fix(anthropic): non-whitespace placeholder for synthesized leading user turn

[acke]

Problem

_ensure_leading_user_turn() in agent/anthropic_adapter.py prepends a synthetic user turn when a compacted message window starts with an assistant message (Anthropic requires messages[0].role == "user"). The placeholder text was a single space " ".

Anthropic's Messages API rejects empty/whitespace-only text content blocks:

HTTP 400 invalid_request_error
messages: text content blocks must contain non-whitespace text

Because messages[0] is re-sent on every turn, any conversation whose compacted window leads with an assistant turn (e.g. after a second context compaction where protect_head has decayed to 0) wedges permanently behind the same 400. Both the primary model and the fallback receive the identical malformed payload and fail identically — a non-retryable client error — so the user only ever sees repeated "model provider failed / switching to fallback" banners with no recovery.

Observed live in a long-lived gateway (Discord) session: every inbound message 400'd across restarts until the placeholder was fixed. It never self-heals because the offending block is structural, not content the user can edit away.

Fix

Use the module's existing _EMPTY_TEXT_PLACEHOLDER ("(empty)") — the same non-whitespace placeholder already used by the sanitize paths (_safe_text, _sanitize_replay_block, _convert_assistant_message). This is a one-line behavioral change; the surrounding whitespace-coercion machinery was already in place (ref #69512) but this synthesis site was missed.

Tests

• New TestEnsureLeadingUserTurn in tests/agent/test_anthropic_whitespace_text_blocks.py (4 cases): prepends when first is assistant, placeholder is non-whitespace, no-op when already user-led, empty list untouched.
• Updated two change-detector assertions in tests/agent/test_anthropic_adapter.py that had frozen the buggy " " value to assert _EMPTY_TEXT_PLACEHOLDER instead.

scripts/run_tests.sh tests/agent/test_anthropic_whitespace_text_blocks.py tests/agent/test_anthropic_adapter.py → 198 passed. (3 unrelated TestRunOauthSetupToken failures are pre-existing on main — a pytest-9 MagicMock/json.loads interaction — and are untouched by this change.)

Ref #69512.

[alt-glitch]

This was generated by AI during triage.

Duplicate of #70081: both current diffs replace the same whitespace-only synthesized leading user turn with a printable sentinel. #70910 remains related because it adds a broader final assembled-request backstop.

[teknium1]

Thanks for isolating the synthesis-site gap. The premise remains valid on current main: agent/anthropic_adapter.py:2613 still inserts " " as the synthetic leading user text block, and convert_messages_to_anthropic() calls that helper on the returned provider message list at agent/anthropic_adapter.py:2674. The adapter already defines the appropriate sentinel at agent/anthropic_adapter.py:1931.

The proposed replacement with _EMPTY_TEXT_PLACEHOLDER is correctly scoped and preserves the existing role-normalization behavior. Current coverage also still exercises the compaction-shaped conversion path, but freezes the buggy literal at tests/agent/test_anthropic_adapter.py:871-883.

The branch predates substantial test pruning after base eb52760564dbba2e5971fa54bd67384e281cd3b8; salvage should reconcile the test hunks against the remaining current test rather than preserve removed fixtures verbatim.

Automated hermes-sweeper review.

[GottZ]

This was generated by AI during triage.

Summary

Four PRs address the same assembly-time Anthropic whitespace failure. #70081, #70159, and #71935 replace only the invalid synthetic leading-user placeholder, while #70910 fixes that root cause with the canonical sentinel and additionally guards pre-shaped message blocks and cache-controlled system blocks in the fully assembled request.

Related pull requests

• #70081 duplicate — (+46/-3) — close as duplicate of #70910: it fixes the synthesized leading-user block but hard-codes "(empty)" instead of using _EMPTY_TEXT_PLACEHOLDER; despite the keep_open review on #70081, #70910 contains the same source fix using the requested canonical constant and covers additional assembly-time blank-block paths.
• #70159 duplicate — (+8/-3) — close as duplicate of #70910: it changes the same synthesized block to "." and updates the existing tests to assert non-whitespace behavior; despite the keep_open review on #70159, #70910 fixes the identical site with _EMPTY_TEXT_PLACEHOLDER and adds coverage for blank pre-shaped and system blocks.
• #70910 best fix — (+165/-3) — keep open with a salvage path: this is the recorded best fix for #70909, using _EMPTY_TEXT_PLACEHOLDER at the proven synthesis site and adding a final assembled-request guard for the independently reproduced pre-shaped-content and cache-controlled-system cases. Preserve the root-cause hunk, the final text-block-only guard, and focused regressions while reconciling the test additions with current main.
• #71935 fixes — (+48/-3) — close as duplicate of #70910: its source hunk correctly uses _EMPTY_TEXT_PLACEHOLDER, but its scope remains limited to the synthesized leading turn. Despite the MAINTAINER-BOT keep_open verdict on #71935, the complete diff of #70910 contains that same fix plus tested request-level protection for two additional blank-block paths, so #71935 has no distinct implementation to salvage.

Duplicates

#70081, #70159, and #71935 duplicate the synthesized-leading-user fix already contained in #70910; #70081 and #71935 are especially close at the source level, while #70159 differs only by choosing "." rather than the canonical placeholder.

Suggested consolidation

Keep #70910 open with a salvage path: retain its canonical _EMPTY_TEXT_PLACEHOLDER root-cause fix, final text-block-only request guard, and focused regressions, rebasing or reconciling the tests against current main as needed. Close #70081, #70159, and #71935 as duplicates of #70910 because their diffs cover only the root-cause subset and add no independent behavior beyond the recorded best fix.

Complex graph

flowchart LR
    classDef open fill:#dbeafe,stroke:#1d4ed8,color:#1e3a8a
    classDef merged fill:#dcfce7,stroke:#15803d,color:#14532d
    classDef closed fill:#e5e7eb,stroke:#6b7280,color:#1f2937
    classDef unverified fill:#f3f4f6,stroke:#9ca3af,color:#374151
    classDef best stroke-width:3px,stroke:#b45309
    classDef target stroke-width:3px,stroke:#4338ca
    I70909(["issue #70909 (open)"])
    subgraph Dup70081 ["PRs duplicating each other"]
        P70081["PR #70081 (open)"]
        P70159["PR #70159 (open)"]
        P71935["PR #71935 (open)"]
    end
    P71935 -->|fixes| I70909
    class I70909 open
    class P70081 open
    class P70159 open
    class P71935 open
    class P71935 target
    click I70909 "https://github.com/NousResearch/hermes-agent/issues/70909"
    click P70081 "https://github.com/NousResearch/hermes-agent/pull/70081"
    click P70159 "https://github.com/NousResearch/hermes-agent/pull/70159"
    click P71935 "https://github.com/NousResearch/hermes-agent/pull/71935"

Loading

Graph: solid arrow = fixes / best fix, dashed arrow = partial or unverified (see edge label); boxed group = PRs duplicating each other; amber border = best fix; indigo border = target; gray node = closed (state tag in the node label).

Cross-PR triage: Reviewed 4 pull requests and 1 issue in this complex. Each diff was read against this issue; Assessment working set: 21 kB of PR diffs, 13 kB of issue/PR text, 4 kB of discussion (8 comments), 8 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.