Centralize context-summary generation without centralizing workflows

Status: accepted

Current decision

Mevedel uses one generator to turn frozen model-visible context into context- summary text, while each consuming workflow retains ownership of source selection, hooks, retries, persistence, injection, and source mutation. The generator supports continuation, handoff, and digest purposes through a shared, validated summary core; only continuation summaries carry actionable next steps, while handoff source material remains evidence beneath a separately supplied authoritative task. Only continuation generation treats an earlier continuation summary as authoritative retained context; handoff generation instead re-filters the parent's complete effective context, including any anchored summary, against the receiving task. Source transcripts are projected as one delimited evidence document with provenance labels rather than replayed provider roles, so user turns, assistant text, and tool results remain untrusted evidence instead of live summarizer instructions. Caller guidance may focus content but cannot override the purpose, structure, or authority contract. This shares prompt and model-request behavior without making non-mutating plan, worktree, or agent handoffs inherit conversation-compaction lifecycle semantics.

Journal digest generation needs the same frozen evidence, model policy, admission, and callback settlement as continuation and handoff generation. That shared behavior is the reason for adding a third purpose rather than a second request implementation. The digest purpose has a separate factual bullet-list prompt and four headings (Done, Learned, Surprised, Unfinished), a 16 KiB visible-output cap, and the ordinary model context reserve during input admission. Explicit server output limits are honored on the realized provider payload; the journal adds no automatic token ceiling. It treats prior summaries as evidence and never turns unfinished state into instructions. Capture jobs, source projection, persistence, deadlines, and retry policy belong to the journal consumer. Existing continuation and handoff contracts remain intact.

Digest policy preserves configured reasoning effort and otherwise uses the provider default. The journal workload is independent of summarization and defaults to balanced. Configured token limits and frozen policy remain authoritative, but the journal neither invents an output-token allowance nor forces reasoning off. Providers lacking token controls retain client byte/deadline guards without a server billing guarantee.

The compaction manual owns evidence selection, generation, and continuation recovery. The memory manual owns journal capture, digest processing, and consolidation configuration.

Rationale and consequences

Sharing frozen-input admission, prompt construction, validation, and callback settlement avoids duplicate model-request implementations. Keeping selection, hooks, retries, and source mutation with consumers prevents a non-mutating handoff from inheriting compaction's lifecycle. Applicable prepared skill bodies remain summary evidence without becoming a new durable skill-authority store.

Decision history

The instruction-delivery audit found that compaction supplied skill names and invocation metadata but omitted the prepared bodies already recorded in live session state. A capped tool result or changed source file could therefore leave the generator without the relevant guidance. The selected records now include the original prepared body and source/conversation attribution as historical evidence. This neither invokes the skill nor proves it remains active; the summary contract retains applicable obligations, corrections, and retirement explicitly. Full bodies remain subject to the existing input-size gate, and cold resume still relies on durable transcript/summary evidence, because invocation records are not a persisted sidecar field. The cold-restore check then exposed the same loss in persisted Skill results: ordinary tool-output truncation removed trailing obligations even though the restored transcript still contained them. The evidence projector now retains complete Skill instruction results, still labelled as tool evidence, without adding a second durable skill-state format. Ordinary tool output remains capped.

The first real-model long-transcript evaluation exhausted the configured DeepSeek model's 4,000-token allowance in reasoning and returned no digest. Prompt shortening alone reproduced that failure. The initial response was to default unspecified digest effort to disabled, or none, when supported, while preserving explicit choices. The September 2026 controlled evaluation below replaces that response: the small shared output budget caused the empty answer, while disabling reasoning could reduce selection quality.

The configured gptel branch also encoded DeepSeek's disabled effort as reasoning_effort: "disabled", which the service rejected. A separate isolated dependency patch omits that field when thinking.type is disabled, following DeepSeek's API contract. That quality evaluation used the dependency patch; its affected checkout required the fix. This historical result does not establish the status of every current gptel installation. mevedel adds no provider adapter or compatibility advice for it.

The subsequent configured-model comparison recovered the buried lesson with GPT-5.6 Sol, which the user selected provisionally. Its Codex OAuth transport removes max_output_tokens; requiring that field rejected every digest before dispatch despite retaining the same input reserve, streaming byte limit, and 120-second journal deadline. The requirement is bounded evidence admission and bounded accepted output, not universal provider support for a token parameter. Generation now accepts providers without that control, omits the unsupported Codex setting, and still clamps every server limit that is present. Client cancellation bounds local processing and accepted output; it cannot guarantee a server-side token or billing ceiling. This limitation is explicit instead of being hidden by a test-only admission exception. Model selection stays in the ordinary configurable workload map, not in the journal implementation.

The September 2026 configuration review exposed unnecessary coupling: tuning journal extraction also changed compaction and handoffs, while tuning memory consolidation also changed Buddy. Shared request mechanics do not require shared model selection. The generator now resolves journal for digests and summarization for continuation/handoff; journal capture freezes journal before dispatch. Consolidation resolves memory instead of buddy. Both new workloads default to balanced and use the existing tier/provider/effort map. Existing queued captures keep their frozen policy. The generator and workflow ownership boundaries remain unchanged.

September 2026 reasoning and output-budget evaluation

A controlled evaluation used six synthetic reusable-knowledge fixtures, two runs per fixture, and six settings through the configured DeepSeek V4.1 Flash and GPT-6 Astra transports (72 measured calls). DeepSeek with reasoning disabled passed 18/24 cases; reasoning high passed 22/24. Failures without reasoning included retaining tracker state, omitting a reference, and inventing an observed outcome. One additional failure was a malformed empty-section bullet; excluding that formatting failure does not reverse the comparison. High with a 4,000-token cap again consumed exactly 4,000 reasoning tokens and returned no visible digest. High without that cap completed every request, although one still omitted a reference. Astra low and high both passed 12/12; that endpoint does not support reasoning off. This small fixture set does not establish a universal optimum or perfect reliability.

The replacement is to remove both the forced reasoning-off default and the journal's automatic 4,000-token cap. Keep explicit user choices, normal context admission, brief-output instructions, 16 KiB accepted text, and the consumer's 120-second deadline. Provider defaults still impose their own limits. Existing captures retain their stored choices; no queued evidence is rewritten. This uses native workload/provider configuration rather than inventing a cross-model reasoning scale. The evaluation's evidence, grading, actual request controls, dependency hashes, limitations and four failed pilot measurements are retained with the repository's working material rather than in this documentation.