Skip to content

Replace board-driven pipelines with agent-invoked how-to guides - #1173

Draft
GiggleLiu wants to merge 29 commits into
mainfrom
agentic-skills
Draft

GiggleLiu wants to merge 29 commits into
mainfrom
agentic-skills

Conversation

@GiggleLiu

@GiggleLiu GiggleLiu commented Sep 25, 2026 •

Copy link
Copy Markdown
Contributor

Summary

Strategy change: the repo no longer relies on the GitHub project board. Pipelines (run/review/auto/final) are removed; agents auto-invoke short how-to-* guides while working, and a human merges.

Skills: 23 → 12 (~6,700 → ~1,500 lines)

  • Guides: how-to-code, how-to-verify, how-to-write-manual, how-to-review, how-to-triage-issue, how-to-ship
  • Tools (trimmed): propose, find-solver, find-problem, dev-setup, release, update-papers

Behavior changes (agreed during design)

  • how-to-verify posts a ## Verification certificate PR comment (verdict, counts, seeds, scripts)
  • how-to-ship requires a fresh-context review subagent (## Agentic Review Report) before merge-ready; the agent pushes (never force) and replies on addressed threads; a human merges
  • Triage: mechanical issue fixes are automatic, substantive ones are discussed; "proposed data types" check dropped
  • Redundancy check uses pred path --limit all (the removed find_dominated_rules is no longer referenced)
  • Plans are uncommitted working notes; make run-plan removed
  • dev-setup confirms once before installing; release (skill + Makefile) requires clean, up-to-date main and make check
  • find-solver/find-problem use pred reduce --via and pred inspect

Infrastructure

  • Deleted: pipeline_board.py, pipeline_worktree.py, project_board_{poll,recover}.py, make_helpers.sh + tests; run-*, board-*, worktree-*, issue-*, pr-* make targets; Codex runner variables
  • Trimmed pipeline_pr.py, pipeline_checks.py (issue-context), pipeline_skill_context.py (review-implementation)
  • community-call.yml: agenda from open PRs (CI green / needs work) + Good issues without a PR, via GITHUB_TOKEN (tested locally against the live repo). The PROJECT_READ_TOKEN secret can be deleted.

Test plan

  • Kept script tests: 47 pass; 1 failure (test_create_pr_uses_gh_pr_create_and_returns_current_context) is pre-existing on main and left unmodified
  • Every make target, scripts/*.py subcommand, pred subcommand, and guide cross-reference named in skills/CLAUDE.md resolves
  • Repo grep shows no references to deleted skills/targets/scripts (except historical comments in two Rust test files)
  • make help and make -n release render correctly
  • Rust sources unchanged; CI make check

Paper and docs fixes (second commit)

  • Reduction-rule pred create commands now use a new rule-spec(ex) helper (Src/k=v/... --to Tgt/k=v/...), which recreates the rule fixture's source instance so the documented --config applies. Before, 24 rule entries created a different instance, MaximumClique and QUBO/f64 failed outright, and one used invalid syntax. Keyed specs reproduce all 289 rule fixtures (positional specs fail on 84, e.g. ambiguous ILP/i64/bool). Verified with make paper and by running every command.
  • The MaximumClique model entry and the MIS/MVC gold-standard entries use problem-spec(x).
  • CLAUDE.md: aggregate mappings register via #[aggregate_reduction]; manual ReductionEntry remains only for Turing and metadata-only edges.

Not changed

  • Global user skills fix-pr / issue-to-pr (in ~/.claude, outside the repo) may compete with how-to-ship here.

🤖 Generated with Claude Code

isPANN and others added 28 commits September 18, 2026 15:54
Use exact integer CVP and a bipartite solver for K2 coloring. Validate derived model state and report reference-solver size and interval failures explicitly. Correct signed vertex-cover weights and prize-collecting Steiner gadget costs, with regression tests and updated proofs.
Register value mappings beside their implementations and share constructed results across witness and aggregate recovery. Enforce decision thresholds, correct reduction edge cases, and verify typed errors and multi-step mappings.
…panic

- QUBO sparse loader reserves the dense matrix fallibly, returning a
  ConstructionError instead of aborting on huge num_vars.
- KColoring -> PartitionIntoCliques maps k = 0 on a nonempty graph to a
  NO instance instead of panicking in PartitionIntoCliques::new.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The project no longer relies on the GitHub project board. Board queues,
card moves, and the pipeline orchestrators are removed; agents instead
pick up concise how-to-* guides while working, and a human merges.

- Add guides: how-to-code, how-to-verify, how-to-write-manual,
  how-to-review, how-to-triage-issue, how-to-ship. They absorb 17 old
  skills and fix stale APIs (transform=, dimensions(), extract_solution
  signature, pred reduce --via, parameters key, removed paper helpers).
- how-to-verify posts a verification certificate on the PR; how-to-ship
  requires a fresh-context review subagent before merge-ready.
- Trim propose, find-solver, find-problem, dev-setup, release,
  update-papers; dev-setup confirms before installing, release is guarded.
- Delete board scripts, worktree helper, make_helpers.sh and the
  run-*/board-*/worktree-* make targets; trim remaining pipeline_*.py
  helpers to the subcommands the guides use.
- make release refuses unless on a clean, up-to-date main and runs
  make check.
- community-call agenda now comes from open PRs and Good issues via
  GITHUB_TOKEN instead of the project board.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@codecov

codecov Bot commented Sep 25, 2026 •

Copy link
Copy Markdown

Codecov Report

❌ Patch coverage is 99.70149% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 96.75%. Comparing base (5a35ce7) to head (8cb216e).

Files with missing lines Patch % Lines
src/models/graph/bounded_diameter_spanning_tree.rs 97.50% 1 Missing ⚠️
Additional details and impacted files
@@            Coverage Diff             @@
##             main    #1173      +/-   ##
==========================================
+ Coverage   96.22%   96.75%   +0.52%     
==========================================
  Files        1070     1069       -1     
  Lines      133605   138852    +5247     
==========================================
+ Hits       128564   134343    +5779     
+ Misses       5041     4509     -532     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:
  • ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
  • 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

@GiggleLiu
GiggleLiu marked this pull request as draft September 25, 2026 05:16
Reduction-rule entries created their source with a bare name or
problem-spec(ex.source), which yields the source *model* example, not
the rule fixture that the following --config refers to (24 entries
produced a different instance; MaximumClique and QUBO/f64 failed
outright; one used invalid 'MVC {weight: One}' syntax).

- Add keyed-spec() and rule-spec(ex) helpers; rule-spec renders
  `Src/k=v/... --to Tgt/k=v/...`. Keyed tokens reproduce all 289 rule
  fixtures; positional specs fail on 84 (e.g. ambiguous ILP/i64/bool).
- Use rule-spec in all 128 rule create commands.
- Use problem-spec(x) in the MaximumClique model entry (its example is
  on the i64 variant) and in the MIS/MVC gold-standard entries.
- how-to-write-manual documents rule-spec.
- CLAUDE.md: aggregate mappings register via #[aggregate_reduction];
  manual ReductionEntry remains only for Turing and metadata-only edges.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants