Dos Self Improve

dos-self-improve — the loop where DOS adjudicates its own improvement

The first self-improving work loop for DOS. It proposes a change, checks it, measures it, and keeps it only if a witness the change's author did not write confirms it improved — the green suite on a clean worktree, the truth syscall, and a strictly-measured metric gain. Everything else is reverted. The keep decision is the kernel's improve verdict, not the loop's opinion of its own work. A run of candidates nothing accepts trips a breaker that hands the judgment back to a human.

This is the dispatch loop (/dos-dispatch-loop) turned inward on a codebase's own source, with the one rule no prior auto-improver enforced:

A self-improving loop's fatal failure mode is grading its own homework — the agent that wrote the change is the same one that says "yes, this is better," so it learns to narrate improvement instead of making it. dos improve closes that hole: the keep-bit is a pure function of facts the loop did not author, so the loop cannot keep a change by claiming it is better. The only path to KEEP is to actually move a metric the environment measures.

The recursive-self-improvement literature names verification as the gating constraint ("requiring verification regimes enabling labs to confirm"; "human judgment on which problems matter remains the bottleneck"). This loop is that verification regime, with the human kept in the loop by construction — the breaker escalates to a person exactly when the loop runs dry of witnessed improvements.

What the kernel decides vs. what you do

The kernel contributes ZERO intelligence to the proposal — only the refusal to keep an unwitnessed one. That asymmetry is the whole design.

Inputs

• --metric <name> (required) — what "improvement" means for THIS workspace, as a non-negative integer the environment measures, higher = better. The reference metric is lint-clean = a large constant minus the dos lint finding count (driving dead policy to zero); other honest metrics: a passing-property-test count, a coverage percent, a negated wall-clock budget. The kernel does not know the unit — it only compares magnitudes (the productivity/efficiency work-unit split).
• --max-cycles <N> (default 5) — the backstop cap (the ITERATION_CAP analogue).
• --max-reverts <N> (default 3) — ESCALATE to a human after this many candidates in a row that nothing accepted (the breaker).
• --lane <name> (optional) — the lane to take; a bare loop auto-picks a free one.

Step 0 — Pre-flight: take a lane, record the GREEN baseline

dos doctor --workspace . --json
dos arbitrate --workspace . --lane <LANE>
RID=$(dos run-id mint dos-self-improve \
      | python -c "import sys,json;print(json.load(sys.stdin)['run_id'])")

RID is the loop's correlation run-id — pass it (with --observe) on every dos improve call below so the kernel records each cycle under THIS loop, and an operator (or DOS) can watch the whole trajectory live with dos observe --loops (docs/383). A long RSI run is otherwise dark while it runs: without --observe the per-cycle verdict evaporates and the only artifact is the final loop record (Step 5).

Then establish the baseline — and this gate is non-negotiable:

python -m pytest -q          # MUST be green to start
<measure the metric>         # e.g. dos lint --workspace . --json | (1000 - finding count)

You cannot measure improvement from a red baseline. If the suite is red, STOP and surface it — fix the suite first (that itself is a separate, ordinary task, not a self-improvement cycle). Record the baseline metric B.

Step 1 — Per cycle: propose ONE candidate in an ISOLATED worktree

The candidate edit is the SELF_MODIFY / global-lane hazard ([[self-modification-hazard]]): editing the kernel's own running path is exactly what the arbiter refuses, and what would let a candidate rewrite the kernel that is adjudicating it. So work in an isolated git worktree, never the live tree:

git worktree add ../_si-candidate HEAD

Spawn ONE subagent with a tight brief, working in that worktree:

Improve exactly ONE thing in this codebase that will move <metric>. Make the SMALLEST diff that moves it. Do not touch tests to make them pass; do not weaken an assertion. Commit your change with a clear subject. Return what you changed and why. If you cannot find a real improvement, say so — do not invent one.

One candidate per cycle keeps the witness attributable (the commit-audit "one commit, one claim, one diff" discipline). If the subagent returns "nothing to improve," the cycle is a skip — not a revert; move on.

Step 2 — Gather the env-authored witnesses (on the worktree)

Every fact the kernel reads is measured by the ENVIRONMENT, never taken from the subagent's word:

# (1) the suite, on the candidate-only worktree — the runner authors the bit
cd ../_si-candidate && python -m pytest -q ; SUITE=$?

# (2) the truth syscall — git ancestry, the oracle authors the bit
dos commit-audit --workspace ../_si-candidate HEAD      # claim vs its own diff
# (and, if the candidate claims a plan phase, dos verify it)

# (3) the metric, re-measured on the worktree AFTER the candidate
<measure the metric on ../_si-candidate>   # → W

# (4) the tokens the subagent spent (the provider usage record) → T

A missing witness is a FAILING witness — if you cannot run the suite or the truth syscall, treat it as red/dirty (fail-safe). Never substitute the subagent's "tests pass" claim for the runner's exit code.

Step 3 — Ask the kernel: KEEP / REVERT / ESCALATE (the kernel decides)

This is the load-bearing step: the decision is a kernel mechanism, not prose.

dos improve --workspace . \
  $( [ "$SUITE" -eq 0 ] && echo --suite-passed ) \
  $( <truth clean> && echo --truth-clean ) \
  --work "$W" --baseline-work "$B" --tokens "$T" \
  --consecutive-reverts "$REVERTS" --max-reverts <N> \
  --narrated "<the subagent's description>" \
  --observe --run-id "$RID" --subject "cycle-$CYCLE" --json

The last line is the observability wiring (docs/383): --observe records this verdict to the journal, --run-id "$RID" groups it with the rest of THIS loop, and --subject "cycle-$CYCLE" tags the iteration — so dos observe --loops can fold the loop's metric curve, breaker distance-to-escalate, and liveness while it runs.

The verdict IS the exit code — branch on it, do not re-read the prose:

• 0 KEEP — suite green, truth clean, W > B (a strict env-measured gain), and not wasteful. The improvement is witnessed. Go to Step 4-KEEP.
• 3 REVERT — either a REGRESSION (suite red / truth dirty — the non-negotiable floor) or a NO-OP (safe but W did not beat B). Go to Step 4-REVERT.
• 4 ESCALATE — the breaker is open (--max-reverts non-keeps in a row). Go to Step 4-ESCALATE and STOP.

The --narrated text is carried for the operator and parsed for nothing — it cannot move REVERT → KEEP (docs/234). Do not try to argue a candidate into a keep; make the metric move or revert it.

Step 4 — Carry out the verdict

• KEEP — merge the worktree commit onto the lane, then RAISE the baseline to W (the ratchet — the next candidate must beat the improved tree, not the original) and reset the revert count to 0:

  git merge --ff-only <candidate-sha>   # or cherry-pick onto the lane
  B=$W ; REVERTS=0
  

• REVERT — discard the worktree candidate; the live tree is untouched. Bump the revert count (REVERTS=$((REVERTS+1))). The baseline is unchanged.

  git worktree remove --force ../_si-candidate
  

• ESCALATE — discard the candidate, then file a human decision and STOP the loop (the loop has run dry of witnessed improvements; a person decides what matters next):

  dos decisions ...   # surface the escalation with the kernel's reason
  

Then re-create a fresh worktree for the next cycle (off the now-possibly-advanced lane HEAD) and return to Step 1, until KEEP/REVERT cycles reach --max-cycles or an ESCALATE stops the loop.

Step 5 — Archive + release

When the loop stops, write a loop record under paths.runs (the per-cycle verdicts, the kept/reverted/skipped counts, the final baseline — the high-water mark that measures how much the loop improved the codebase, and the stop reason), and release the lane lease. Commit any kept improvements with a generic subject read from config — no hardcoded prefix.

Watch it live (any time during the run). Because each cycle was --observed under $RID, an operator or a supervising agent can fold the loop's trajectory from the kernel's own verdict stream — no need to ask the loop how it is doing (docs/383):

dos observe --loops              # every RSI loop: iteration, metric curve, breaker, liveness
dos observe --loops --run "$RID" # THIS loop's full per-cycle curve

A loop reads ACTIVE (a fresh cycle landed), STALLED (no cycle within the window — wedged), or ESCALATED (the breaker handed back to a human). The metric high-water and the breaker N/max distance-to-escalate come straight from the env-authored verdicts, never the candidate's self-assessment.

Why this is the honest version of recursive self-improvement

• The keep-bit reads four env-authored facts — the suite exit, the truth syscall, the metric before, the metric after — and zero loop-authored bytes. A loop that learns to write "great improvement" in every commit gains exactly zero keep-probability, because the claim is not in the decision.
• The loop's default action on an unwitnessed candidate is undo, not keep — the abstention-first discipline at loop scale.
• The breaker escalates to a human precisely when the loop cannot find a witnessed improvement — the RSI literature's irreducible bottleneck, made a kernel rule rather than a hope.

What this skill deliberately does NOT do (no silent gap)

• No open-ended autonomy. The bound is the breaker + the cap; the loop terminates into a human decision, by construction. It does not run unbounded.
• No metric the loop can game. The metric must be measured by the environment on the worktree, not reported by the subagent. A metric the subagent computes about itself is back to grading its own homework — log and refuse such a metric.
• No test-weakening. A candidate that makes the suite pass by deleting an assertion REGRESSES the truth syscall / will be caught by a metric that counts real coverage; the brief forbids it and the green-suite floor is on the FULL suite, not a subset.

Anti-patterns

• ❌ Believing the subagent's "it's better" — re-measure every fact; only dos improve over env-authored witnesses decides KEEP.
• ❌ Editing the live tree — always work in an isolated worktree; the kernel adjudicating a candidate must not be the kernel being rewritten by it.
• ❌ Measuring improvement from a red baseline — fix the suite first; you cannot measure a gain you cannot see.
• ❌ Re-running a candidate the kernel reverted with a better-sounding commit message — the message is not in the decision; make the metric move.