toasterson/akh-medu
Watch
1
0
Fork
You've already forked akh-medu
 0
Code Issues 12 Pull requests 42 Projects Releases 3 Packages 8 Wiki Activity Actions 1

WIP: grug-train: fail closed when a checkpoint's positional encoding does not match the model it is loaded into #358

Draft
toasterson wants to merge 4 commits from claude/wi-01a07aaa-grug-train-fail-closed-when-a-checkpoint into main
pull from: claude/wi-01a07aaa-grug-train-fail-closed-when-a-checkpoint
merge into: toasterson:main
Conversation 0 Commits 4 Files changed 3 +499 -1
toasterson commented 2026-09-07 07:04:36 +00:00
Owner
Copy link

Anima work item 01a07aaa-9f87-71a2-b531-e702746d8ecf.

Found by the ADR 104 cross-commit checkpoint compatibility check (2026-09-07). Since 1fc01b74 TinyTransformer.position_embedding is Option<Embedding<B>>. Burn's Option<T>::load_record short-circuits when self.num_params() == 0: constructing the model as PositionalEncoding::Rope (position_embedding = None) and loading a checkpoint written by a Learned model silently discards the embedding weights — no error, no panic. Matching-config loads are bitwise exact in both directions (verified with NamedMpkFileRecorder via save_checkpoint/load_model_only).

Task: in crates/grug-train load_checkpoint / load_model_only (checkpoint.rs) cross-check the artifact's recorded KnowledgeModelShape.positional_encoding (foundation-model.json / run.json / model artifact JSON) against the model config being loaded into, and return a typed miette diagnostic (new GrugError variant, with help text naming both encodings) on mismatch. Add a test that a Learned checkpoint refused by a Rope model errors rather than loading with dropped weights. Keep with_serving_length (Rope-only) working.

Anima work item `01a07aaa-9f87-71a2-b531-e702746d8ecf`. Found by the ADR 104 cross-commit checkpoint compatibility check (2026-09-07). Since 1fc01b74 `TinyTransformer.position_embedding` is `Option<Embedding<B>>`. Burn's `Option<T>::load_record` short-circuits when `self.num_params() == 0`: constructing the model as `PositionalEncoding::Rope` (position_embedding = None) and loading a checkpoint written by a `Learned` model silently discards the embedding weights — no error, no panic. Matching-config loads are bitwise exact in both directions (verified with NamedMpkFileRecorder<FullPrecisionSettings> via save_checkpoint/load_model_only). Task: in crates/grug-train `load_checkpoint` / `load_model_only` (checkpoint.rs) cross-check the artifact's recorded `KnowledgeModelShape.positional_encoding` (foundation-model.json / run.json / model artifact JSON) against the model config being loaded into, and return a typed miette diagnostic (new `GrugError` variant, with help text naming both encodings) on mismatch. Add a test that a Learned checkpoint refused by a Rope model errors rather than loading with dropped weights. Keep `with_serving_length` (Rope-only) working.
- train_rule: --warmup-steps/--lr-min-ratio, linear warmup + cosine decay,
  resume-safe via start_step; constant-LR repro exact when 0/1.0
  (the 50M LR-instability lesson, previously a dead WARMUP_STEPS var)
- run-350m.sh passes the new flags; 11m-repro pinned to exact old behavior
- training/README: quick start checks out main (develop lacks training/)
- data/phase51/character: kata-drills.jsonl (240 White/Yellow drills, 8 kata,
  en/de/fr) + sparring-green.jsonl (90 adversarial dialogues, 6 adversaries)
  in the four-field lineage format — seed slice for the character-SFT stage
- docs: ADR 099 (comm-intel in the agentic loop, amends ADR 044) +
  BabyLLM dojo curriculum manual v0.1

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
p350m run 1 died at step 880: healthy ~0.5 loss -> sudden single-step NaN at
lr 4.4e-5 (mid-warmup, clipping active) -- a poisoned batch in the
deterministic stream, not LR instability. train_step now reads the loss
before the optimizer step and returns untouched on non-finite; train_rule
counts skips and aborts only past 50.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Run 2 falsified the poisoned-batch theory: from step ~879 every batch was
non-finite (51 skips), last finite loss 7.37 vs 0.52 at step 870 -- the
weights died, not the data. Mechanism: task floors at ~0.5 by step 600,
warmup keeps raising LR while Adam's second moment decays, one harder batch
spikes the loss, a gradient overflows, Norm(1.0) clip divides by inf -> NaN
weights in one update.

- train_step_guarded(loss_cap): skip optimizer step when loss non-finite OR
  above cap; train_step delegates with INFINITY (7 callers unchanged)
- train_rule: EMA(0.98) loss tracking, cap = max(4x ema, ema+2) past step
  100, skip counter aborts past 500
- 350m.env: peak LR 1e-4 -> 5e-5 (the plan's LR-AT-SCALE caveat, applied at
  350M scale)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Literature-backed (GLM-130B: collapse lags grad-norm spikes by a few steps;
OPT: final-layer activation norms; Adafactor RMS(update) gauge): loss is the
lagging symptom, grad norm the leading one. Run 2 confirmed locally — finite-
loss steps degraded weights before the loss guard could fire.

- train_step_guarded gains grad_norm_cap, returns (model, loss, grad_norm);
  computes global L2 grad norm via ModuleVisitor over GradientsParams and
  skips optim.step on non-finite or over-cap norm. Burn's Norm(1.0) clip is
  per-parameter-tensor (verified in burn-optim source), so inf-norm-divide
  was reachable through it; the global gate closes that.
- train_rule: gnorm EMA gate (5x ema past step 100), gnorm in step lines
- monitor_precursors.py: gnorm foreshock rules (WARN x3, CRIT x8/non-finite),
  backward compatible with gnorm-less logs

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- n=512 @ ckpt 28000: in-dist ans 0.928 / chain 0.935 (11M gate 0.918);
  len-gen chain 0.000 at 31x capacity — length wall is not capacity
- four-run instability saga documented: per-tensor Norm clip turns inf
  grad-norm into NaN weights; global grad-norm gate refused 501 poisoned
  updates over 17.1h with zero damage
- forensics: truncation FALSIFIED (max 677 tok vs 1024 ctx), batch-content
  logit divergence NOT SUPPORTED (trigger ~= control ~= |logit| 25);
  epoch-periodic storms = data-position x optimizer-state interaction
- committed forensic tools (forensic_batches, probe_logits) + final
  monitor_precursors.py calibration (resolved-storm downgrade, two-strike)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- training/character/: corpus converter (kata+sparring -> mlx-lm chat
  format, lineage fields held out for Black belt), uv+mlx-lm LoRA pipeline
  (Qwen3-0.6B-4bit QLoRA, rank 16, mask-prompt), README
- first belt exam on M2/16GB (9m46s train, 1.74GB peak): FR witness/silence
  form PASSED, EN yes-and half-formed, DE redirect-refusal FAILED (trunk
  lacks German at 0.6B-4bit; adapter overfit past iter ~300)
- manual §9: High Grug arm (kata as language-neutral HG traces) +
  shared-library organ architecture (adapters as .so, engine captain as
  linker, tokenizer as ABI)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
dump_rule_eval (holdout prompts as JSONL, byte-identical to train_rule's
streams) + score_rule_jsonl (score external model outputs with the standard
select/bind/fire/chain scorer) + MLX runner. First sitting: Qwen3-0.6B-4bit
base and white-belt adapters all score strict-zero — base fails on output
grammar (engages content in prose), adapters fail via persona-register
leakage onto out-of-domain input (adapters must be situationally linked,
never resident: validates ADR 099 captain-as-linker for organ composition).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Few-shot K=3 + lenient answer scoring: generic Qwen3-0.6B still fails the
deduction task (lenient 0.375/0.266, below majority baseline; label-copying
not derivation) -> the 350M organ's chain 0.935 is learned, non-latent
capability. White-belt adapter under few-shot: persona leakage worsens
(64/64 len-gen outputs never answer) -> adapters situationally linked only,
never resident (ADR 099 captain-as-linker validated empirically). Belt exam:
fr silence PASS, en yes-and half-formed, de redirect-refusal FAIL (trunk +
overfit). v1 fixes recorded in the exam doc.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Records the Phase 51 P4 design work: adaptive exam controller + native
text/vision exam generation, the tool-native learning model, the compact
conv vision-stem experiment, and the certified native-OCR runtime, plus
the review threads and dojo manual updates that shaped them.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adds the data-side building blocks for Phase 51 P4: foundation-corpus
import/prepare, knowledge workflow generation, tool-contract corpus +
curriculum construction, the adaptive exam controller, and the native
vision exam/audit generators, alongside supporting example binaries.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adds foundation-model, knowledge-model, and native-OCR/vision training
and runtime paths, plus multimodal support, certification binaries
(certify_knowledge_model, certify_multimodal_model), and matching eval
tooling (eval_knowledge_exam, eval_native_ocr, eval_tool_contract_exam,
vision stem/token-budget probes and overfit smokes).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Wires the in-process Grug BabyLLM models into the agent runtime:
GrugChatBackend selection alongside Ollama/Candle, the study-note tool
for tool-native learning capture, and the contract-drift journal that
tracks tool-contract corpus drift over time. Includes the workspace
Cargo.toml/Cargo.lock bump for the new grug-data/grug-train surface
and probe examples for the exported learning contracts and Grug
bundle format.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adds the operational scripts and artifacts around the Phase 51 P4 run:
tool-native adaptive-loop/probe/release-gate supervisors, native-OCR
offline and cross-host gates plus the offline Dockerfile, certified
BabyLLM install/publish/HF-commit scripts with their test suite, the
foundation fetch script and config, and the vision-v1/adaptive-exams
result bundles and HF metadata/schema/README artifacts produced by the
validated run.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Reformats benches/, benchmarks/harness-core/, crates/akh-egraph-bench/,
crates/akhomed-store/, crates/anima-ai/, crates/heka/, crates/seshat/,
top-level examples/, and tests/ to the current rustfmt output
(import ordering, line wrapping) picked up incidentally by the Phase 51
cargo fmt pass — no behavioral changes. Also adds
crates/seshat/examples/export_mcp_tools.rs, a small new example that
prints seshd's live MCP tool descriptors as JSON without a DB
connection.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The post-foundation, adaptive-probe and adaptive-loop supervisors hardcoded
the v1 artifact names and the seven-family v2 knowledge contract. A recovery
lane beside frozen v1 evidence needs fresh dir/status names and the original
five-family v1 corpus. All new flags default to the previous hardcoded
values; legacy launch/loop configs resume via setdefault backfill.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The exclusive-create validate log collided with the frozen v1 lane's log
when running an r2 lane in the same base directory.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Lets a v2-contract lane train from a freshly generated curriculum beside
the frozen 2026-09-04 v1 corpus; default unchanged.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The v2 knowledge-workflow corpus measures p95=2315 / max=3374 tokens; at
context 2048 the trainer compacted 4,701 records (13%), dropping the
retrieval evidence the targets must quote. The model memorized the
corrupted views (loss 4e-4) and failed the intact closed-loop sentinels
100/100 on exactly the long note-revision families, which the adaptive
controller correctly confirmed across two windows and stopped. The 350m
preset already carries 4096 for this stated reason; the small rung now
matches. Costs ~1M position-embedding params (34.5M -> 35.5M, inside the
25-45M weight-class test) and a foundation retrain.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The post-foundation supervisor reads the trainer's train.jsonl while it is
being appended; a partially flushed last line raised JSONDecodeError and
killed the supervisor mid-wait. A torn tail now reads as None (still
writing); post-exit validation still fails closed on real mismatches.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
At 4096 context a batch of 8 with a ~3.4k-token record exceeds the 80GB
A100 (CubeCL allocation failure -> corrupted gradients -> NaN at step 239
of the first 4k SFT). Default stays 8; the 4k lane passes --batch-size 4.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The v2 note-writing targets (ingest_textbook ~790 chars, revise_note ~660
with a 64-hex revision token, distill_note ~525) exceed a 128-token
generation budget, so the probe/loop/gate evaluators truncated exactly the
families that scored 0.0 regardless of learning. The loop and release gate
already expose --max-new-tokens; the probe hardcoded 128. Default unchanged.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Documents the 2026-09-05/06 recovery of the tool-native pipeline: the real
cause of the 09-04 stall (RunDescriptor schema skew), the four independent
blockers found and fixed on the way (dead v1 corpus lane, 2048-context
compaction, 128-token generation budget, 4k-context OOM), the definitive
sentinel numbers (five of seven knowledge families 0.78-1.00 at n=100), and
the single residual failure mode (one-character drops copying doubled-char
hex identifiers). Pod stopped, never deleted; artifacts on the volume.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Decisions: D1 copy-faithful grammar-constrained decoding shared by runtime and
evaluators; D2 identifier-aware BPE pre-tokenization; D3 RoPE + chunked
attention + checkpointing knob, akh_tool_small_8k; D4 100k raw window rejected
for this rung (effective context via ADR 065/102); D5 capability ladder and
one-epoch-to-floor learning schedule + length_generalization sentinel; D6
serving budget 768.

Anima plan 019f90cb-2e66-7f92-ba88-a4060be97c6e, WIs 01a07a7e-6333…/7ac6…/
9836…/b81d…/de30…; both docs ingested into Seshat.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
--ordinary-learning-rate-scale (default 1.0, (0,1]) and --ordinary-warmup-steps
(default: run.json). Effective values recorded per window; both persisted in the
loop config for --resume identical-args checks with setdefault backfill for
legacy configs. The 2026-09-06 lane's second ordinary epoch at full SFT LR
regressed three healthy families.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
PreTokenization { ByteLevel, HexAwareByteLevel { min_hex_run } } (default
HexAware(8)) built from the tokenizers crate's closed primitive set: a single
fancy-regex Split whose alternation isolates a hex char iff it lies inside a run
of >= N hex chars (fixed-length lookaround per offset), then Digits(individual),
then ByteLevel. A 64-hex BLAKE3 token is exactly 64 tokens; 'cafe added a bad
decade' splits byte-identically to plain ByteLevel; decimal digits split
individually (Llama rule). Saved tokenizer JSON is self-describing, so every
existing artifact loads and encodes unchanged (fixture-pinned).
train_with_pretokenization() added; train() keeps its signature and delegates.
The train_foundation flags/manifest fields land with the D3 commit (shared file).

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Eighth KnowledgeScenario: recall/navigation episodes whose evidence (the note
id from study_note_list) sits a controlled distance back in the prompt. Distance
is built from schema-valid padding tool results about other notes/documents
(study_note_list / corpus_documents / memory_recall rotating), measured on the
real rendered SFT prompt until bucket_tokens x 3.5 chars is reached; each episode
records LengthBucketRecord { bucket_tokens, prompt_chars }. Buckets default
[2048, 4096, 8192, 12288]; --length-buckets on prepare/training/exam/sibling.
Every padding call re-validates against the ToolContractSnapshot in tests;
same-seed replay is byte-identical and lands within +/-15% of target.

Generator version bumped v2 -> v3: appending a scenario changes which family a
given (seed, index) produces for the other seven, so v2 curricula are not
reproducible from this binary (episode seeds are domain-separated by version).
Probe/loop default families now list eight; pass --knowledge-families with the
seven when probing a 4096-context checkpoint.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The tool-native lane measured the residual failure once the schema-name
grammar is right: the 35M model drops/corrupts a character copying
identifier-shaped values (64-hex revision tokens, note ids) from context,
which the grader's grounding rule already required to be verbatim. This
moves that rule into the decoder.

- grug-data/src/constrain.rs: ValueConstraint (Enum/Copy/PatternFallback/
  Free) derived per-property from JSON-Schema (enum, pattern, fixed
  length); CopyTrie for O(len) per-token legality checks independent of
  candidate count; ValueConstraintTable precomputes constraints once per
  call from the rendered context. GrammarState's Colon/ValueStart/Str now
  carry the parameter name so the decoder knows which property's value it
  is typing. RuntimeCallGrammar/RuntimeCallState wrap the runtime's
  `<tool_call>\n...\n</tool_call>` envelope: free text (Answer/Gap) stays
  fully unconstrained, structure+copy constraints apply only once the tag
  is committed to. ConstrainedDecodeState is the single seam the decode
  loop drives one token at a time so structural and value-copy legality
  can't drift apart. build_decode_report recovers a DecodeReport
  (per-field copied/fallback/candidate-count) from the finished text.

- grug-train/src/generate.rs: generate_from_prompt_ids_cached_constrained
  applies the mask to the KV-cached decode loop; falls back to the plain
  cached path byte-for-byte when no grammar is supplied.

- grug-train/src/knowledge_runtime.rs: KnowledgeModel::generate_turn(_with_
  system) now decodes constrained by default (DecodingMode::Constrained),
  scanning the *compacted* prompt text (what the model actually sees) for
  copy candidates; generate_turn_with_system_mode(..., DecodingMode::Plain)
  reproduces the pre-D1 decoder for evaluators that need old numbers.

- eval_knowledge_exam.rs / eval_tool_contract_exam.rs (the closed-loop
  scorer and all-tool matrix): route through the same decoder via
  --decoding {constrained,plain} (default constrained), with optional
  --decode-reports-out for forensic review.

- chat_backend.rs: surfaces the DecodeReport via tracing::debug; bumps
  AKH_GRUG_MAX_NEW_TOKENS's default 384 -> 768 (both read sites) per the
  measured >=640-token note-write requirement (tool-native-4k-lane-
  20260906-RESULT.md §4).

Left out: crates/grug-train/src/multimodal_runtime.rs's parallel
generate_turn_with_system (vision runtime) keeps its pre-D1 decoder
unconstrained -- out of scope for this pass. grug-train/src/eval.rs's
Step A/B research harness (untagged ToolCallGrammar over the synthetic
procedural catalog) is untouched; it measures a different, already-shipped
experiment, not the deployed runtime.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
PositionalEncoding { Learned (default), Rope { theta, scale } } via Burn 0.21
RotaryEncoding; TinyTransformer.position_embedding becomes Option<Embedding>
(None under Rope). attention_chunk: Option<usize> computes attention in query
blocks (memory batch x heads x chunk x L) for forward and prefill; decode is
already O(L). Presets akh_tool_small_8k / akh_tool_350m_8k (8192, Rope
theta=10000, chunk 1024); KnowledgeModelSize small_8k / 350m_8k; LR defaults
in train_foundation/train_knowledge extended. with_serving_length(new_max,
scale) re-instantiates a Rope model at a longer max_seq_len (scale = NTK-style
base adjustment theta*scale, since Burn's RotaryEncoding indexes integer
positions); Learned refuses with GrugError::LearnedPositionsCannotExtend.
KnowledgeModelShape gains positional_encoding {kind: learned|rope,...} and
attention_chunk, both #[serde(default)] so every existing artifact loads;
grad_checkpointing cargo feature selects Autodiff<_, BalancedCheckpointing>.

Also lands the D2 flags in train_foundation: --bpe-pretokenizer
{byte-level,hex-aware} (default hex-aware), --bpe-min-hex-run 8, recorded in
TokenCacheManifest and run.json with serde defaults.

Tests: chunked == exact (incl. L not divisible by chunk), Rope prefill/decode
parity, serde defaults + old-fixture load; 111/111 grug-train lib tests.
Pod-side still owed: 8192 memory/throughput smoke with and without
grad_checkpointing; loading the pre-existing 4096 Learned checkpoint with this
binary (Option<Embedding> record shape).

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
docs(phase51): P5 plan — checkpoint compat verified; Option<Embedding> silent-drop risk recorded
Some checks failed
CI / check-akh-medu (pull_request) Failing after 12s
Details
CI / docker-seshd (pull_request) Failing after 1s
Details
CI / publish-chart (pull_request) Failing after 0s
Details
eval-smoke / eval-smoke (pull_request) Failing after 1s
Details
CI / test-akh-medu (pull_request) Failing after 10m57s
Details
CI / check-seshat (pull_request) Successful in 13m26s
Details
7cbaefc263
 
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
fix(grug-train): reject positional checkpoint mismatch
Some checks failed
CI / test-session-platform (pull_request) Has been cancelled
Details
CI / test-sokar (pull_request) Has been cancelled
Details
CI / check-seshat (pull_request) Has been cancelled
Details
CI / check-akh-medu (pull_request) Has been cancelled
Details
CI / test-akh-medu (pull_request) Has been cancelled
Details
CI / docker-seshd (pull_request) Has been cancelled
Details
CI / docker-sokard (pull_request) Has been cancelled
Details
CI / docker-akh-transcripts (pull_request) Has been cancelled
Details
CI / publish-chart (pull_request) Has been cancelled
Details
3a100339a4
fix(grug-train): treat legacy checkpoints as learned
All checks were successful
CI / test-sokar (pull_request) Successful in 10m0s
Details
CI / test-session-platform (pull_request) Successful in 14m13s
Details
CI / docker-seshd (pull_request) Successful in 8m33s
Details
CI / test-akh-medu (pull_request) Successful in 11m25s
Details
CI / check-akh-medu (pull_request) Successful in 13m43s
Details
CI / publish-chart (pull_request) Successful in 6m25s
Details
CI / docker-akh-transcripts (pull_request) Successful in 9m37s
Details
CI / check-seshat (pull_request) Successful in 17m47s
Details
CI / docker-sokard (pull_request) Successful in 16m42s
Details
749fda38b3
fix(grug-train): bind checkpoint positional encoding to artifacts
Some checks failed
CI / test-sokar (pull_request) Successful in 19m53s
Details
CI / check-akh-medu (pull_request) Successful in 23m13s
Details
CI / publish-chart (pull_request) Successful in 6m13s
Details
CI / test-session-platform (pull_request) Successful in 47m39s
Details
CI / docker-seshd (pull_request) Successful in 1h5m46s
Details
CI / test-akh-medu (pull_request) Successful in 1h13m55s
Details
CI / check-seshat (pull_request) Successful in 1h25m40s
Details
CI / docker-akh-transcripts (pull_request) Successful in 1h12m44s
Details
CI / docker-sokard (pull_request) Failing after 3h23m28s
Details
CI / platform-images-home (pull_request) Successful in 18m52s
Details
CI / platform-images-core (pull_request) Successful in 20m12s
Details
CI / platform-images-daemons (pull_request) Successful in 1h11m38s
Details
CI / platform-images-fleet (pull_request) Has been skipped
Details
CI / platform-images-collect (pull_request) Has been skipped
Details
360b4f8e13
Some checks failed
CI / test-sokar (pull_request) Successful in 19m53s
Details
CI / check-akh-medu (pull_request) Successful in 23m13s
Details
CI / publish-chart (pull_request) Successful in 6m13s
Details
CI / test-session-platform (pull_request) Successful in 47m39s
Details
CI / docker-seshd (pull_request) Successful in 1h5m46s
Details
CI / test-akh-medu (pull_request) Successful in 1h13m55s
Details
CI / check-seshat (pull_request) Successful in 1h25m40s
Details
CI / docker-akh-transcripts (pull_request) Successful in 1h12m44s
Details
CI / docker-sokard (pull_request) Failing after 3h23m28s
Details
CI / platform-images-home (pull_request) Successful in 18m52s
Details
CI / platform-images-core (pull_request) Successful in 20m12s
Details
CI / platform-images-daemons (pull_request) Successful in 1h11m38s
Details
CI / platform-images-fleet (pull_request) Has been skipped
Details
CI / platform-images-collect (pull_request) Has been skipped
Details
This pull request is marked as a work in progress.
This branch is out-of-date with the base branch
View command line instructions

Checkout

From your project repository, check out a new branch and test the changes.
git fetch -u origin claude/wi-01a07aaa-grug-train-fail-closed-when-a-checkpoint:claude/wi-01a07aaa-grug-train-fail-closed-when-a-checkpoint
git switch claude/wi-01a07aaa-grug-train-fail-closed-when-a-checkpoint

Merge

Merge the changes and update on Forgejo.

Warning: The "Autodetect manual merge" setting is not enabled for this repository, you will have to mark this pull request as manually merged afterwards.

git switch main
git merge --no-ff claude/wi-01a07aaa-grug-train-fail-closed-when-a-checkpoint
git switch claude/wi-01a07aaa-grug-train-fail-closed-when-a-checkpoint
git rebase main
git switch main
git merge --ff-only claude/wi-01a07aaa-grug-train-fail-closed-when-a-checkpoint
git switch claude/wi-01a07aaa-grug-train-fail-closed-when-a-checkpoint
git rebase main
git switch main
git merge --no-ff claude/wi-01a07aaa-grug-train-fail-closed-when-a-checkpoint
git switch main
git merge --squash claude/wi-01a07aaa-grug-train-fail-closed-when-a-checkpoint
git switch main
git merge --ff-only claude/wi-01a07aaa-grug-train-fail-closed-when-a-checkpoint
git switch main
git merge claude/wi-01a07aaa-grug-train-fail-closed-when-a-checkpoint
git push origin main
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels   
follow-up
Documented operational follow-up

No labels
follow-up
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
toasterson/akh-medu!358
No description provided.