fix(blackbox): single global fixed char->token divisor 4.2 -> 4.5

[Kyzcreig]

Per Ace: use ONE divisor for both gateways (consistency), bump fixed 4.2->4.5 for a better system-prompt estimate; non-fixed stays 3.5.

Why 4.5: live post-deploy density measurement — Apollo/default ~4.36 ch/tok, Aegis ~4.83. 4.5 is the first single divisor landing BOTH inside the [0.90,1.05] acceptance band: Apollo 4.5/4.36=1.03 (3% under, safe), Aegis 4.5/4.83=0.93 (7% over, safe). 4.2 left Aegis at ~0.87 (just under band).

Constant + rationale comment + default-fallback test updated. 28/28 composition tests pass.

[github-actions]

🔎 Lint report: fix/fixed-divisor-4.5 vs origin/main

ruff

Total: 0 on HEAD, 0 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 0 pre-existing issues carried over.

ty (type checker)

Total: 9932 on HEAD, 9932 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 5134 pre-existing issues carried over.

Diagnostics are surfaced as warnings — this check never fails the build.

[greptile-apps]

Greptile Summary

Bumps the default fixed char→token divisor from 4.2 to 4.5 (COMPOSITION_CHARS_PER_TOKEN_FIXED) and updates the rationale comment and test assertions to match. The change is grounded in live post-deploy density measurements showing 4.5 brings both the Apollo (~4.36 ch/tok) and Aegis (~4.83 ch/tok) gateways inside the [0.90, 1.05] acceptance band, where 4.2 left Aegis at ~0.87 (below band).

• agent/model_metadata.py: Default return value changed from 4.2 to 4.5; surrounding comment rewritten with per-gateway acceptance-band math and "one global divisor by fiat" policy note.
• tests/test_request_composition.py: test_fixed_divisor_default_and_out_of_range_fallback assertions updated from 4.2 to 4.5; two inline comments in sibling tests (~4.2 section header and # ~1690 at 4.2 annotation) were not updated and are now stale.

Confidence Score: 4/5

Safe to merge — the core constant change is correct and well-tested; two stale inline comments in the test file are the only loose ends.

The production change is a single numeric constant update with sound empirical backing and a clamp/env-override path that is untouched. The test logic is correct (it reads mm.COMPOSITION_CHARS_PER_TOKEN_FIXED rather than a hardcoded literal). Two inline comments in the test file still cite the old 4.2 value — a minor inconsistency but not a correctness risk.

tests/test_request_composition.py — two inline comments (line 248 section header and line 296 per-line annotation) still reference the old 4.2 divisor and should be updated.

Important Files Changed

Filename	Overview
agent/model_metadata.py	Default fixed char→token divisor bumped from 4.2 to 4.5; comment updated with live measurement data and acceptance-band rationale. Logic is correct and clamp/env-override path is unchanged.
tests/test_request_composition.py	Default-value assertion and fallback-loop assertion updated from 4.2 to 4.5. Two inline comments (section header and per-line annotation) still reference the old 4.2 value and were missed in the sweep.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["_composition_chars_per_token_fixed()"] --> B{"HERMES_COMPOSITION_CHARS_PER_TOKEN_FIXED set?"}
    B -- Yes --> C{"2.0 ≤ val ≤ 8.0?"}
    C -- Yes --> D["Return env value"]
    C -- No --> E["Return 4.5 (default)"]
    B -- No --> E
    E --> F["COMPOSITION_CHARS_PER_TOKEN_FIXED = 4.5"]
    F --> G["Fixed buckets: sys_tokens, tool_schema_tokens\n÷ 4.5 (was 4.2)"]
    F --> H["Non-fixed buckets: history, tool_args, tool_result\n÷ 3.5 (unchanged)"]

Loading

Comments Outside Diff (2)

1. tests/test_request_composition.py, line 248 (link)

   The section header comment still references the old divisor value ~4.2; it should reflect the new default of 4.5 to stay consistent with the constant change.

   Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

2. tests/test_request_composition.py, line 296 (link)

   The inline annotation still reads # ~1690 at 4.2. With the new default of 4.5, math.ceil(7100 / 4.5) is 1578, so the comment is stale and could mislead anyone debugging the bound check.

_{Reviews (1): Last reviewed commit: "fix(blackbox): single global fixed char-..." | Re-trigger Greptile}