mechanical-compiler/docs/PROCESS.md

PROCESS.md

How work gets done on the Mechanical Compiler.

Created	2026-08-17
Scope	All work on a Mechanical Compiler instance
Companions	ENVIRONMENT.md, STAGING-STATE.md, FAILURES.md, ROADMAP.md

---

0. Why this document exists

The four companion documents describe what the environment is. None of them describes how work gets done in it — where files land, who runs what, how a change travels from a conversation into a running container.

Until now that gap was filled by restating the mechanics inside each work order. That worked while the orders were written consecutively by one author. It does not survive a handoff: an assistant reading ENVIRONMENT.md learns how the host is built and nothing about how to operate on it.

Read this before your first command.

---

1. Who does what

Three parties, and confusing them is the most common way to waste a session.

Party	Has	Does
Operator (CIVICVS)	A shell on srv-b and a file manager	Runs every command. Uploads files. Decides.
Assistant	This conversation	Writes commands, reads output, writes code and documents
Architect	This conversation, in a different mode	Sets specification, accepts work, maintains the documents

The constraint that shapes everything

The operator has a shell on srv-b and a browser-based file manager. Nothing else.

No workstation git client. No direct container shell. No IDE. No ability to scp. Every file arrives by upload to a folder on srv-b; every command runs in that one shell.

An assistant that assumes otherwise produces instructions the operator cannot execute. This has already happened once.

What follows from it

• Files reach a container as: upload to srv-b → pct push or expand and copy → pct exec to act on them.
• srv-b is the only place a human types. Containers are reached through pct exec, never by SSH from elsewhere.
• This is a legitimate deployment path, not a workaround. Do not design around a git client that does not exist.

---

2. Two modes of work, and they are not the same

Applying the wrong one is slow in one direction and dangerous in the other.

Infrastructure mode

For anything that changes the host, the containers, the network, or a service that other work depends on.

• One command group at a time
• Read-only before write
• Record a failure before correcting it, in FAILURES.md format
• Smallest corrective experiment — not the one that also fixes three adjacent worries
• Preserve a rollback copy before editing any configuration file
• Prove the negative: assert the forbidden path fails, not that the interface is gone
• Assert health after a host reboot, never before

This is deliberate ceremony. Each step is irreversible or expensive to undo, and the failure log is a deliverable in its own right — it is the primary input to the eventual production automation.

Work orders 001 through 004 were all infrastructure mode.

Development mode

For anything inside the repository: code, tests, documents, fixtures.

• Iterate freely. Edit, run the tests, edit again.
• No command-group ceremony. A test-fix-rerun loop is not a provisioning step.
• Failures during a normal red-green loop are not FAILURES.md entries.
• Commit when something works, not when something is finished.

The distinction is reversibility. A wrong iptables rule can strand the operator. A wrong line of Python fails a test and gets deleted. Ceremony appropriate to the first is obstruction applied to the second.

Which mode am I in?

Ask: if this is wrong, what does it cost?

Cost	Mode
A rebooted host, a lost shell, a broken service	Infrastructure
A red test	Development

If genuinely unsure, use infrastructure mode. Being slow is recoverable.

---

3. Getting files into the environment

The path

assistant produces a tarball
  -> operator downloads it
  -> operator uploads it to /root/incoming on srv-b   (file manager)
  -> operator expands it on srv-b
  -> pct push, or expand-then-copy, into CT 100
  -> pct exec to act on it

REQ — /root/incoming on srv-b is the landing area. One directory, always the same, so nothing has to be remembered between sessions.

REQ — Files landing in a container must end up owned by mechcomp. pct push writes as root; fix ownership immediately afterwards. See F-008 — repository operations run as the service user, and a root-owned file inside the tree causes exactly that failure later.

REQ (F-037) — The chown must name the directory the files landed in, not only the files. A tar stream rooted at . carries an entry for the destination directory itself, and tar x as root rewrites that directory's ownership. Naming only the payload subdirectories leaves the repository root root:root, and git then refuses the worktree with the F-008 message — which invites the F-008 mistake as its own remedy. Fix the owner. Never add safe.directory.

REQ (F-037) — Verification that depends on the repository being intact must run after the landing, and a landing step must not be able to destroy the check that would have caught it. A git diff placed before the step that broke git reports nothing wrong and looks like a pass.

REQ — An assistant delivering files states, in this order: what the archive contains, where it expands, what it overwrites, and how to verify it landed correctly. "Overwrites nothing" is a claim that must be checked, not assumed.

Delivering a whole tree

Prefer one tarball that expands over the existing tree. Individual file paths are error-prone to transcribe through a file manager.

Name the payload's top-level directories explicitly, when creating the archive and when extracting it: tar cf - -C <dir> src tests, never -C <dir> .. The second form is what produced F-037. For a single file, pct push is simpler and cannot reproduce it at all.

Delivering a single small file

For anything that fits comfortably on screen, a heredoc in the srv-b shell is faster than an upload, and leaves the content visible in the session transcript where it can be checked.

---

3a. The delivery chain, end to end

The whole path from an assistant producing something to it existing in Gitea. It has been explained conversationally to every new assistant for months and was written in no document until now, which is the reason this section exists.

Nothing but the operator writes. The assistant reads Gitea over MCP, which is read-only, and builds files in its own sandbox. The operator is the only party with write access anywhere in the chain. That is deliberate, and it is section 1 restated: the operator executes, and the assistant never assumes a command succeeded.

The steps

assistant builds the artifact in its sandbox
  -> states sha256, size, contents, where it lands, what it overwrites
  -> operator downloads it and uploads to /root/incoming on srv-b
  -> assistant gives ONE command group
  -> operator pastes it into the srv-b shell, pastes output back
  -> repeat until landed, verified, pushed

Three groups, not one

REQ — landing, verification and commit are separate command groups, because section 4 requires a suite result the assistant has actually seen before a commit exists. A single group that lands and commits cannot satisfy that.

Group	Does	Ends when
1	pct push or expand-and-copy, fix ownership, confirm checksum, git status	the file is in place and owned by mechcomp
2	make verify-oracle, make test	the assistant has read a real result
3	git add, git commit, git push, verify	git status -sb shows no ahead marker

Idioms that need no shell

pct exec runs no shell (section 9). These forms avoid needing one:

• make -C /var/www/mechcomp test — never cd ... && make test
• git -C /var/www/mechcomp status — never cd ... && git status
• runuser -u mechcomp -- <cmd> — never su

REQ — a commit message is parsed by the host shell before pct sees it. Use repeated -m flags, and no dollar sign, backtick or exclamation mark anywhere in the text. A heredoc does not survive pct exec.

State which failures are passes

REQ — an instruction set says what the expected output is, including where an error is the correct result. The group-1 check that a target does not yet exist fails with No such file or directory when the delivery is correct; an operator who reads that as a fault stops a working delivery. An operator who cannot tell success from failure cannot report usefully (section 9).

Amending a large document

REQ — do not ship a rewritten file to change a few passages of a large one. Ship an anchored replacement script that validates every anchor matches exactly once and writes nothing if any does not.

A rewrite regenerates the whole document from the assistant's reading of it, and a transcription error is silent. FAILURES.md is 52 kB of correct prose whose value depends on never being rewritten, and the same hazard applies to any document large enough that nobody will diff it carefully.

This is the HANDOFF.md problem one level down: a document replaced rather than amended loses content, and nothing reports what left.

---

4. The repository

Gitea is the source of truth. https://gitea.barternetwork.us/TheRON/mechanical-compiler

CT 100 holds a clone at /var/www/mechcomp, owned by mechcomp.

Two directions, both valid

Gitea → container. git pull inside CT 100, as mechcomp. Use this whenever the change already exists in Gitea.

Upload → container → Gitea. Land the files, verify they work, then commit and push from inside CT 100. Use this when the assistant produced something new.

The second direction is the normal one for assistant-produced work, because the operator has no git client outside the environment.

REQ (F-008) — Every git command runs as mechcomp. Never as root, and never add a root safe.directory exception — it would mask every future instance of the same mistake.

What must be true before a push

• The oracle verifies: python3 fixtures/strap-beam-8.0.0/make_fixtures.py --verify
• The test suite runs, with a result the assistant has actually seen
• git status shows nothing unexpected — particularly not venv/

The first and third are one check, not two (F-033). --verify recomputes the hash from the document it reads and compares it against the value stored inside that same document, so a wholly regenerated oracle is self-consistent and passes. It proves internal integrity, never identity with the committed bytes. Only git status can tell you which oracle is present. Dropping either as redundant removes the only check that catches a substituted oracle.

The two interpreters are deliberate, and the Makefile is right. make verify-oracle runs python3 because make_fixtures.py imports nothing outside the standard library — that is what makes the oracle check trustworthy on any machine at any time, and it must not be made to depend on a deployment. make test runs venv/bin/python because the suite imports mechcomp, Shapely and pytest, none of which system python3 has (F-036). Neither is a mistake to be tidied into consistency.

Recording provenance

Artifacts are attributed to the commit that produced them, not to a timestamp or a file copy. When an artifact is generated, the commit SHA of the tree that produced it is part of its record. This is why the transport does not matter but the SHA does.

---

5. The development loop

Inside CT 100, as mechcomp, from /var/www/mechcomp:

make deps        once, and after any requirements change
make test        the loop
make verify-oracle   before any commit that touches fixtures

The oracle is the gate

fixtures/strap-beam-8.0.0/ holds 123 frozen cases. It is the acceptance criterion for the port, and it is not negotiable.

The ten rejected cases are part of the contract. A port that accepts them is wrong however good its numbers are elsewhere. Reproducing geometry is straightforward; keeping the constraint that made the geometry trustworthy is the actual work.

The oracle is never edited to make a test pass. If the port disagrees with it, the port is wrong until proven otherwise. If the oracle is genuinely wrong, that is a finding, it is recorded, and it is regenerated inside the pinned toolchain — never hand-edited.

Test discipline

A test must distinguish failure of the tool from failure of the thing under test (F-027). Every negative assertion first establishes that it could have observed the positive case. A test whose failure mode is indistinguishable from success manufactures confidence.

Prove a new harness by making it fail deliberately before trusting it green.

---

6. Work orders

A work order is an infrastructure-mode instruction set. Development work does not need one.

When to write one

• The work changes the host, containers, network, or a shared service
• The work has an acceptance criterion that can be stated in advance
• The work will be executed by someone other than the author

Required sections

Section	Purpose
Scope	In and out. Explicit "do not touch" list.
Success criteria	Strict, testable, stated before the work
Known state	So the operator does not re-derive what is already recorded
Command groups	One at a time, read-only first
Acceptance	What must be true to close
Reporting	What comes back to the architect

Executing one

• One group at a time. Paste output back before the next.
• Record a failure before correcting it.
• A defect in the work order is a finding, not something to work around silently. F-024 and F-028 are work-order defects, logged as such.
• "The problem is elsewhere and here is the evidence" is a complete outcome. Do not manufacture a local workaround for a remote problem.

---

7. When something goes wrong

Is it a FAILURES.md entry?

Yes if it is a surprise about the environment: something behaved differently from what the specification or a work order said, and a future automation writer would get it wrong the same way.

No if it is an ordinary development failure: a red test, a syntax error, a wrong algorithm. Those are the loop working.

Entry format

### F-nnn — one-line summary
Host or container. Phase.
**Observed:** verbatim where possible
**Cause:** proven, or explicitly "unproven"
**Correction:** the smallest change that fixed it
**Consequence:** what the specification or automation must do differently

An entry with no Consequence is either not understood or not worth recording.

"Unproven" is a respectable result. F-006, F-012 and F-022 are carried open and the log is better for it. Never invent a cause to close an entry, and never close one on a change that merely coincided with the symptom disappearing.

Escalating

Stop and ask the operator when:

• A decision is needed that is not the assistant's to make
• The work requires changing infrastructure outside this project
• A REQ in the specification cannot be satisfied
• The specification and observed reality disagree

The last one is not a blocker — reality wins, and the specification is corrected. But it is always worth saying out loud.

---

8. Session handoff

Sessions end abruptly. Assume the next assistant has these documents and nothing else.

Before a session ends

• Anything working is committed and pushed
• Anything learned is in FAILURES.md
• Anything changed on a host is in STAGING-STATE.md
• Any open decision is in STAGING-STATE.md section 6

Starting a session

Read in this order:

1. PROCESS.md — this document
2. STAGING-STATE.md — what is true right now
3. FAILURES.md — what has already gone wrong
4. HANDOFF.md — where the last session left off
5. DIVERGENCES.md — what is required and not met
6. ENVIRONMENT.md — the specification
7. ROADMAP.md — where it is going

HANDOFF.md was created after this list was written and was missing from it for three weeks, while its own section 0 carried a different list missing ENVIRONMENT.md. Two reading orders, neither complete, each looking authoritative. If you add a document, add it here.

If writing provisioning automation, read FAILURES.md before the specification. Every entry is something a script written from the specification alone would have got wrong.

Authority when documents disagree

1. STAGING-STATE.md — factual, wins on what is true
2. FAILURES.md — evidence from contact with real hosts
3. DIVERGENCES.md — where a requirement stands and is not met
4. ENVIRONMENT.md — the specification, corrected when proven wrong
5. ROADMAP.md — sequence

The rule underneath the list: a document that describes something yields to the thing it describes; a document that prescribes something does not. STAGING-STATE.md describes a host, so the host wins. deploy/README.md prescribes a host configuration, so it does not — and it says so itself. Both are correct, and the apparent contradiction dissolves once the two kinds are named.

A permanent deviation is not a deviation. It is the specification. Promote it and delete the exception.

That applies to facts about a host, not to requirements the code has not met. A REQ is never lowered to match what was built. It stands, the gap is recorded in DIVERGENCES.md, and the correction is owed by the code. Promoting a divergence would make every document true by construction and worthless — a specification that agrees with whatever exists specifies nothing.

---

9. Instructions the operator can actually run

REQ — pct exec runs no shell. A glob, redirect, pipe or && in a pct exec command is expanded by the host shell against the host's filesystem, and the container receives whatever literal survives. Wrap them: pct exec 101 -- sh -c '...'. An unwrapped glob produces an error that describes the container rather than the invocation, which is the same misdirection as F-035 and F-036 — three instances now, all of them the tool being invoked wrongly and the message naming the wrong subject.

This section exists because it has already gone wrong.

REQ — One command group per message. Wait for output.

REQ — Commands run in the srv-b shell, or via pct exec from it. Nothing assumes a tool the operator does not have.

REQ — A group is copy-pasteable as a block, with echo markers separating sections so the output can be read back.

REQ — State what the expected output is. An operator who cannot tell success from failure cannot report usefully.

Do not deliver a numbered multi-stage procedure spanning several systems and expect it to be executed. It will not be, and the parts that are will not be separable in the output.

Do not assume the operator will improvise the missing step. If a command needs a directory to exist, create it in the same group.

---

9a. The baseline check

ct-baseline.sh is the executable definition of the container standard on this host. Read-only, runnable at any time, exits non-zero on divergence.

Run it:

• before starting work on a container;
• after any change to a container or to host firewall rules;
• after any host reboot;
• before considering backup work (see ENVIRONMENT.md §15 gate 4).

A property not checked by it is not part of the standard. If something should be uniform across containers, add it to the script. If it should not, leave it out and stop worrying about the difference. That is the whole point: it converts an endless comparison into a pass or a fail.

It covers containers belonging to more than one project, and encodes host requirements neither project owns alone. Treat it as host property.

---

10. Current mode

Infrastructure: complete and accepted, and now includes public ingress — WORK-ORDER-004-public-ingress.md, closed 2026-09-11. See STAGING-STATE.md for what is true on the host, and DIVERGENCES.md for where the running instance and the specification disagree.

Backup: deliberately postponed. Nothing exists yet whose loss would cost more than an afternoon; the documents are in Gitea. This changes when artifacts/ stops being empty.

Container baseline: established, and not re-run since the host last changed. See DIV-006.

Development: the Shapely port is complete. The compiler builds cross sections, reports them, renders SVG, exports STL, and serves a composer over HTTP. ROADMAP.md section 4 carries what comes next.

This section no longer records a commit, a version or a test count. Those are reported by git rev-parse HEAD, git describe --tags and make test, and every previous attempt to hold them current in prose went stale — the last one claimed the repository was still at its seed commit while the work it described had been finished for three weeks. A document states what no command can report: a decision, a constraint, a reason, a hazard.

From here the mode is development unless the work touches the host, the containers, or a shared service.