Darkwood · Symfony · Flow · context

Nolife Tokens

Shrinking LLM context
without destroying recoverability

Symfony + Flow experiment

The problem

Tool

↓

50 KB stdout

↓

LLM context

logs

JSON telemetry

git diffs

repo trees

compiler output

test output

Large context windows make it possible to send everything.
They do not make it wise.

Hypothesis

Do more deterministic work
before the LLM.

Before

Tool
→ raw output
→ LLM

After

Tool
→ deterministic reducer
→ useful context
→ LLM

PHP handles

deduplication · classification · filtering · hashing · recovery

These tasks do not require a model.

content/nolife-tokens

A laboratory, not a framework

PHP 8.5

Symfony 8.1

Console

darkwood/flow

Not used

No embeddings
No vector DB
No MCP
No Redis
No LangChain

Pipeline

Load → Classify
→ Measure → Optimize
→ Measure → Evaluate

Content is not all the same

Different noise. Different limits.

Logs

[INFO] started
[INFO] started
[INFO] started
[ERROR] Payment failed

↓

[INFO] started ×3
[ERROR] Payment failed

JSON

id
status
error
telemetry…

↓

id
status
error

Git diff

- security check
+ …

Changed lines are
often already
the signal.

Natural compression limits differ by type.

TOKEN_PIPELINE_POC · measured

Volume vs signal

log · −75.4%

raw

4839

opt

1190

diff · −24.0%

raw

3186

opt

2421

JSON · −99.0%

raw

6780

opt

67

Case	Raw*	Opt*	Saved	Signal
log	4,839	1,190	75.4%	PASS
diff	3,186	2,421	24.0%	PASS
JSON	6,780	67	99.0%	PASS

* estimated tokens = bytes / 4 — not provider-billed

Token reduction without signal retention is vanity.

The gap

REMOVE FROM CONTEXT
≠
DESTROY INFORMATION

Bad

raw
→ compress
→ information gone

Better

raw
       ↘ stored omission
compress
   ↓
visible context

Next milestone: REVERSIBLE_CONTEXT_POC

Reversible references

Structure stays. Bulk leaves.

Original

{
  "id": 42,
  "status": "failed",
  "error": "…",
  "telemetry": { "…6000 tok…" }
}

Optimized

{
  "id": 42,
  "status": "failed",
  "error": "…",
  "telemetry": "#ref:ctx_31154f"
}

#ref:ctx_31154f

→

var/context/…json

→

exact bytes

Marker cost ≈ 4 estimated tokens

Flow pipeline

Two invariants

1Load

2Classify

3MeasureRaw

4Optimize + store refs

5MeasureVisible

6Signal check

7Recovery check

SIGNAL = PASS

RECOVERY = PASS

bin/console tokens:show-ref ctx_31154f

REVERSIBLE_CONTEXT_POC · measured

Out of context ≠ gone

Case	Raw*	Visible*	Referenced*	Saved	Refs
log coarse	4,839	1,204	3,671	75.1%	1
log fine	4,839	1,202	3,670	75.2%	3
JSON	6,780	146	6,191	97.8%	9
diff	3,186	2,434	749	23.6%	1

* estimated (bytes / 4) · SIGNAL PASS · RECOVERY PASS on all rows

6,191 estimated JSON tokens moved out of active context.
Not destroyed. Recoverable.

Granularity experiment · logs

Granularity has a cost

Coarse

1 ref
4 marker tokens
1,204 visible

Fine

3 refs
12 marker tokens
1,202 visible

3× more references for ~2 estimated tokens saved.
For this workload: coarse > fine

Workload-dependent — structured JSON can benefit from finer, positional refs.

For coding agents

Better context. Same agents.

Tool output → PHP reducer

↓

Useful context → LLM

↓ needs more?

#ref expansion → raw source

git diff

logs

tests / JSON / trees

Not another agent interface — better context for the ones we already use.
Next: progressive disclosure (L0 overview → expand #ref → raw).

What is the minimum context the model needs
to make the correct decision?