Skip to content

ROADMAP §1: floor wall-clock measured gap (26m → <1min, resolve-phase incrementality) - #5711

Merged
briansrls merged 1 commit into
mainfrom
roadmap/ci-floor-incrementality
Jun 24, 2026
Merged

briansrls merged 1 commit into
mainfrom
roadmap/ci-floor-incrementality

Conversation

@gunbai-bot

@gunbai-bot gunbai-bot Bot commented Jun 24, 2026

Copy link
Copy Markdown
Contributor

Operator asked why CI is slow and noted the ~26m floor should be <1min. This adds the measured grounding to ROADMAP §1.

What I measured

The single gunbc ci step is 26m (run-time, not queue). Step-level profile:

  • ~518s resolves all 870 witnesses cold — a full recompute every run, regardless of what the PR changed
  • ~275s a second run_discovery_corpus pass (within-run double-resolve)
  • ~360s effect-bound shell gates · ~134s seed compile

On a one-file PR, ~99% of that resolve recomputes witnesses whose inputs did not change — the §2 redundancy, un-inhabited.

The change (roadmap only)

Adds a §1 Floor wall-clock group with the measured numbers + the <1min target, and names three levers the roadmap did not have explicitly:

  1. resolve-phase affected-set pruning — the floor's existing selection (Widen the rust gate by construction (§1): invert the hand-picked 3-filter allowlist (29 of 792 green tests) → run-all-unless-#[ignore]d-with-written-reason; add CI-coverage-completeness so a new test is covered by default (fail-closed); measure CI-time impact before committing the full set #5427) shrinks compile/test but not the 518s resolve; extend it across discovery→resolve so unchanged witnesses skip
  2. kill the within-run double-resolve (~275s, no cross-run cache needed)
  3. profile the 518s for the known resolver pathology (budget_roster ~450x its twin)

Re-frames §2 P3 (resolve-cache, labeled "core ask, ~18%"): that is the warm lever and was measured net-negative for cold-dominated CI. The <1min target rides incrementality (don't resolve what didn't change), upstream of caching. Flagged for operator review — does not delete the §2 item.

ROADMAP.md regenerated from roadmap_authority.dag (drift-gate green; verified regen == committed).

🤖 Generated with Claude Code

… incrementality)

Profiled the CI floor: the single `gunbc ci` step is 26m, of which ~518s
re-resolves ALL 870 witnesses cold every run, plus a second discovery-corpus
pass (~275s) — a full recompute regardless of what the PR changed. On a
one-file PR ~99% of that resolve is redundant (§2).

Adds the measured grounding to §1 and names three levers the roadmap did not
have explicitly: resolve-phase affected-set pruning (the existing selection
items applied to the resolve phase, not just compile/test), killing the
within-run double-resolve, and profiling the 518s resolve for the known
resolver pathology. Re-frames §2 P3 (resolve-cache "core ask"): that is the
WARM ~18% lever, measured net-negative for cold-dominated CI; the <1min target
rides incrementality (don't resolve what didn't change), not caching.

ROADMAP.md regenerated from the authority (drift-gate green).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@gunbai-bot gunbai-bot Bot mentioned this pull request Jun 24, 2026
6 tasks
@briansrls
briansrls merged commit 56a30ad into main Jun 24, 2026
2 checks passed
@briansrls
briansrls deleted the roadmap/ci-floor-incrementality branch June 24, 2026 13:20
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.

1 participant