kami 04336308f5 feat(inference): clamp max_tokens to context-window runway (#291) ...

The stage completion cap (e.g. 24_576) is a ceiling, not a promise. When the
rendered prompt + tool schemas fill most of the model window, sending the raw
cap makes llama.cpp truncate the prompt from the left — the traced Gemma4
32_767-token blow-up. Compute effectiveMaxTokens = min(cap, contextSize -
promptTokens - toolSchemas - templateOverhead - reserve), counting prompt/tool
tokens with the model's own tokenizer (char/4 fallback). Live-path only;
deterministic replay never calls infer, so no event recording needed.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

2026-07-20 18:11:43 +04:00

.correx

feat(validation): auto build-gate on terminal freestyle stage (Gate 4 floor)

2026-07-07 01:01:12 +04:00

apps

feat(stage-workingset): durable, replay-safe stage working set (slices 1-4)

2026-07-20 14:22:12 +04:00

core

feat(context): overlay sibling-stage writes onto repo retrieval

2026-07-20 14:38:34 +04:00

docs

wip(freestyle/acr): grounding & edit-tool fixes + ACR-compiler experiment

2026-07-19 01:20:37 +04:00

examples

fix(toolintent): manifest gate defers to task affected_paths; scaffold-glob guidance

2026-07-19 17:32:48 +04:00

gradle/wrapper

epic-12: after epic audit and init commit

2026-05-17 03:19:39 +04:00

infrastructure

feat(inference): clamp max_tokens to context-window runway (#291)

2026-07-20 18:11:43 +04:00

interfaces

docs(dox): build out the DOX AGENTS.md hierarchy

2026-06-28 20:43:27 +04:00

plugins

docs(dox): build out the DOX AGENTS.md hierarchy

2026-06-28 20:43:27 +04:00

prompts

fix: workflow inference chain — surface llama-server errors, user-turn-last, bundle-relative prompts

2026-06-01 23:23:29 +04:00

qa

docs(qa): mark F-002/F-003/F-004/F-006/F-016 fixed in audit report

2026-06-10 11:14:09 +04:00

scripts

Implement closed-loop workspace follow-ups

2026-07-15 23:48:12 +04:00

sidecars/llmlingua

fix(context,tools): context hygiene + tool ergonomics from freestyle QA

2026-07-02 00:44:04 +04:00

testing

feat(context): overlay sibling-stage writes onto repo retrieval

2026-07-20 14:38:34 +04:00

.correxignore

feat(context): reasoning-thread continuity, L3 doc filtering, and gate hardening

2026-07-10 11:13:43 +04:00

.gitignore

feat(validation): auto build-gate on terminal freestyle stage (Gate 4 floor)

2026-07-07 01:01:12 +04:00

AGENTS.md

feat(orchestration): plan grounding, positive-pattern mining, stage-prompt audition rig

2026-07-16 14:55:56 +04:00

BACKLOG.md

merge: integrate feat/backlog-burndown into master

2026-06-26 11:30:41 +00:00

build.gradle

refactor: decomposition WIP (orchestrator/server/execution-plan)

2026-07-15 13:17:01 +04:00

detekt.yml

chore(detekt): ratchet maxIssues 120→90 after orchestrator decomposition

2026-07-13 12:09:41 +04:00

gradle.properties

epic-12: after epic audit and init commit

2026-05-17 03:19:39 +04:00

gradlew

epic-12: after epic audit and init commit

2026-05-17 03:19:39 +04:00

gradlew.bat

epic-12: after epic audit and init commit

2026-05-17 03:19:39 +04:00

RETRO.md

merge: integrate feat/backlog-burndown into master

2026-06-26 11:30:41 +00:00

settings.gradle

feat(stage-workingset): durable, replay-safe stage working set (slices 1-4)

2026-07-20 14:22:12 +04:00

tool-result-schema.md

refactor(detekt): extract magic-number constants, wrap long lines, drop legit suppressions

2026-07-12 23:12:28 +04:00

S

Description

No description provided

4.9 MiB

Languages

Kotlin 88.4%

Go 11.4%

Python 0.2%