCP chips on the probe result page (#245)

Changed

PR preview databases switched from prod-clone (--with-data) to data-less Supabase branches populated by a deterministic seed (backend/preview_seed.py): reflection-driven, idempotent and convergent (upsert + full-PK reconcile), spanning both Alembic stacks without ORM imports; seed-gate CI job validates schema/seed drift on every PR; preview branch readiness deadline cut from 20 min to 5 min (#243)
Harbor dependency bumped to 0.13.1 (from 0.8.0), adding the glm-claude-code agent (z.ai base URL, ZAI_API_KEY auth, recommended streaming env, Claude Code version pin 2.1.167), closed-internet IPv4 fixes for api.z.ai, and harbor-framework v0.13.1 (#247)
Settings now auto-loads .env.local layered over .env for local backend development; later file wins on duplicate keys, exported env vars still outrank both (#230)

Fixed

Probe result page now shows agent transcript and verifier output for cloud trials; cloud runs do not inline _artifacts into trial.result, so a new GET /trials/{id}/probe-artifacts endpoint and BFF proxy download artifacts on demand from object storage and cache them for finished trials (#227)
Cloud probe summary no longer fails with auth errors (PermissionDeniedError / TypeError); the probe analyzer now always runs on the direct Anthropic API (ANTHROPIC_API_KEY) instead of Bedrock; Bedrock inference-profile model IDs are normalized to their plain API form via new to_anthropic_api_model_id() helper (#236)
Cloud probe summary NameError fixed: extract_probe_artifacts and run_probe_analyzer imports were missing from trial_handler.py, causing every cloud probe run's inline summary to fail with NameError: name 'extract_probe_artifacts' is not defined (#232)
Experiment page no longer returns 500 MissingGreenlet errors for tasks with trials; TrialModel.harbor_config and TrialModel.is_probe added to the compact-trials load_only allowlist so deferred JSONB columns are not lazy-loaded outside the async greenlet (#228, #235)
Preview branches stuck in RESTORE_FAILED, INIT_FAILED, or PAUSE_FAILED states are now detected immediately and torn down for recreation instead of polling until the 20-minute deadline; deadline timeouts also trigger delete-and-recreate; retry budget raised to 3; seed reclaims clerk_org_id from pre-existing rows on reused branches to avoid duplicate-key errors (#224)

[2026-06-09]

Added

GLM / z.ai routing for the claude-code harness: GLM models (zai/glm-x-preview[1m], bare glm-..., or z-ai//z.ai/ prefixes) canonicalize to a zai/<id> id so they get their own zai provider and zai/<id> queue bucket instead of inheriting claude-code's fixed Bedrock provider/queue (keeping GLM trials from contending with Bedrock traffic for concurrency slots); harbor_runner injects the z.ai Anthropic-skin env (ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN=${ZAI_API_KEY}, model + size aliases, z.ai's recommended long-context settings) and blanks the ambient Bedrock/Anthropic credentials so the z.ai route wins
oddish probe CLI command with full cloud probe agent support: probe presets stored in Postgres with CRUD backend endpoints and UI on task pages; probe trials bypass strict task.toml timeout validation and use a capped 30-minute agent timeout; local and cloud runners share the same probe implementation (#218)

Changed

Experiment trials table on the experiment page now defaults to A→Z task name sort instead of insertion order; the sort header toggle still cycles through all options (#220)

Fixed

Probe UI no longer shows "Failed to fetch" CORS errors; browser fetches for probe presets and sweep submissions moved from direct cross-origin backend calls to the same-origin Next.js BFF proxy (/api/probe-presets, /api/tasks/sweep), removing CORS as a failure point; reverts the hardcoded prod-origin CORS allowlist from #221 since it is no longer needed (#222)
Probe trials on tasks whose task.toml omits agent timeout settings no longer hard-fail validation; a shared PROBE_AGENT_TIMEOUT_SEC constant (default 30 min, overridable via ODDISH_PROBE_AGENT_TIMEOUT_SEC) is applied by both local and cloud runners (#225)
PR lineage badges now render consistently across dashboard, experiment page, task detail header, and task browser cards; fixed blank badge on the experiment page caused by TaskModel.link missing from the compact_trials load_only set; fixed missing badge on task cards when the PR URL lived only in github_meta.pr_url; added shared taskPrUrl(link, github_meta) resolver in lib/utils.ts; dashboard PR column now falls back to the link column when github_meta is absent (#197, #223)
Fresh-database alembic upgrade head no longer fails with duplicate column or foreign-key errors; three incremental migrations (add_column last_activity_at, fk_tasks_current_version_id, fk_trials_experiment_id) are now idempotent; a merge migration joins the two divergent heads into a single chain (#215)
Resolved Alembic double-head in the oddish migration chain caused by probe_presets_001 and dispatch_log branching from the same migration tip; probe_presets_001 re-parented onto c1d2e3f4a5b6 to restore linear history (#219)

Removed

Repository dispatch emitter (github/dispatch.py) and experiment_dispatch_log table dropped; the repository_dispatch event path requires Contents:write permission that the production token lacks (#211)

[2026-06-08]

Added

Fire a repository_dispatch webhook to consumer repos when all tasks in an experiment reach a terminal state; an experiment_dispatch_log table provides idempotent single-fire semantics; dispatch target and event type are read from github_meta.dispatch on task tags; gated by GITHUB_DISPATCH_ALLOWED_REPOS allowlist and authenticated via GITHUB_DISPATCH_TOKEN (falls back to GITHUB_TOKEN) (#206)

Changed

Bake a per-model ODDISH_MODEL_CONCURRENCY_OVERRIDES default into the Modal deploy that raises the google/gemini-3.5-flash queue-key concurrency lease to 128 (up from the 48 default); operators can still override the whole JSON via the env var / oddish-prod secret (#213)
Raise ODDISH_DEFAULT_MODEL_CONCURRENCY fallback from 32 to 48, increasing per-model queue-key concurrency in the Modal runtime without changing the per-poll spawn cap (#208)
Experiment trials table tooltip for truncated task names now shows the full task name instead of the generic "View task files" label, with responsive max-width and word-break styling for long names (#207)

[2026-06-07]

Changed

Automated daily changelog updated with entries for 2026-06-06 changes (#202)

[2026-06-06]

Added

Trial detail panel now shows a "Sandbox" button linking to the Daytona dashboard when a trial has an associated Daytona worker job; provider and external_id fields exposed in the worker job API response to enable this (#190)

Fixed

Codex workers running against Azure OpenAI endpoints no longer fail with 302 errors from the websocket Responses route; a new AzureCompatibleCodex runner disables the unified_exec websocket transport and injects an HTTP-only OpenAI-compatible provider config; trajectory is recovered from stdout JSONL as a fallback when the Codex session file is sparse (#193)
enqueue_analysis_worker_job now skips enqueueing analysis for trials with no stored result (neither S3 key nor local path), immediately marking analysis FAILED instead of burning all 6 retries on a doomed job; a staleness-gated cleanup backstop finalizes ANALYZING tasks with no live trials and cancels their dangling queued ANALYSIS worker jobs (#196)
Stuck-ANALYZING cleanup pass rescoped to correctly target tasks whose live trials have analysis_status = NULL (analysis was never enqueued) rather than tasks with no live trials at all; NULL analysis statuses are now marked terminal so maybe_start_verdict_stage can advance the task to VERDICT_PENDING instead of leaving it indefinitely blocked; tasks with no live trials are still finalized FAILED (#200)
Worker containers now drain short-job queues by claiming and running multiple jobs back-to-back on their held slot until the queue empties or a wall-clock budget (ODDISH_MODAL_WORKER_BATCH_BUDGET_SECONDS, default 300s) expires; lifts utilization for analysis (~54s), verdict (~9s), and nop-oracle (~46s) queues toward 100% without changing global spawn rates or concurrency limits; long agent trials exceed the budget on the first job and continue to run one-per-container (#201)

[2026-06-05]

Added

oddish cancel gains --analysis and --verdict flags to cancel active analysis or verdict jobs independently without stopping unrelated trials; new API endpoints POST /tasks/{task_id}/analysis/cancel, POST /tasks/{task_id}/verdict/cancel, and POST /trials/{trial_id}/analysis/cancel; dashboard adds per-trial, bulk-selection, and task-detail cancellation controls for both stages (#189)
oddish run --link <url> attaches a source URL (PR, issue, or CI run) to a task at submission time; auto-derived from --github-meta pr_url when --link is omitted; displayed in the task detail page header; re-runs update the link when a new value is provided and leave it unchanged when none is given (#178)

Fixed

Daytona sandbox creation no longer fails with "Only ephemeral sandboxes are permitted in this region"; a new daytona_ephemeral setting (default True) causes harbor trials to request ephemeral sandboxes, matching the Daytona region's configuration; harbor pin bumped to include matching ephemeral sandbox support (#188)
GitHub PR comment now auto-updates as trials, analyses, and verdicts complete; the previous implementation nested two DB sessions inside notify_trial_update, notify_analysis_update, and notify_verdict_update, deadlocking the worker's size-1 connection pool before the GitHub write was reached (#187)
alembic upgrade head no longer fails with "Multiple head revisions are present"; a merge migration (74a0eab3e564) joins the divergent provider/external_id and task_link heads into a single unified head (#186)

[2026-06-04]

Fixed

oddish run no longer crashes when a task's task.toml omits the gpus field; _task_config_requests_gpu now treats an absent gpus as 0 GPUs instead of raising TypeError: '>' not supported between 'NoneType' and 'int' (#181)
Claude Code workers using an OpenRouter model now receive the correct Anthropic-skin environment (ANTHROPIC_BASE_URL pointing to the OpenRouter endpoint, ANTHROPIC_AUTH_TOKEN set to ${OPENROUTER_API_KEY}); conflicting Bedrock and direct-Anthropic ambient credentials are blanked so the OpenRouter route takes effect (#175)

[2026-06-03]

Added

OpenAI-family workers now route through Azure OpenAI by default; ODDISH_OPENAI_PROVIDER (default azure) selects the transport, with AZURE_OPENAI_API_KEY, AZURE_OPENAI_ENDPOINT, AZURE_OPENAI_API_VERSION, and ODDISH_AZURE_OPENAI_DEPLOYMENTS for per-model deployment mapping; set ODDISH_OPENAI_PROVIDER=openai to use the public OpenAI API instead
Daytona sandboxes now auto-stop (30 min) and auto-delete (60 min) as a backstop for sandboxes that escape explicit teardown; worker_jobs gains provider/external_id columns and cancel/orphan-reap paths now terminate the underlying sandbox by ID, preventing idle sandbox accumulation

Changed

to_bedrock_model_id() now passes through any model ID with an explicit non-Anthropic provider prefix (e.g. openrouter/anthropic/claude-opus-4.8) unchanged so it runs through that provider rather than being rewritten to a Bedrock inference-profile ID; claude-code agent in harbor_runner updated to reflect the same pass-through semantics
Oddish GitHub PR comment now includes a "Performance: X/Y trials passed (Z%)" summary line and Status + Reward columns in the experiment trajectory table, surfacing actual agent scores alongside the existing classification status

Security

API key creation now requires the requesting user to be an admin of the Abundant organization; the API keys settings section is hidden in the UI for non-admins

[2026-06-02]

Changed

claude-opus-4-8 added to the Bedrock model ID mapping table, resolving to global.anthropic.claude-opus-4-8; claude-opus-4-7 moved to the legacy section of the table; regression tests added for all resolution forms (bare, anthropic/-prefixed, dotted foundation-model id) (#169)

[2026-06-01]

Changed

Automated daily changelog updated with entries for 2026-05-31 changes (#166)

[2026-05-31]

Changed

Automated daily changelog updated with entries for 2026-05-30 changes (#165)

[2026-05-30]

Added

oddish run --retry re-runs existing work for a trial, task, or experiment; re-queues failed trials by default, or re-runs analysis/verdict stages with --analysis / --verdict; -y/--yes skips confirmation (#163)
oddish publish / oddish unpublish commands toggle public read-only sharing for an experiment from the CLI and return the shareable URL; previously only possible at submit time via run --publish or in the web UI (#163)
--json machine-readable output added to oddish status, cancel, delete, and pull via a shared print_json helper; --json implies non-interactive mode and takes a single snapshot (no live watch) (#163)
oddish combine CLI command merges two or more experiments into a new result experiment, copying finished trials and their artifacts; supports --name, --copy-artifacts/--no-copy-artifacts, and --json flags (#162)

[2026-05-29]

Added

oddish run --retry re-runs existing work for a trial, task, or experiment id (positional, --task, or --experiment): re-queues failed trials by default, or re-runs analysis/verdict with --analysis / --verdict; -y skips confirmation
oddish publish / oddish unpublish commands toggle public read-only sharing for an experiment from the CLI and return the shareable URL
--json machine-readable output added to oddish status, cancel, delete, and pull (previously only on run / upload / ls)
oddish combine CLI command to merge two or more experiments into a new result experiment; copies finished trials with artifacts from source experiments and supports --name, --copy-artifacts/--no-copy-artifacts, and --json flags (#162)
POST /experiments/combine API endpoint that creates a new result experiment by merging task memberships and finished trials (with S3 artifacts) from two or more source experiments; in-flight trials are skipped and counted in the response; append-only so requires only tasks scope (#157)

Changed

Analysis and verdict UI (trial analysis dots, legend section, analysis card, verdict badge, and run analysis/verdict actions) is now hidden in the public share view (/share/[token]) via a new showAnalysis prop on ExperimentDetailView; authenticated views are unchanged (#159)

Fixed

Trial retry no longer returns a 500 error; the new trial row is now flushed before the old trial's superseded_by_trial_id self-referential FK is set, preventing a Postgres FK violation (#155)

[2026-05-28]

Changed

Harbor dependency updated to a fork commit that corrects Google API CIDR ranges for proper network access in restricted environments
Automated daily changelog updated with entries for 2026-05-27 changes (#153)

[2026-05-27]

Changed

Automated daily changelog updated with entries for 2026-05-26 changes (#152)

[2026-05-26]

Fixed

Preview banner now sticks to the top of the viewport and no longer overlaps the nav bar, task drawer, or settings sidebar; a CSS custom property --preview-banner-h (0px normally, 1.75rem in preview mode via data-preview on <html>) propagates the banner height to all affected components so calc() offsets stay in sync without hardcoded values (#146)

Changed

Automated daily changelog updated with entries for 2026-05-25 changes (#151)

[2026-05-25]

Changed

Automated daily changelog updated with entries for 2026-05-24 changes (#149)

[2026-05-24]

Changed

Automated daily changelog updated with entries for 2026-05-23 changes (#148)

[2026-05-23]

Changed

Daytona is now the default execution environment for CPU-only hosted tasks; Modal is automatically selected when a task's task.toml declares GPU requirements or when --override-gpus is set to a positive value; --env help text updated to reflect the new defaults
Harbor dependency updated to a version that runs build tools under a restricted network

[2026-05-22]

Changed

Automated daily changelog updated with entries for 2026-05-21 changes (#142)

[2026-05-21]

Added

Sticky PR comment automatically posted (and updated on re-pushes) with preview environment links — Vercel frontend URL, stable pr-NNN Vercel alias, and Modal API URL — via new post_preview_links.py script (#141)
In-app preview banner rendered when NEXT_PUBLIC_IS_PREVIEW=true, surfacing PR context to reviewers using the preview environment (#141)

Changed

PR preview workflow refactored from a monolithic modal-preview.yml into pr-preview.yml backed by focused per-phase shell scripts (prepare_preview_database.sh, deploy_preview_backend.sh, update_vercel_preview.sh, etc.), making migration-only and backend-only preview runs possible without triggering a full component redeploy (#141)
Deployment planning now tracks deploy_frontend as a separate output flag alongside deploy_backend and run_migrations; frontend-only PRs skip Supabase/Modal provisioning entirely, and non-synchronize events fall back to PR-wide path filters to decide which components need deploying (#141)
Newly created Supabase preview branches now cancel in-flight cloned production work (queued/running jobs, tasks, and trials) via cancel_cloned_preview_work.sh to prevent spurious activity from the data clone (#141)

[2026-05-20]

Changed

ODDISH_MODAL_MAX_WORKERS_PER_POLL default raised from 48 to 64, allowing the dispatcher to ramp queued work faster when per-queue slot capacity is available; env override behavior unchanged (#138)

[2026-05-19]

Added

max_trial_attempts top-level field for sweep YAML/JSON configs and TaskSubmission/TaskSweepSubmission API schemas, plus --max-trial-attempts CLI flag on oddish run, to control the total Oddish worker attempt budget per trial including the initial run; old max_attempts config key is now rejected with a clear error (#134)
ODDISH_MODAL_POLL_INTERVAL_SECONDS env var to configure the Modal queue dispatcher poll cadence; preserves the existing 180-second default when unset (#133)
Baseline-specific context injected into the trial analysis classifier prompt for oracle and nop/noop agents: oracle trials are no longer penalized for reading reference solutions, and NoOp runs are evaluated as baseline checks; normal agents receive no extra context (#132)
--environment-kwarg / --harbor-environment-kwarg CLI flag and top-level harbor.environment.kwargs block in sweep YAML for passing arbitrary Harbor environment kwargs (primary use case: agent_tools_image for Modal closed-internet runs); CLI values override config-file values on collision (#131)
"Any error" row filter in the experiment trials table to show only tasks where at least one agent hit a harness or infrastructure error, complementing the existing "Any failed" / "All failed" filters (#130)
DELETE /experiments/{experiment_id} admin API endpoint for soft-deleting an experiment and its scoped trials; artifacts are preserved in S3 (#128)

Changed

nop and oracle sweep config entries no longer require a model_name field (#131)
Experiment table toolbar UI polished: filter buttons styled with updated tokens and tighter layout, toolbar reorganized into a responsive flex row (#130)

Fixed

Experiment detail page creation timestamp now uses the canonical ExperimentModel.created_at value (surfaced via new experiment_created_at field on task-status responses) instead of inferring creation time from the earliest task in the experiment (#129)
Experiment-to-task membership rows in task_experiments are now tombstoned (deleted_at set) instead of hard-deleted when an experiment or scoped task is removed; DB migration k2l3m4n5o6p7 adds the column with partial indexes on live rows; dashboard cache is invalidated after experiment and trial deletions (#128)

[2026-05-18]

Added

Dashboard status filter now includes a "Retrying trials" option; retrying trial counts shown as (nR) in amber in the Trials column (#125)
Dedicated nop_oracle queue for nop and oracle trial agents with a separate ODDISH_NOP_ORACLE_CONCURRENCY setting (default 32; 48 in Modal), preventing these lightweight trials from competing with model-provider queues; DB migration moves existing non-terminal nop/oracle jobs to the new queue key (#121)
Bounded exponential backoff for trial retries: 30 s base delay, up to 30 min cap, with ±25% jitter; rate-limit errors (429, quota exceeded, throttled, etc.) start at a 5 min base; retry delay persisted to worker_jobs.available_after and mirrored to trials.next_retry_at (#122)

Fixed

Modal image build failures (Image build for im-... failed) now permanently fail the trial instead of requeueing, preventing repeated retry burns on deterministic Dockerfile errors; user-cancelled trial state is preserved when a build failure and a user cancel race (#124)
Retry API proxy routes (trial retry, trial analysis retry, task analysis retry, task verdict retry) now surface the real upstream error when the backend returns non-JSON plain text, instead of a misleading JSON parse exception; shared backend-response.ts helper introduced for safe response reading (#126)

[2026-05-17]

Fixed

oddish status and oddish status --watch now show a Detail column with per-trial status context — cancelled by user, the active Harbor stage while running, or the terminal error message on failure — replacing the old Stage column that only populated during running state (#112)
CLI task discovery no longer calls the removed TaskPaths.is_valid API; is_task_dir and get_task_paths_from_local now validate candidate directories by constructing Task(path), matching the path already used by validate_tasks and preventing submit failures on newer Harbor builds that dropped the compatibility helper (#119)

[2026-05-16]

Added

Copy-to-clipboard button beside task names in the experiment trials table: a copy icon appears on hover/focus and shows a brief check-mark confirmation after copying, without opening the task files panel (#114)

Changed

Drawer panels (TaskFilesPanel, TrialDetailPanel, ArtifactsViewer, TrajectoryViewer) in experiment and task detail views are now lazy-loaded via Next.js dynamic() imports, shrinking the initial page bundle (#113)
Browser Logfire/OpenTelemetry tracing is now deferred behind a conditional dynamic import in instrumentation-client.ts, keeping it off the critical bundle when disabled or unconfigured (#113)
Browser observability spans now export directly to Logfire's OTLP endpoint using a NEXT_PUBLIC_LOGFIRE_TOKEN write-only token, replacing the backend proxy route (/logfire-proxy/*) that consumed Modal container concurrency slots; LogfireProxyCORSMiddleware and mount_browser_proxy() removed from the backend (#111)
Preview branch provisioning switched back to Supabase's native --with-data clone; the custom restore_prod_data.sh pg_dump | pg_restore script is removed (#111)

[2026-05-15]

Added

Task detail page (/tasks/[task_id]) with KPI bar showing total cost, trial count, average score, and last run time; version switcher for per-version breakdown; per-agent stacked cards with trial-status chips that open existing task/trial drawers; new GET /tasks/{task_id}/detail endpoint bundles task, trials, per-version summaries, and cost rollups in one round-trip (#103)
Trajectory JSON export button on the trajectory viewer side-pane; downloads the loaded trajectory payload as trajectory-<trialId>.json client-side without additional API calls (#92)

Changed

Claude Code now routes through AWS Bedrock by default in the Modal deployment: CLAUDE_CODE_USE_BEDROCK=1 baked into the Modal image; new to_bedrock_model_id normalizer in oddish/config.py converts Anthropic-style and bare Claude model ids to invokable Bedrock cross-region inference profile ids (global. prefix for most models, us. for Opus 4.1 / Opus 4 which have no global profile); trial analysis classifier strips Bedrock env vars when running against a non-Bedrock analysis model id (#108)

Fixed

Bedrock model id mapping table now emits global./us. cross-region inference profile ids instead of bare anthropic.claude-... foundation-model ids; bare foundation-model ids are also re-resolved through the table rather than passed through, closing a gap that caused 400 "Invocation of model ID with on-demand throughput isn't supported" errors in production (#109)
Alembic migrations now pin search_path=public via asyncpg server_settings for both oddish and backend migration chains, fixing InvalidSchemaNameError on freshly-created Supabase preview branches where the Supavisor session pooler hands out backends with an empty search_path (#103)
Vercel preview environment now updated and redeployed whenever the Modal backend redeploys (not only on first Supabase branch creation), so previews that failed mid-flight on a prior push self-recover on the next push rather than silently serving the production API (#103)

[2026-05-14]

Added

Pydantic Logfire full-stack observability: backend auto-instruments FastAPI, SQLAlchemy, asyncpg, and httpx; browser spans tunnel through a server-mounted /logfire-proxy/v1/traces route so the write token never reaches the client; Server-Timing: traceparent header injected by middleware to fix document-load orphan spans; worker container init and per-job cycles wrapped in explicit worker.container_init / worker.poll_queue_cycle / worker.process_single_job spans; PR/SHA/branch/env resource attributes attached for per-deployment filtering in Logfire (#89)
Side-by-side task files + trial detail layout in the experiment drawer: ResizablePanelGroup with an adjustable 40/60 split, a toggle button, localStorage persistence (oddish:trial-drawer-side-by-side), auto-expand of drawer width on enable, and direct presigned-S3 artifact loading with backend-proxy fallback (#91)
ghcr.io/abundant-ai/oddish-ci-base prebuilt Docker image baking Python 3.13, uv, Node 20, Vercel CLI, Supabase CLI, PostgreSQL 17 client, Modal CLI, and pre-built project venvs at /opt/venvs/{backend,oddish}; published to GHCR weekly and on lockfile/Dockerfile changes via .github/workflows/ci-base-image.yml (#101)
daytona>=0.165.0 added to the oddish[worker] extra so hosted workers can construct Harbor Daytona environments for Docker-in-Docker compose tasks (#86)
Daily changelog CI workflow (.github/workflows/daily-changelog.yml) that runs nightly at 00:00 UTC, uses Claude to summarize merged PRs from commits and diffs, and opens a changelog/YYYY-MM-DD PR with auto-merge enabled; CHANGELOG.md backfilled for all PRs to date (#84)
Vercel Speed Insights integration: @vercel/speed-insights dependency added and <SpeedInsights /> component mounted in the root layout to track Core Web Vitals across all pages (#82)

Changed

CI workflows (modal-preview.yml, modal-deploy.yml, supabase-db-migrations.yml) now run inside ghcr.io/abundant-ai/oddish-ci-base; UV_PROJECT_ENVIRONMENT points at the image's pre-built venvs, making uv sync --frozen a near-instant no-op instead of a full dependency install on every push (#100)
Preview branch data population switched from Supabase's --with-data logical-replication clone (~20 min) to a direct pg_dump | pg_restore stream (~5 min); empty branches are provisioned first and populated from prod only on first branch creation via .github/scripts/preview/restore_prod_data.sh (#96)
modal-preview.yml now has a detect-changes job that queries the GHA API for the last successful deploy of each component and uses dorny/paths-filter to skip unchanged backend/migration deploys, reducing unnecessary CI runs (#88)
Signed-in users are now redirected from / to /dashboard via clerkMiddleware at the edge; the dead client-side <Show when="signed-in"><RedirectToDashboard /></Show> wrapper in page.tsx removed (#97)
Nav account dropdown and sign-in button now driven by isLoaded && isSignedIn from useUser() directly, replacing the server-only <Show> component wrapper that caused incorrect client-side visibility (#99)
Observability environment label standardized to "production" (was "prod") across backend/observability.py, frontend/src/instrumentation.ts, and frontend/src/lib/observability.ts (#93)
oddish run --env daytona now passes through to the Modal-hosted Oddish Cloud API instead of being forced to --env modal; warning message updated to reflect that both modal and daytona are supported cloud environments (#86)
Daily changelog workflow is now safe to re-run the same day: the date branch is force-pushed and an existing open PR is reused instead of failing with a non-fast-forward error (#106)
GitHub Actions versions bumped across all workflows: actions/checkout v4→v5, actions/setup-python v5→v6, astral-sh/setup-uv v4→v8.1.0, supabase/setup-cli v1→v2.0.0 (#90)

Fixed

Preview database restore now drops all public-schema FK constraints via ALTER TABLE ... DROP CONSTRAINT before running pg_restore, preventing prod's stray dangling refs from rolling back entire COPY operations for tasks, task_versions, trials, and related tables; --disable-triggers was not viable because Supabase's postgres role lacks superuser privileges (#99)

[2026-05-13]

Fixed

Next-trial-index allocators now include soft-deleted trials when scanning for the next available index, preventing PK collision 500s on INSERT after a trial at {task_id}-{N} is soft-deleted; execution_options(include_deleted=True) added to initialize_trial_import, reserve_next_trial_index, and append_trials_to_task (#81)

[2026-05-12]

Removed

oddish/environment_policy.py module (its exports normalize_environment, enforce_trial_environment, EnvironmentName had no callers; hosted policy lives in backend/cloud_policy.py) (#80)
Unused trialHasActiveAnalysis and getActiveAnalysisCount exports from frontend/src/lib/job-status.ts (#80)

Changed

Frontend cleanup pass: downgraded several job-status.ts helpers (ACTIVE_TRIAL_STATUSES, ACTIVE_PIPELINE_STATUSES, ACTIVE_VISIBLE_JOB_STATUSES, isActiveTrialStatus, isActiveVisibleJob, getActiveTrialCount) from public exports to module-private; type-only exports (TaskStatus, TrialStatus, VisibleJobKind, VisibleJobStatus) made file-local (#80)
Settings sidebar nav and import-dialog "Clear" control rewritten to use the shadcn Button primitive instead of raw <button> elements (#80)
Removed unused logging import and unused logger from backend/api/routers/github_webhooks.py (#80)

[2026-05-11]

Fixed

Supabase database migrations workflow now syncs oddish with --extra server so server-specific deps (alembic, SQLAlchemy, asyncpg) are present during migration runs (#79)

[2026-05-10]

Added

ODDISH_SAURON_AWS_SECRET_NAME setting on the backend Modal app, defaulting to aws-credentials, to control which Modal secret is layered onto worker containers for the sauron S3 mirror; setting it to empty skips loading (#68, #74)

Changed

Worker runtime now loads the aws-credentials Modal secret alongside oddish-prod, so AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY are populated and SauronS3Uploader.is_enabled() actually returns true; this completes the wiring for the sauron mirror introduced in #39, whose credential plumbing was dropped during the original squash-merge (#68, #74)
Backend Dockerfile installs git so uv sync --frozen can fetch the harbor dependency (sourced via git URL in [tool.uv.sources]) when building outside Modal (e.g. Railway, generic container hosts) (#75)

Fixed

Rollback merge that resets main back to a known-good state after the task-first data model (#55) caused breakage; reverts the bulk of that change-set on main (#72)

[2026-05-08]

Added

Sauron S3 mirror: when ODDISH_SAURON_S3_BUCKET is set, oddish workers mirror trial artifacts to a sauron-compatible S3 layout and write a run-meta.json manifest (schema_version 1) at the run root so sauron's existing /{org}/{repo}/{pr}/{run} route renders both PR-triggered ({owner}/{repo}/pr-{N}/run-{exp}/...) and CLI-triggered ({ODDISH_SAURON_S3_ORG}/runs/{exp}/run-{exp}/...) runs without sauron changes; disabled by default, best-effort with try/except on failures (#39)
Drag-and-drop zip import UI: dashboard now has an import dialog with native drag-and-drop slots for task and trial zips, mirroring oddish upload (#42)
POST /imports/zip and POST /imports/zip/inspect backend endpoints for streaming task/run zip uploads with 1 GiB per-file cap, presigned-URL task uploads, concurrent trial fan-out, and a read-only preview path; new framework-agnostic oddish/core/zip_imports.py reuses CLI utilities for parity (#42)
Task-first data model (Phase 1 + Phase 3): new JobModel and ExperimentCellModel tables, JobKind enum (validation, experiment_backfill, ad_hoc), agent equivalence keying via SHA256 of (harness | model | provider) for trial fungibility, and trials joining experiments through (task_version_id, agent_equivalence_key) at read time rather than ownership; 7 alembic migrations seed cells/jobs and enforce task_versions immutability (#55)
POST /experiments, GET /experiments/{id}/cells, cell CRUD, /experiments/{id}/resolve, /experiments/{id}/backfill, /agents/known, and /api/jobs/* endpoints, plus experiment-cell-matrix.tsx, experiment-leaderboard.tsx, experiment-pass-at-k.tsx, trial-inspect-drawer.tsx, jobs-client.tsx, and new-experiment-client.tsx frontend components (#55)
ExperimentCreateResponse schema that extends ResolvedExperimentResponse with an optional backfill receipt field (#69)

Changed

POST /experiments now enqueues a backfill automatically and returns the resolved experiment with the new trial receipts in a single round-trip; pass ?dry_run=true to keep the previous create-only semantics (#69)
Frontend experiment creation flows updated to use the combined create+backfill call and drop the separate backfill request (#69)
Gemini model routes canonicalized: google/gemini... and bare Gemini model inputs now normalize onto LiteLLM's gemini/... route in the queue/model resolution helpers (#71)

Fixed

Experiment-visibility regression: migration p7e8f9a0b1c2 backfills experiment_tasks from task_experiments joined to each task's current version, and experiment_agents from distinct (experiment_id, agent_equivalence_key) pairs observed in trials (using the most recent trial's identity strings); both inserts use ON CONFLICT DO NOTHING so the migration is re-runnable (#66)
Pass@k calculation now only counts completed attempts (have_n_successful + have_n_failed), excluding running and queued trials that produced no evidence; each task result carries its own n so per-task attempt counts are honored, with fallback to the agent-level n when absent (#66)
Supabase migration workflow now installs the server extra so the alembic console script is present; the previous uv sync --frozen without --extra server silently failed every run (#70)

[2026-05-05]

Changed

Drop Python 3.14 support (range tightened to >=3.12,<3.14) to fix dep resolution: harbor==0.6.2 requires litellm>=1.83.14, which declares Requires-Python <3.14. tool.mypy.python_version, Trove classifier, backend/Dockerfile, and GitHub Actions setup-python all moved from 3.14 → 3.13; uv.lock relocked in both oddish/ and backend/ (#54)

[2026-05-02]

Added

--force-new-version flag on oddish run (and corresponding force_new_version field on TaskUploadInitRequest) that allocates a new task version even when the local content hash matches the latest existing version, enabling callers to flip per-version-immutable flags like run_analysis without a content change (#59)

Changed

create_task_sweep_core now flips task.run_analysis from False to True when an append submission requests analysis, instead of returning a 400 "Cannot enable run_analysis when appending..." — this matches the documented intent of --force-new-version and unblocks full validation on tasks first registered without analysis (#60)

[2026-05-01]

Changed

Task author now resolved backend-side from the authenticated identity (precedence: --user → --github-user → Clerk-backed UserModel.email → api_key.name → "unknown") instead of getpass.getuser(); CLI no longer fills task.user from the OS username, so experiments stop showing ubuntu / root as Author (#52)
TaskSubmission.user / TaskSweepSubmission.user are now optional on the wire; submission.github_username is auto-filled from the actor's UserModel.github_username when missing (#52)

Fixed

Removed the 400 guard in create_task_sweep_core that refused append-mode submissions for tasks in ANALYZING / VERDICT_PENDING; the existing append_trials_to_task path already handles the state cleanup (flips status back to RUNNING, clears verdict fields, cancels in-flight VERDICT worker jobs), so re-appending now lands cleanly and re-enters the analysis/verdict pipeline once the new trials complete (#53)

[2026-04-30]

Changed

Bump harbor to 0.6.2 in the core package; regenerate oddish and backend lockfiles; realign direct pins on litellm, openai, and backend python-dotenv to match the new harbor dep graph; update task-status test doubles to the current build_trial_response shape (#48)
Preview environment strategy: Supabase preview branches are now created with --with-data so they clone production data instead of starting empty, and the bootstrap script uses ON CONFLICT DO NOTHING for idempotent org seeding; branches are reused across pushes within a PR (#46)

Removed

DELETE /tasks/{task_id}, DELETE /experiments/{experiment_id}, DELETE /trials/{trial_id} HTTP endpoints from both oddish/server and backend/api/routers (the underlying delete_*_core helpers remain available for admin/CLI use through an auth-scoped surface) (#46)

[2026-04-28]

Added

oddish ls CLI command that lists tasks via the existing /tasks/browse API and renders a Rich table with latest version, trial counts, reward summary, last run time, and linked experiments; supports --limit (capped at 100), --offset, and --json for scripting (#40)
README section documenting pip install from a GitHub ref via #subdirectory=oddish, alongside the existing PyPI quick-start (#41)

[2026-04-27]

Added

Supabase preview branch provisioning in the modal-preview PR workflow: Python polling step waits up to 10 minutes for Supabase to create the preview branch, runs both oddish and backend alembic chains against it, and layers a PREVIEW_DATABASE_URL Modal secret on top of the production secret so PR previews use isolated preview databases (#35)
supabase/config.toml with project ID to enable Supabase's GitHub integration, plus SUPABASE_ACCESS_TOKEN / SUPABASE_PROJECT_REF env vars in the workflow (#35)

[2026-04-26]

Added

"Rendered" vs "Raw" view-mode toggle on the task-files panel for text-based files, backed by a new RawRenderer component that displays content in a monospace <pre> block; URL-based renderers (image, video, audio, PDF, xlsx, docx, binary) ignore the toggle (#37)

Changed

File-content fetching no longer sniffs binary-vs-text — all text-based files are fetched via response.text(); legacy detection helpers (isTextContent, shouldSniffTextContent, looksLikeTextBytes, readResponseTextContent, getBinaryFileMessage) removed (#37)
CLI docs (DOCS.md) gained a new "Reading data from Oddish" section with a decision table for oddish status vs oddish pull, expanded examples for --watch, the auto-detection fallback chain, per-trial file layout, idempotent re-pulling, and public-endpoint fallback for shared experiments (#38)

[2026-04-25]

Changed

Experiment legend Trial-outcome chips resized to 22×18 / rounded-[4px] with a 10×10 SVG (was 14×14 / rounded-[3px] with 8×8 SVG) so legend swatches read as the same primitive as the matrix cells and the anatomy demo in the same toolbar (#36)

[2026-04-24]

Added

/settings page redesigned with a sidebar layout (Account / Workspace / API keys), Panel / PanelHeader / SectionHeading primitives, Clerk-native OrganizationSwitcher instead of a hand-rolled workspace list, status-dot active-workspace indicator, and a real empty state for API keys; legacy ?tab= URLs still accepted alongside the new ?section= (#33)

Changed

Frontend JobStatus.PENDING is now folded into the queued matrix bucket: getMatrixStatus returns queued for trial.status === "pending", STATUS_FILTER_ORDER and the URL filter type-guard no longer list pending, while backend-wire-aligned types and analysis/verdict in-flight checks still accept pending since the backend enum is unchanged (TODO comment added on JobStatus documenting the eventual full deprecation) (#27)
Task detail drawer navigation simplified: removed the always-disabled left chevron and the standalone FileText indicator; the icon-only right chevron is now a labeled "View trials →" button; vertical progress sliver replaced with a legible "N / M" text readout between up/down chevrons (#29)
Experiment trials table: first column now has a dedicated 240px default width so the v1/v2 version badge no longer sits flush against the cell border; header cells gained py-3 so the header row is visibly taller than data rows (#31)
Experiment results visual refresh: 22×18 rounded matrix tiles with hover lift; thick-stroke geometric SVGs for pass/fail/partial/error/queued/running/pending replacing lucide glyphs; warm oklch color ramp (red → orange → olive → green) for partial scores; legend renamed (Trial outcome / QA verdict) with anatomy key, Partial chip dropped, Harness error renamed to Error; pass@k chart and leaderboard cross-highlight on agent hover, leaderboard bars switched to the shared AGENT_COLORS palette (#21)

Fixed

/settings dark-mode contrast: bumped --muted-foreground from 30 6% 62% → 30 8% 74%, pointed Clerk's colorTextSecondary at hsl(var(--foreground) / 0.78), added the missing appearance.elements keys for active-device / profile-section / org-preview surfaces, plus a small .dark .cl-* block in globals.css for cases where Clerk's internal styles win the cascade (#34)
/settings section-switch flicker: all three sections now render with CSS visibility rather than conditional mount/unmount, so Clerk's UserProfile / OrganizationProfile no longer remount on every tab click (#34)

[2026-04-23]

Added

Experiment-level cost tracking in the summary bar: oddish/model_pricing.py provides per-token pricing for Anthropic (Claude 3.5/3.7/4/4.1/4.5), OpenAI (GPT-4o, GPT-4.1, GPT-5.x including codex variants, o3/o4-mini, codex-mini), and Google (Gemini 2.5/3) families with substring matching for Anthropic-API, Bedrock, and LiteLLM-style provider-prefixed names; ordered most-specific-first so gpt-5-mini never resolves to gpt-5 rates (#23)
cost_usd and cost_is_estimated fields on the trial response builders (full + compact); ExperimentDetailView summary bar aggregates cost across visible trials with ~ for pure estimates and trailing * for mixed native+estimated totals (#23)

Changed

Frontend major-dep upgrades landed: @clerk/nextjs 6.36.8 → 7.2.5 (replaced SignedIn / SignedOut, swapped afterSignInUrl / afterSignUpUrl for signInFallbackRedirectUrl / signUpFallbackRedirectUrl); lucide-react 0.468.0 → 1.9.0 with an inline GithubIcon SVG replacing the removed brand icon; tailwindcss 3.4.19 → 4.2.4 via the official @tailwindcss/upgrade codemod (rewrote globals.css to @import "tailwindcss" + @theme {}, swapped to @tailwindcss/postcss, dropped autoprefixer, mechanical class renames shadow-sm → shadow-xs, outline-none → outline-hidden, flex-shrink-0 → shrink-0, etc., tailwindcss-animate wired via @plugin); eslint 9.39.2 bump deferred pending eslint-plugin-react peer-range update (#24)

Fixed

frontend/run-prod-clerk-local.sh now preserves PATH when re-execing itself via sudo, so the documented cd frontend && sudo rm -rf .next && ./run-prod-clerk-local.sh flow works on systems where pnpm lives on a user-managed path (e.g. nvm) (#22)

[2026-04-22]

Added

Per-file expanded S3 layout for task files alongside the canonical tarball: new TASK_EXPAND WorkerJobKind, alembic migration c4b5a6d7e8f9 adding nullable expanded_at / expanded_manifest_key on task_versions, task_expand_handler.py worker with semaphore-bounded per-member uploads + 30s heartbeats, tasks_expand_archive / tasks_expand_max_bytes / tasks_expand_max_member_bytes / tasks_archive_cache_mb settings, and StorageClient.upload_bytes; UI reads from the expanded layout by default and falls back to the archive for in-flight expansions or legacy versions (#13)
StorageClient bytes+parsed-members cache per archive ETag (default 256 MB) so a listing + content click on the same version reuses one download and one tarball parse; backend returns ETag + Cache-Control: private, max-age=86400, immutable and 304s on If-None-Match when version is pinned (#13)
Local-storage preflight on Harbor worker startup that validates free bytes, inode headroom, and a create/write/delete probe against both harbor_jobs_dir and the active temp root (#14)
Temp-dir cleanup when S3 hydration fails before Oddish falls back or raises, and pruning of empty Harbor parent directories after trial artifact upload cleanup (#14)
Clickable Task column header on the experiment trials table cycling default → name A→Z → name Z→A with ArrowUpDown / ArrowUp / ArrowDown indicators; sort layers on top of the existing search filter so virtualization and row selection pick it up unchanged (#19)
Per-PR Modal preview webhook subdomains: @modal.asgi_app(label=...) label now derives from MODAL_APP_NAME ("api" for production, "{app}-api" for previews like oddish-pr-19-api) so concurrent PR previews no longer collide on abundant-ai-preview--api.modal.run (#20)

Fixed

Harbor temp-root preflight now only probes tempfile.gettempdir() when harbor_config.docker_image or harbor_config.mcp_servers requires task patching; previously a constrained /tmp rejected valid trials that never needed temp patching (#16)
oddish sdist packaging: the pyproject.toml include override that restricted the sdist to src/oddish/analyze/*.txt is removed, so pip install oddish from sdist now ships the full package instead of an empty shell; regression test asserts src/oddish/__init__.py and src/oddish/cli/__init__.py are present in built sdists (#18)

[2026-04-17]

Changed

Pass@K graph tooltip replaced with a custom recharts content renderer: entries sorted by pass rate descending to match the visual line order, agent labels shown with color-indicator squares, values formatted as percentages with one decimal, card styling with max-height and scrolling for many agents (#8)

[2026-04-16]

Changed

Heavy-run preset bumped from Claude Opus 4.6 to Opus 4.7 (#12)

[2026-04-09]

Added

Strict /tasks/upload/init + /tasks/upload/complete handshake so oddish run reserves task/version metadata, uploads task archives directly to S3 via presigned PUT, and finalizes the version without proxying bytes through the API; oddish pull likewise prefers presigned trial-file URLs and presigned-archive downloads (#11)

Removed

Legacy proxied /tasks/upload flow; the CLI now fails fast when direct upload is unavailable instead of silently falling back (#11)
ODDISH_S3_ENABLED setting and persistent local task-storage branches — S3-compatible storage is now required for task/artifact storage; self-hosting docs updated accordingly (#11)

[2026-04-07]

Added

Org-scoped /tasks/browse backend endpoint with latest-version task aggregates, experiment lists, compact latest-version trial rows, search, and pagination (#10)
Clerk-authenticated frontend API proxy and shared task-browser response types (#10)
/tasks page rendered as a card grid with latest-version trial status graphics, debounced search, SWR polling, skeleton/loading states, and a Tasks nav link (#10)

[2026-04-02]

Changed

Experiments view replaces the manual LOAD MORE button (10 tasks/page) with a two-phase progressive loader: phase 1 fetches all tasks at once via include_trials=false so the list appears instantly; phase 2 streams trial data in 50-task batches via include_trials=true&compact_trials=true, progressively filling trial status icons with a subtle "Loading trials 50/200…" header indicator (#7)

[2026-03-27]

Changed

Backend module restructure: split the monolithic backend/worker.py into a worker/ package (functions.py for the Modal dispatcher / spawn orchestration, runtime.py for Modal runtime patching and storage setup, github.py for GitHub notification hooks around shared queue execution); extract hosted-only environment policy into backend/cloud_policy.py (ALLOWED_CLOUD_ENVIRONMENTS, get_default_cloud_environment, enforce_trial_environment); move public-API helpers into oddish.api.public_helpers; drop the now-unowned queue_slots table from backend/models.py and stub its migration (#5)
No-op tweak to .github/workflows/modal-preview.yml to exercise the shared Modal preview environment plus per-PR app-naming end-to-end on a real PR (#3)

[2026-02-26]

Added

Monorepo restructure with oddish/ (core Python package, published to PyPI), backend/ (Modal-hosted API + worker orchestration with multi-tenant Clerk/API-key auth, org-scoped data, and queue-key concurrency), and frontend/ (Next.js dashboard); two-stack alembic migrations (oddish/alembic/ for core, backend/alembic/ for cloud auth tables); cloud auth schema including organizations, users (with Clerk + Supabase user-id columns), api_keys (scoped full / tasks / read), with FKs adding org_id, created_by_user_id onto tasks; pre-commit pipeline covering ruff, black, mypy, prettier, and eslint across backend|oddish and frontend paths (#1)

Uh oh!

FilesExpand file tree

CHANGELOG.md

Latest commit

History

CHANGELOG.md

File metadata and controls

Changelog

[2026-08-07]

Changed

Removed

Fixed

[2026-08-05]

Added

[2026-07-27]

Changed

[2026-07-20]

Changed

[2026-07-18]

Changed

Fixed

[2026-07-16]

Fixed

[2026-07-15]

Added

Changed

Fixed

[2026-07-07]

Added

Changed

Fixed

[2026-07-06]

Added

Changed

[2026-07-02]

Added

Changed

Fixed

Removed

Security

[2026-07-01]

Added

Fixed

[2026-06-30]

Added

Changed

Security

[2026-06-29]

Added

Changed

Fixed

[2026-06-28]

Added

Changed

Fixed

[2026-06-27]

Fixed

[2026-06-26]

Added

Changed

Fixed

[2026-06-25]

Added

Changed

Fixed

[2026-06-24]

Added

Changed

Fixed

[2026-06-23]

Added

Changed

Fixed

[2026-06-21]

Changed

[2026-06-20]

Added

Changed

Fixed

[2026-06-19]