#617 M3: teach the strict take-after-flatMap trap — limitations, prompt v0.16.6, footgun row, changelog - #675
Merged
Conversation
…pt v0.16.6, footgun row, changelog M3 of m-take-flatmap-peak-memory. The fix built in v0.10.0 (d41e438) and exposed by M1/M2 becomes discoverable — the missing step that is the whole reason #617 exists. - docs/docs/reference/limitations.md: canonical entry with the V1/V2 (425 MB/81.9s -> 89 MB/0.06s) and V25/V26 tables, the corrected cost model (peak = source residency + largest single f(x) + n retained outputs), the --max-recursion-depth anti-pattern, and the bound-it-inside-f guidance. - docs/LIMITATIONS.md: summary row. - prompts/v0.16.6.md + cmd/ailang/prompts/v0.16.6.md: new version (copy of v0.16.5 plus a 6-line Common-Mistakes block), byte-identical across both trees, recorded with its real sha256 in both versions.json, "active" bumped in both. v0.16.2/v0.16.5 untouched, so pinned eval baselines still resolve byte-identical. - internal/diag/footguns.md: row 22, status `inventoried`; blind-spot column records both known false negatives (value-level aliases V10; allocating-f take-map V25/V26). Header repointed off the frozen v0.16.2 per D-1. - changelogs/v0.18-current.md: entry with the release-evidence table. Per the plan's D-1 correction, which supersedes the design doc's AC-5: the doc names the FROZEN prompts/v0.16.2.md; teaching into it would have left the shipped prompt at zero mentions — recreating #617's own shipped-but-unreachable failure one layer up. Verified by the controller outside the executor sandbox: - AC-5: takeFlatMap AND takeMap both >=1 in all five files (baseline 0/0, same-path non-empty controls firing); active == v0.16.6 in both trees; recorded hash == file sha256 in both; cross-tree diff -q rc=0. - Reachability (the check #617 itself failed): the built binary's `prompt --source=embedded` actually serves the new teaching (takeFlatMap=1, takeMap=1) with the v0.16.5 control content intact (toInts=2), active resolving to v0.16.6. - The recorded hash is a REAL gate, not decoration: corrupting it (mutation asserted LANDED by sha256) reds TestAILANGPromptLoading and TestPromptDisambiguation with `hash mismatch for "v0.16.6"` — the right mechanism, not just a red exit code. Restored from a cp backup, byte-identical. - Gates: check-changelog, fmt-check, verify-examples (188 modules, 1 pre-existing missing-on-disk, unchanged), test-stdlib-ail, lint, and the AC-7 builtins arm all rc=0. Co-Authored-By: codex gpt-5.6-sol Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The evaluator found the changelog shipped only the takeFlatMap (V1/V2) pair. The plan's AC-3b requires /usr/bin/time -l on BOTH rewritten repros and names both targets (425 MB/81.9s and 559 MB/18.78s -> 101 MB/0.08s); reproduced first-party against the plan text before acting. Adds the V25/V26 rows plus the V7 caveat that a non-allocating scalar f does not amplify, which is why the map half is conditional rather than a blanket rule. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
sunholo-voight-kampff
added a commit
that referenced
this pull request
Aug 12, 2026
…ne died with a clean rc=0 (#677) Records for iteration 185: charter STATUS stamp (rotation invariant held at 3, line-count assertion passed), mission log entry 187, dashboard refreshed, #617 queue row updated to M3 LANDED (merge ebbc5a7, PR #675). Credits iteration 184, which ran and left zero records: it spawned the codex executor, announced a wait and exited rc=0 six minutes later. Standing rule 7, 4th instance (159/167/176/184), and the first confirmed under bg-wait-ceiling=0ms where the rule's own grep-tell is blind. Raised to Mark as D-11, because the codex lane structurally mandates a background spawn and no wording change closes that. Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
M3 (TEACH) of
m-take-flatmap-peak-memory/ #617. M1 and M2 are already on dev. M4 is out of scope.The fix built in v0.10.0 (
d41e43894) and exposed by M1/M2 becomes discoverable — the missing step that is the whole reason #617 exists.What landed
docs/docs/reference/limitations.md— canonical entry with the V1/V2 and V25/V26 tables, the corrected cost model (peak = source residency + largest singlef(x)+nretained outputs), the--max-recursion-depthanti-pattern, and the bound-it-inside-fguidance.docs/LIMITATIONS.md— summary row.v0.16.6in bothprompts/andcmd/ailang/prompts/, byte-identical, real sha256 recorded in bothversions.json,activebumped in both.v0.16.2/v0.16.5untouched, so pinned eval baselines still resolve byte-identical.internal/diag/footguns.md— row 22 (inventoried), blind-spot column recording both known false negatives (value-level aliases V10; allocating-ftake-map V25/V26). Header repointed off the frozenv0.16.2.changelogs/v0.18-current.md— AC-3b release-evidence table, both pairs.Per the plan's D-1 correction, which supersedes the design doc's AC-5: the doc names the frozen
prompts/v0.16.2.md. Teaching into it would have left the shipped prompt at zero mentions — recreating #617's own shipped-but-unreachable failure one layer up.Verification (controller, outside the executor sandbox)
takeFlatMapandtakeMapboth ≥1 in all five files (baseline 0/0, same-path non-empty controls firing);active == v0.16.6in both trees; recorded hash == file sha256 in both; cross-treediff -qrc=0.take(n, flatMap(f, xs))cannot bound memory under strict evaluation — a written-as-deliberate memory cap that never capped, and OOM'd the host twice #617 itself failed: the built binary'sprompt --source=embeddedactually serves the new teaching (takeFlatMap=1,takeMap=1) with the v0.16.5 control content intact (toInts=2), active resolving tov0.16.6.TestAILANGPromptLoadingandTestPromptDisambiguationwithhash mismatch for "v0.16.6"— the right mechanism, not just a red exit code. Restored from acpbackup, byte-identical.check-changelog,fmt-check,verify-examples(188 modules, 1 pre-existing missing-on-disk, unchanged),test-stdlib-ail,lint, and the AC-7 builtins arm all rc=0.Evaluator (sonnet, distinct provider from the codex executor): PASS 91/100, zero blocking. It independently re-ran all five named targets and refuted none, reproducing all four memory arms itself. Its one real finding — the AC-3b table shipped only the
takeFlatMaphalf — was reproduced against the plan text first-party and fixed in the second commit.🤖 Generated with Claude Code