Skip to content

Verify: config and doctor disagreements #468

Closed

Description

@possiblyneal

Part of #448

Question

Do doctor and the daemon disagree on provider detection, does doctor prompt in a non-interactive shell, do OpenRouter suffixes break the dimension lookup, does status mislabel embeddings, and does MAX_TOKENS 4096 truncate Summaries? Also the repo cleanups. Upstream: rohitg00#1572, rohitg00#1586, rohitg00#1671, rohitg00#1494, rohitg00#1488, rohitg00#1587, rohitg00#1739, rohitg00#1798.

Activity

  1. self-assigned this
    on Nov 18, 2026
  2. possiblyneal commented on Nov 18, 2026

    @possiblyneal
    OwnerAuthor

    Verdict #468 (worktree tmp/wt/verify-468, base fdc3d45)

    Fix class: none of these change Recall/Injection/stored content; no eval:suite needed.

    rohitg00#1572 placeholder key picks a provider (config)

    APPLY. Repro: test/config-placeholder-keys.test.ts fails 1/1 on base (anthropic chosen for sk-ant-...). Fix: src/utils/placeholder-key.ts, used by detectProvider/detectLlmProviderKind. Reach: Operator keys are real; latent. Class 2: no.

    rohitg00#1586 gemini/google placeholder key

    APPLY. Same test file; src/providers/index.ts picks first non-placeholder of GEMINI_API_KEY/GOOGLE_API_KEY. Reach: latent. Class 2: no (provider selection only).

    rohitg00#1671 doctor: OPENAI_API_KEY_FOR_LLM=false and agent-sdk

    APPLY. test/cli-doctor-fixes.test.ts: 2 tests fail on base. Fix: realProviderKeys honours FOR_LLM=false; no-llm-provider-key passes with AGENTMEMORY_ALLOW_AGENT_SDK=true. Reach: diagnostics output only. Class 2: no.

    rohitg00#1494 doctor prompts in non-TTY/CI

    APPLY. doctorPromptsAllowed test fails on base. runDoctor reports only without a TTY or under CI unless --all. Reach: scripted runs. Class 2: no.

    rohitg00#1488 status prints bm25-only

    REJECT. src/triggers/api.ts:272 sends "embeddings"|"none" and src/cli.ts:561,815 compare to "embeddings"; no repro. Retry if: status output diverges from /health provider field.

    rohitg00#1739 cleanups

    WONT-FIX/NOT APPLIED. package.json: dotenv is a dependency but never imported in src/scripts (removing it is a trivial follow-up, not a defect); migrate script points at dist/functions/migrate.js and src/functions/migrate.ts exists, so the script is not broken (claim does not hold here). ROADMAP, Node 20 EOL, --legacy-peer-deps: upstream housekeeping; fork floor is already Node >=22.13. Retry if: a dependency audit or install failure names dotenv or the migrate script.

    rohitg00#1798 FallbackChainProvider lacks describeImage

    APPLY. test/fallback-chain.test.ts fails 1/1 on base. Fix: describeImage defined only when some provider has it, tried in order. Reach: zero today (Operator's OpenAI-compatible provider has no describeImage; only anthropic.ts does). Class 2: no.

    Also

    embedding dimensions: model:variant suffix stripped in lookupModelDimensions; test fails 1/1 on base; latent.

    rohitg00#1587 (raise default MAX_TOKENS 4096 -> 16384): REJECT, no reach

    • Live copy: 1,801 Summaries, none end mid-sentence; estimated output tokens p50 287, p90 761, p99 2,405, max 6,082. The stray tags found while checking (e.g. </color> inside concepts) are parser leakage, not truncation: filed as Session Summaries store stray XML tags in their fields #472.
    • src/config.ts:93 defaults MAX_TOKENS to 4096 only when the env key is absent. ~/.agentmemory/.env sets MAX_TOKENS (key name read, value not), so the default never applies to the Operator's daemon.
    • A cut-off reply is already reported, not silent: src/providers/openai.ts:217 throws on finish_reason: length with no content, naming MAX_TOKENS.
    • Retry if: the Operator drops MAX_TOKENS from .env, or summarize/compress failures are attributed to finish_reason: length.

    Bar

    • typecheck 0 errors; npm test 3,124 passed / 1 skipped; eval:gate and eval:gate:rules pass; eval:suite fast vs base-fdc3d459: every metric Δ 0.000.
    • Branch verify/468-config-doctor, uncommitted, worktree tmp/wt/verify-468.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions