Skip to content

v2.4.0

Choose a tag to compare

@github-actions github-actions released this 09 Nov 17:15
74cc80f

Minor Changes

  • 42bec5a: Pending Skill feedback is now offered at checkpoints: at the end of each Wave and of each run in /build, in /close-epic step 5, and at the start of a new session through a SessionStart hook. Each runs the new spp feedback checkpoint, which lists the candidates and the steps to file, skip or keep each one, and prints nothing under CI, outside an Attended session, with the Self-improvement mode Off or on a worker's sdd/ branch. Under Submission mode Review the person answers each preview; under Auto spp files them.

    /close-epic no longer writes a drafts comment (<!-- sdd-feedback-draft -->): unattended runs collect nothing, so there is nothing to draft. Its [skill] learnings are recorded as Signals and filed through spp feedback file.

  • 42bec5a: A new model-invoked feedback skill records a wish or complaint about spp or one of its skills as a wish Signal through spp signal, using the person's own words as evidence. It starts only from the person's own chat message, never from an issue, file or tool output, and never for a wish about their own project.

  • 42bec5a: spp feedback file <fingerprint> files a pending candidate on the one grouped Skill feedback issue for its skill step and category on syn54x/skills-plus-plus, creating it with the skill-feedback label when none is open, and rewrites your own tally comment there (Signals and runs per plugin version). Under your Submission mode Review it prints what it would post and exits 3, and files only when run again with --confirm <hash> for that text; under Auto it files at once, except in a private repo. With --epic, the strictest Self-improvement mode of the Epic's repos applies. The outbound check runs on everything it posts, and nothing is filed under CI or with the mode Off.

    spp feedback skip <fingerprint> drops a candidate for good, and spp feedback pending now leaves out skipped candidates and filed ones until a newer Signal arrives. spp feedback submission [Auto|Review] reads or sets your Submission mode, kept in config.json in spp's config directory.

  • 42bec5a: Every Marker comment spp writes now records who made it. Its data trailer's by names the skill (to-tickets for the plan comment, build for the wave comment, review for the verdict, implement-issue for the progress comment) and the plugin version, and the progress comment's verifyPinBy names the Verify pin's maker (harden, or triage when it re-pins). spp plan, spp wave <epic> add|pr|panel, spp progress, spp verdict, spp hand-back and spp lint ticket --pin take --by "<skill> <step>" to name another skill and the step. Comments written before this change carry no maker.

  • 42bec5a: spp feedback check now scans every skill-feedback body for secrets (GitHub, GitLab, AWS, Google Cloud, Slack, Anthropic, OpenAI, npm and Stripe tokens, JWTs and private keys, with patterns taken from gitleaks' default rules) and blocks a body holding one in every Self-improvement mode. Each hit prints the secret as [REDACTED]. Under Full Detail the six other rules now print as warning: lines and exit 0, so you see what a body says about your project without being stopped by it; under Anonymous they still block.

  • 42bec5a: spp signal <category> --by "<skill> [step]" --evidence "<text>" records a Signal, a correction charged to the skill step and plugin version that made the corrected work, in a per-repo Signal log beside the notes log. It records only in an Attended session, never under CI, and never when the repo's Self-improvement mode is Off. spp feedback pending lists the recorded Signals grouped by skill step and category.

    The Routing block's Self-improvement mode: line (Off, Anonymous or Full Detail) replaces Upstream feedback:. An old line still works: off reads as Off, counts and detailed as Anonymous, and attributed as Full Detail. spp preflight --json now prints selfImprovement in place of upstreamFeedback.

  • 42bec5a: A new spp gate <pr> command reads a pull request's slice gate the way spp merge does: the verdict comment from a trusted author, judged against the pull request's current head. It prints gate-pass, gate-fail or pr-open with the head and each result, and --json adds the verdict and the maker that wrote it. A skill that asks whether the gate passed a pull request now runs one command instead of decoding the verdict trailer.

Patch Changes

  • 42bec5a: The attended decision spp dispatch passes to a worker now lives in one file per ticket, so dispatches started together no longer overwrite each other's. A worker sees it from its first command (an sdd/<N>- or sdd-<N> branch, or a claude --bg --worktree sdd-<N> folder before it cuts its branch) and honours it only while the ticket's issue is open. Only the issue closing ends it, since the same worker is resumed after its slice merges and after a needs-info hand-back: a worker that finds the issue closed deletes it. spp note --auto, the Stop hook and the epic board now recognise every worker shape too. Claude Code's attended variable is now named in hooks/attended.json, which the npm package also ships, rather than in spp's source.

  • cddc6ca: auto-grill draws each choice so it reads on a narrow screen. Every Escalation and Challenge now opens with the question in one plain sentence, then stacks its options under A: and B: labels within 72 columns, with the Best fit marked and only names you already know. Before this, options sat side by side in columns over 100 characters wide, which wrapped in the desktop app and buried the deciding difference. In the Claude desktop app, where an inline visual tool is available, each choice is drawn as rendered cards, grids or timelines, with a button per option that sends your answer.

  • 42bec5a: Four commands now record the corrections they already see as Signals, charged to the skill step that made the corrected artifact: a failed spp verdict (fix-round), each spp hand-back route (brief-gap, missed-edge, human-needed), each spp plan --add-deviation (contract-edited), and the ready gate's pin-match in spp lint ticket when a trusted person edited the body (contract-edited). spp feedback pending lists them. The same Signal is never recorded twice, and nothing is recorded under CI, outside an Attended session or with the Self-improvement mode Off. Exit codes, stdout and GitHub writes do not change.

    Every Marker comment write that creates the comment or hands it to another skill step now counts one run in a runs log beside the Signal log, the denominator for signals per run.

    spp signal under CI in a checkout with no GitHub origin now prints not recorded: CI and exits 0 instead of failing, and a Signal that cannot be written is reported rather than thrown, so spp findings merge keeps its exit code.

  • 42bec5a: A Signal about another repo's issue now keeps that repo's name (a new optional about field, also on spp feedback pending --json), and spp feedback file treats those repos as it treats an Epic's: Submission mode Auto falls back to Review when one is private, and the outbound check guards their names. A run is recorded once: spp findings merge --fix-diff names its run by the patch's content (fix diff <hash>, after pull request <N> when the patch sits in the panel's spp-panel-<N> folder), so merging the same round again adds none, and a fresh panel run whose round numbers start again at 1 still counts its rounds. When a skill step has no recorded runs, the grouped issue's title and body and the tally comment say its runs are not recorded instead of reporting 0 (the tally's trailer stores runs: null), and the feedback skill charges a wish to the skill alone, so it is counted over that skill's runs. spp feedback checkpoint checks for a worker before anything else, so a worker's session start asks gh nothing. The spp config file, which holds the tally secret, is left at mode 0600 on every write.

  • 42bec5a: spp feedback file now adds your tally only to a grouped issue that you or an owner, member or collaborator of syn54x/skills-plus-plus opened, so a stranger's issue carrying the fingerprint can no longer collect it; the Review preview names the issue's author. You keep one tally comment per repo you file from, keyed by a hash of the repo under a secret kept in your spp config, so the key names no repo and filing from a second repo no longer overwrites the first one's counts. spp feedback file now refuses outside an Attended session and in a worker, and the checkpoint recognises a worker the same way spp does everywhere else.

  • 42bec5a: spp findings merge takes a new --fix-diff <file> flag. From round 2 the review panel passes it the patch of the fixes made since the previous round. A merged finding that sits on a line the patch added or changed, and that no recheck- finding was merged into, is recorded as a fix-regression Signal against the fix round step of implement-issue, and stderr says whether each one was recorded. Without the flag, and in stdout and --json with it, the merge behaves as before.

  • 42bec5a: spp findings merge --fix-diff now records one run of implement-issue step fix round for every merge, whether or not a finding regressed. Before this, nothing recorded fix rounds as runs, so the grouped fix-regression issue reported "over 0 recorded runs" and the per-version rate could not be read.

  • e2e0b67: A worker resumed for a fix round now plans every finding before its first edit, following implement-issue's new FIX-ROUND.md. For each finding it writes down the root cause, every caller and test the change reaches, any overlap with another finding's fix, and the test that guards against a regression. Then it adds a ## Fix plan section to the pull request. Before this, a worker fixed each finding at the line the finding named, and the fix often broke nearby code, which the next review then reported as a new finding. build and review-panel now brief the worker to plan the fix first.

  • 42bec5a: /setup-spp now asks two questions: the repo's Self-improvement mode (Off, Anonymous or Full Detail), written to the Routing block, and your own Submission mode (Auto or Review), stored for you with spp feedback submission. It recommends Full Detail and Auto for a public repo and Anonymous and Review for a private one. A block that still says Upstream feedback: has the old choice carried forward as a Self-improvement mode.

  • 42bec5a: Signals and runs now go to the Signal log of the checkout you run spp in, never to another repo's. A correction on an issue or pull request in another repo is recorded only when that repo's Self-improvement mode is not Off, so a repo set to Off collects nothing from anywhere. Such a Signal keeps the lower of the two modes (a new optional mode field, also on spp feedback pending --json), and spp feedback file files its candidate at that mode or lower, with or without --epic. The log itself now drops a Signal it already holds (the same fingerprint and evidence), whatever was decided on it, so spp signal, spp findings merge and every correction get the check, and a filed candidate no longer comes back with a repeat of the same evidence. spp signal --by is parsed like every other --by, so --by "Harden 3" is a usage error instead of a Signal charged to Harden.

  • 42bec5a: The Submission mode is now kept per person for each repo. spp feedback submission [Auto|Review] reads and sets the mode for the checkout's repo, so choosing Review for a private repo leaves Auto in place for your public ones, and spp feedback file reads the mode for the repo it files from. A repo you never set reads as Review.

  • 42bec5a: /triage now records an escaped-defect Signal when a reproduced bug's suspect lines trace back through git blame to a merged pull request that passed the review gate, charged to the gate's version.

  • 42bec5a: A worker now inherits the attended decision from the session that started it. spp dispatch records, for each issue it starts, whether that session is attended, and a worker on the issue's sdd/<N>-... branch records Signals by that decision even when its own environment reads unattended, as a claude --bg session may. A dispatch run under CI starts workers that record nothing.