Skip to content

PDD 0.1.26

Choose a tag to compare

View all tags
@Serhan-Asad Serhan-Asad released this 08 Nov 17:39
alpha-v0.1.26
b46348e
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

PDD 0.1.26 release notes

Release: https://github.com/promptdriven/pdd/releases/tag/alpha-v0.1.26
Release PR: #3077.

Summary

0.1.26 makes Autopilot fit the machine it runs on, recover by itself, and ask people less. Heavy checks now share the machine fairly. Runs continue after a restart and after an idle chat. Machine problems retry by themselves, and finished, replaced or closed chats are cleaned up.

Install or update

npm install -g github:promptdriven/pdd
npx skills add promptdriven/pdd -g
pdd --version   # prints 0.1.26

Update the command and the skills together, then start a new agent chat. Running Autopilot? Update every machine on a repository, then run pdd autopilot start again. Since 0.1.25, open runs keep their claims when Autopilot starts again.

Autopilot fits the machine

  • Check slots from the machine. Heavy checks (full test suites, full builds) wait in one queue per machine. Autopilot calculates the slots from the cores and memory at each tick, so a bigger or smaller VM gets more or fewer slots by itself. Each check gets its share of the cores, so a test tool starts fewer workers. A run that waits says "waiting for a check slot", and the stuck clock pauses. On a laptop or desktop, Autopilot uses at most half of the cores. pdd autopilot start --checks N sets the slots by hand. (#3056, #3067)
  • One interpreter queue per machine. Workflow chats on one machine share one queue for interpreter checks. (#3066)
  • A health check with no AI. Autopilot checks agent limits, disk and memory before it starts work. (#3048)

Autopilot recovers by itself

  • Machine problems retry. Autopilot sorts a stuck cause into machine, interpreter or work. A machine cause retries automatically when the machine is healthy again, and it does not count as a repair. A trusted person's retry (removing pdd:stuck, or a retry comment) wakes a stuck run at the next tick. (#3064)
  • Stopped and idle chats continue. After a restart, Autopilot resumes runs whose chat stopped and tells them to continue. A chat that is idle in a working step gets a bounded nudge. A chat that waits on purpose (a person gate, the interpreter, a check slot) is left alone. (#3068, #3072)
  • One base branch. Automatic runs use autopilot.base in pdd.json, or else the repository's default branch, and never take a base from documents. A base that is behind is refused before a merge. (#3071)
  • Quieter alerts. No repeated alert for each stuck review, and no infrastructure noise in the Autopilot channel. (#3057)
  • A cost ledger. pdd records the cost of each issue: tokens, CI minutes, GitHub calls, person time and the outcome. (#3061)

Sorting and grouping issues

  • Autopilot sorts issues (still real, owner, duplicate, root cause), routes them, and runs one global fix for each group, including issues that touch several areas. (#3053)

Workflow chats

  • Finished, replaced and closed chats are cleaned up. pdd closes chats that it no longer needs and marks them, so an automatic resume never reopens them. pdd workflow tidy closes finished areas, and it keeps the Herdr pane count under its limit. An agent in a replaced run is told why it stopped. A merged run no longer blocks a new run on the same issue. (#2709, #2929, #2930, #2931)
  • A launch profile for each run. A run has a launch profile (agent, account folder, model, Herdr agent name). A person or Autopilot can change it for the next chat, for example to move a run to another account when the first one reaches its limit. (#2575, #2713, #2819, #2820, #2907)

Checks for this release

  • Runner-VM full suite on the release head: 7,247 of 7,248 pass (1 skipped, 0 failed). GitHub CI tests passed on a re-run; the one failure of the first run was a timing race in a test, filed as #3089.
  • GitHub CI tests and CodeQL: green on the release head.
  • Each change had fresh independent reviews, and the release had independent integration reviews with no remaining must-fix.

Known limits

  • A run cannot yet change its PR 1 prompt after its code step started (#2897, in progress).
  • After a machine restart, a run may need its finished Herdr workspace closed before its next area chat starts (#3084, fix ready for the next release).
  • A sleeping laptop pauses Autopilot. Use an always-on machine for 24/7 work.

Questions or problems: https://github.com/promptdriven/pdd/issues

Assets 2
Loading