What's Changed
Version bumped from 1.0.55 to 1.0.56
Recent Changes
- Bump version to 1.0.56 (#8880) (0bfb831)
- Copilot: send the current screenshot sharp on every call, keep history small (SKY-18247) (#8905) (4e533b5)
- Split growth-funnel activation into real, Copilot-test and editor-debug runs (#8903) (a041c5e)
- Show where an agent's actions go, block by block, on the agent page (#8902) (34a47d5)
- Docs: sync with production (2026-10-13) (#8901) (6eb128f)
- Disable Copilot New chat while history loads (#SKY-16435) (#8900) (f400e73)
- Stop the Task V3 answer check from holding radio and checkbox picks (SKY-18212) (#8899) (b172e91)
- Keep a fetched OTP code until an action types it and enforce the expected type on the webhook parse (SKY-18274) (#8898) (c3e9a23)
- Stop job-recipe Apply login at authentication and bind the verification-code identifier per block (SKY-18278) (#8897) (c67df09)
- Ask for the country code when a phone field reads a national number as international (SKY-18276) (#8896) (cdae5a1)
- Move four DataDragon schedules off Redshift maintenance and the 09:00 ET pile-up (SKY-18010) (#8895) (8d4c881)
- Refuse out-of-credit work at run start, code blocks, cached steps and SDK actions (SKY-18216) (#8894) (fcf65e0)
- A2A updates user on progress (#8893) (edcded4)
- Invoice enterprise job recipe runs from a RECIPE_JOBS pricing row (SKY-18158) (#8892) (af35b88)
- Stop a cancelled driver call being dropped un-awaited before it starts (#8891) (c14c1e6)
- Cache the fixed v1 action prompt on Claude models (SKY-18088) (#8890) (54e7ef5)
- Give the v1 Haiku 5.5 variant a Bedrock fallback (SKY-18088) (#8889) (878a5b9)
- Raise the file-parser LLM call deadline from 30 s to 250 s (#8888) (95fc8a7)
- Record base fingerprints for Copilot proposals (#8886) (84c5086)
- security: fix 6 skyvern-ts client npm vulnerabilities (SKY-18245) (#8883) (e4a0d1b)
- security: fix 27 skyvern-frontend npm vulnerabilities (SKY-18245) (#8882) (cace66a)
- Log an archived S3 object download at warning without a traceback (#8884) (e22ecd6)
- Add ENABLE_URL_VALIDATION kill switch for the SSRF host checks (#8881) (bef8a5e)
- Remove internal working files from docs/ (#8878) (624cc73)
- Retry inconsistent staging workflow run reads (#8871) (744e7c7)
- Settle the page before Task V3 captures its post-action screenshot (SKY-18163) (#8870) (f5f3f28)
- Check a Task V3 job application's legal-standing answers against the data before the form moves on (#8869) (069f87e)
- Add Claude Haiku 5.5 as an LLM_NAME variant for v1 (SKY-18088) (#8868) (57890df)
- Task V3: log where a navigation dead end's URL came from (SKY-16206) (#8867) (f045e06)
- Complete runs that publish the generated file they promised (SKY-18053) (#8866) (0b68d4d)
- Stop the Copilot secret screen flagging a quoted saved-login name (SKY-17984) (#8865) (3a84431)
- Task V3: say the page did not change in the repeated-action verdict only when its fingerprint never moved (SKY-17960) (#8864) (94c7ee1)
- Stop re-clicking a failed download in the same document (#8863) (4536fea)
- Point the Terms model-training opt-out to an enterprise agreement and sales@skyvern.com (SKY-17923) (#8862) (5cd25fc)
- Re-read the run's engines after the actions before generating a script (SKY-18016 follow-up) (#8861) (7de63c0)
- Task V3: refuse a vetoed finish and retry, failing on the 5th veto as v1 does (SKY-15649) (#8860) (7e423f7)
- Revert the one-look Task V3 terminated-verdict hold (SKY-18113) (#8858) (7ba6bcb)
- Flip the Task V3 A/B coin before the code-mode decision; v3 runs never load or generate scripts (SKY-18016) (#8857) (b0a79af)
- Log expected page-reload and missing-actions step failures at warning (#8855) (1d849bf)
- Shallow-fetch history for changed-file CI jobs (#8854) (a1ccbcd)
- Clean up owned auth credentials when fixture reset is blocked (SKY-17591) (#8853) (196ddd2)
- Let a download filename keep the site's own name with {{ original_filename }} (SKY-17905) (#8852) (c93e211)
- Freeze the import-time heap once pytest collection finishes (#8851) (093ba64)
- Commit a workflow version and its definition in one transaction (SKY-17857) (#8850) (54121a8)
- Find a credential by its exact ID in credential search (SKY-17991) (#8848) (8bb860a)
- fix: make Windows CLI tests portable (#8788) (1face0f)
- Update API specifications with fern api update (#8846) (7dd6e90)
- Document who owns a customer issue once it reaches Follow-up (SKY-17954) (#8847) (669a159)
- Log queue wait on the workflow webhook activity (#8845) (d961e3f)
- Use the dedicated outcome Judge key before reserving eval sources (SKY-17591) (#8844) (4e9d1aa)
- Task V3: let an action block land its click after a hover, a precursor or a failed click (#8843) (990abdb)
- Let the GCP deploy scripts build as a named service account (SKY-16553) (#8842) (99faea7)
- Redesign the low-credit banner as a single meter row (SKY-17994) (#8841) (5368676)
- Add a daily DataDragon product usage report for #growth (SKY-17914) (#8839) (c4bbb3d)
- Commit a value into a click-to-edit numeric grid cell through its editor in Task V3 (#8838) (6d0eddd)
- Stop counting a sequential chain's few-second hand-off as a stuck run (SKY-17780) (#8830) (ecec477)
- Regenerate the Python and TS SDK clients with Fern (#8760) (74f07a6)
- Fail a workflow run within seconds when its worker container restarts (SKY-17683) (#8828) (44de8cd)
- Finish a swept workflow run even when its child cascade fails (SKY-17793) (#8827) (e3a4b83)
- Bind Copilot evaluation identity to the API environment (#8826) (3c5ef48)
- Ask for helpful/not-helpful feedback after Record Browser generates steps (SKY-16738) (#8825) (6c1ec08)
- Classify a site rate-limit or 403 page as bot protection, not an AI model error (#8824) (4b95f97)
- Commit filtered ui-select options by clicking the exact row (#8823) (0053ea1)
- Make the VNC backoff test wait for React to commit instead of relying on timer luck (#8821) (e5e9887)
- Take the Temporal describe off browser-session close, action_logs and GET reads (SKY-17573) (#8820) (1f20ed7)
- Stop replacing Copilot answers that show an example block when nothing was proposed (SKY-17769) (#8819) (8c93f8d)
- Make Copilot's tool descriptions name only what Copilot can call (SKY-17760) (#8818) (f0efb64)
- Log live-stream outcomes for persistent-session and vendor live views (SKY-16903) (#8817) (2d95a50)
- Keep Copilot in scope when earlier turns drift, and add account management to its scope (SKY-17731) (#8814) (ddff084)
- Recover a verification code glued to the next word when parsing TOTP emails (SKY-17543) (#8813) (55bea7f)
- Hold the Task V3 credential submit until 45 s into the block, behind a per-run flag, and delete the pointer treatment (#8812) (f928fcb)
- Ask the goal judge before a single-action block completes on its step cap (#8811) (5c623a8)
- Add a disabled synthetic auth fixture on the staging API origin (#8810) (aed3c6b)
- Drop post-commit refresh reads from artifact, task, TOTP and browser-session inserts (SKY-17572) (#8808) (34bc418)
- Medwig/sky 12782 casa deps (#8807) (65026db)
- Medwig/sky 16553 agent waits for runs (#8806) (facec04)
- Turn on the Copilot runner lane in production as a 27 + 10 split of the sandbox node (SKY-17745) (#8805) (90ed94a)
- Stop another session's short secret from redacting unrelated log paths, status codes and durations (SKY-17742) (#8804) (ba32bfa)
- Refuse a run whose parameters cannot be used before writing it, so a 400 leaves no failed run (SKY-17733) (#8803) (be9d666)
- Never pair a failed code block's page URL with another page's title (SKY-17711) (#8802) (e964995)
- Test a restored Copilot proposal on its stored bytes, with no re-save needed (SKY-17668) (#8801) (778ae69)
- Tell Copilot the saved-block download call in the browser-code description (SKY-17671) (#8800) (a710ee5)
- Hand a stored run-scoped sign-in link to the Task V3 code tool (SKY-17544) (#8799) (4e60ca4)
- Log a rejected text-input target at warning, not error (#8798) (76e7eb2)
- Label a run stuck on an unsolved captcha as bot protection, not a browser error, when its browser closes afterward (SKY-17725) (#8797) (7a4f0f1)
- Add a new-orgs O2 drilldown to the infra scorecard: where new orgs' run start wait goes (SKY-17306) (#8795) (74d0e26)
- Read Bitwarden credentials with fewer CLI processes and keep the session key out of errors (SKY-17721) (#8794) (253f0e4)
- Wait for a closing session's archive before saving a browser profile from it (SKY-17719) (#8793) (ce1b943)
- Log where a step is waiting when it runs past 20 minutes (SKY-17603) (#8791) (0b02150)
- Replace stale Copilot cases within the existing eval harness (#8790) (19bf399)
- Stop telling Task V3 to dismiss a layer when the covering element is not one (SKY-17651) (#8787) (60c58cc)
- Put back date segments a write moved, and never report a corrupted date as ok (#8786) (eb6c5d0)
- Log when Task V3's select_combobox leaves an already-selected value as is (#8785) (b66cf1a)
- Reserve 500m CPU per secure CodeBlock runner pod and run 37 replicas on the same node (SKY-17700) (#8784) (4ef05fa)
- Show 'Browser session expired' with the fix, not 'Browser error', when a run's session had already ended (SKY-17600) (#8783) (ba92e3e)
- Say a bank payment is still clearing and send canceled subscriptions to checkout (SKY-17328) (#8782) (ff9bae8)
- Cache Copilot's Claude prompts (SKY-17680) (#8780) (b81be31)
- Fix run-scoped file URI resolution (SKY-16915) (#8778) (f09b973)
- Back off failed plan-tier reads (SKY-7437) (#8777) (ba27d80)
- Let users sign in themselves from the Copilot credential card (SKY-17622) (#8776) (8308cc2)
- Release the Copilot save hold on an interrupted row the server marks final (SKY-17648) (#8775) (be9f868)
- Offer Skyvern 1.0 in every workflow's engine picker, and honor it on older workflows only when a person picked it (#8774) (da8260c)
- Add serial workflow run groups that submit, recover, and report up to 25 runs as one operation (SKY-17274) (#8773) (0ca8f24)
- Let Task V3's finish settle check pass a page with a ticking countdown, clock or headline (#8772) (830fe1b)
- Tell every Task V3 extraction-only block to report what the page shows, and remove its experiment flag (#8771) (d962f9f)
- Keep Task V3 observe readable when a page defines its own toJSON (#8770) (84e59bc)
- Remove the Task V3 humanized-input experiment, keeping one plain input module (#8769) (29a04e9)
- Route every eligible Lever run to the dedicated lane and stop routing in staging (SKY-16970) (#8768) (902c716)
- Stop Task V3 menu picks landing on a shifted or guessed row number (SKY-17529) (#8767) (e0aade0)
- Name an icon-only close on a covering layer, and stop telling the model the layer has no controls (#8766) (7d4b26a)
- Remove the killed Task V3 customer-precedence treatment and its switch (#8765) (d7daecd)
- Audit credential, folder, schedule, tag and script writes, and cap how long an audit write can hold a request (#8764) (d1f395f)
- Resolve the run type with one task_runs read in GET /v1/runs/{id}/timeline (SKY-17567) (#8763) (fa84028)
- Show only failing tests' output in OSS CI logs (#8666) (a6d352a)
Full Changelog: v1.0.55...v1.0.56
What's Changed
- Show only failing tests' output in OSS CI logs by @wintonzheng in #8666
- Resolve the run type with one task_runs read in GET /v1/runs/{id}/timeline (SKY-17567) by @wintonzheng in #8763
- Audit credential, folder, schedule, tag and script writes, and cap how long an audit write can hold a request by @AronPerez in #8764
- Remove the killed Task V3 customer-precedence treatment and its switch by @pedrohsdb in #8765
- Name an icon-only close on a covering layer, and stop telling the model the layer has no controls by @pedrohsdb in #8766
- Stop Task V3 menu picks landing on a shifted or guessed row number (SKY-17529) by @pedrohsdb in #8767
- Route every eligible Lever run to the dedicated lane and stop routing in staging (SKY-16970) by @wintonzheng in #8768
- Remove the Task V3 humanized-input experiment, keeping one plain input module by @pedrohsdb in #8769
- Keep Task V3 observe readable when a page defines its own toJSON by @pedrohsdb in #8770
- Tell every Task V3 extraction-only block to report what the page shows, and remove its experiment flag by @pedrohsdb in #8771
- Let Task V3's finish settle check pass a page with a ticking countdown, clock or headline by @pedrohsdb in #8772
- Add serial workflow run groups that submit, recover, and report up to 25 runs as one operation (SKY-17274) by @andrewneilson in #8773
- Offer Skyvern 1.0 in every workflow's engine picker, and honor it on older workflows only when a person picked it by @pedrohsdb in #8774
- Release the Copilot save hold on an interrupted row the server marks final (SKY-17648) by @andrewneilson in #8775
- Let users sign in themselves from the Copilot credential card (SKY-17622) by @andrewneilson in #8776
- Back off failed plan-tier reads (SKY-7437) by @AronPerez in #8777
- Fix run-scoped file URI resolution (SKY-16915) by @AronPerez in #8778
- Cache Copilot's Claude prompts (SKY-17680) by @andrewneilson in #8780
- Say a bank payment is still clearing and send canceled subscriptions to checkout (SKY-17328) by @wintonzheng in #8782
- Show 'Browser session expired' with the fix, not 'Browser error', when a run's session had already ended (SKY-17600) by @wintonzheng in #8783
- Reserve 500m CPU per secure CodeBlock runner pod and run 37 replicas on the same node (SKY-17700) by @wintonzheng in #8784
- Log when Task V3's select_combobox leaves an already-selected value as is by @pedrohsdb in #8785
- Put back date segments a write moved, and never report a corrupted date as ok by @pedrohsdb in #8786
- Stop telling Task V3 to dismiss a layer when the covering element is not one (SKY-17651) by @andrewneilson in #8787
- Replace stale Copilot cases within the existing eval harness by @andrewneilson in #8790
- Log where a step is waiting when it runs past 20 minutes (SKY-17603) by @wintonzheng in #8791
- Wait for a closing session's archive before saving a browser profile from it (SKY-17719) by @wintonzheng in #8793
- Read Bitwarden credentials with fewer CLI processes and keep the session key out of errors (SKY-17721) by @wintonzheng in #8794
- Add a new-orgs O2 drilldown to the infra scorecard: where new orgs' run start wait goes (SKY-17306) by @wintonzheng in #8795
- Label a run stuck on an unsolved captcha as bot protection, not a browser error, when its browser closes afterward (SKY-17725) by @wintonzheng in #8797
- Log a rejected text-input target at warning, not error by @wintonzheng in #8798
- Hand a stored run-scoped sign-in link to the Task V3 code tool (SKY-17544) by @wintonzheng in #8799
- Tell Copilot the saved-block download call in the browser-code description (SKY-17671) by @andrewneilson in #8800
- Test a restored Copilot proposal on its stored bytes, with no re-save needed (SKY-17668) by @andrewneilson in #8801
- Never pair a failed code block's page URL with another page's title (SKY-17711) by @andrewneilson in #8802
- Refuse a run whose parameters cannot be used before writing it, so a 400 leaves no failed run (SKY-17733) by @wintonzheng in #8803
- Stop another session's short secret from redacting unrelated log paths, status codes and durations (SKY-17742) by @wintonzheng in #8804
- Turn on the Copilot runner lane in production as a 27 + 10 split of the sandbox node (SKY-17745) by @wintonzheng in #8805
- Medwig/sky 16553 agent waits for runs by @wintonzheng in #8806
- Medwig/sky 12782 casa deps by @wintonzheng in #8807
- Drop post-commit refresh reads from artifact, task, TOTP and browser-session inserts (SKY-17572) by @wintonzheng in #8808
- Add a disabled synthetic auth fixture on the staging API origin by @andrewneilson in #8810
- Ask the goal judge before a single-action block completes on its step cap by @pedrohsdb in #8811
- Hold the Task V3 credential submit until 45 s into the block, behind a per-run flag, and delete the pointer treatment by @pedrohsdb in #8812
- Recover a verification code glued to the next word when parsing TOTP emails (SKY-17543) by @wintonzheng in #8813
- Keep Copilot in scope when earlier turns drift, and add account management to its scope (SKY-17731) by @andrewneilson in #8814
- Log live-stream outcomes for persistent-session and vendor live views (SKY-16903) by @wintonzheng in #8817
- Make Copilot's tool descriptions name only what Copilot can call (SKY-17760) by @andrewneilson in #8818
- Stop replacing Copilot answers that show an example block when nothing was proposed (SKY-17769) by @andrewneilson in #8819
- Take the Temporal describe off browser-session close, action_logs and GET reads (SKY-17573) by @wintonzheng in #8820
- Make the VNC backoff test wait for React to commit instead of relying on timer luck by @AronPerez in #8821
- Commit filtered ui-select options by clicking the exact row by @LawyZheng in #8823
- Classify a site rate-limit or 403 page as bot protection, not an AI model error by @wintonzheng in #8824
- Ask for helpful/not-helpful feedback after Record Browser generates steps (SKY-16738) by @cindehaa in #8825
- Bind Copilot evaluation identity to the API environment by @andrewneilson in #8826
- Finish a swept workflow run even when its child cascade fails (SKY-17793) by @AronPerez in #8827
- Fail a workflow run within seconds when its worker container restarts (SKY-17683) by @AronPerez in #8828
- Regenerate the Python and TS SDK clients with Fern by @wintonzheng in #8760
- Stop counting a sequential chain's few-second hand-off as a stuck run (SKY-17780) by @AronPerez in #8830
- Commit a value into a click-to-edit numeric grid cell through its editor in Task V3 by @pedrohsdb in #8838
- Add a daily DataDragon product usage report for #growth (SKY-17914) by @wintonzheng in #8839
- Redesign the low-credit banner as a single meter row (SKY-17994) by @wintonzheng in #8841
- Let the GCP deploy scripts build as a named service account (SKY-16553) by @wintonzheng in #8842
- Task V3: let an action block land its click after a hover, a precursor or a failed click by @pedrohsdb in #8843
- Use the dedicated outcome Judge key before reserving eval sources (SKY-17591) by @andrewneilson in #8844
- Log queue wait on the workflow webhook activity by @cindehaa in #8845
- Document who owns a customer issue once it reaches Follow-up (SKY-17954) by @AronPerez in #8847
- Update API specifications with fern api update by @wintonzheng in #8846
- fix: make Windows CLI tests portable by @sardortaylakov in #8788
- Find a credential by its exact ID in credential search (SKY-17991) by @andrewneilson in #8848
- Commit a workflow version and its definition in one transaction (SKY-17857) by @AronPerez in #8850
- Freeze the import-time heap once pytest collection finishes by @wintonzheng in #8851
- Let a download filename keep the site's own name with {{ original_filename }} (SKY-17905) by @AronPerez in #8852
- Clean up owned auth credentials when fixture reset is blocked (SKY-17591) by @andrewneilson in #8853
- Shallow-fetch history for changed-file CI jobs by @AronPerez in #8854
- Log expected page-reload and missing-actions step failures at warning by @wintonzheng in #8855
- Flip the Task V3 A/B coin before the code-mode decision; v3 runs never load or generate scripts (SKY-18016) by @pedrohsdb in #8857
- Revert the one-look Task V3 terminated-verdict hold (SKY-18113) by @AronPerez in #8858
- Task V3: refuse a vetoed finish and retry, failing on the 5th veto as v1 does (SKY-15649) by @pedrohsdb in #8860
- Re-read the run's engines after the actions before generating a script (SKY-18016 follow-up) by @pedrohsdb in #8861
- Point the Terms model-training opt-out to an enterprise agreement and sales@skyvern.com (SKY-17923) by @wintonzheng in #8862
- Stop re-clicking a failed download in the same document by @LawyZheng in #8863
- Task V3: say the page did not change in the repeated-action verdict only when its fingerprint never moved (SKY-17960) by @pedrohsdb in #8864
- Stop the Copilot secret screen flagging a quoted saved-login name (SKY-17984) by @andrewneilson in #8865
- Complete runs that publish the generated file they promised (SKY-18053) by @andrewneilson in #8866
- Task V3: log where a navigation dead end's URL came from (SKY-16206) by @pedrohsdb in #8867
- Add Claude Haiku 5.5 as an LLM_NAME variant for v1 (SKY-18088) by @cindehaa in #8868
- Check a Task V3 job application's legal-standing answers against the data before the form moves on by @pedrohsdb in #8869
- Settle the page before Task V3 captures its post-action screenshot (SKY-18163) by @AronPerez in #8870
- Retry inconsistent staging workflow run reads by @AronPerez in #8871
- Remove internal working files from docs/ by @nickjtch-skyvern in #8878
- Add ENABLE_URL_VALIDATION kill switch for the SSRF host checks by @wintonzheng in #8881
- Log an archived S3 object download at warning without a traceback by @wintonzheng in #8884
- security: fix 27 skyvern-frontend npm vulnerabilities (SKY-18245) by @wintonzheng in #8882
- security: fix 6 skyvern-ts client npm vulnerabilities (SKY-18245) by @wintonzheng in #8883
- Record base fingerprints for Copilot proposals by @AronPerez in #8886
- Raise the file-parser LLM call deadline from 30 s to 250 s by @wintonzheng in #8888
- Give the v1 Haiku 5.5 variant a Bedrock fallback (SKY-18088) by @cindehaa in #8889
- Cache the fixed v1 action prompt on Claude models (SKY-18088) by @cindehaa in #8890
- Stop a cancelled driver call being dropped un-awaited before it starts by @wintonzheng in #8891
- Invoice enterprise job recipe runs from a RECIPE_JOBS pricing row (SKY-18158) by @wintonzheng in #8892
- A2A updates user on progress by @wintonzheng in #8893
- Refuse out-of-credit work at run start, code blocks, cached steps and SDK actions (SKY-18216) by @andrewneilson in #8894
- Move four DataDragon schedules off Redshift maintenance and the 09:00 ET pile-up (SKY-18010) by @wintonzheng in #8895
- Ask for the country code when a phone field reads a national number as international (SKY-18276) by @pedrohsdb in #8896
- Stop job-recipe Apply login at authentication and bind the verification-code identifier per block (SKY-18278) by @AronPerez in #8897
- Keep a fetched OTP code until an action types it and enforce the expected type on the webhook parse (SKY-18274) by @AronPerez in #8898
- Stop the Task V3 answer check from holding radio and checkbox picks (SKY-18212) by @wintonzheng in #8899
- Disable Copilot New chat while history loads (#SKY-16435) by @AronPerez in #8900
- Docs: sync with production (2026-10-13) by @wintonzheng in #8901
- Show where an agent's actions go, block by block, on the agent page by @AronPerez in #8902
- Split growth-funnel activation into real, Copilot-test and editor-debug runs by @AronPerez in #8903
- Copilot: send the current screenshot sharp on every call, keep history small (SKY-18247) by @andrewneilson in #8905
- Bump version to 1.0.56 by @wintonzheng in #8880
New Contributors
- @nickjtch-skyvern made their first contribution in #8878
Full Changelog: v1.0.55...v1.0.56