Configuration

l0-compressor is configured by command-line flags. An optional per-command config file fills in values you did not set on the command line (see Config file below) — there is still no required config, and explicit flags always win.

Tuning Parameters

Parameter	Default	Description
--head	30	Lines to keep from the start of output
--tail	30	Lines to keep from the end of output
--tail-error	120	Tail lines when exit code is non-zero
--threshold	100	Minimum lines before truncation kicks in
--no-squelch	false	Disable the clean-success squelch (keep the full tail on clean exits)
--no-auto	false	Disable adaptive parameter auto-tuning
--auto-floor	10	Safety floor limit under --auto
--auto-ceiling	1000	Max ceiling limit under --auto
--token-factor	4	Divisor for token count estimation

Choosing Values

For build tools (cargo, npm, gradle): defaults work well. Build output is typically headers + progress + summary.

For test runners (pytest, jest, cargo test): consider --tail-error 200 to capture full stack traces on failure.

For log inspection (docker logs, journalctl): consider --head 10 --tail 80 to prioritize recent entries.

Clean-success squelch

On a zero exit with no error/warning signal anywhere in the stream, the displayed tail is trimmed further — halved and floored at 5 lines (--tail 30 shows 15, a tail already ≤ 5 is left untouched, never expanded). The middle of a clean build/test/install is almost always progress noise that the agent does not need to confirm success. Always preserved: the head (command echo) and the final summary line.

The moment any signal appears — error, warn, fail, exception, panic, traceback, fatal (the same keyword set as --only-errors) — or the command exits non-zero, the full tail is restored. Failures are never squelched. The truncation banner reports the actual tail rendered (e.g. 30 head + 15 tail), not the configured cap.

On by default. Disable per-run with --no-squelch, or per-command with squelch = false in the config file.

Config file (optional)

Drop a file in $XDG_CONFIG_HOME/l0-compressor/ (or ~/.config/l0-compressor/) to set per-command defaults without recompiling and without per-tool parsers. l0-compressor auto-detects, in this order, config.{json,toml,yaml,yml,conf,ini} — transparent multi-format with zero extra dependencies (JSON is parsed strictly by serde; TOML/YAML/INI share a small flat parser, since the schema is flat):

# config.toml
[defaults]
recover = true

[cargo]
tail_error = 300
head = 50

[git]
head = 10
tail = 40

The same configuration in JSON or YAML:

{ "defaults": { "recover": true },
  "commands": { "cargo": { "tail_error": 300, "head": 50 }, "git": { "head": 10, "tail": 40 } } }

defaults:
  recover: true
cargo:
  tail_error: 300
  head: 50
git:
  head: 10
  tail: 40

• Tunable keys (all optional): head, tail, tail_error, threshold, only_errors, recover, squelch.
• A section names a command; [defaults] (or [*]) apply to every command, and a per-command section layers on top (command wins field-by-field).
• Commands are matched by resolved name, so sh -c "cargo test" matches the cargo block — the same name used by metrics and auto-tuning.
• Precedence: an explicit CLI flag > config file > built-in default. Auto-tuning then adjusts from that resolved base.
• A missing/unreadable file is silently ignored; a malformed file is ignored with a single stderr note (unless --quiet). Unknown keys are skipped, so a config written for a newer l0-compressor won't break an older binary.

Parameter Auto-tuning (Enabled by Default)

By default, l0-compressor automatically tunes head, tail, and tail_error per bucket — (cmd, args_hash) — based on the execution history in the local metrics log. Pass --no-auto to disable. Each bucket carries its own learning, so e.g. curl https://api.openai.com and curl https://example.com don't pollute each other.

Rules

Six rules feed off the same per-bucket history. The first one whose trigger matches wins (top-to-bottom precedence).

1. expand_tail_err — anti-loop backoff. If the last F consecutive bucket records were failures with output (exit_code != 0 && lines_raw > 0), scale tail_error by (1 + F), capped by --auto-ceiling (default 1000). Failing records with lines_raw == 0 (e.g. grep "no match", find "not found") break the streak rather than feed it — their failure mode isn't one extra tail of error context can help with. Records with --stats --json track this distinction under the noisy counter.

2. decay_moderate — 3-4 consecutive truncated successes. Shrink head and tail by 20% (floored at --auto-floor, default 10).

3. decay_strong — 5+ consecutive truncated successes. Shrink head and tail by 40% (same floor).

4. recover_defaults — the un-ratchet. Every other rule only moves head/tail down and tail_error up, compounding across runs. After 5 consecutive clean (successful, non-truncated) runs, a bucket sitting below its configured base restores head/tail to base — unless the tune was seeded by proactive_shrink, whose evidence a clean streak confirms — and a tail_error expanded above base returns to base once the failures stop. Both restores happen in one firing. Recovery is strictly per (cmd, args_hash) bucket: the same bucket must start producing clean runs (e.g. the same cat <file> over a now-smaller file), not merely the same command with different args.

5. decay_steady — window-adaptive. If the bucket has ≥20 records, all successful, and ≥80% of the last 20 are truncated, shrink head and tail by 30%. Catches the steady-state pattern that "5 consecutive" misses when occasional non-truncated runs break the streak.

6. proactive_shrink — long clean histories. If the bucket has ≥20 records, all clean (success + not truncated), and max(lines_raw) + 5 is at most half the current head + tail budget, set head = max(lines_raw) + 5 and tail = default_tail / 4. Max-based, so it never introduces a new truncation vs. observed history.

A trigger whose numeric result equals the seeded values (floor/ceiling pinned) is not a firing: no event is recorded and nothing is persisted.

Persistence (tuned.jsonl)

Each time a rule fires (with a real change to head/tail/tail_error), the result is upserted into $XDG_DATA_HOME/l0-compressor/tuned.jsonl, keyed by (cmd, args_hash) — the file is compacted on write to one line per bucket, and entries older than 30 days are pruned (an expired tune also stops seeding runs). The next run of the same bucket starts from the saved tune instead of the CLI defaults, so the decay/shrink rules compound:

run 4:  decay_moderate from defaults (30, 30)   → head=24 tail=24 saved
run 5:  decay_moderate from cached  (24, 24)    → head=19 tail=19 saved
run 6:  decay_strong   from cached  (19, 19)    → head=11 tail=11 saved
run 7:  decay_strong   from cached  (11, 11)    → head=10 tail=10  (floor hit)

The floor (--auto-floor) stops further compounding — and once the workload changes, recover_defaults walks the bucket back to its base. Persistence is best-effort: a missing or unreadable tuned.jsonl silently degrades to the no-persistence behavior. --reset-stats deletes the file along with the metrics (or delete it by hand to reset all learned tunes).

Diagnostic print

When a rule changes the params, l0-compressor prints a single note to stderr (silenced by --quiet), e.g.

l0-compressor: auto-tuning: 2 consecutive failures detected, expanding tail_error to 720

The same firings are aggregated under the AUTO-TUNING section of l0-compressor --stats (and the auto_tuning block of --stats --json).

Environment Variables

Variable	Purpose
XDG_DATA_HOME	Override metrics directory (default: ~/.local/share/l0-compressor/)
XDG_CONFIG_HOME	Override config directory (default: ~/.config/l0-compressor/)
HOME	Used if the XDG variables are not set
L0_COMPRESSOR_GUARD	Force the safety guard on (1/true/on) or off (0/false/off); otherwise auto-detected
L0_COMPRESSOR_NO_TELEMETRY	If truthy, skip all writes to metrics.jsonl / tuned.jsonl for the invocation
NO_COLOR	If set (any value), disables ANSI color in --stats / --doctor
FORCE_COLOR / CLICOLOR_FORCE	Force color on even when stdout is not a TTY (CI captures, screenshots)

The pre-rename L0_CACHE_GUARD / L0_CACHE_NO_TELEMETRY names are still read as deprecated fallbacks; they will be removed at the next major version.

By default --stats and --doctor emit color only when stdout is an interactive terminal, so piping or redirecting them yields clean, escape-free text.

If neither XDG_DATA_HOME nor HOME is set (containers, cron), l0-compressor falls back to /etc/passwd lookup.

Metrics Location

Metrics are stored at $XDG_DATA_HOME/l0-compressor/metrics.jsonl (or ~/.local/share/l0-compressor/metrics.jsonl by default).

The file auto-rotates at 10 MB. The previous file is kept as metrics.jsonl.old. When rotating, entries older than 30 days are automatically filtered out of the .old file to keep disk usage in check.