Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion .devin-plugin/plugin.json
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
{
"name": "pstack",
"version": "0.9.2",
"version": "0.11.0",
"platforms": ["local"],
"description": "Rigorous agent workflows for Devin CLI. Write less, higher quality code.",
"homepage": "https://github.com/athul-ramesh/ptstack-devin",
Expand Down
16 changes: 14 additions & 2 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -22,6 +22,18 @@ The 0.2.0 release was verified against Devin CLI 3000.11.3 (9c803229faa4), local

Avoid installing two copies with the same plugin name. Devin may source-qualify inventory names while loading skills from an older copy; verify the path returned by the skill tool.

## Using 0.11.0

This release ports upstream pstack 0.15.43 (`e559e62`). Fourteen principle leaves, unslop, and technical-writing have shorter bodies. Opening a PR, Babysit, and the multi-phase plan are shorter. The plan no longer runs `check-plan.mjs`. Reflect reads the transcript itself and waits for approval. Architect is a short brief and sends a large or contested design to arena. A standing program routes to `/pstack:figure-it-out`. Every `run_subagent` task starts by pointing at `skills/poteto-mode/worker.md`. Whether a child follows that pointer is untested.

`/pstack:setup-pstack` writes `# defaults: 3` and drops the architect and reflect roles. A sheet that already has `# defaults: 2` keeps its choices. Profile pins stay `gpt-6-1-sol-medium`, `glm-5-3-flash-high`, and `swe`. Devin's docs still do not publish `claude-opus-5-5-xhigh-fast`, `claude-fable-5-2-high`, or `gemini-4-carbon-high`, so those pins and the three effort budgets were not added. See the [0.15.43 port audit](docs/pstack-0.15.43-port-audit.md).

## Using 0.10.0

`/pstack:why` is a short cited read of commits and PR discussion. `/pstack:how` is one read-only `subagent_explore` pass. Poteto-mode traces the code itself. Teach and poteto-help still offer both skills. Recall starts one investigator per git or `gh` source.

`/pstack:setup-pstack` writes `# defaults: 2`. An older sheet drops the retired how and why roles. A `bug-fix` or `hillclimb` row that is still the old `pstack:poteto-agent-sonnet` default moves to `pstack:poteto-agent-opus` when `opus` is available. Any other choice stays. Profile `model:` pins are unchanged. Devin's docs do not publish the Cursor slugs `claude-fable-5-2-*`, `claude-opus-5-5-*`, or `gemini-4-carbon-*`, so those pins were not added. See the [0.15.42 port audit](docs/pstack-0.15.42-port-audit.md).

## Using 0.9.2

Ask `/pstack:poteto-help` with a setup or workflow question. This explicit-user skill answers with one suggested prompt and a source link. It does not start a task for a help question. A request to do work follows poteto-mode instead.
Expand All @@ -30,7 +42,7 @@ Invoke [`/pstack:ketchup`](skills/ketchup/SKILL.md) to catch up on this chat sin

The [usage guide](docs/usage-guide.md) covers prompting, investigation, prototypes before plans, real behavior verification, benchmark evidence, repeated-mistake enforcement, and bounded in-session delivery. The help skill includes [prompting guidance](skills/poteto-help/references/prompting.md) and [copyable recipes](skills/poteto-help/references/recipes.md).

Invoke [`/pstack:sharpen`](skills/sharpen/SKILL.md) on a prompt before you send it. The agent says what a top professional would add, the raw material and methods they would use, and an outside check, then hands back the rewritten prompt with your choices left open. It does not do the task. It runs only when requested.
Invoke [`/pstack:sharpen`](skills/sharpen/SKILL.md) on a prompt before you send it. The agent says what a top professional would add, the raw material and methods they would use, and an outside check, then hands back the rewritten prompt. The rewrite asks the agent to prototype a choice an experiment can settle and recommend from what it observed. Product and preference choices stay with you. It does not do the task. It runs only when requested.

Invoke [`/pstack:create-formal-verification-skill`](skills/create-formal-verification-skill/SKILL.md) when tests cannot reach the case that breaks a race, retry, state machine, parser, or unsafe code. It verifies one property with a tool that fits and runs on your machine, shows the check failing on a broken variant, and writes a project skill under `.devin/skills/` to recheck it. [`/pstack:maintain-formal-verification-skill`](skills/maintain-formal-verification-skill/SKILL.md) brings that verification back in line after the covered code changes. This plugin bundles no verifier; the agent asks before installing one.

Expand All @@ -44,7 +56,7 @@ Release 0.8.0 selectively ports upstream pstack 0.15.33 through 0.15.35 to `ab1f
- Invoke `/pstack:benchmark-checklist` before trusting a measured speedup or regression. Comparisons require at least five interleaved runs per side, median and range, correct outputs, and proof that the timed work happened. A requested ballpark permits one labeled run with errors and actual work still checked.
- Run `/pstack:setup-pstack` without losing your configured role lists or notes. Explicit role changes use repeatable `--role "role=profile[,profile...]"` arguments to its installed `write-models.mjs` helper. Repeated profiles remain separate fresh-context perspectives.
- Existing performance workflows now use the measurement checklist and try avoiding work before optimizing its cost. Follow-ups after a completed worker use a fresh context; permission repair retains the same worker.
- Multi-phase plans must pass the installed `check-plan.mjs` helper. Verification records the exact code-ready SHA and two independent audit focuses. CLI evidence can be a transcript; screenshots require a UI adapter.
- Multi-phase plans are a checklist with one section per PR. As of 0.11.0 they do not run `check-plan.mjs`. CLI evidence can be a transcript; screenshots require a UI adapter.
- Decision logs retain their history and identify separate runs. PR descriptions now include both What changed and Scope, with deliberate exclusions stated explicitly.

Profile pins and merge authority are unchanged. Native verification confirmed routing, both new slash commands, foreground workers, separate sequential passes, and background collection. Narrowly approved background permissions and same-worker permission repair remain untested.
Expand Down
2 changes: 1 addition & 1 deletion agents/comment-sicko/AGENT.md
Original file line number Diff line number Diff line change
Expand Up @@ -27,7 +27,7 @@ That list is my only leash. When I am not sure a keep clause applies, the commen

`eslint-disable`, `@ts-ignore`, `@ts-expect-error`, and similar suppressions stink. Read the surrounding code. If the rule catches real bugs or protects correctness or safety, kill the suppression and mark the exact guilty symbol `MUST KILL`.

`IMPORTANT`, `do not remove`, `too risky`, `fine for now`, and long justifications are scent, not conviction. Before judging, I read nearby code. Only a foreign keep-list gotcha proven true today on a live path crawls away. Our-code surprises die with the reshape flag above. Doubt after the hunt is meat.
`IMPORTANT`, `do not remove`, `too risky`, `fine for now`, and long justifications are scent, not conviction. Before judging, I read nearby code. If the claim is not obvious there, I trace the named symbol with `read` and `grep`. I have no `exec`, so I do not claim to have read the PR. Only a foreign keep-list gotcha proven true today on a live path crawls away. Our-code surprises die with the reshape flag above. Doubt after the hunt is meat.

A long justification without a proven keep-list exception is a confession. Kill it. Never polish meat into a shorter alibi. Mark the exact guilty symbol `MUST KILL`. My kill ends there. I do not touch the code.

Expand Down
48 changes: 48 additions & 0 deletions docs/pstack-0.15.42-port-audit.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,48 @@
# Port pstack 0.15.42 to Devin

Audited on 2026-11-15 on top of `main` at `6e060b5`. PRs #11 (`ab1f702`, 0.8.0) and #13 (`c742dd6`, 0.9.1) are merged. Plugin version is 0.10.0.

## Source and coverage

Upstream range: [`c742dd6...bf0c7cd`](https://github.com/cursor/plugins/compare/c742dd61706fbacc1b95f1a20b7b988cf096ce6c...bf0c7cd80abdbf7fe47b634978a7d1b4071f71a7) (`pstack/`, 0.15.36 to 0.15.42). The squash is [bf0c7cd](https://github.com/cursor/plugins/commit/bf0c7cd80abdbf7fe47b634978a7d1b4071f71a7) (PR #1236). `harness/pstack/claude-code` is the Claude Code mirror and was not ported.

| Upstream change | Devin disposition |
| --- | --- |
| Sharpen's last line asks the agent to prototype a choice an experiment can settle and recommend from what it observed. Product and preference choices stay with the user. | Already present in 0.9.2. README and the usage guide now describe that line. |
| Reflect synthesizer drift example is a function rename. | Adapted. The example is `readConfig` to `loadSettings` in `config.ts`. |
| Why is one short file. Investigators, synthesizer, model routing, output format, and `references/` are deleted. | Adapted. The body is one paragraph plus a Devin line to use `exec` for git and `gh`. It names no source catalog. The reference files are deleted. |
| Recall starts its own investigator per source. | Adapted. One `pstack:panel-opus` investigator each for git history and, when `gh` is authenticated, pull requests and issues. `subagent_explore` is not used, because that profile cannot run `gh`. `run_subagent` has no readonly flag, so the brief forbids writes. |
| Blast radius and teach stop pointing at removed why parts. | Adapted. Blast radius pulls the PR with git and `gh` and cites `file:line` itself. Teach still runs how and why, and no longer asks to narrow a sweep why no longer has. |
| How is one file that starts one read-only subagent. | Adapted. The profile is the documented built-in `subagent_explore`. Explorer, explainer, and critic prompts are deleted. |
| Poteto-mode no longer routes to how. Investigation, Feature, Bug fix, Hillclimb, Perf issue, Refactoring, architect, no-comments, and comment-sicko trace the code. | Adapted in those files. Comment-sicko traces with `read` and `grep` only. Its `allowed-tools` list has no `exec`, so it does not claim to have read the PR. |
| Teach and poteto-help keep how. | Kept. Help, prompting, and the usage guide describe the shorter skills. |
| Setup drops the two why roles and the two how roles. | Adapted. `how explorer`, `how explainer`, `why investigators`, and `why synthesizer` are dropped. `how critics` is dropped with them because the one-file how has no critic panel. |
| New default models, Fable as its own family and the same vendor as Claude, Grok 4.7 no longer a default, and a `# defaults: 2` marker that migrates old defaults. | Partial. See Models below. The marker is written. An unmarked sheet migrates `bug-fix` and `hillclimb` only when the row is still exactly `pstack:poteto-agent-sonnet` and `opus` is available. Other values stay. |
| Statechart SANY error text and parent-border Mermaid arrows. | Ported onto the checker this repo already ships. The drawing and the 40-line SANY section follow the upstream tests. |
| Benny automation tweaks. | Skipped. Benny is not part of this port. |

## Models

Checked [Devin CLI models](https://docs.devin.ai/cli/models), [Desktop models](https://docs.devin.ai/desktop/models) (pricing text dated 2026-11-09), and [Subagents](https://docs.devin.ai/cli/subagents) on 2026-11-15.

The CLI documents short aliases `opus`, `sonnet`, `swe`, `codex`, and `gemini`. A profile `model:` field takes an alias or an id the account allows. Desktop pricing names Claude Fable 5.2 with Low, Medium, High, XHigh, Max, and Fast Mode, and it names Gemini 4 Argon. It does not publish `claude-fable-5-2-max`, `claude-opus-5-5-xhigh`, `claude-opus-5-5-high-fast`, or `gemini-4-carbon-high`. The CLI page does not list `fable` as a short alias. Opus 5.5 is not named.

Shipped pins stay `gpt-6-1-sol-medium` (`opus` and `codex`), `glm-5-3-flash-high` (`sonnet`), and `swe`. No new profile was added.

Role routing that maps onto those pins:

- Feature, refactoring, bug-fix, perf-issue, and hillclimb use `pstack:poteto-agent-opus`. Bug-fix and hillclimb used to default to `sonnet`.
- Strongest judgment stays `pstack:panel-opus`. There is no documented Fable pin to put above it.
- Swarm workers and reflect tooling stay `pstack:panel-sonnet`. Upstream moved them to `claude-opus-5-5-high-fast`. This port has no second opus pin for that speed.
- Arena runners, architect runners, and interrogate reviewers stay `opus`, `swe`, and `sonnet`. The cross-judge pool stays `opus` and `swe`, compared by `model:` pin. Upstream's vendor rule (Fable and other Claude models are one vendor) is not applied, because no shipped pin is a documented Fable id or a Claude Opus 5.5 id.
- The setup budget ladder, `inherit-parent`, and `auto` stay out. `run_subagent` takes a profile, not an effort or a speed.

Grok 4.7 was never a default in this repo. The migration recognizes the old Devin profile rows, not a `grok-*` slug.

## Checks

- `node --test tests/*.test.mjs`: 56 passed.
- `git diff --check`: passes.
- `npm test --prefix skills/statechart/scripts` with TLC 1.7.4 on `TLA2TOOLS`: 373 passed, none skipped. That run includes the SANY line-and-column cases and the parent-border chart tests.

Native Devin activation of 0.10.0 is untested. `subagent_explore` as the how profile is taken from the subagents page and has not been run in this session.
47 changes: 47 additions & 0 deletions docs/pstack-0.15.43-port-audit.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,47 @@
# Port pstack 0.15.43 to Devin

Audited on 2026-11-16 on top of `port/pstack-bf0c7cd` (`da5a6fa`, plugin 0.10.0). That branch is open as PR #15 and is not merged. Plugin version is 0.11.0.

## Source and coverage

Upstream commit [`e559e62`](https://github.com/cursor/plugins/commit/e559e62b729e040186bc9d573e71255e5a96f50d) (PR [#1249](https://github.com/cursor/plugins/pull/1249), pstack 0.15.43). The `pstack/` tree at its parent matches [`bf0c7cd`](https://github.com/cursor/plugins/commit/bf0c7cd80abdbf7fe47b634978a7d1b4071f71a7). `harness/pstack/claude-code` is the generated Claude Code copy and was not ported.

Checked [Devin CLI models](https://docs.devin.ai/cli/models), [Desktop models](https://docs.devin.ai/desktop/models), [Subagents](https://docs.devin.ai/cli/subagents), [Creating skills](https://docs.devin.ai/cli/extensibility/skills/creating-skills), and [Devin Cloud](https://docs.devin.ai/cli/cloud) on 2026-11-16.

| Upstream change | Devin disposition |
| --- | --- |
| Fourteen principle bodies shortened. Frontmatter unchanged. | Adapted. Bodies match the short text. Frontmatter stays `triggers: model`. Type System Discipline names `/pstack:typescript-best-practices`. |
| Unslop and technical-writing shortened. Rule 14 stays "no mid-sentence colon". | Adapted. Local-only unslop rules 1, 2, 4, 6, and 21 and the soul section were not in upstream 0.15.42 and go with the shorter catalog. Technical writing invokes `/pstack:unslop`. |
| Opening a PR cut to three paragraphs. | Adapted. Still user-requested only, `gh` only, `/pstack:deslop`, and never merge. Origin CLI is not in the Devin docs, so it stays out. |
| Babysit cut to six rules. | Adapted. Same three modes, conflict-thread-CI order, no rebase, skeptical Bugbot triage, no merge. Status is `gh pr view` and `gh pr checks`. No bundled watcher and no Origin CLI. This session only. |
| Multi-phase plan cut to five steps. `check-plan.mjs` and its test deleted. | Adapted. The plan file path stays under `docs/` or `~/.config/devin/pstack/`. Live checks use `exec` or name a missing adapter. `tests/plan-validator.test.mjs` is deleted with the script. |
| Reflect is one short file. It reads the transcript, waits for approval, and spawns no subagents. Reviewer files deleted. | Adapted. The destination rule stays `.devin/skills/` and `~/.config/devin/skills/`. It does not mention `plugins/cache/`. |
| Setup drops reflect roles. | Adapted. `reflect tooling` and `reflect judgment, divergent, synthesizer` are retired. |
| Architect is a short brief and routes a large or contested design to arena. Runner references and the architect runners role are deleted. | Adapted. The brief calls `/pstack:arena`. Feature, Refactoring, Bug fix, poteto-help, and the usage guide stop describing parallel design exploration and the checkpoint option. |
| Arena runner and cross-judge fallback chains trimmed. | Partial. This port never had the slug fallback chains. Phase A now drops an unregistered profile and keeps at least two runners. The judge still compares `model:` pins. |
| Poteto-mode shortened. Orchestrate deleted. Program-scale work routes to figure-it-out. Twenty-two playbooks. | Adapted. The Devin skill stays the routing document. Orchestrate is deleted. Figure-it-out keeps one owner per PR, a verifier on a different pin, and a first unit before fan-out. Playbook count here is 19. Shipping, autopilot-full, and autopilot-stack stay cut, which is why this is not 22. `skip:` is gone. Cursor's todolist tool is not in Devin's documented tools, so the checklist remains. |
| New `worker.md`. The first line of every kickoff points at it. | Adapted and uncertain. The pointer is the first line of a `run_subagent` `task`. See Worker below. |
| Opus 5.5 xhigh-fast default, Fable 5.2 exception-only, cap at xhigh, `# defaults: 3`, three budgets. | Partial. See Models below. The marker is `# defaults: 3`. Pins and budgets are unchanged. |

## Models

The CLI still documents short aliases `opus`, `sonnet`, `swe`, `codex`, and `gemini`. Desktop pricing names Claude Fable 5.2 with Low, Medium, High, XHigh, Max, and Fast Mode. It does not publish `claude-opus-5-5-xhigh-fast`, `claude-fable-5-2-high`, or `gemini-4-carbon-high`. The CLI page does not list `fable`. Opus 5.5 is not named. `run_subagent` takes a profile, not a model or an effort.

Shipped pins stay `gpt-6-1-sol-medium`, `glm-5-3-flash-high`, and `swe`. No new profile was added.

- Code, prose, and judgment stay on the existing `opus` profiles. There is no second pin for xhigh-fast versus high-fast.
- Swarm workers stay `pstack:panel-sonnet`.
- The arena cross-judge pool stays `opus` and `swe`, compared by `model:` pin. Upstream dropped Fable from a three-seat pool this port never shipped.
- A sheet with neither `# defaults: 3` nor `# defaults: 2` still migrates an exact old `bug-fix` or `hillclimb` row from `pstack:poteto-agent-sonnet` to `pstack:poteto-agent-opus` when `opus` is available. A sheet that already has `# defaults: 2` keeps that choice and only loses the newly retired roles. The writer then stores `# defaults: 3`.
- The three budgets are not written. They change an effort token on a Cursor slug. A Devin alias has no such token.

## Worker

`skills/poteto-mode/worker.md` is ported and marked uncertain in the file. Devin documents subagents as separate conversations with a read tool, and it withholds `ask_user_question` from them. Cloud sessions are separate VMs started with `devin --cloud`, not Cursor Project workers. Nothing in the docs says a child loads a file named in the first line of its task. The instruction is shipped. Following it is untested.

## Checks

- `node --test tests/*.test.mjs` on 2026-11-16: 47 tests passed, 0 failed. That command is the repo's structural check. It is not a native Devin run.
- `git diff --check` passes.

Native Devin activation of 0.11.0 is untested. The worker.md pointer is untested.
Loading