Stage E: fix deferred-write content generation + jitter repair retries

[Anmolnoor]

Context

Two robustness fixes surfaced by running the report flow against live Ollama Cloud models (gemini / kimi / qwen3.5). The crash fixes (A/B) and questioning (C/D) all worked; these address quality/robustness gaps that only appear against a real model.

Note: based on stage-d/out-of-scope-escalation; merge order is #3 → #4 → #5 → #6 → this. GitHub retargets as bases merge.

1. Deferred-write body came out wrapped in a JSON plan

When Stage B materialized a content_brief via the text call, _generate_file_body replayed the planning conversation (prior plans as assistant turns). That primed the model to keep "planning" — it returned another {actions: [...]} object, so the written file was a JSON blob with the real markdown buried in a content field (and stats were vague/hallucinated because the data arrived as planning turns, not clean reference data).

• Reframe the call: drop the assistant-plan turns; pass gathered data as a single plain reference block with a hard "you are a file-content writer, not a planner — output raw bytes only" instruction.
• _unwrap_generated_file_body defensive net: if the model still returns an AssistantPlan-shaped blob, extract the write action's content. Plain text and legitimate JSON files (no actions) pass through untouched.

2. Plan repair retries were deterministic

At temperature=0, a model that emits malformed/empty JSON reproduces it identically on retry — wasting the attempt (exactly what kimi did, breaking the same brace twice). ProviderPrompt gains an optional temperature; the planner sets 0.4 on repair retries (first attempt stays deterministic at 0). Ollama honors it for both JSON and text calls.

Verification

Re-ran the original report request end-to-end against qwen3.5:397b-cloud: the report now writes as clean markdown starting with #, using real GitHub values (31 followers, account created 2019, etc.) instead of a JSON wrapper with invented stats.

Tests

4 new + 1 extended (401 total, ruff clean): _unwrap_generated_file_body extracts wrapped content and leaves plain/real-JSON untouched; the orchestrator unwraps a plan-wrapped generated body; the Ollama adapter honors a prompt temperature; the repair retry carries temperature 0.4.

🤖 Generated with Claude Code

[Anmolnoor]

Merged into main via commit 27c97b0 (stages A–G landed as one merge). See the summary on #9.