memory backpass: analysis calls run pi with tools off #280

Merged

vnprc merged 1 commit from agent/backpass-analysis-no-tools into master

2026-10-20 05:02:17 +01:00

allod-agent commented

2026-10-19 06:28:02 +01:00

Member

The built-in pi seat now runs the bulk and fallback analysis calls with pi -p --no-tools; the synthesis call and its fix call keep tools on because they edit the worktree. Nothing else changes. What proves it: the first real allod memory backpass --report-only run on a dev VM showed the analysis call opening with "Let me read the memory files first as instructed", because pi -p loads the machine's harness instruction file and, with tools on, follows its read-memory-at-start step instead of answering; the built-in seat test now records the stub pi's argv and asserts --no-tools on every analysis call and its absence on the synthesis call. If it is wrong: revert the one commit; an explicit --bulk-seat wrapper is unaffected either way.

Refs allod/strategy#55

Risk

R1. One flag on one built-in invocation, with a test on the argv.

Validation

gofmt -l cmd/allod empty; go vet ./cmd/allod/... clean
go test ./cmd/allod ok; go test -tags site,secret ./cmd/allod ok, rerun by the integrator
nix build .#checks.x86_64-linux.go-checks passes (worker's run)

The built-in `pi` seat now runs the bulk and fallback analysis calls with `pi -p --no-tools`; the synthesis call and its fix call keep tools on because they edit the worktree. Nothing else changes. What proves it: the first real `allod memory backpass --report-only` run on a dev VM showed the analysis call opening with "Let me read the memory files first as instructed", because `pi -p` loads the machine's harness instruction file and, with tools on, follows its read-memory-at-start step instead of answering; the built-in seat test now records the stub pi's argv and asserts `--no-tools` on every analysis call and its absence on the synthesis call. If it is wrong: revert the one commit; an explicit `--bulk-seat` wrapper is unaffected either way. Refs allod/strategy#55 ## Risk R1. One flag on one built-in invocation, with a test on the argv. ## Validation - `gofmt -l cmd/allod` empty; `go vet ./cmd/allod/...` clean - `go test ./cmd/allod` ok; `go test -tags site,secret ./cmd/allod` ok, rerun by the integrator - `nix build .#checks.x86_64-linux.go-checks` passes (worker's run)

allod-agent added 1 commit

2026-10-19 06:28:02 +01:00

memory: run the built-in pi seat's analysis calls with tools off 70ff57d0cf

A real run showed analysis wandering into reading the harness's memory
files instead of answering the prompt, with tools on. The built-in seat
now takes a role: bulk/fallback run 'pi -p --no-tools', synthesis (and
its fix call) keep tools on for editing the worktree.