Use the same logged per-node normalization for an empty rendering as for incapable routes and broken request series: empty each active later system node, then empty the first node. Clearing only the latest node exposes stale earlier instructions, including converter-demoted user text. Dormant empty tails neither override effective text nor produce repeated replacements.
Keep the cleared structural head and ordinary restoration semantics: a capable continuing series may append new nonempty instructions, while incapable or broken-series requests refill the head. No initial empty-head creation change, deletion event, or rollback machinery.
Tests: capable/plain clearing after three active prompt versions fails before the fix and passes afterward; repeated clear and seeded resume remain empty without duplicate system events, restoration includes only new instructions, pi-ai conversion has no stale prompt, and source-event replay reconstructs every request. Final 395 loop tests pass with exact agent.ts/runtime-context.ts coverage at 100% all metrics; source tsc and focused lint pass. Fix new test discriminant narrowing discovered by focused compiler audit. Update EN/ZH consumers and owning rationale, seven pairs verified; parent owns broad gates and generated catalog/SDK/GUI updates.
Resolve agent/request and prepareCall inside the accepted open step before committing system and user input. Preserve prompt-assembly model selection, bind capability and dispatch to the same prepared adapter, then log the envelope and derive the request synchronously. Preparation cancellation leaves a balanced empty step with no admitted input.
Normalize incapable-route retained prompt versions with logged per-node empty replacements and a current head, including unchanged rendering and resumed history. Dormant empty tails do not override the effective prompt. No new event or delete operation.
Tests: 389 agent-loop cases pass with exact agent.ts/runtime-context.ts coverage at 100% for all metrics; tsc -b packages/core/agent-loop/tsconfig.json passes. Seven actual-loop MockAdapter cases exercise pi-ai conversion, source-event reconstruction, both route transitions, resume, cancellation barriers, and bound selection. Stale-capability negative control fails all three transition regressions. Update EN/ZH request visibility, lifecycle generator, and owning rationale. Broad catalogs, recorded SDK/GUI and aggregate gates remain parent-owned.
Consolidate the existing in-history capability, loop, UI and artifacts into one baseline. Preserve the reviewed tree so each independently verified correction has a subsequent rationale-rich commit.
The session already derives the protected system head as a Message, and both adapters accept leading system history. Passing it through a separate SummarizationInput.system string unnecessarily flattens that value and rebuilds the same wire message in the adapter. Prepend the derived head to messages and remove textContent, the separate field, and GenerateOptions.system plumbing from the summarizer.
Keep range selection, shadowed seq accounting, session head protection, routed tools, image references, target policy, and the model-visible compaction instruction unchanged. Empty-content heads still derive to null and contribute no request message, but their surface node remains protected. Update subclass consumers/tests, EN/ZH package and subsystem prose, and the existing system-prompt surface owning note with refreshed pairing records.
Evidence: pnpm exec vitest run packages/compaction/compaction-basic/tests packages/llm/llm-deepseek/tests/serialize.spec.ts packages/llm/llm-pi-ai/tests/context.spec.ts --coverage --coverage.include='packages/compaction/compaction-basic/src/region.ts' --coverage.include='packages/compaction/compaction-basic/src/summarizer.ts' passed 203 tests in 6 files; both changed sources have 100% statements, branches, functions, and lines. Region-to-default-summarizer cases pin exact prefix and tools for nonempty Unicode/multiline, empty, and absent heads. DeepSeek JSON byte equality and pi-ai context equality pin leading-message vs separate-system equivalence on text and image-capable conversion paths.
pnpm run doc-sync passed all 33 gates including doc-typecheck, documentation build, translation pairing and model-experience checks. git diff --check passed. Own dependencies installed with pnpm install --frozen-lockfile. An initial test iteration used a nonexistent ctx.dispose teardown on the in-memory fixture; corrected to its existing fixture lifecycle and reran successfully. No runtime/model behavior, normalizer marker, main worktree, push, or rebase changes.
Consolidate the representation-change PR and its rebase reconciliations into one baseline. Preserve the exact tree and keep the in-history feature in the dependent PR. Follow-up fixes remain separate.
Session format v2 embeds each attempt's compact stream in
assistant/message and assistant/attempt, but Host and client consumers
still expanded it into per-member TimedStreamChunk arrays and did
per-member work; expandAssistantStream materializes the full array before
find/toReversed/break can answer. Session Stats (the projection phase of
every Session open), the token meter's usage and provider-assembly folds,
the subagent output fold, and the Session Controller image lookup still
paid O(members) allocation and time per settlement.
The Chat and Trajectory definitions were already settled from
message.content on master; the remaining per-member folds stay.
dsh-llm now exports record-level readers (first token, visible content,
visible text, last raw chunk of a type, raw chunks of a type, joined
text, run-aware assembly, per-run first-token/first-visible times) that
scan the compact records once with early exit. Session Stats reads
assistantStreamFirstTokenTime, the token meter reads
lastAssistantStreamChunk(stream, 'usage') and assembles through
assembleAssistantStream, the subagent output fold appends
joinAssistantStreamText, and the Session Controller scans
assistantStreamChunks(stream, 'block-end').
expandAssistantStream is deliberately not memoized: retaining expansions
costs roughly ten times the compact stream for the Session's lifetime.
It remains the validating path at durable boundaries.
Synthetic 200-turn v0 migration benchmark, median of five: first-open
projection 28.0 ms -> 5.4 ms, first-open total 76.9 -> 50.0 ms, peak RSS
137.2 -> 94.9 MB; reopen projection 17.8 -> 5.6 ms; all phase budgets and
the 128 MB heap constraint keep passing.
Run the dsh-archive-agent-notes audit over every active Agent Note on
current master, judging each record by whether its rationale still guides
work rather than by size or age.
- Archive 453 implemented bilingual triplets (417,882 English words):
completed UI chrome, narrow adapters, closed bug fixes, implementation
walkthroughs whose package READMEs, docs pages, generators, or successor
notes now carry the useful behavior, and 51 records fully superseded by
a later active note. Keep 201 implemented notes whose ownership rules,
negative guarantees, durable or wire semantics, security rules,
reintroduction conditions, or still-tempting rejected alternatives
remain useful.
- Reject 7 proposals whose premise is gone or whose work shipped in
amended form under other records; delete 2 rejected notes that no
longer prevent a plausible mistake.
- Retarget every remaining inbound link to the archived path, and repair
active prose that named an archived record as the owner of a live fact:
parenthetical citations drop, ownership sentences redirect to the
README, docs page, or active note that states the fact, and history
citations say so. Chinese files link the English archived path because
the pairing gate treats the frozen tree as outside the bilingual corpus.
- Seal 1,359 new frozen artifacts; existing seals are unchanged and
outbound links from archived notes are neither inspected nor repaired.
- Regenerate docs/config-catalog.md after the hook-bridge comment edits
shifted two source line numbers.
- gen-cordis-catalog / gen-cordis-inspect-catalog / gen-client-catalog /
gen-config-catalog / gen-doc-graphs / gen-module-graph outputs pick up
the converged Remote vocabulary (gateway/* codes, RemoteError JSDoc)
and the dsh-util-time package; the zh sides of the three
English-generated pages follow the same line-number shifts.
- the message-feedback protocol golden records the accepted wire
change: a boundary-validation failure now reports
gateway/input-invalid with structured details instead of a bare
internal code.
A max-tokens response that included a tool call persisted assembler-transformed
content next to replay metadata projected from the untransformed native message,
so the next request died in history reconstruction with INVALID_REPLAY_STATE and
the session stayed permanently stuck.
Write side: the finish chunk's replayState becomes a typed ReplayEnvelope —
opaque response-level metadata plus optional per-block entries aligned with the
emitted block sequence. BlockAssembler computes one keep/drop decision for
blocks and entries together, so stored metadata always describes stored content
and retained blocks keep their signatures. pi-ai splits its state into a
version-2 response half and per-block signature entries.
Read side: durable content is authoritative. toPiAssistant degrades any
unusable state — foreign kind, other versions (including the flat v1 form
already on disk), malformed metadata, or content/block mismatches — to the
existing provider-neutral conversion with an onReplayDegrade diagnostic instead
of failing the request, which un-bricks sessions poisoned before this change.
Covered by assembler and replay unit tests, an agent-loop continuation
regression, keyless real-composition continuation tests (native pruned-envelope
replay and legacy flat-state degrade), and the authored keyless snapshot
scenario max-tokens-continue through the assembled ACP app.
Correct the reworked English wording in the CLI profile section, architecture map, and llm-streaming cross-reference, then bring every edited pair's Chinese side along and re-record the pairing hashes.
Review round findings:
- Clear the streaming attempt before the request-error waterfall: a cancel
landing during recovery (typically the llm/retry backoff, after clients
reset the streamed rendering) must not resurrect the failed stream's
prefix. Provider failures commit nothing, now including that window.
- Record interrupted: true on the cancellation-finalized assistant/message.
The chat projection keeps the settled prefix classified as interrupted
(Stopped chip, restored web goldens), and request inspection leaves the
request uncompleted so the step boundary classifies it as before.
- Pin the recovery-window and retry-discard semantics with content-bearing
failed streams in cancel.spec; update the ACP late-end expectation to the
finalized-prefix transcript.
- Mention interruptedBlocks() in the assembler stream-lifecycle JSDoc.
A turn cancelled mid-stream dropped everything the interrupted step had
streamed: chunks stayed in the log for replay, but no assistant/message
joined the surface, so the next request carried nothing the user had
watched stream. Follow-ups after cancel could not connect and forks
inherited the gap.
Keep the streaming attempt alive across the step's request loop and, when
an abort escapes with the attempt uncommitted, finalize its user-visible
prefix as the step's ordinary assistant/message citing the logged chunk
seqs. BlockAssembler.interruptedBlocks() owns the safe subset next to the
max-tokens rule: closed and open text/reasoning blocks with content, in
stream order; tool calls, empty blocks, and unknown open blocks drop.
Retry clears the attempt first, so an abort after llm/retry finalizes
nothing from the reset stream. Tool-phase cancellation and provider
failures keep their existing shapes.
Yichen Jiang