[piyushmakhija28]

Both found by reading the live run's log, not the code.

1. A function that has been dead for three months, and lied about it

synthesize_prompt_with_flow_data loaded a PromptGenerator by dynamic import from:

scripts/architecture/03-execution-system/00-prompt-generation/prompt-generation-policy.py

That file was deleted on 2026-05-04 in 683c23a. The import is the function's first statement, so it has raised on every run since, and the ninety lines of flow-data assembly after it were unreachable the whole time. Removed rather than left reading as live code — 97 lines out, 31 in.

The except-branch was not a fallback, it was a false report. It returned state[USER_MESSAGE] under the key synthesized_prompt, and flow_trace_converter writes that key to current-synthesis.txt and prints:

Synthesis: Generated (N chars)

So for three months the trace claimed a synthesis had been produced while writing out the raw task, unchanged. Returning empty lets the consumer's own if synthesized_prompt: guard skip the file — nothing was synthesised, so nothing is reported as synthesised. The metadata names the commit that removed the generator, so the next reader does not have to bisect for it.

The per-run stderr [WARNING] Prompt synthesis failed: goes with it. It announced a permanent condition as though it were an incident, on every run — and it was the only line in an otherwise clean log that looked like something had gone wrong today.

2. docs/policies/ was never a mirror of ~/.claude/policies/

CLAUDE.md claimed it was, and claimed the same for docs/standards/ ↔ ~/.claude/rules/. Measured on 2026-08-08, the policy trees had drifted apart by 25 files:

In the repo, absent from the machine	18
On the machine, in no repository at all	7

Several of the 7 are for 02-plan-mode/ and 05-skill-agent-selection/ — features v1.15.2 purged.

The more useful correction is that they were never meant to match. docs/policies/ is documentation of pipeline behaviour (quality-gate-policy.md, pr-code-review-policy.md) and nothing loads it. ~/.claude/policies/ is the runtime corpus get_policies_dir() reads, of which the standards selector consumes only the 02-standards-system/ tier and the language-mapped files.

"Mirror" invited the assumption that editing one affected the other, which it never did in either direction. The note also records that the 7 machine-only files are unversioned — if that machine is lost, so are they.

Verified on a live run

• [WARNING] Prompt synthesis failed — gone
• Synthesis: Generated — gone, and no current-synthesis.txt written
• Step 1 still emits its 333,165-char prompt
• 176 tests across every suite touching the orchestrator, synthesis or flow-trace; all six gates

One scare, recorded because it looked like a regression

The run immediately after this change hit the script's 300-second orchestration cap, where earlier runs took 46 s and 53 s. It was not this change: the 167.9-second gap was inside context_sync/session_loader, which this diff does not touch, scanning 684 MB across 47 session directories and 1,099 files accumulated by the day's test runs. A re-run finished in 99 s.

The 300 s cap and the unbounded session history are a real pairing and will bite again — but not in this commit.

🤖 Generated with Claude Code