Repository navigation
sync: upstream pstack 0.15.36 (c742dd6) - #26
Conversation
Port pstack from cursor/plugins 99b609a (0.15.32) to c742dd61706fbacc1b95f1a20b7b988cf096ce6c. Co-authored-by: Kaitian Xie <caelaxie@users.noreply.github.com>
|
harness proof. codex npm isolated required skills present ( resolved from that installed tree:
ci: https://github.com/pstack-ports/codex-pstack/actions/runs/40618783764 |
caelaxie
left a comment
There was a problem hiding this comment.
verdict: changes requested (comment review, same-account limit).
checked against cursor/plugins c742dd6, 465 files:
- 406 new statechart files byte-identical to upstream. all 33 removals (orch scripts, old statechart scripts, interrogate references) are gone upstream too.
- every changed line in the 26 modified files matches the upstream 99b609a..c742dd6 change, except the usual codex swaps (
~/.agents/pstack-models.md,.agents/skills/links, installed swarm path,$pstack:names, cursor-team-kit mentions dropped). setup-pstack and arena take the new cross-judge sentences verbatim. - plugin.json and UPSTREAM.md at 0.15.36 / c742dd6. no
.cursorfiles, no cursor-team-kit mentions, deslop/control-cli/control-ui untouched. - install ci green on this head; that's the codex proof. the body says harness proof is in a comment, but there's no comment on the pr. either post it or drop the line.
one fix, inline: upstream orchestrate.md line 5 now keeps program state in "the Project tracker (notes.md) and the Project store". the adapter maps the Project store but not the Project tracker, so a codex coordinator has no idea where notes.md lives. map it in the same sentence. no issue to link, so no Closes/Refs needed.
| ## Durable state and recovery | ||
|
|
||
| Where upstream says the agent's store, use `${CODEX_HOME:-~/.codex}/pstack/<workspace-key>/`. Derive a stable workspace key from the primary Git worktree's directory name plus the first 12 hex characters of the SHA-256 of its canonical absolute path encoded as UTF-8, so linked worktrees share state and unrelated repositories do not collide. Outside Git, use the canonical workspace path. Record the resolved absolute directory in the plan and every delegate brief. Plans go under its `docs/`; orchestration uses `orchestrate/<project-slug>/`. Pass that orchestration directory as `--store <absolute-path>` or `ORCH_STORE` on every `orch` call. Keep this state outside the plugin cache and the target repository. | ||
| Where upstream says the agent's store or the Project store, use `${CODEX_HOME:-~/.codex}/pstack/<workspace-key>/`. Derive a stable workspace key from the primary Git worktree's directory name plus the first 12 hex characters of the SHA-256 of its canonical absolute path encoded as UTF-8, so linked worktrees share state and unrelated repositories do not collide. Outside Git, use the canonical workspace path. Record the resolved absolute directory in the plan and every delegate brief. Plans go under its `docs/`. Keep this state outside the plugin cache and the target repository. |
There was a problem hiding this comment.
upstream orchestrate.md line 5 says "Keep the program's state in the Project tracker (notes.md) and the Project store." this maps the store but not the tracker. smallest fix: name the tracker here too, e.g. "Where upstream says the agent's store or the Project store, use .... The Project tracker is notes.md in that directory." nothing else in the adapter needs to change.
Co-authored-by: Kaitian Xie <caelaxie@users.noreply.github.com>
|
harness proof for isolated required skills present (
ci: https://github.com/pstack-ports/codex-pstack/actions/runs/40619257318 |
caelaxie
left a comment
There was a problem hiding this comment.
verdict: pass, merging (comment review, same account as author).
re-review of 8dcf6d2 against my review on 68aea6c:
- the only commit since 68aea6c is 8dcf6d2. it changes one line in
.agents/skills/poteto-mode/references/codex.md, adding "The Project tracker isnotes.mdin that same directory." after the store mapping. that's the fix i asked for. orchestrate.md keeps upstream's line 5 verbatim, and the adapter now maps both the tracker and the store. - the proof comment exists now (#26 (comment)), so the body's pointer resolves.
- proof on this sha:
Plugin installationrun 40619257318 (pull_request, head 8dcf6d2) is green. it runscodex plugin marketplace addon the checkout, thencodex plugin add pstack@codex-pstack. it asserts poteto-mode, setup-pstack, deslop, control-cli, and control-ui are installed, and that every.agents/file is byte-identical in the installed copy. the comment shows the same steps locally on codex-cli 0.165.0, plusskills/listresolving$pstack:poteto-modeand orchestrate from the installed tree.codex execstops at a 401 with no login. that's the accepted install + list + installed-copy precedent. - not a draft, mergeable (clean). the content findings from my last review still hold: upstream verbatim apart from codex swaps, 0.15.36, no
.cursorpaths, no cursor-team-kit mentions, and deslop/control-cli/control-ui embedded.
sync pstack from 99b609a (0.15.32) to c742dd61706fbacc1b95f1a20b7b988cf096ce6c (0.15.36).
upstream commits, oldest first:
deslop, control-cli, and control-ui did not change between those commits.
codex adaptations reapplied on the new text:
.agents/skills/.disable-model-invocationstays out of frontmatter.agents/openai.yamlstill setsallow_implicit_invocation: false. each skill still starts by reading the codex adapter.~/.cursor/rules/pstack-models.mdcstays~/.agents/pstack-models.md..cursor/skills/stays.agents/skills/. guide links stay under.agents/skills/.../because skills live under.agents/. the install paragraph stays the codex one. the built-in loop and create-skill bullet stays the codex host bullet.git show origin/main:pstack/skills/...lines stay. the adapter already maps them.harness proof is in a comment.